ubifs: Set s_uuid in super block to support ima/evm uuid options
[linux-2.6-microblaze.git] / mm / hugetlb.c
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Generic hugetlb support.
4  * (C) Nadia Yvette Chambers, April 2004
5  */
6 #include <linux/list.h>
7 #include <linux/init.h>
8 #include <linux/mm.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/compiler.h>
17 #include <linux/cpuset.h>
18 #include <linux/mutex.h>
19 #include <linux/memblock.h>
20 #include <linux/sysfs.h>
21 #include <linux/slab.h>
22 #include <linux/sched/mm.h>
23 #include <linux/mmdebug.h>
24 #include <linux/sched/signal.h>
25 #include <linux/rmap.h>
26 #include <linux/string_helpers.h>
27 #include <linux/swap.h>
28 #include <linux/swapops.h>
29 #include <linux/jhash.h>
30 #include <linux/numa.h>
31 #include <linux/llist.h>
32 #include <linux/cma.h>
33
34 #include <asm/page.h>
35 #include <asm/pgalloc.h>
36 #include <asm/tlb.h>
37
38 #include <linux/io.h>
39 #include <linux/hugetlb.h>
40 #include <linux/hugetlb_cgroup.h>
41 #include <linux/node.h>
42 #include <linux/userfaultfd_k.h>
43 #include <linux/page_owner.h>
44 #include "internal.h"
45
46 int hugetlb_max_hstate __read_mostly;
47 unsigned int default_hstate_idx;
48 struct hstate hstates[HUGE_MAX_HSTATE];
49
50 #ifdef CONFIG_CMA
51 static struct cma *hugetlb_cma[MAX_NUMNODES];
52 #endif
53 static unsigned long hugetlb_cma_size __initdata;
54
55 /*
56  * Minimum page order among possible hugepage sizes, set to a proper value
57  * at boot time.
58  */
59 static unsigned int minimum_order __read_mostly = UINT_MAX;
60
61 __initdata LIST_HEAD(huge_boot_pages);
62
63 /* for command line parsing */
64 static struct hstate * __initdata parsed_hstate;
65 static unsigned long __initdata default_hstate_max_huge_pages;
66 static bool __initdata parsed_valid_hugepagesz = true;
67 static bool __initdata parsed_default_hugepagesz;
68
69 /*
70  * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
71  * free_huge_pages, and surplus_huge_pages.
72  */
73 DEFINE_SPINLOCK(hugetlb_lock);
74
75 /*
76  * Serializes faults on the same logical page.  This is used to
77  * prevent spurious OOMs when the hugepage pool is fully utilized.
78  */
79 static int num_fault_mutexes;
80 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
81
82 /* Forward declaration */
83 static int hugetlb_acct_memory(struct hstate *h, long delta);
84
85 static inline bool subpool_is_free(struct hugepage_subpool *spool)
86 {
87         if (spool->count)
88                 return false;
89         if (spool->max_hpages != -1)
90                 return spool->used_hpages == 0;
91         if (spool->min_hpages != -1)
92                 return spool->rsv_hpages == spool->min_hpages;
93
94         return true;
95 }
96
97 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
98 {
99         spin_unlock(&spool->lock);
100
101         /* If no pages are used, and no other handles to the subpool
102          * remain, give up any reservations based on minimum size and
103          * free the subpool */
104         if (subpool_is_free(spool)) {
105                 if (spool->min_hpages != -1)
106                         hugetlb_acct_memory(spool->hstate,
107                                                 -spool->min_hpages);
108                 kfree(spool);
109         }
110 }
111
112 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
113                                                 long min_hpages)
114 {
115         struct hugepage_subpool *spool;
116
117         spool = kzalloc(sizeof(*spool), GFP_KERNEL);
118         if (!spool)
119                 return NULL;
120
121         spin_lock_init(&spool->lock);
122         spool->count = 1;
123         spool->max_hpages = max_hpages;
124         spool->hstate = h;
125         spool->min_hpages = min_hpages;
126
127         if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
128                 kfree(spool);
129                 return NULL;
130         }
131         spool->rsv_hpages = min_hpages;
132
133         return spool;
134 }
135
136 void hugepage_put_subpool(struct hugepage_subpool *spool)
137 {
138         spin_lock(&spool->lock);
139         BUG_ON(!spool->count);
140         spool->count--;
141         unlock_or_release_subpool(spool);
142 }
143
144 /*
145  * Subpool accounting for allocating and reserving pages.
146  * Return -ENOMEM if there are not enough resources to satisfy the
147  * request.  Otherwise, return the number of pages by which the
148  * global pools must be adjusted (upward).  The returned value may
149  * only be different than the passed value (delta) in the case where
150  * a subpool minimum size must be maintained.
151  */
152 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
153                                       long delta)
154 {
155         long ret = delta;
156
157         if (!spool)
158                 return ret;
159
160         spin_lock(&spool->lock);
161
162         if (spool->max_hpages != -1) {          /* maximum size accounting */
163                 if ((spool->used_hpages + delta) <= spool->max_hpages)
164                         spool->used_hpages += delta;
165                 else {
166                         ret = -ENOMEM;
167                         goto unlock_ret;
168                 }
169         }
170
171         /* minimum size accounting */
172         if (spool->min_hpages != -1 && spool->rsv_hpages) {
173                 if (delta > spool->rsv_hpages) {
174                         /*
175                          * Asking for more reserves than those already taken on
176                          * behalf of subpool.  Return difference.
177                          */
178                         ret = delta - spool->rsv_hpages;
179                         spool->rsv_hpages = 0;
180                 } else {
181                         ret = 0;        /* reserves already accounted for */
182                         spool->rsv_hpages -= delta;
183                 }
184         }
185
186 unlock_ret:
187         spin_unlock(&spool->lock);
188         return ret;
189 }
190
191 /*
192  * Subpool accounting for freeing and unreserving pages.
193  * Return the number of global page reservations that must be dropped.
194  * The return value may only be different than the passed value (delta)
195  * in the case where a subpool minimum size must be maintained.
196  */
197 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
198                                        long delta)
199 {
200         long ret = delta;
201
202         if (!spool)
203                 return delta;
204
205         spin_lock(&spool->lock);
206
207         if (spool->max_hpages != -1)            /* maximum size accounting */
208                 spool->used_hpages -= delta;
209
210          /* minimum size accounting */
211         if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
212                 if (spool->rsv_hpages + delta <= spool->min_hpages)
213                         ret = 0;
214                 else
215                         ret = spool->rsv_hpages + delta - spool->min_hpages;
216
217                 spool->rsv_hpages += delta;
218                 if (spool->rsv_hpages > spool->min_hpages)
219                         spool->rsv_hpages = spool->min_hpages;
220         }
221
222         /*
223          * If hugetlbfs_put_super couldn't free spool due to an outstanding
224          * quota reference, free it now.
225          */
226         unlock_or_release_subpool(spool);
227
228         return ret;
229 }
230
231 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
232 {
233         return HUGETLBFS_SB(inode->i_sb)->spool;
234 }
235
236 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
237 {
238         return subpool_inode(file_inode(vma->vm_file));
239 }
240
241 /* Helper that removes a struct file_region from the resv_map cache and returns
242  * it for use.
243  */
244 static struct file_region *
245 get_file_region_entry_from_cache(struct resv_map *resv, long from, long to)
246 {
247         struct file_region *nrg = NULL;
248
249         VM_BUG_ON(resv->region_cache_count <= 0);
250
251         resv->region_cache_count--;
252         nrg = list_first_entry(&resv->region_cache, struct file_region, link);
253         list_del(&nrg->link);
254
255         nrg->from = from;
256         nrg->to = to;
257
258         return nrg;
259 }
260
261 static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg,
262                                               struct file_region *rg)
263 {
264 #ifdef CONFIG_CGROUP_HUGETLB
265         nrg->reservation_counter = rg->reservation_counter;
266         nrg->css = rg->css;
267         if (rg->css)
268                 css_get(rg->css);
269 #endif
270 }
271
272 /* Helper that records hugetlb_cgroup uncharge info. */
273 static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg,
274                                                 struct hstate *h,
275                                                 struct resv_map *resv,
276                                                 struct file_region *nrg)
277 {
278 #ifdef CONFIG_CGROUP_HUGETLB
279         if (h_cg) {
280                 nrg->reservation_counter =
281                         &h_cg->rsvd_hugepage[hstate_index(h)];
282                 nrg->css = &h_cg->css;
283                 /*
284                  * The caller will hold exactly one h_cg->css reference for the
285                  * whole contiguous reservation region. But this area might be
286                  * scattered when there are already some file_regions reside in
287                  * it. As a result, many file_regions may share only one css
288                  * reference. In order to ensure that one file_region must hold
289                  * exactly one h_cg->css reference, we should do css_get for
290                  * each file_region and leave the reference held by caller
291                  * untouched.
292                  */
293                 css_get(&h_cg->css);
294                 if (!resv->pages_per_hpage)
295                         resv->pages_per_hpage = pages_per_huge_page(h);
296                 /* pages_per_hpage should be the same for all entries in
297                  * a resv_map.
298                  */
299                 VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h));
300         } else {
301                 nrg->reservation_counter = NULL;
302                 nrg->css = NULL;
303         }
304 #endif
305 }
306
307 static void put_uncharge_info(struct file_region *rg)
308 {
309 #ifdef CONFIG_CGROUP_HUGETLB
310         if (rg->css)
311                 css_put(rg->css);
312 #endif
313 }
314
315 static bool has_same_uncharge_info(struct file_region *rg,
316                                    struct file_region *org)
317 {
318 #ifdef CONFIG_CGROUP_HUGETLB
319         return rg && org &&
320                rg->reservation_counter == org->reservation_counter &&
321                rg->css == org->css;
322
323 #else
324         return true;
325 #endif
326 }
327
328 static void coalesce_file_region(struct resv_map *resv, struct file_region *rg)
329 {
330         struct file_region *nrg = NULL, *prg = NULL;
331
332         prg = list_prev_entry(rg, link);
333         if (&prg->link != &resv->regions && prg->to == rg->from &&
334             has_same_uncharge_info(prg, rg)) {
335                 prg->to = rg->to;
336
337                 list_del(&rg->link);
338                 put_uncharge_info(rg);
339                 kfree(rg);
340
341                 rg = prg;
342         }
343
344         nrg = list_next_entry(rg, link);
345         if (&nrg->link != &resv->regions && nrg->from == rg->to &&
346             has_same_uncharge_info(nrg, rg)) {
347                 nrg->from = rg->from;
348
349                 list_del(&rg->link);
350                 put_uncharge_info(rg);
351                 kfree(rg);
352         }
353 }
354
355 static inline long
356 hugetlb_resv_map_add(struct resv_map *map, struct file_region *rg, long from,
357                      long to, struct hstate *h, struct hugetlb_cgroup *cg,
358                      long *regions_needed)
359 {
360         struct file_region *nrg;
361
362         if (!regions_needed) {
363                 nrg = get_file_region_entry_from_cache(map, from, to);
364                 record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg);
365                 list_add(&nrg->link, rg->link.prev);
366                 coalesce_file_region(map, nrg);
367         } else
368                 *regions_needed += 1;
369
370         return to - from;
371 }
372
373 /*
374  * Must be called with resv->lock held.
375  *
376  * Calling this with regions_needed != NULL will count the number of pages
377  * to be added but will not modify the linked list. And regions_needed will
378  * indicate the number of file_regions needed in the cache to carry out to add
379  * the regions for this range.
380  */
381 static long add_reservation_in_range(struct resv_map *resv, long f, long t,
382                                      struct hugetlb_cgroup *h_cg,
383                                      struct hstate *h, long *regions_needed)
384 {
385         long add = 0;
386         struct list_head *head = &resv->regions;
387         long last_accounted_offset = f;
388         struct file_region *rg = NULL, *trg = NULL;
389
390         if (regions_needed)
391                 *regions_needed = 0;
392
393         /* In this loop, we essentially handle an entry for the range
394          * [last_accounted_offset, rg->from), at every iteration, with some
395          * bounds checking.
396          */
397         list_for_each_entry_safe(rg, trg, head, link) {
398                 /* Skip irrelevant regions that start before our range. */
399                 if (rg->from < f) {
400                         /* If this region ends after the last accounted offset,
401                          * then we need to update last_accounted_offset.
402                          */
403                         if (rg->to > last_accounted_offset)
404                                 last_accounted_offset = rg->to;
405                         continue;
406                 }
407
408                 /* When we find a region that starts beyond our range, we've
409                  * finished.
410                  */
411                 if (rg->from >= t)
412                         break;
413
414                 /* Add an entry for last_accounted_offset -> rg->from, and
415                  * update last_accounted_offset.
416                  */
417                 if (rg->from > last_accounted_offset)
418                         add += hugetlb_resv_map_add(resv, rg,
419                                                     last_accounted_offset,
420                                                     rg->from, h, h_cg,
421                                                     regions_needed);
422
423                 last_accounted_offset = rg->to;
424         }
425
426         /* Handle the case where our range extends beyond
427          * last_accounted_offset.
428          */
429         if (last_accounted_offset < t)
430                 add += hugetlb_resv_map_add(resv, rg, last_accounted_offset,
431                                             t, h, h_cg, regions_needed);
432
433         VM_BUG_ON(add < 0);
434         return add;
435 }
436
437 /* Must be called with resv->lock acquired. Will drop lock to allocate entries.
438  */
439 static int allocate_file_region_entries(struct resv_map *resv,
440                                         int regions_needed)
441         __must_hold(&resv->lock)
442 {
443         struct list_head allocated_regions;
444         int to_allocate = 0, i = 0;
445         struct file_region *trg = NULL, *rg = NULL;
446
447         VM_BUG_ON(regions_needed < 0);
448
449         INIT_LIST_HEAD(&allocated_regions);
450
451         /*
452          * Check for sufficient descriptors in the cache to accommodate
453          * the number of in progress add operations plus regions_needed.
454          *
455          * This is a while loop because when we drop the lock, some other call
456          * to region_add or region_del may have consumed some region_entries,
457          * so we keep looping here until we finally have enough entries for
458          * (adds_in_progress + regions_needed).
459          */
460         while (resv->region_cache_count <
461                (resv->adds_in_progress + regions_needed)) {
462                 to_allocate = resv->adds_in_progress + regions_needed -
463                               resv->region_cache_count;
464
465                 /* At this point, we should have enough entries in the cache
466                  * for all the existings adds_in_progress. We should only be
467                  * needing to allocate for regions_needed.
468                  */
469                 VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress);
470
471                 spin_unlock(&resv->lock);
472                 for (i = 0; i < to_allocate; i++) {
473                         trg = kmalloc(sizeof(*trg), GFP_KERNEL);
474                         if (!trg)
475                                 goto out_of_memory;
476                         list_add(&trg->link, &allocated_regions);
477                 }
478
479                 spin_lock(&resv->lock);
480
481                 list_splice(&allocated_regions, &resv->region_cache);
482                 resv->region_cache_count += to_allocate;
483         }
484
485         return 0;
486
487 out_of_memory:
488         list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
489                 list_del(&rg->link);
490                 kfree(rg);
491         }
492         return -ENOMEM;
493 }
494
495 /*
496  * Add the huge page range represented by [f, t) to the reserve
497  * map.  Regions will be taken from the cache to fill in this range.
498  * Sufficient regions should exist in the cache due to the previous
499  * call to region_chg with the same range, but in some cases the cache will not
500  * have sufficient entries due to races with other code doing region_add or
501  * region_del.  The extra needed entries will be allocated.
502  *
503  * regions_needed is the out value provided by a previous call to region_chg.
504  *
505  * Return the number of new huge pages added to the map.  This number is greater
506  * than or equal to zero.  If file_region entries needed to be allocated for
507  * this operation and we were not able to allocate, it returns -ENOMEM.
508  * region_add of regions of length 1 never allocate file_regions and cannot
509  * fail; region_chg will always allocate at least 1 entry and a region_add for
510  * 1 page will only require at most 1 entry.
511  */
512 static long region_add(struct resv_map *resv, long f, long t,
513                        long in_regions_needed, struct hstate *h,
514                        struct hugetlb_cgroup *h_cg)
515 {
516         long add = 0, actual_regions_needed = 0;
517
518         spin_lock(&resv->lock);
519 retry:
520
521         /* Count how many regions are actually needed to execute this add. */
522         add_reservation_in_range(resv, f, t, NULL, NULL,
523                                  &actual_regions_needed);
524
525         /*
526          * Check for sufficient descriptors in the cache to accommodate
527          * this add operation. Note that actual_regions_needed may be greater
528          * than in_regions_needed, as the resv_map may have been modified since
529          * the region_chg call. In this case, we need to make sure that we
530          * allocate extra entries, such that we have enough for all the
531          * existing adds_in_progress, plus the excess needed for this
532          * operation.
533          */
534         if (actual_regions_needed > in_regions_needed &&
535             resv->region_cache_count <
536                     resv->adds_in_progress +
537                             (actual_regions_needed - in_regions_needed)) {
538                 /* region_add operation of range 1 should never need to
539                  * allocate file_region entries.
540                  */
541                 VM_BUG_ON(t - f <= 1);
542
543                 if (allocate_file_region_entries(
544                             resv, actual_regions_needed - in_regions_needed)) {
545                         return -ENOMEM;
546                 }
547
548                 goto retry;
549         }
550
551         add = add_reservation_in_range(resv, f, t, h_cg, h, NULL);
552
553         resv->adds_in_progress -= in_regions_needed;
554
555         spin_unlock(&resv->lock);
556         VM_BUG_ON(add < 0);
557         return add;
558 }
559
560 /*
561  * Examine the existing reserve map and determine how many
562  * huge pages in the specified range [f, t) are NOT currently
563  * represented.  This routine is called before a subsequent
564  * call to region_add that will actually modify the reserve
565  * map to add the specified range [f, t).  region_chg does
566  * not change the number of huge pages represented by the
567  * map.  A number of new file_region structures is added to the cache as a
568  * placeholder, for the subsequent region_add call to use. At least 1
569  * file_region structure is added.
570  *
571  * out_regions_needed is the number of regions added to the
572  * resv->adds_in_progress.  This value needs to be provided to a follow up call
573  * to region_add or region_abort for proper accounting.
574  *
575  * Returns the number of huge pages that need to be added to the existing
576  * reservation map for the range [f, t).  This number is greater or equal to
577  * zero.  -ENOMEM is returned if a new file_region structure or cache entry
578  * is needed and can not be allocated.
579  */
580 static long region_chg(struct resv_map *resv, long f, long t,
581                        long *out_regions_needed)
582 {
583         long chg = 0;
584
585         spin_lock(&resv->lock);
586
587         /* Count how many hugepages in this range are NOT represented. */
588         chg = add_reservation_in_range(resv, f, t, NULL, NULL,
589                                        out_regions_needed);
590
591         if (*out_regions_needed == 0)
592                 *out_regions_needed = 1;
593
594         if (allocate_file_region_entries(resv, *out_regions_needed))
595                 return -ENOMEM;
596
597         resv->adds_in_progress += *out_regions_needed;
598
599         spin_unlock(&resv->lock);
600         return chg;
601 }
602
603 /*
604  * Abort the in progress add operation.  The adds_in_progress field
605  * of the resv_map keeps track of the operations in progress between
606  * calls to region_chg and region_add.  Operations are sometimes
607  * aborted after the call to region_chg.  In such cases, region_abort
608  * is called to decrement the adds_in_progress counter. regions_needed
609  * is the value returned by the region_chg call, it is used to decrement
610  * the adds_in_progress counter.
611  *
612  * NOTE: The range arguments [f, t) are not needed or used in this
613  * routine.  They are kept to make reading the calling code easier as
614  * arguments will match the associated region_chg call.
615  */
616 static void region_abort(struct resv_map *resv, long f, long t,
617                          long regions_needed)
618 {
619         spin_lock(&resv->lock);
620         VM_BUG_ON(!resv->region_cache_count);
621         resv->adds_in_progress -= regions_needed;
622         spin_unlock(&resv->lock);
623 }
624
625 /*
626  * Delete the specified range [f, t) from the reserve map.  If the
627  * t parameter is LONG_MAX, this indicates that ALL regions after f
628  * should be deleted.  Locate the regions which intersect [f, t)
629  * and either trim, delete or split the existing regions.
630  *
631  * Returns the number of huge pages deleted from the reserve map.
632  * In the normal case, the return value is zero or more.  In the
633  * case where a region must be split, a new region descriptor must
634  * be allocated.  If the allocation fails, -ENOMEM will be returned.
635  * NOTE: If the parameter t == LONG_MAX, then we will never split
636  * a region and possibly return -ENOMEM.  Callers specifying
637  * t == LONG_MAX do not need to check for -ENOMEM error.
638  */
639 static long region_del(struct resv_map *resv, long f, long t)
640 {
641         struct list_head *head = &resv->regions;
642         struct file_region *rg, *trg;
643         struct file_region *nrg = NULL;
644         long del = 0;
645
646 retry:
647         spin_lock(&resv->lock);
648         list_for_each_entry_safe(rg, trg, head, link) {
649                 /*
650                  * Skip regions before the range to be deleted.  file_region
651                  * ranges are normally of the form [from, to).  However, there
652                  * may be a "placeholder" entry in the map which is of the form
653                  * (from, to) with from == to.  Check for placeholder entries
654                  * at the beginning of the range to be deleted.
655                  */
656                 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
657                         continue;
658
659                 if (rg->from >= t)
660                         break;
661
662                 if (f > rg->from && t < rg->to) { /* Must split region */
663                         /*
664                          * Check for an entry in the cache before dropping
665                          * lock and attempting allocation.
666                          */
667                         if (!nrg &&
668                             resv->region_cache_count > resv->adds_in_progress) {
669                                 nrg = list_first_entry(&resv->region_cache,
670                                                         struct file_region,
671                                                         link);
672                                 list_del(&nrg->link);
673                                 resv->region_cache_count--;
674                         }
675
676                         if (!nrg) {
677                                 spin_unlock(&resv->lock);
678                                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
679                                 if (!nrg)
680                                         return -ENOMEM;
681                                 goto retry;
682                         }
683
684                         del += t - f;
685                         hugetlb_cgroup_uncharge_file_region(
686                                 resv, rg, t - f, false);
687
688                         /* New entry for end of split region */
689                         nrg->from = t;
690                         nrg->to = rg->to;
691
692                         copy_hugetlb_cgroup_uncharge_info(nrg, rg);
693
694                         INIT_LIST_HEAD(&nrg->link);
695
696                         /* Original entry is trimmed */
697                         rg->to = f;
698
699                         list_add(&nrg->link, &rg->link);
700                         nrg = NULL;
701                         break;
702                 }
703
704                 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
705                         del += rg->to - rg->from;
706                         hugetlb_cgroup_uncharge_file_region(resv, rg,
707                                                             rg->to - rg->from, true);
708                         list_del(&rg->link);
709                         kfree(rg);
710                         continue;
711                 }
712
713                 if (f <= rg->from) {    /* Trim beginning of region */
714                         hugetlb_cgroup_uncharge_file_region(resv, rg,
715                                                             t - rg->from, false);
716
717                         del += t - rg->from;
718                         rg->from = t;
719                 } else {                /* Trim end of region */
720                         hugetlb_cgroup_uncharge_file_region(resv, rg,
721                                                             rg->to - f, false);
722
723                         del += rg->to - f;
724                         rg->to = f;
725                 }
726         }
727
728         spin_unlock(&resv->lock);
729         kfree(nrg);
730         return del;
731 }
732
733 /*
734  * A rare out of memory error was encountered which prevented removal of
735  * the reserve map region for a page.  The huge page itself was free'ed
736  * and removed from the page cache.  This routine will adjust the subpool
737  * usage count, and the global reserve count if needed.  By incrementing
738  * these counts, the reserve map entry which could not be deleted will
739  * appear as a "reserved" entry instead of simply dangling with incorrect
740  * counts.
741  */
742 void hugetlb_fix_reserve_counts(struct inode *inode)
743 {
744         struct hugepage_subpool *spool = subpool_inode(inode);
745         long rsv_adjust;
746
747         rsv_adjust = hugepage_subpool_get_pages(spool, 1);
748         if (rsv_adjust) {
749                 struct hstate *h = hstate_inode(inode);
750
751                 hugetlb_acct_memory(h, 1);
752         }
753 }
754
755 /*
756  * Count and return the number of huge pages in the reserve map
757  * that intersect with the range [f, t).
758  */
759 static long region_count(struct resv_map *resv, long f, long t)
760 {
761         struct list_head *head = &resv->regions;
762         struct file_region *rg;
763         long chg = 0;
764
765         spin_lock(&resv->lock);
766         /* Locate each segment we overlap with, and count that overlap. */
767         list_for_each_entry(rg, head, link) {
768                 long seg_from;
769                 long seg_to;
770
771                 if (rg->to <= f)
772                         continue;
773                 if (rg->from >= t)
774                         break;
775
776                 seg_from = max(rg->from, f);
777                 seg_to = min(rg->to, t);
778
779                 chg += seg_to - seg_from;
780         }
781         spin_unlock(&resv->lock);
782
783         return chg;
784 }
785
786 /*
787  * Convert the address within this vma to the page offset within
788  * the mapping, in pagecache page units; huge pages here.
789  */
790 static pgoff_t vma_hugecache_offset(struct hstate *h,
791                         struct vm_area_struct *vma, unsigned long address)
792 {
793         return ((address - vma->vm_start) >> huge_page_shift(h)) +
794                         (vma->vm_pgoff >> huge_page_order(h));
795 }
796
797 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
798                                      unsigned long address)
799 {
800         return vma_hugecache_offset(hstate_vma(vma), vma, address);
801 }
802 EXPORT_SYMBOL_GPL(linear_hugepage_index);
803
804 /*
805  * Return the size of the pages allocated when backing a VMA. In the majority
806  * cases this will be same size as used by the page table entries.
807  */
808 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
809 {
810         if (vma->vm_ops && vma->vm_ops->pagesize)
811                 return vma->vm_ops->pagesize(vma);
812         return PAGE_SIZE;
813 }
814 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
815
816 /*
817  * Return the page size being used by the MMU to back a VMA. In the majority
818  * of cases, the page size used by the kernel matches the MMU size. On
819  * architectures where it differs, an architecture-specific 'strong'
820  * version of this symbol is required.
821  */
822 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
823 {
824         return vma_kernel_pagesize(vma);
825 }
826
827 /*
828  * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
829  * bits of the reservation map pointer, which are always clear due to
830  * alignment.
831  */
832 #define HPAGE_RESV_OWNER    (1UL << 0)
833 #define HPAGE_RESV_UNMAPPED (1UL << 1)
834 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
835
836 /*
837  * These helpers are used to track how many pages are reserved for
838  * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
839  * is guaranteed to have their future faults succeed.
840  *
841  * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
842  * the reserve counters are updated with the hugetlb_lock held. It is safe
843  * to reset the VMA at fork() time as it is not in use yet and there is no
844  * chance of the global counters getting corrupted as a result of the values.
845  *
846  * The private mapping reservation is represented in a subtly different
847  * manner to a shared mapping.  A shared mapping has a region map associated
848  * with the underlying file, this region map represents the backing file
849  * pages which have ever had a reservation assigned which this persists even
850  * after the page is instantiated.  A private mapping has a region map
851  * associated with the original mmap which is attached to all VMAs which
852  * reference it, this region map represents those offsets which have consumed
853  * reservation ie. where pages have been instantiated.
854  */
855 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
856 {
857         return (unsigned long)vma->vm_private_data;
858 }
859
860 static void set_vma_private_data(struct vm_area_struct *vma,
861                                                         unsigned long value)
862 {
863         vma->vm_private_data = (void *)value;
864 }
865
866 static void
867 resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,
868                                           struct hugetlb_cgroup *h_cg,
869                                           struct hstate *h)
870 {
871 #ifdef CONFIG_CGROUP_HUGETLB
872         if (!h_cg || !h) {
873                 resv_map->reservation_counter = NULL;
874                 resv_map->pages_per_hpage = 0;
875                 resv_map->css = NULL;
876         } else {
877                 resv_map->reservation_counter =
878                         &h_cg->rsvd_hugepage[hstate_index(h)];
879                 resv_map->pages_per_hpage = pages_per_huge_page(h);
880                 resv_map->css = &h_cg->css;
881         }
882 #endif
883 }
884
885 struct resv_map *resv_map_alloc(void)
886 {
887         struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
888         struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
889
890         if (!resv_map || !rg) {
891                 kfree(resv_map);
892                 kfree(rg);
893                 return NULL;
894         }
895
896         kref_init(&resv_map->refs);
897         spin_lock_init(&resv_map->lock);
898         INIT_LIST_HEAD(&resv_map->regions);
899
900         resv_map->adds_in_progress = 0;
901         /*
902          * Initialize these to 0. On shared mappings, 0's here indicate these
903          * fields don't do cgroup accounting. On private mappings, these will be
904          * re-initialized to the proper values, to indicate that hugetlb cgroup
905          * reservations are to be un-charged from here.
906          */
907         resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);
908
909         INIT_LIST_HEAD(&resv_map->region_cache);
910         list_add(&rg->link, &resv_map->region_cache);
911         resv_map->region_cache_count = 1;
912
913         return resv_map;
914 }
915
916 void resv_map_release(struct kref *ref)
917 {
918         struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
919         struct list_head *head = &resv_map->region_cache;
920         struct file_region *rg, *trg;
921
922         /* Clear out any active regions before we release the map. */
923         region_del(resv_map, 0, LONG_MAX);
924
925         /* ... and any entries left in the cache */
926         list_for_each_entry_safe(rg, trg, head, link) {
927                 list_del(&rg->link);
928                 kfree(rg);
929         }
930
931         VM_BUG_ON(resv_map->adds_in_progress);
932
933         kfree(resv_map);
934 }
935
936 static inline struct resv_map *inode_resv_map(struct inode *inode)
937 {
938         /*
939          * At inode evict time, i_mapping may not point to the original
940          * address space within the inode.  This original address space
941          * contains the pointer to the resv_map.  So, always use the
942          * address space embedded within the inode.
943          * The VERY common case is inode->mapping == &inode->i_data but,
944          * this may not be true for device special inodes.
945          */
946         return (struct resv_map *)(&inode->i_data)->private_data;
947 }
948
949 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
950 {
951         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
952         if (vma->vm_flags & VM_MAYSHARE) {
953                 struct address_space *mapping = vma->vm_file->f_mapping;
954                 struct inode *inode = mapping->host;
955
956                 return inode_resv_map(inode);
957
958         } else {
959                 return (struct resv_map *)(get_vma_private_data(vma) &
960                                                         ~HPAGE_RESV_MASK);
961         }
962 }
963
964 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
965 {
966         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
967         VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
968
969         set_vma_private_data(vma, (get_vma_private_data(vma) &
970                                 HPAGE_RESV_MASK) | (unsigned long)map);
971 }
972
973 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
974 {
975         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
976         VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
977
978         set_vma_private_data(vma, get_vma_private_data(vma) | flags);
979 }
980
981 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
982 {
983         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
984
985         return (get_vma_private_data(vma) & flag) != 0;
986 }
987
988 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
989 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
990 {
991         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
992         if (!(vma->vm_flags & VM_MAYSHARE))
993                 vma->vm_private_data = (void *)0;
994 }
995
996 /* Returns true if the VMA has associated reserve pages */
997 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
998 {
999         if (vma->vm_flags & VM_NORESERVE) {
1000                 /*
1001                  * This address is already reserved by other process(chg == 0),
1002                  * so, we should decrement reserved count. Without decrementing,
1003                  * reserve count remains after releasing inode, because this
1004                  * allocated page will go into page cache and is regarded as
1005                  * coming from reserved pool in releasing step.  Currently, we
1006                  * don't have any other solution to deal with this situation
1007                  * properly, so add work-around here.
1008                  */
1009                 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
1010                         return true;
1011                 else
1012                         return false;
1013         }
1014
1015         /* Shared mappings always use reserves */
1016         if (vma->vm_flags & VM_MAYSHARE) {
1017                 /*
1018                  * We know VM_NORESERVE is not set.  Therefore, there SHOULD
1019                  * be a region map for all pages.  The only situation where
1020                  * there is no region map is if a hole was punched via
1021                  * fallocate.  In this case, there really are no reserves to
1022                  * use.  This situation is indicated if chg != 0.
1023                  */
1024                 if (chg)
1025                         return false;
1026                 else
1027                         return true;
1028         }
1029
1030         /*
1031          * Only the process that called mmap() has reserves for
1032          * private mappings.
1033          */
1034         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1035                 /*
1036                  * Like the shared case above, a hole punch or truncate
1037                  * could have been performed on the private mapping.
1038                  * Examine the value of chg to determine if reserves
1039                  * actually exist or were previously consumed.
1040                  * Very Subtle - The value of chg comes from a previous
1041                  * call to vma_needs_reserves().  The reserve map for
1042                  * private mappings has different (opposite) semantics
1043                  * than that of shared mappings.  vma_needs_reserves()
1044                  * has already taken this difference in semantics into
1045                  * account.  Therefore, the meaning of chg is the same
1046                  * as in the shared case above.  Code could easily be
1047                  * combined, but keeping it separate draws attention to
1048                  * subtle differences.
1049                  */
1050                 if (chg)
1051                         return false;
1052                 else
1053                         return true;
1054         }
1055
1056         return false;
1057 }
1058
1059 static void enqueue_huge_page(struct hstate *h, struct page *page)
1060 {
1061         int nid = page_to_nid(page);
1062         list_move(&page->lru, &h->hugepage_freelists[nid]);
1063         h->free_huge_pages++;
1064         h->free_huge_pages_node[nid]++;
1065         SetHPageFreed(page);
1066 }
1067
1068 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
1069 {
1070         struct page *page;
1071         bool nocma = !!(current->flags & PF_MEMALLOC_NOCMA);
1072
1073         list_for_each_entry(page, &h->hugepage_freelists[nid], lru) {
1074                 if (nocma && is_migrate_cma_page(page))
1075                         continue;
1076
1077                 if (PageHWPoison(page))
1078                         continue;
1079
1080                 list_move(&page->lru, &h->hugepage_activelist);
1081                 set_page_refcounted(page);
1082                 ClearHPageFreed(page);
1083                 h->free_huge_pages--;
1084                 h->free_huge_pages_node[nid]--;
1085                 return page;
1086         }
1087
1088         return NULL;
1089 }
1090
1091 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
1092                 nodemask_t *nmask)
1093 {
1094         unsigned int cpuset_mems_cookie;
1095         struct zonelist *zonelist;
1096         struct zone *zone;
1097         struct zoneref *z;
1098         int node = NUMA_NO_NODE;
1099
1100         zonelist = node_zonelist(nid, gfp_mask);
1101
1102 retry_cpuset:
1103         cpuset_mems_cookie = read_mems_allowed_begin();
1104         for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1105                 struct page *page;
1106
1107                 if (!cpuset_zone_allowed(zone, gfp_mask))
1108                         continue;
1109                 /*
1110                  * no need to ask again on the same node. Pool is node rather than
1111                  * zone aware
1112                  */
1113                 if (zone_to_nid(zone) == node)
1114                         continue;
1115                 node = zone_to_nid(zone);
1116
1117                 page = dequeue_huge_page_node_exact(h, node);
1118                 if (page)
1119                         return page;
1120         }
1121         if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1122                 goto retry_cpuset;
1123
1124         return NULL;
1125 }
1126
1127 static struct page *dequeue_huge_page_vma(struct hstate *h,
1128                                 struct vm_area_struct *vma,
1129                                 unsigned long address, int avoid_reserve,
1130                                 long chg)
1131 {
1132         struct page *page;
1133         struct mempolicy *mpol;
1134         gfp_t gfp_mask;
1135         nodemask_t *nodemask;
1136         int nid;
1137
1138         /*
1139          * A child process with MAP_PRIVATE mappings created by their parent
1140          * have no page reserves. This check ensures that reservations are
1141          * not "stolen". The child may still get SIGKILLed
1142          */
1143         if (!vma_has_reserves(vma, chg) &&
1144                         h->free_huge_pages - h->resv_huge_pages == 0)
1145                 goto err;
1146
1147         /* If reserves cannot be used, ensure enough pages are in the pool */
1148         if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
1149                 goto err;
1150
1151         gfp_mask = htlb_alloc_mask(h);
1152         nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1153         page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1154         if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
1155                 SetHPageRestoreReserve(page);
1156                 h->resv_huge_pages--;
1157         }
1158
1159         mpol_cond_put(mpol);
1160         return page;
1161
1162 err:
1163         return NULL;
1164 }
1165
1166 /*
1167  * common helper functions for hstate_next_node_to_{alloc|free}.
1168  * We may have allocated or freed a huge page based on a different
1169  * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1170  * be outside of *nodes_allowed.  Ensure that we use an allowed
1171  * node for alloc or free.
1172  */
1173 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1174 {
1175         nid = next_node_in(nid, *nodes_allowed);
1176         VM_BUG_ON(nid >= MAX_NUMNODES);
1177
1178         return nid;
1179 }
1180
1181 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1182 {
1183         if (!node_isset(nid, *nodes_allowed))
1184                 nid = next_node_allowed(nid, nodes_allowed);
1185         return nid;
1186 }
1187
1188 /*
1189  * returns the previously saved node ["this node"] from which to
1190  * allocate a persistent huge page for the pool and advance the
1191  * next node from which to allocate, handling wrap at end of node
1192  * mask.
1193  */
1194 static int hstate_next_node_to_alloc(struct hstate *h,
1195                                         nodemask_t *nodes_allowed)
1196 {
1197         int nid;
1198
1199         VM_BUG_ON(!nodes_allowed);
1200
1201         nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1202         h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1203
1204         return nid;
1205 }
1206
1207 /*
1208  * helper for free_pool_huge_page() - return the previously saved
1209  * node ["this node"] from which to free a huge page.  Advance the
1210  * next node id whether or not we find a free huge page to free so
1211  * that the next attempt to free addresses the next node.
1212  */
1213 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1214 {
1215         int nid;
1216
1217         VM_BUG_ON(!nodes_allowed);
1218
1219         nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1220         h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1221
1222         return nid;
1223 }
1224
1225 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask)           \
1226         for (nr_nodes = nodes_weight(*mask);                            \
1227                 nr_nodes > 0 &&                                         \
1228                 ((node = hstate_next_node_to_alloc(hs, mask)) || 1);    \
1229                 nr_nodes--)
1230
1231 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask)            \
1232         for (nr_nodes = nodes_weight(*mask);                            \
1233                 nr_nodes > 0 &&                                         \
1234                 ((node = hstate_next_node_to_free(hs, mask)) || 1);     \
1235                 nr_nodes--)
1236
1237 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1238 static void destroy_compound_gigantic_page(struct page *page,
1239                                         unsigned int order)
1240 {
1241         int i;
1242         int nr_pages = 1 << order;
1243         struct page *p = page + 1;
1244
1245         atomic_set(compound_mapcount_ptr(page), 0);
1246         atomic_set(compound_pincount_ptr(page), 0);
1247
1248         for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1249                 clear_compound_head(p);
1250                 set_page_refcounted(p);
1251         }
1252
1253         set_compound_order(page, 0);
1254         page[1].compound_nr = 0;
1255         __ClearPageHead(page);
1256 }
1257
1258 static void free_gigantic_page(struct page *page, unsigned int order)
1259 {
1260         /*
1261          * If the page isn't allocated using the cma allocator,
1262          * cma_release() returns false.
1263          */
1264 #ifdef CONFIG_CMA
1265         if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order))
1266                 return;
1267 #endif
1268
1269         free_contig_range(page_to_pfn(page), 1 << order);
1270 }
1271
1272 #ifdef CONFIG_CONTIG_ALLOC
1273 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1274                 int nid, nodemask_t *nodemask)
1275 {
1276         unsigned long nr_pages = 1UL << huge_page_order(h);
1277         if (nid == NUMA_NO_NODE)
1278                 nid = numa_mem_id();
1279
1280 #ifdef CONFIG_CMA
1281         {
1282                 struct page *page;
1283                 int node;
1284
1285                 if (hugetlb_cma[nid]) {
1286                         page = cma_alloc(hugetlb_cma[nid], nr_pages,
1287                                         huge_page_order(h), true);
1288                         if (page)
1289                                 return page;
1290                 }
1291
1292                 if (!(gfp_mask & __GFP_THISNODE)) {
1293                         for_each_node_mask(node, *nodemask) {
1294                                 if (node == nid || !hugetlb_cma[node])
1295                                         continue;
1296
1297                                 page = cma_alloc(hugetlb_cma[node], nr_pages,
1298                                                 huge_page_order(h), true);
1299                                 if (page)
1300                                         return page;
1301                         }
1302                 }
1303         }
1304 #endif
1305
1306         return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1307 }
1308
1309 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1310 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1311 #else /* !CONFIG_CONTIG_ALLOC */
1312 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1313                                         int nid, nodemask_t *nodemask)
1314 {
1315         return NULL;
1316 }
1317 #endif /* CONFIG_CONTIG_ALLOC */
1318
1319 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1320 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1321                                         int nid, nodemask_t *nodemask)
1322 {
1323         return NULL;
1324 }
1325 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1326 static inline void destroy_compound_gigantic_page(struct page *page,
1327                                                 unsigned int order) { }
1328 #endif
1329
1330 static void update_and_free_page(struct hstate *h, struct page *page)
1331 {
1332         int i;
1333         struct page *subpage = page;
1334
1335         if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1336                 return;
1337
1338         h->nr_huge_pages--;
1339         h->nr_huge_pages_node[page_to_nid(page)]--;
1340         for (i = 0; i < pages_per_huge_page(h);
1341              i++, subpage = mem_map_next(subpage, page, i)) {
1342                 subpage->flags &= ~(1 << PG_locked | 1 << PG_error |
1343                                 1 << PG_referenced | 1 << PG_dirty |
1344                                 1 << PG_active | 1 << PG_private |
1345                                 1 << PG_writeback);
1346         }
1347         VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1348         VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page);
1349         set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1350         set_page_refcounted(page);
1351         if (hstate_is_gigantic(h)) {
1352                 /*
1353                  * Temporarily drop the hugetlb_lock, because
1354                  * we might block in free_gigantic_page().
1355                  */
1356                 spin_unlock(&hugetlb_lock);
1357                 destroy_compound_gigantic_page(page, huge_page_order(h));
1358                 free_gigantic_page(page, huge_page_order(h));
1359                 spin_lock(&hugetlb_lock);
1360         } else {
1361                 __free_pages(page, huge_page_order(h));
1362         }
1363 }
1364
1365 struct hstate *size_to_hstate(unsigned long size)
1366 {
1367         struct hstate *h;
1368
1369         for_each_hstate(h) {
1370                 if (huge_page_size(h) == size)
1371                         return h;
1372         }
1373         return NULL;
1374 }
1375
1376 static void __free_huge_page(struct page *page)
1377 {
1378         /*
1379          * Can't pass hstate in here because it is called from the
1380          * compound page destructor.
1381          */
1382         struct hstate *h = page_hstate(page);
1383         int nid = page_to_nid(page);
1384         struct hugepage_subpool *spool = hugetlb_page_subpool(page);
1385         bool restore_reserve;
1386
1387         VM_BUG_ON_PAGE(page_count(page), page);
1388         VM_BUG_ON_PAGE(page_mapcount(page), page);
1389
1390         hugetlb_set_page_subpool(page, NULL);
1391         page->mapping = NULL;
1392         restore_reserve = HPageRestoreReserve(page);
1393         ClearHPageRestoreReserve(page);
1394
1395         /*
1396          * If HPageRestoreReserve was set on page, page allocation consumed a
1397          * reservation.  If the page was associated with a subpool, there
1398          * would have been a page reserved in the subpool before allocation
1399          * via hugepage_subpool_get_pages().  Since we are 'restoring' the
1400          * reservation, do not call hugepage_subpool_put_pages() as this will
1401          * remove the reserved page from the subpool.
1402          */
1403         if (!restore_reserve) {
1404                 /*
1405                  * A return code of zero implies that the subpool will be
1406                  * under its minimum size if the reservation is not restored
1407                  * after page is free.  Therefore, force restore_reserve
1408                  * operation.
1409                  */
1410                 if (hugepage_subpool_put_pages(spool, 1) == 0)
1411                         restore_reserve = true;
1412         }
1413
1414         spin_lock(&hugetlb_lock);
1415         ClearHPageMigratable(page);
1416         hugetlb_cgroup_uncharge_page(hstate_index(h),
1417                                      pages_per_huge_page(h), page);
1418         hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
1419                                           pages_per_huge_page(h), page);
1420         if (restore_reserve)
1421                 h->resv_huge_pages++;
1422
1423         if (HPageTemporary(page)) {
1424                 list_del(&page->lru);
1425                 ClearHPageTemporary(page);
1426                 update_and_free_page(h, page);
1427         } else if (h->surplus_huge_pages_node[nid]) {
1428                 /* remove the page from active list */
1429                 list_del(&page->lru);
1430                 update_and_free_page(h, page);
1431                 h->surplus_huge_pages--;
1432                 h->surplus_huge_pages_node[nid]--;
1433         } else {
1434                 arch_clear_hugepage_flags(page);
1435                 enqueue_huge_page(h, page);
1436         }
1437         spin_unlock(&hugetlb_lock);
1438 }
1439
1440 /*
1441  * As free_huge_page() can be called from a non-task context, we have
1442  * to defer the actual freeing in a workqueue to prevent potential
1443  * hugetlb_lock deadlock.
1444  *
1445  * free_hpage_workfn() locklessly retrieves the linked list of pages to
1446  * be freed and frees them one-by-one. As the page->mapping pointer is
1447  * going to be cleared in __free_huge_page() anyway, it is reused as the
1448  * llist_node structure of a lockless linked list of huge pages to be freed.
1449  */
1450 static LLIST_HEAD(hpage_freelist);
1451
1452 static void free_hpage_workfn(struct work_struct *work)
1453 {
1454         struct llist_node *node;
1455         struct page *page;
1456
1457         node = llist_del_all(&hpage_freelist);
1458
1459         while (node) {
1460                 page = container_of((struct address_space **)node,
1461                                      struct page, mapping);
1462                 node = node->next;
1463                 __free_huge_page(page);
1464         }
1465 }
1466 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1467
1468 void free_huge_page(struct page *page)
1469 {
1470         /*
1471          * Defer freeing if in non-task context to avoid hugetlb_lock deadlock.
1472          */
1473         if (!in_task()) {
1474                 /*
1475                  * Only call schedule_work() if hpage_freelist is previously
1476                  * empty. Otherwise, schedule_work() had been called but the
1477                  * workfn hasn't retrieved the list yet.
1478                  */
1479                 if (llist_add((struct llist_node *)&page->mapping,
1480                               &hpage_freelist))
1481                         schedule_work(&free_hpage_work);
1482                 return;
1483         }
1484
1485         __free_huge_page(page);
1486 }
1487
1488 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1489 {
1490         INIT_LIST_HEAD(&page->lru);
1491         set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1492         hugetlb_set_page_subpool(page, NULL);
1493         set_hugetlb_cgroup(page, NULL);
1494         set_hugetlb_cgroup_rsvd(page, NULL);
1495         spin_lock(&hugetlb_lock);
1496         h->nr_huge_pages++;
1497         h->nr_huge_pages_node[nid]++;
1498         ClearHPageFreed(page);
1499         spin_unlock(&hugetlb_lock);
1500 }
1501
1502 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1503 {
1504         int i;
1505         int nr_pages = 1 << order;
1506         struct page *p = page + 1;
1507
1508         /* we rely on prep_new_huge_page to set the destructor */
1509         set_compound_order(page, order);
1510         __ClearPageReserved(page);
1511         __SetPageHead(page);
1512         for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1513                 /*
1514                  * For gigantic hugepages allocated through bootmem at
1515                  * boot, it's safer to be consistent with the not-gigantic
1516                  * hugepages and clear the PG_reserved bit from all tail pages
1517                  * too.  Otherwise drivers using get_user_pages() to access tail
1518                  * pages may get the reference counting wrong if they see
1519                  * PG_reserved set on a tail page (despite the head page not
1520                  * having PG_reserved set).  Enforcing this consistency between
1521                  * head and tail pages allows drivers to optimize away a check
1522                  * on the head page when they need know if put_page() is needed
1523                  * after get_user_pages().
1524                  */
1525                 __ClearPageReserved(p);
1526                 set_page_count(p, 0);
1527                 set_compound_head(p, page);
1528         }
1529         atomic_set(compound_mapcount_ptr(page), -1);
1530         atomic_set(compound_pincount_ptr(page), 0);
1531 }
1532
1533 /*
1534  * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1535  * transparent huge pages.  See the PageTransHuge() documentation for more
1536  * details.
1537  */
1538 int PageHuge(struct page *page)
1539 {
1540         if (!PageCompound(page))
1541                 return 0;
1542
1543         page = compound_head(page);
1544         return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1545 }
1546 EXPORT_SYMBOL_GPL(PageHuge);
1547
1548 /*
1549  * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1550  * normal or transparent huge pages.
1551  */
1552 int PageHeadHuge(struct page *page_head)
1553 {
1554         if (!PageHead(page_head))
1555                 return 0;
1556
1557         return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR;
1558 }
1559
1560 /*
1561  * Find and lock address space (mapping) in write mode.
1562  *
1563  * Upon entry, the page is locked which means that page_mapping() is
1564  * stable.  Due to locking order, we can only trylock_write.  If we can
1565  * not get the lock, simply return NULL to caller.
1566  */
1567 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
1568 {
1569         struct address_space *mapping = page_mapping(hpage);
1570
1571         if (!mapping)
1572                 return mapping;
1573
1574         if (i_mmap_trylock_write(mapping))
1575                 return mapping;
1576
1577         return NULL;
1578 }
1579
1580 pgoff_t __basepage_index(struct page *page)
1581 {
1582         struct page *page_head = compound_head(page);
1583         pgoff_t index = page_index(page_head);
1584         unsigned long compound_idx;
1585
1586         if (!PageHuge(page_head))
1587                 return page_index(page);
1588
1589         if (compound_order(page_head) >= MAX_ORDER)
1590                 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1591         else
1592                 compound_idx = page - page_head;
1593
1594         return (index << compound_order(page_head)) + compound_idx;
1595 }
1596
1597 static struct page *alloc_buddy_huge_page(struct hstate *h,
1598                 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1599                 nodemask_t *node_alloc_noretry)
1600 {
1601         int order = huge_page_order(h);
1602         struct page *page;
1603         bool alloc_try_hard = true;
1604
1605         /*
1606          * By default we always try hard to allocate the page with
1607          * __GFP_RETRY_MAYFAIL flag.  However, if we are allocating pages in
1608          * a loop (to adjust global huge page counts) and previous allocation
1609          * failed, do not continue to try hard on the same node.  Use the
1610          * node_alloc_noretry bitmap to manage this state information.
1611          */
1612         if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1613                 alloc_try_hard = false;
1614         gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1615         if (alloc_try_hard)
1616                 gfp_mask |= __GFP_RETRY_MAYFAIL;
1617         if (nid == NUMA_NO_NODE)
1618                 nid = numa_mem_id();
1619         page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
1620         if (page)
1621                 __count_vm_event(HTLB_BUDDY_PGALLOC);
1622         else
1623                 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1624
1625         /*
1626          * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1627          * indicates an overall state change.  Clear bit so that we resume
1628          * normal 'try hard' allocations.
1629          */
1630         if (node_alloc_noretry && page && !alloc_try_hard)
1631                 node_clear(nid, *node_alloc_noretry);
1632
1633         /*
1634          * If we tried hard to get a page but failed, set bit so that
1635          * subsequent attempts will not try as hard until there is an
1636          * overall state change.
1637          */
1638         if (node_alloc_noretry && !page && alloc_try_hard)
1639                 node_set(nid, *node_alloc_noretry);
1640
1641         return page;
1642 }
1643
1644 /*
1645  * Common helper to allocate a fresh hugetlb page. All specific allocators
1646  * should use this function to get new hugetlb pages
1647  */
1648 static struct page *alloc_fresh_huge_page(struct hstate *h,
1649                 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1650                 nodemask_t *node_alloc_noretry)
1651 {
1652         struct page *page;
1653
1654         if (hstate_is_gigantic(h))
1655                 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1656         else
1657                 page = alloc_buddy_huge_page(h, gfp_mask,
1658                                 nid, nmask, node_alloc_noretry);
1659         if (!page)
1660                 return NULL;
1661
1662         if (hstate_is_gigantic(h))
1663                 prep_compound_gigantic_page(page, huge_page_order(h));
1664         prep_new_huge_page(h, page, page_to_nid(page));
1665
1666         return page;
1667 }
1668
1669 /*
1670  * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1671  * manner.
1672  */
1673 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1674                                 nodemask_t *node_alloc_noretry)
1675 {
1676         struct page *page;
1677         int nr_nodes, node;
1678         gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1679
1680         for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1681                 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
1682                                                 node_alloc_noretry);
1683                 if (page)
1684                         break;
1685         }
1686
1687         if (!page)
1688                 return 0;
1689
1690         put_page(page); /* free it into the hugepage allocator */
1691
1692         return 1;
1693 }
1694
1695 /*
1696  * Free huge page from pool from next node to free.
1697  * Attempt to keep persistent huge pages more or less
1698  * balanced over allowed nodes.
1699  * Called with hugetlb_lock locked.
1700  */
1701 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1702                                                          bool acct_surplus)
1703 {
1704         int nr_nodes, node;
1705         int ret = 0;
1706
1707         for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1708                 /*
1709                  * If we're returning unused surplus pages, only examine
1710                  * nodes with surplus pages.
1711                  */
1712                 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1713                     !list_empty(&h->hugepage_freelists[node])) {
1714                         struct page *page =
1715                                 list_entry(h->hugepage_freelists[node].next,
1716                                           struct page, lru);
1717                         list_del(&page->lru);
1718                         h->free_huge_pages--;
1719                         h->free_huge_pages_node[node]--;
1720                         if (acct_surplus) {
1721                                 h->surplus_huge_pages--;
1722                                 h->surplus_huge_pages_node[node]--;
1723                         }
1724                         update_and_free_page(h, page);
1725                         ret = 1;
1726                         break;
1727                 }
1728         }
1729
1730         return ret;
1731 }
1732
1733 /*
1734  * Dissolve a given free hugepage into free buddy pages. This function does
1735  * nothing for in-use hugepages and non-hugepages.
1736  * This function returns values like below:
1737  *
1738  *  -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1739  *          (allocated or reserved.)
1740  *       0: successfully dissolved free hugepages or the page is not a
1741  *          hugepage (considered as already dissolved)
1742  */
1743 int dissolve_free_huge_page(struct page *page)
1744 {
1745         int rc = -EBUSY;
1746
1747 retry:
1748         /* Not to disrupt normal path by vainly holding hugetlb_lock */
1749         if (!PageHuge(page))
1750                 return 0;
1751
1752         spin_lock(&hugetlb_lock);
1753         if (!PageHuge(page)) {
1754                 rc = 0;
1755                 goto out;
1756         }
1757
1758         if (!page_count(page)) {
1759                 struct page *head = compound_head(page);
1760                 struct hstate *h = page_hstate(head);
1761                 int nid = page_to_nid(head);
1762                 if (h->free_huge_pages - h->resv_huge_pages == 0)
1763                         goto out;
1764
1765                 /*
1766                  * We should make sure that the page is already on the free list
1767                  * when it is dissolved.
1768                  */
1769                 if (unlikely(!HPageFreed(head))) {
1770                         spin_unlock(&hugetlb_lock);
1771                         cond_resched();
1772
1773                         /*
1774                          * Theoretically, we should return -EBUSY when we
1775                          * encounter this race. In fact, we have a chance
1776                          * to successfully dissolve the page if we do a
1777                          * retry. Because the race window is quite small.
1778                          * If we seize this opportunity, it is an optimization
1779                          * for increasing the success rate of dissolving page.
1780                          */
1781                         goto retry;
1782                 }
1783
1784                 /*
1785                  * Move PageHWPoison flag from head page to the raw error page,
1786                  * which makes any subpages rather than the error page reusable.
1787                  */
1788                 if (PageHWPoison(head) && page != head) {
1789                         SetPageHWPoison(page);
1790                         ClearPageHWPoison(head);
1791                 }
1792                 list_del(&head->lru);
1793                 h->free_huge_pages--;
1794                 h->free_huge_pages_node[nid]--;
1795                 h->max_huge_pages--;
1796                 update_and_free_page(h, head);
1797                 rc = 0;
1798         }
1799 out:
1800         spin_unlock(&hugetlb_lock);
1801         return rc;
1802 }
1803
1804 /*
1805  * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1806  * make specified memory blocks removable from the system.
1807  * Note that this will dissolve a free gigantic hugepage completely, if any
1808  * part of it lies within the given range.
1809  * Also note that if dissolve_free_huge_page() returns with an error, all
1810  * free hugepages that were dissolved before that error are lost.
1811  */
1812 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1813 {
1814         unsigned long pfn;
1815         struct page *page;
1816         int rc = 0;
1817
1818         if (!hugepages_supported())
1819                 return rc;
1820
1821         for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1822                 page = pfn_to_page(pfn);
1823                 rc = dissolve_free_huge_page(page);
1824                 if (rc)
1825                         break;
1826         }
1827
1828         return rc;
1829 }
1830
1831 /*
1832  * Allocates a fresh surplus page from the page allocator.
1833  */
1834 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
1835                 int nid, nodemask_t *nmask)
1836 {
1837         struct page *page = NULL;
1838
1839         if (hstate_is_gigantic(h))
1840                 return NULL;
1841
1842         spin_lock(&hugetlb_lock);
1843         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1844                 goto out_unlock;
1845         spin_unlock(&hugetlb_lock);
1846
1847         page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1848         if (!page)
1849                 return NULL;
1850
1851         spin_lock(&hugetlb_lock);
1852         /*
1853          * We could have raced with the pool size change.
1854          * Double check that and simply deallocate the new page
1855          * if we would end up overcommiting the surpluses. Abuse
1856          * temporary page to workaround the nasty free_huge_page
1857          * codeflow
1858          */
1859         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1860                 SetHPageTemporary(page);
1861                 spin_unlock(&hugetlb_lock);
1862                 put_page(page);
1863                 return NULL;
1864         } else {
1865                 h->surplus_huge_pages++;
1866                 h->surplus_huge_pages_node[page_to_nid(page)]++;
1867         }
1868
1869 out_unlock:
1870         spin_unlock(&hugetlb_lock);
1871
1872         return page;
1873 }
1874
1875 static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
1876                                      int nid, nodemask_t *nmask)
1877 {
1878         struct page *page;
1879
1880         if (hstate_is_gigantic(h))
1881                 return NULL;
1882
1883         page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1884         if (!page)
1885                 return NULL;
1886
1887         /*
1888          * We do not account these pages as surplus because they are only
1889          * temporary and will be released properly on the last reference
1890          */
1891         SetHPageTemporary(page);
1892
1893         return page;
1894 }
1895
1896 /*
1897  * Use the VMA's mpolicy to allocate a huge page from the buddy.
1898  */
1899 static
1900 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
1901                 struct vm_area_struct *vma, unsigned long addr)
1902 {
1903         struct page *page;
1904         struct mempolicy *mpol;
1905         gfp_t gfp_mask = htlb_alloc_mask(h);
1906         int nid;
1907         nodemask_t *nodemask;
1908
1909         nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1910         page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
1911         mpol_cond_put(mpol);
1912
1913         return page;
1914 }
1915
1916 /* page migration callback function */
1917 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
1918                 nodemask_t *nmask, gfp_t gfp_mask)
1919 {
1920         spin_lock(&hugetlb_lock);
1921         if (h->free_huge_pages - h->resv_huge_pages > 0) {
1922                 struct page *page;
1923
1924                 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1925                 if (page) {
1926                         spin_unlock(&hugetlb_lock);
1927                         return page;
1928                 }
1929         }
1930         spin_unlock(&hugetlb_lock);
1931
1932         return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
1933 }
1934
1935 /* mempolicy aware migration callback */
1936 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
1937                 unsigned long address)
1938 {
1939         struct mempolicy *mpol;
1940         nodemask_t *nodemask;
1941         struct page *page;
1942         gfp_t gfp_mask;
1943         int node;
1944
1945         gfp_mask = htlb_alloc_mask(h);
1946         node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1947         page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask);
1948         mpol_cond_put(mpol);
1949
1950         return page;
1951 }
1952
1953 /*
1954  * Increase the hugetlb pool such that it can accommodate a reservation
1955  * of size 'delta'.
1956  */
1957 static int gather_surplus_pages(struct hstate *h, long delta)
1958         __must_hold(&hugetlb_lock)
1959 {
1960         struct list_head surplus_list;
1961         struct page *page, *tmp;
1962         int ret;
1963         long i;
1964         long needed, allocated;
1965         bool alloc_ok = true;
1966
1967         needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1968         if (needed <= 0) {
1969                 h->resv_huge_pages += delta;
1970                 return 0;
1971         }
1972
1973         allocated = 0;
1974         INIT_LIST_HEAD(&surplus_list);
1975
1976         ret = -ENOMEM;
1977 retry:
1978         spin_unlock(&hugetlb_lock);
1979         for (i = 0; i < needed; i++) {
1980                 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
1981                                 NUMA_NO_NODE, NULL);
1982                 if (!page) {
1983                         alloc_ok = false;
1984                         break;
1985                 }
1986                 list_add(&page->lru, &surplus_list);
1987                 cond_resched();
1988         }
1989         allocated += i;
1990
1991         /*
1992          * After retaking hugetlb_lock, we need to recalculate 'needed'
1993          * because either resv_huge_pages or free_huge_pages may have changed.
1994          */
1995         spin_lock(&hugetlb_lock);
1996         needed = (h->resv_huge_pages + delta) -
1997                         (h->free_huge_pages + allocated);
1998         if (needed > 0) {
1999                 if (alloc_ok)
2000                         goto retry;
2001                 /*
2002                  * We were not able to allocate enough pages to
2003                  * satisfy the entire reservation so we free what
2004                  * we've allocated so far.
2005                  */
2006                 goto free;
2007         }
2008         /*
2009          * The surplus_list now contains _at_least_ the number of extra pages
2010          * needed to accommodate the reservation.  Add the appropriate number
2011          * of pages to the hugetlb pool and free the extras back to the buddy
2012          * allocator.  Commit the entire reservation here to prevent another
2013          * process from stealing the pages as they are added to the pool but
2014          * before they are reserved.
2015          */
2016         needed += allocated;
2017         h->resv_huge_pages += delta;
2018         ret = 0;
2019
2020         /* Free the needed pages to the hugetlb pool */
2021         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
2022                 int zeroed;
2023
2024                 if ((--needed) < 0)
2025                         break;
2026                 /*
2027                  * This page is now managed by the hugetlb allocator and has
2028                  * no users -- drop the buddy allocator's reference.
2029                  */
2030                 zeroed = put_page_testzero(page);
2031                 VM_BUG_ON_PAGE(!zeroed, page);
2032                 enqueue_huge_page(h, page);
2033         }
2034 free:
2035         spin_unlock(&hugetlb_lock);
2036
2037         /* Free unnecessary surplus pages to the buddy allocator */
2038         list_for_each_entry_safe(page, tmp, &surplus_list, lru)
2039                 put_page(page);
2040         spin_lock(&hugetlb_lock);
2041
2042         return ret;
2043 }
2044
2045 /*
2046  * This routine has two main purposes:
2047  * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2048  *    in unused_resv_pages.  This corresponds to the prior adjustments made
2049  *    to the associated reservation map.
2050  * 2) Free any unused surplus pages that may have been allocated to satisfy
2051  *    the reservation.  As many as unused_resv_pages may be freed.
2052  *
2053  * Called with hugetlb_lock held.  However, the lock could be dropped (and
2054  * reacquired) during calls to cond_resched_lock.  Whenever dropping the lock,
2055  * we must make sure nobody else can claim pages we are in the process of
2056  * freeing.  Do this by ensuring resv_huge_page always is greater than the
2057  * number of huge pages we plan to free when dropping the lock.
2058  */
2059 static void return_unused_surplus_pages(struct hstate *h,
2060                                         unsigned long unused_resv_pages)
2061 {
2062         unsigned long nr_pages;
2063
2064         /* Cannot return gigantic pages currently */
2065         if (hstate_is_gigantic(h))
2066                 goto out;
2067
2068         /*
2069          * Part (or even all) of the reservation could have been backed
2070          * by pre-allocated pages. Only free surplus pages.
2071          */
2072         nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2073
2074         /*
2075          * We want to release as many surplus pages as possible, spread
2076          * evenly across all nodes with memory. Iterate across these nodes
2077          * until we can no longer free unreserved surplus pages. This occurs
2078          * when the nodes with surplus pages have no free pages.
2079          * free_pool_huge_page() will balance the freed pages across the
2080          * on-line nodes with memory and will handle the hstate accounting.
2081          *
2082          * Note that we decrement resv_huge_pages as we free the pages.  If
2083          * we drop the lock, resv_huge_pages will still be sufficiently large
2084          * to cover subsequent pages we may free.
2085          */
2086         while (nr_pages--) {
2087                 h->resv_huge_pages--;
2088                 unused_resv_pages--;
2089                 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
2090                         goto out;
2091                 cond_resched_lock(&hugetlb_lock);
2092         }
2093
2094 out:
2095         /* Fully uncommit the reservation */
2096         h->resv_huge_pages -= unused_resv_pages;
2097 }
2098
2099
2100 /*
2101  * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2102  * are used by the huge page allocation routines to manage reservations.
2103  *
2104  * vma_needs_reservation is called to determine if the huge page at addr
2105  * within the vma has an associated reservation.  If a reservation is
2106  * needed, the value 1 is returned.  The caller is then responsible for
2107  * managing the global reservation and subpool usage counts.  After
2108  * the huge page has been allocated, vma_commit_reservation is called
2109  * to add the page to the reservation map.  If the page allocation fails,
2110  * the reservation must be ended instead of committed.  vma_end_reservation
2111  * is called in such cases.
2112  *
2113  * In the normal case, vma_commit_reservation returns the same value
2114  * as the preceding vma_needs_reservation call.  The only time this
2115  * is not the case is if a reserve map was changed between calls.  It
2116  * is the responsibility of the caller to notice the difference and
2117  * take appropriate action.
2118  *
2119  * vma_add_reservation is used in error paths where a reservation must
2120  * be restored when a newly allocated huge page must be freed.  It is
2121  * to be called after calling vma_needs_reservation to determine if a
2122  * reservation exists.
2123  */
2124 enum vma_resv_mode {
2125         VMA_NEEDS_RESV,
2126         VMA_COMMIT_RESV,
2127         VMA_END_RESV,
2128         VMA_ADD_RESV,
2129 };
2130 static long __vma_reservation_common(struct hstate *h,
2131                                 struct vm_area_struct *vma, unsigned long addr,
2132                                 enum vma_resv_mode mode)
2133 {
2134         struct resv_map *resv;
2135         pgoff_t idx;
2136         long ret;
2137         long dummy_out_regions_needed;
2138
2139         resv = vma_resv_map(vma);
2140         if (!resv)
2141                 return 1;
2142
2143         idx = vma_hugecache_offset(h, vma, addr);
2144         switch (mode) {
2145         case VMA_NEEDS_RESV:
2146                 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2147                 /* We assume that vma_reservation_* routines always operate on
2148                  * 1 page, and that adding to resv map a 1 page entry can only
2149                  * ever require 1 region.
2150                  */
2151                 VM_BUG_ON(dummy_out_regions_needed != 1);
2152                 break;
2153         case VMA_COMMIT_RESV:
2154                 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2155                 /* region_add calls of range 1 should never fail. */
2156                 VM_BUG_ON(ret < 0);
2157                 break;
2158         case VMA_END_RESV:
2159                 region_abort(resv, idx, idx + 1, 1);
2160                 ret = 0;
2161                 break;
2162         case VMA_ADD_RESV:
2163                 if (vma->vm_flags & VM_MAYSHARE) {
2164                         ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2165                         /* region_add calls of range 1 should never fail. */
2166                         VM_BUG_ON(ret < 0);
2167                 } else {
2168                         region_abort(resv, idx, idx + 1, 1);
2169                         ret = region_del(resv, idx, idx + 1);
2170                 }
2171                 break;
2172         default:
2173                 BUG();
2174         }
2175
2176         if (vma->vm_flags & VM_MAYSHARE)
2177                 return ret;
2178         else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
2179                 /*
2180                  * In most cases, reserves always exist for private mappings.
2181                  * However, a file associated with mapping could have been
2182                  * hole punched or truncated after reserves were consumed.
2183                  * As subsequent fault on such a range will not use reserves.
2184                  * Subtle - The reserve map for private mappings has the
2185                  * opposite meaning than that of shared mappings.  If NO
2186                  * entry is in the reserve map, it means a reservation exists.
2187                  * If an entry exists in the reserve map, it means the
2188                  * reservation has already been consumed.  As a result, the
2189                  * return value of this routine is the opposite of the
2190                  * value returned from reserve map manipulation routines above.
2191                  */
2192                 if (ret)
2193                         return 0;
2194                 else
2195                         return 1;
2196         }
2197         else
2198                 return ret < 0 ? ret : 0;
2199 }
2200
2201 static long vma_needs_reservation(struct hstate *h,
2202                         struct vm_area_struct *vma, unsigned long addr)
2203 {
2204         return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2205 }
2206
2207 static long vma_commit_reservation(struct hstate *h,
2208                         struct vm_area_struct *vma, unsigned long addr)
2209 {
2210         return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2211 }
2212
2213 static void vma_end_reservation(struct hstate *h,
2214                         struct vm_area_struct *vma, unsigned long addr)
2215 {
2216         (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2217 }
2218
2219 static long vma_add_reservation(struct hstate *h,
2220                         struct vm_area_struct *vma, unsigned long addr)
2221 {
2222         return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2223 }
2224
2225 /*
2226  * This routine is called to restore a reservation on error paths.  In the
2227  * specific error paths, a huge page was allocated (via alloc_huge_page)
2228  * and is about to be freed.  If a reservation for the page existed,
2229  * alloc_huge_page would have consumed the reservation and set
2230  * HPageRestoreReserve in the newly allocated page.  When the page is freed
2231  * via free_huge_page, the global reservation count will be incremented if
2232  * HPageRestoreReserve is set.  However, free_huge_page can not adjust the
2233  * reserve map.  Adjust the reserve map here to be consistent with global
2234  * reserve count adjustments to be made by free_huge_page.
2235  */
2236 static void restore_reserve_on_error(struct hstate *h,
2237                         struct vm_area_struct *vma, unsigned long address,
2238                         struct page *page)
2239 {
2240         if (unlikely(HPageRestoreReserve(page))) {
2241                 long rc = vma_needs_reservation(h, vma, address);
2242
2243                 if (unlikely(rc < 0)) {
2244                         /*
2245                          * Rare out of memory condition in reserve map
2246                          * manipulation.  Clear HPageRestoreReserve so that
2247                          * global reserve count will not be incremented
2248                          * by free_huge_page.  This will make it appear
2249                          * as though the reservation for this page was
2250                          * consumed.  This may prevent the task from
2251                          * faulting in the page at a later time.  This
2252                          * is better than inconsistent global huge page
2253                          * accounting of reserve counts.
2254                          */
2255                         ClearHPageRestoreReserve(page);
2256                 } else if (rc) {
2257                         rc = vma_add_reservation(h, vma, address);
2258                         if (unlikely(rc < 0))
2259                                 /*
2260                                  * See above comment about rare out of
2261                                  * memory condition.
2262                                  */
2263                                 ClearHPageRestoreReserve(page);
2264                 } else
2265                         vma_end_reservation(h, vma, address);
2266         }
2267 }
2268
2269 struct page *alloc_huge_page(struct vm_area_struct *vma,
2270                                     unsigned long addr, int avoid_reserve)
2271 {
2272         struct hugepage_subpool *spool = subpool_vma(vma);
2273         struct hstate *h = hstate_vma(vma);
2274         struct page *page;
2275         long map_chg, map_commit;
2276         long gbl_chg;
2277         int ret, idx;
2278         struct hugetlb_cgroup *h_cg;
2279         bool deferred_reserve;
2280
2281         idx = hstate_index(h);
2282         /*
2283          * Examine the region/reserve map to determine if the process
2284          * has a reservation for the page to be allocated.  A return
2285          * code of zero indicates a reservation exists (no change).
2286          */
2287         map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2288         if (map_chg < 0)
2289                 return ERR_PTR(-ENOMEM);
2290
2291         /*
2292          * Processes that did not create the mapping will have no
2293          * reserves as indicated by the region/reserve map. Check
2294          * that the allocation will not exceed the subpool limit.
2295          * Allocations for MAP_NORESERVE mappings also need to be
2296          * checked against any subpool limit.
2297          */
2298         if (map_chg || avoid_reserve) {
2299                 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2300                 if (gbl_chg < 0) {
2301                         vma_end_reservation(h, vma, addr);
2302                         return ERR_PTR(-ENOSPC);
2303                 }
2304
2305                 /*
2306                  * Even though there was no reservation in the region/reserve
2307                  * map, there could be reservations associated with the
2308                  * subpool that can be used.  This would be indicated if the
2309                  * return value of hugepage_subpool_get_pages() is zero.
2310                  * However, if avoid_reserve is specified we still avoid even
2311                  * the subpool reservations.
2312                  */
2313                 if (avoid_reserve)
2314                         gbl_chg = 1;
2315         }
2316
2317         /* If this allocation is not consuming a reservation, charge it now.
2318          */
2319         deferred_reserve = map_chg || avoid_reserve || !vma_resv_map(vma);
2320         if (deferred_reserve) {
2321                 ret = hugetlb_cgroup_charge_cgroup_rsvd(
2322                         idx, pages_per_huge_page(h), &h_cg);
2323                 if (ret)
2324                         goto out_subpool_put;
2325         }
2326
2327         ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2328         if (ret)
2329                 goto out_uncharge_cgroup_reservation;
2330
2331         spin_lock(&hugetlb_lock);
2332         /*
2333          * glb_chg is passed to indicate whether or not a page must be taken
2334          * from the global free pool (global change).  gbl_chg == 0 indicates
2335          * a reservation exists for the allocation.
2336          */
2337         page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2338         if (!page) {
2339                 spin_unlock(&hugetlb_lock);
2340                 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2341                 if (!page)
2342                         goto out_uncharge_cgroup;
2343                 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2344                         SetHPageRestoreReserve(page);
2345                         h->resv_huge_pages--;
2346                 }
2347                 spin_lock(&hugetlb_lock);
2348                 list_add(&page->lru, &h->hugepage_activelist);
2349                 /* Fall through */
2350         }
2351         hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2352         /* If allocation is not consuming a reservation, also store the
2353          * hugetlb_cgroup pointer on the page.
2354          */
2355         if (deferred_reserve) {
2356                 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
2357                                                   h_cg, page);
2358         }
2359
2360         spin_unlock(&hugetlb_lock);
2361
2362         hugetlb_set_page_subpool(page, spool);
2363
2364         map_commit = vma_commit_reservation(h, vma, addr);
2365         if (unlikely(map_chg > map_commit)) {
2366                 /*
2367                  * The page was added to the reservation map between
2368                  * vma_needs_reservation and vma_commit_reservation.
2369                  * This indicates a race with hugetlb_reserve_pages.
2370                  * Adjust for the subpool count incremented above AND
2371                  * in hugetlb_reserve_pages for the same page.  Also,
2372                  * the reservation count added in hugetlb_reserve_pages
2373                  * no longer applies.
2374                  */
2375                 long rsv_adjust;
2376
2377                 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2378                 hugetlb_acct_memory(h, -rsv_adjust);
2379                 if (deferred_reserve)
2380                         hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
2381                                         pages_per_huge_page(h), page);
2382         }
2383         return page;
2384
2385 out_uncharge_cgroup:
2386         hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2387 out_uncharge_cgroup_reservation:
2388         if (deferred_reserve)
2389                 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
2390                                                     h_cg);
2391 out_subpool_put:
2392         if (map_chg || avoid_reserve)
2393                 hugepage_subpool_put_pages(spool, 1);
2394         vma_end_reservation(h, vma, addr);
2395         return ERR_PTR(-ENOSPC);
2396 }
2397
2398 int alloc_bootmem_huge_page(struct hstate *h)
2399         __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2400 int __alloc_bootmem_huge_page(struct hstate *h)
2401 {
2402         struct huge_bootmem_page *m;
2403         int nr_nodes, node;
2404
2405         for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2406                 void *addr;
2407
2408                 addr = memblock_alloc_try_nid_raw(
2409                                 huge_page_size(h), huge_page_size(h),
2410                                 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
2411                 if (addr) {
2412                         /*
2413                          * Use the beginning of the huge page to store the
2414                          * huge_bootmem_page struct (until gather_bootmem
2415                          * puts them into the mem_map).
2416                          */
2417                         m = addr;
2418                         goto found;
2419                 }
2420         }
2421         return 0;
2422
2423 found:
2424         BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2425         /* Put them into a private list first because mem_map is not up yet */
2426         INIT_LIST_HEAD(&m->list);
2427         list_add(&m->list, &huge_boot_pages);
2428         m->hstate = h;
2429         return 1;
2430 }
2431
2432 static void __init prep_compound_huge_page(struct page *page,
2433                 unsigned int order)
2434 {
2435         if (unlikely(order > (MAX_ORDER - 1)))
2436                 prep_compound_gigantic_page(page, order);
2437         else
2438                 prep_compound_page(page, order);
2439 }
2440
2441 /* Put bootmem huge pages into the standard lists after mem_map is up */
2442 static void __init gather_bootmem_prealloc(void)
2443 {
2444         struct huge_bootmem_page *m;
2445
2446         list_for_each_entry(m, &huge_boot_pages, list) {
2447                 struct page *page = virt_to_page(m);
2448                 struct hstate *h = m->hstate;
2449
2450                 WARN_ON(page_count(page) != 1);
2451                 prep_compound_huge_page(page, huge_page_order(h));
2452                 WARN_ON(PageReserved(page));
2453                 prep_new_huge_page(h, page, page_to_nid(page));
2454                 put_page(page); /* free it into the hugepage allocator */
2455
2456                 /*
2457                  * If we had gigantic hugepages allocated at boot time, we need
2458                  * to restore the 'stolen' pages to totalram_pages in order to
2459                  * fix confusing memory reports from free(1) and another
2460                  * side-effects, like CommitLimit going negative.
2461                  */
2462                 if (hstate_is_gigantic(h))
2463                         adjust_managed_page_count(page, pages_per_huge_page(h));
2464                 cond_resched();
2465         }
2466 }
2467
2468 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2469 {
2470         unsigned long i;
2471         nodemask_t *node_alloc_noretry;
2472
2473         if (!hstate_is_gigantic(h)) {
2474                 /*
2475                  * Bit mask controlling how hard we retry per-node allocations.
2476                  * Ignore errors as lower level routines can deal with
2477                  * node_alloc_noretry == NULL.  If this kmalloc fails at boot
2478                  * time, we are likely in bigger trouble.
2479                  */
2480                 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
2481                                                 GFP_KERNEL);
2482         } else {
2483                 /* allocations done at boot time */
2484                 node_alloc_noretry = NULL;
2485         }
2486
2487         /* bit mask controlling how hard we retry per-node allocations */
2488         if (node_alloc_noretry)
2489                 nodes_clear(*node_alloc_noretry);
2490
2491         for (i = 0; i < h->max_huge_pages; ++i) {
2492                 if (hstate_is_gigantic(h)) {
2493                         if (hugetlb_cma_size) {
2494                                 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
2495                                 goto free;
2496                         }
2497                         if (!alloc_bootmem_huge_page(h))
2498                                 break;
2499                 } else if (!alloc_pool_huge_page(h,
2500                                          &node_states[N_MEMORY],
2501                                          node_alloc_noretry))
2502                         break;
2503                 cond_resched();
2504         }
2505         if (i < h->max_huge_pages) {
2506                 char buf[32];
2507
2508                 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2509                 pr_warn("HugeTLB: allocating %lu of page size %s failed.  Only allocated %lu hugepages.\n",
2510                         h->max_huge_pages, buf, i);
2511                 h->max_huge_pages = i;
2512         }
2513 free:
2514         kfree(node_alloc_noretry);
2515 }
2516
2517 static void __init hugetlb_init_hstates(void)
2518 {
2519         struct hstate *h;
2520
2521         for_each_hstate(h) {
2522                 if (minimum_order > huge_page_order(h))
2523                         minimum_order = huge_page_order(h);
2524
2525                 /* oversize hugepages were init'ed in early boot */
2526                 if (!hstate_is_gigantic(h))
2527                         hugetlb_hstate_alloc_pages(h);
2528         }
2529         VM_BUG_ON(minimum_order == UINT_MAX);
2530 }
2531
2532 static void __init report_hugepages(void)
2533 {
2534         struct hstate *h;
2535
2536         for_each_hstate(h) {
2537                 char buf[32];
2538
2539                 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2540                 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2541                         buf, h->free_huge_pages);
2542         }
2543 }
2544
2545 #ifdef CONFIG_HIGHMEM
2546 static void try_to_free_low(struct hstate *h, unsigned long count,
2547                                                 nodemask_t *nodes_allowed)
2548 {
2549         int i;
2550
2551         if (hstate_is_gigantic(h))
2552                 return;
2553
2554         for_each_node_mask(i, *nodes_allowed) {
2555                 struct page *page, *next;
2556                 struct list_head *freel = &h->hugepage_freelists[i];
2557                 list_for_each_entry_safe(page, next, freel, lru) {
2558                         if (count >= h->nr_huge_pages)
2559                                 return;
2560                         if (PageHighMem(page))
2561                                 continue;
2562                         list_del(&page->lru);
2563                         update_and_free_page(h, page);
2564                         h->free_huge_pages--;
2565                         h->free_huge_pages_node[page_to_nid(page)]--;
2566                 }
2567         }
2568 }
2569 #else
2570 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2571                                                 nodemask_t *nodes_allowed)
2572 {
2573 }
2574 #endif
2575
2576 /*
2577  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
2578  * balanced by operating on them in a round-robin fashion.
2579  * Returns 1 if an adjustment was made.
2580  */
2581 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2582                                 int delta)
2583 {
2584         int nr_nodes, node;
2585
2586         VM_BUG_ON(delta != -1 && delta != 1);
2587
2588         if (delta < 0) {
2589                 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2590                         if (h->surplus_huge_pages_node[node])
2591                                 goto found;
2592                 }
2593         } else {
2594                 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2595                         if (h->surplus_huge_pages_node[node] <
2596                                         h->nr_huge_pages_node[node])
2597                                 goto found;
2598                 }
2599         }
2600         return 0;
2601
2602 found:
2603         h->surplus_huge_pages += delta;
2604         h->surplus_huge_pages_node[node] += delta;
2605         return 1;
2606 }
2607
2608 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2609 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
2610                               nodemask_t *nodes_allowed)
2611 {
2612         unsigned long min_count, ret;
2613         NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
2614
2615         /*
2616          * Bit mask controlling how hard we retry per-node allocations.
2617          * If we can not allocate the bit mask, do not attempt to allocate
2618          * the requested huge pages.
2619          */
2620         if (node_alloc_noretry)
2621                 nodes_clear(*node_alloc_noretry);
2622         else
2623                 return -ENOMEM;
2624
2625         spin_lock(&hugetlb_lock);
2626
2627         /*
2628          * Check for a node specific request.
2629          * Changing node specific huge page count may require a corresponding
2630          * change to the global count.  In any case, the passed node mask
2631          * (nodes_allowed) will restrict alloc/free to the specified node.
2632          */
2633         if (nid != NUMA_NO_NODE) {
2634                 unsigned long old_count = count;
2635
2636                 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2637                 /*
2638                  * User may have specified a large count value which caused the
2639                  * above calculation to overflow.  In this case, they wanted
2640                  * to allocate as many huge pages as possible.  Set count to
2641                  * largest possible value to align with their intention.
2642                  */
2643                 if (count < old_count)
2644                         count = ULONG_MAX;
2645         }
2646
2647         /*
2648          * Gigantic pages runtime allocation depend on the capability for large
2649          * page range allocation.
2650          * If the system does not provide this feature, return an error when
2651          * the user tries to allocate gigantic pages but let the user free the
2652          * boottime allocated gigantic pages.
2653          */
2654         if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
2655                 if (count > persistent_huge_pages(h)) {
2656                         spin_unlock(&hugetlb_lock);
2657                         NODEMASK_FREE(node_alloc_noretry);
2658                         return -EINVAL;
2659                 }
2660                 /* Fall through to decrease pool */
2661         }
2662
2663         /*
2664          * Increase the pool size
2665          * First take pages out of surplus state.  Then make up the
2666          * remaining difference by allocating fresh huge pages.
2667          *
2668          * We might race with alloc_surplus_huge_page() here and be unable
2669          * to convert a surplus huge page to a normal huge page. That is
2670          * not critical, though, it just means the overall size of the
2671          * pool might be one hugepage larger than it needs to be, but
2672          * within all the constraints specified by the sysctls.
2673          */
2674         while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2675                 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2676                         break;
2677         }
2678
2679         while (count > persistent_huge_pages(h)) {
2680                 /*
2681                  * If this allocation races such that we no longer need the
2682                  * page, free_huge_page will handle it by freeing the page
2683                  * and reducing the surplus.
2684                  */
2685                 spin_unlock(&hugetlb_lock);
2686
2687                 /* yield cpu to avoid soft lockup */
2688                 cond_resched();
2689
2690                 ret = alloc_pool_huge_page(h, nodes_allowed,
2691                                                 node_alloc_noretry);
2692                 spin_lock(&hugetlb_lock);
2693                 if (!ret)
2694                         goto out;
2695
2696                 /* Bail for signals. Probably ctrl-c from user */
2697                 if (signal_pending(current))
2698                         goto out;
2699         }
2700
2701         /*
2702          * Decrease the pool size
2703          * First return free pages to the buddy allocator (being careful
2704          * to keep enough around to satisfy reservations).  Then place
2705          * pages into surplus state as needed so the pool will shrink
2706          * to the desired size as pages become free.
2707          *
2708          * By placing pages into the surplus state independent of the
2709          * overcommit value, we are allowing the surplus pool size to
2710          * exceed overcommit. There are few sane options here. Since
2711          * alloc_surplus_huge_page() is checking the global counter,
2712          * though, we'll note that we're not allowed to exceed surplus
2713          * and won't grow the pool anywhere else. Not until one of the
2714          * sysctls are changed, or the surplus pages go out of use.
2715          */
2716         min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2717         min_count = max(count, min_count);
2718         try_to_free_low(h, min_count, nodes_allowed);
2719         while (min_count < persistent_huge_pages(h)) {
2720                 if (!free_pool_huge_page(h, nodes_allowed, 0))
2721                         break;
2722                 cond_resched_lock(&hugetlb_lock);
2723         }
2724         while (count < persistent_huge_pages(h)) {
2725                 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2726                         break;
2727         }
2728 out:
2729         h->max_huge_pages = persistent_huge_pages(h);
2730         spin_unlock(&hugetlb_lock);
2731
2732         NODEMASK_FREE(node_alloc_noretry);
2733
2734         return 0;
2735 }
2736
2737 #define HSTATE_ATTR_RO(_name) \
2738         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2739
2740 #define HSTATE_ATTR(_name) \
2741         static struct kobj_attribute _name##_attr = \
2742                 __ATTR(_name, 0644, _name##_show, _name##_store)
2743
2744 static struct kobject *hugepages_kobj;
2745 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2746
2747 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2748
2749 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2750 {
2751         int i;
2752
2753         for (i = 0; i < HUGE_MAX_HSTATE; i++)
2754                 if (hstate_kobjs[i] == kobj) {
2755                         if (nidp)
2756                                 *nidp = NUMA_NO_NODE;
2757                         return &hstates[i];
2758                 }
2759
2760         return kobj_to_node_hstate(kobj, nidp);
2761 }
2762
2763 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2764                                         struct kobj_attribute *attr, char *buf)
2765 {
2766         struct hstate *h;
2767         unsigned long nr_huge_pages;
2768         int nid;
2769
2770         h = kobj_to_hstate(kobj, &nid);
2771         if (nid == NUMA_NO_NODE)
2772                 nr_huge_pages = h->nr_huge_pages;
2773         else
2774                 nr_huge_pages = h->nr_huge_pages_node[nid];
2775
2776         return sysfs_emit(buf, "%lu\n", nr_huge_pages);
2777 }
2778
2779 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2780                                            struct hstate *h, int nid,
2781                                            unsigned long count, size_t len)
2782 {
2783         int err;
2784         nodemask_t nodes_allowed, *n_mask;
2785
2786         if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2787                 return -EINVAL;
2788
2789         if (nid == NUMA_NO_NODE) {
2790                 /*
2791                  * global hstate attribute
2792                  */
2793                 if (!(obey_mempolicy &&
2794                                 init_nodemask_of_mempolicy(&nodes_allowed)))
2795                         n_mask = &node_states[N_MEMORY];
2796                 else
2797                         n_mask = &nodes_allowed;
2798         } else {
2799                 /*
2800                  * Node specific request.  count adjustment happens in
2801                  * set_max_huge_pages() after acquiring hugetlb_lock.
2802                  */
2803                 init_nodemask_of_node(&nodes_allowed, nid);
2804                 n_mask = &nodes_allowed;
2805         }
2806
2807         err = set_max_huge_pages(h, count, nid, n_mask);
2808
2809         return err ? err : len;
2810 }
2811
2812 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2813                                          struct kobject *kobj, const char *buf,
2814                                          size_t len)
2815 {
2816         struct hstate *h;
2817         unsigned long count;
2818         int nid;
2819         int err;
2820
2821         err = kstrtoul(buf, 10, &count);
2822         if (err)
2823                 return err;
2824
2825         h = kobj_to_hstate(kobj, &nid);
2826         return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2827 }
2828
2829 static ssize_t nr_hugepages_show(struct kobject *kobj,
2830                                        struct kobj_attribute *attr, char *buf)
2831 {
2832         return nr_hugepages_show_common(kobj, attr, buf);
2833 }
2834
2835 static ssize_t nr_hugepages_store(struct kobject *kobj,
2836                struct kobj_attribute *attr, const char *buf, size_t len)
2837 {
2838         return nr_hugepages_store_common(false, kobj, buf, len);
2839 }
2840 HSTATE_ATTR(nr_hugepages);
2841
2842 #ifdef CONFIG_NUMA
2843
2844 /*
2845  * hstate attribute for optionally mempolicy-based constraint on persistent
2846  * huge page alloc/free.
2847  */
2848 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2849                                            struct kobj_attribute *attr,
2850                                            char *buf)
2851 {
2852         return nr_hugepages_show_common(kobj, attr, buf);
2853 }
2854
2855 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2856                struct kobj_attribute *attr, const char *buf, size_t len)
2857 {
2858         return nr_hugepages_store_common(true, kobj, buf, len);
2859 }
2860 HSTATE_ATTR(nr_hugepages_mempolicy);
2861 #endif
2862
2863
2864 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2865                                         struct kobj_attribute *attr, char *buf)
2866 {
2867         struct hstate *h = kobj_to_hstate(kobj, NULL);
2868         return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages);
2869 }
2870
2871 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2872                 struct kobj_attribute *attr, const char *buf, size_t count)
2873 {
2874         int err;
2875         unsigned long input;
2876         struct hstate *h = kobj_to_hstate(kobj, NULL);
2877
2878         if (hstate_is_gigantic(h))
2879                 return -EINVAL;
2880
2881         err = kstrtoul(buf, 10, &input);
2882         if (err)
2883                 return err;
2884
2885         spin_lock(&hugetlb_lock);
2886         h->nr_overcommit_huge_pages = input;
2887         spin_unlock(&hugetlb_lock);
2888
2889         return count;
2890 }
2891 HSTATE_ATTR(nr_overcommit_hugepages);
2892
2893 static ssize_t free_hugepages_show(struct kobject *kobj,
2894                                         struct kobj_attribute *attr, char *buf)
2895 {
2896         struct hstate *h;
2897         unsigned long free_huge_pages;
2898         int nid;
2899
2900         h = kobj_to_hstate(kobj, &nid);
2901         if (nid == NUMA_NO_NODE)
2902                 free_huge_pages = h->free_huge_pages;
2903         else
2904                 free_huge_pages = h->free_huge_pages_node[nid];
2905
2906         return sysfs_emit(buf, "%lu\n", free_huge_pages);
2907 }
2908 HSTATE_ATTR_RO(free_hugepages);
2909
2910 static ssize_t resv_hugepages_show(struct kobject *kobj,
2911                                         struct kobj_attribute *attr, char *buf)
2912 {
2913         struct hstate *h = kobj_to_hstate(kobj, NULL);
2914         return sysfs_emit(buf, "%lu\n", h->resv_huge_pages);
2915 }
2916 HSTATE_ATTR_RO(resv_hugepages);
2917
2918 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2919                                         struct kobj_attribute *attr, char *buf)
2920 {
2921         struct hstate *h;
2922         unsigned long surplus_huge_pages;
2923         int nid;
2924
2925         h = kobj_to_hstate(kobj, &nid);
2926         if (nid == NUMA_NO_NODE)
2927                 surplus_huge_pages = h->surplus_huge_pages;
2928         else
2929                 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2930
2931         return sysfs_emit(buf, "%lu\n", surplus_huge_pages);
2932 }
2933 HSTATE_ATTR_RO(surplus_hugepages);
2934
2935 static struct attribute *hstate_attrs[] = {
2936         &nr_hugepages_attr.attr,
2937         &nr_overcommit_hugepages_attr.attr,
2938         &free_hugepages_attr.attr,
2939         &resv_hugepages_attr.attr,
2940         &surplus_hugepages_attr.attr,
2941 #ifdef CONFIG_NUMA
2942         &nr_hugepages_mempolicy_attr.attr,
2943 #endif
2944         NULL,
2945 };
2946
2947 static const struct attribute_group hstate_attr_group = {
2948         .attrs = hstate_attrs,
2949 };
2950
2951 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2952                                     struct kobject **hstate_kobjs,
2953                                     const struct attribute_group *hstate_attr_group)
2954 {
2955         int retval;
2956         int hi = hstate_index(h);
2957
2958         hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2959         if (!hstate_kobjs[hi])
2960                 return -ENOMEM;
2961
2962         retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2963         if (retval) {
2964                 kobject_put(hstate_kobjs[hi]);
2965                 hstate_kobjs[hi] = NULL;
2966         }
2967
2968         return retval;
2969 }
2970
2971 static void __init hugetlb_sysfs_init(void)
2972 {
2973         struct hstate *h;
2974         int err;
2975
2976         hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2977         if (!hugepages_kobj)
2978                 return;
2979
2980         for_each_hstate(h) {
2981                 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2982                                          hstate_kobjs, &hstate_attr_group);
2983                 if (err)
2984                         pr_err("HugeTLB: Unable to add hstate %s", h->name);
2985         }
2986 }
2987
2988 #ifdef CONFIG_NUMA
2989
2990 /*
2991  * node_hstate/s - associate per node hstate attributes, via their kobjects,
2992  * with node devices in node_devices[] using a parallel array.  The array
2993  * index of a node device or _hstate == node id.
2994  * This is here to avoid any static dependency of the node device driver, in
2995  * the base kernel, on the hugetlb module.
2996  */
2997 struct node_hstate {
2998         struct kobject          *hugepages_kobj;
2999         struct kobject          *hstate_kobjs[HUGE_MAX_HSTATE];
3000 };
3001 static struct node_hstate node_hstates[MAX_NUMNODES];
3002
3003 /*
3004  * A subset of global hstate attributes for node devices
3005  */
3006 static struct attribute *per_node_hstate_attrs[] = {
3007         &nr_hugepages_attr.attr,
3008         &free_hugepages_attr.attr,
3009         &surplus_hugepages_attr.attr,
3010         NULL,
3011 };
3012
3013 static const struct attribute_group per_node_hstate_attr_group = {
3014         .attrs = per_node_hstate_attrs,
3015 };
3016
3017 /*
3018  * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
3019  * Returns node id via non-NULL nidp.
3020  */
3021 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3022 {
3023         int nid;
3024
3025         for (nid = 0; nid < nr_node_ids; nid++) {
3026                 struct node_hstate *nhs = &node_hstates[nid];
3027                 int i;
3028                 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3029                         if (nhs->hstate_kobjs[i] == kobj) {
3030                                 if (nidp)
3031                                         *nidp = nid;
3032                                 return &hstates[i];
3033                         }
3034         }
3035
3036         BUG();
3037         return NULL;
3038 }
3039
3040 /*
3041  * Unregister hstate attributes from a single node device.
3042  * No-op if no hstate attributes attached.
3043  */
3044 static void hugetlb_unregister_node(struct node *node)
3045 {
3046         struct hstate *h;
3047         struct node_hstate *nhs = &node_hstates[node->dev.id];
3048
3049         if (!nhs->hugepages_kobj)
3050                 return;         /* no hstate attributes */
3051
3052         for_each_hstate(h) {
3053                 int idx = hstate_index(h);
3054                 if (nhs->hstate_kobjs[idx]) {
3055                         kobject_put(nhs->hstate_kobjs[idx]);
3056                         nhs->hstate_kobjs[idx] = NULL;
3057                 }
3058         }
3059
3060         kobject_put(nhs->hugepages_kobj);
3061         nhs->hugepages_kobj = NULL;
3062 }
3063
3064
3065 /*
3066  * Register hstate attributes for a single node device.
3067  * No-op if attributes already registered.
3068  */
3069 static void hugetlb_register_node(struct node *node)
3070 {
3071         struct hstate *h;
3072         struct node_hstate *nhs = &node_hstates[node->dev.id];
3073         int err;
3074
3075         if (nhs->hugepages_kobj)
3076                 return;         /* already allocated */
3077
3078         nhs->hugepages_kobj = kobject_create_and_add("hugepages",
3079                                                         &node->dev.kobj);
3080         if (!nhs->hugepages_kobj)
3081                 return;
3082
3083         for_each_hstate(h) {
3084                 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
3085                                                 nhs->hstate_kobjs,
3086                                                 &per_node_hstate_attr_group);
3087                 if (err) {
3088                         pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
3089                                 h->name, node->dev.id);
3090                         hugetlb_unregister_node(node);
3091                         break;
3092                 }
3093         }
3094 }
3095
3096 /*
3097  * hugetlb init time:  register hstate attributes for all registered node
3098  * devices of nodes that have memory.  All on-line nodes should have
3099  * registered their associated device by this time.
3100  */
3101 static void __init hugetlb_register_all_nodes(void)
3102 {
3103         int nid;
3104
3105         for_each_node_state(nid, N_MEMORY) {
3106                 struct node *node = node_devices[nid];
3107                 if (node->dev.id == nid)
3108                         hugetlb_register_node(node);
3109         }
3110
3111         /*
3112          * Let the node device driver know we're here so it can
3113          * [un]register hstate attributes on node hotplug.
3114          */
3115         register_hugetlbfs_with_node(hugetlb_register_node,
3116                                      hugetlb_unregister_node);
3117 }
3118 #else   /* !CONFIG_NUMA */
3119
3120 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3121 {
3122         BUG();
3123         if (nidp)
3124                 *nidp = -1;
3125         return NULL;
3126 }
3127
3128 static void hugetlb_register_all_nodes(void) { }
3129
3130 #endif
3131
3132 static int __init hugetlb_init(void)
3133 {
3134         int i;
3135
3136         BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE <
3137                         __NR_HPAGEFLAGS);
3138
3139         if (!hugepages_supported()) {
3140                 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
3141                         pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
3142                 return 0;
3143         }
3144
3145         /*
3146          * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists.  Some
3147          * architectures depend on setup being done here.
3148          */
3149         hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
3150         if (!parsed_default_hugepagesz) {
3151                 /*
3152                  * If we did not parse a default huge page size, set
3153                  * default_hstate_idx to HPAGE_SIZE hstate. And, if the
3154                  * number of huge pages for this default size was implicitly
3155                  * specified, set that here as well.
3156                  * Note that the implicit setting will overwrite an explicit
3157                  * setting.  A warning will be printed in this case.
3158                  */
3159                 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
3160                 if (default_hstate_max_huge_pages) {
3161                         if (default_hstate.max_huge_pages) {
3162                                 char buf[32];
3163
3164                                 string_get_size(huge_page_size(&default_hstate),
3165                                         1, STRING_UNITS_2, buf, 32);
3166                                 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
3167                                         default_hstate.max_huge_pages, buf);
3168                                 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
3169                                         default_hstate_max_huge_pages);
3170                         }
3171                         default_hstate.max_huge_pages =
3172                                 default_hstate_max_huge_pages;
3173                 }
3174         }
3175
3176         hugetlb_cma_check();
3177         hugetlb_init_hstates();
3178         gather_bootmem_prealloc();
3179         report_hugepages();
3180
3181         hugetlb_sysfs_init();
3182         hugetlb_register_all_nodes();
3183         hugetlb_cgroup_file_init();
3184
3185 #ifdef CONFIG_SMP
3186         num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
3187 #else
3188         num_fault_mutexes = 1;
3189 #endif
3190         hugetlb_fault_mutex_table =
3191                 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
3192                               GFP_KERNEL);
3193         BUG_ON(!hugetlb_fault_mutex_table);
3194
3195         for (i = 0; i < num_fault_mutexes; i++)
3196                 mutex_init(&hugetlb_fault_mutex_table[i]);
3197         return 0;
3198 }
3199 subsys_initcall(hugetlb_init);
3200
3201 /* Overwritten by architectures with more huge page sizes */
3202 bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
3203 {
3204         return size == HPAGE_SIZE;
3205 }
3206
3207 void __init hugetlb_add_hstate(unsigned int order)
3208 {
3209         struct hstate *h;
3210         unsigned long i;
3211
3212         if (size_to_hstate(PAGE_SIZE << order)) {
3213                 return;
3214         }
3215         BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
3216         BUG_ON(order == 0);
3217         h = &hstates[hugetlb_max_hstate++];
3218         h->order = order;
3219         h->mask = ~(huge_page_size(h) - 1);
3220         for (i = 0; i < MAX_NUMNODES; ++i)
3221                 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
3222         INIT_LIST_HEAD(&h->hugepage_activelist);
3223         h->next_nid_to_alloc = first_memory_node;
3224         h->next_nid_to_free = first_memory_node;
3225         snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
3226                                         huge_page_size(h)/1024);
3227
3228         parsed_hstate = h;
3229 }
3230
3231 /*
3232  * hugepages command line processing
3233  * hugepages normally follows a valid hugepagsz or default_hugepagsz
3234  * specification.  If not, ignore the hugepages value.  hugepages can also
3235  * be the first huge page command line  option in which case it implicitly
3236  * specifies the number of huge pages for the default size.
3237  */
3238 static int __init hugepages_setup(char *s)
3239 {
3240         unsigned long *mhp;
3241         static unsigned long *last_mhp;
3242
3243         if (!parsed_valid_hugepagesz) {
3244                 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
3245                 parsed_valid_hugepagesz = true;
3246                 return 0;
3247         }
3248
3249         /*
3250          * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
3251          * yet, so this hugepages= parameter goes to the "default hstate".
3252          * Otherwise, it goes with the previously parsed hugepagesz or
3253          * default_hugepagesz.
3254          */
3255         else if (!hugetlb_max_hstate)
3256                 mhp = &default_hstate_max_huge_pages;
3257         else
3258                 mhp = &parsed_hstate->max_huge_pages;
3259
3260         if (mhp == last_mhp) {
3261                 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
3262                 return 0;
3263         }
3264
3265         if (sscanf(s, "%lu", mhp) <= 0)
3266                 *mhp = 0;
3267
3268         /*
3269          * Global state is always initialized later in hugetlb_init.
3270          * But we need to allocate >= MAX_ORDER hstates here early to still
3271          * use the bootmem allocator.
3272          */
3273         if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
3274                 hugetlb_hstate_alloc_pages(parsed_hstate);
3275
3276         last_mhp = mhp;
3277
3278         return 1;
3279 }
3280 __setup("hugepages=", hugepages_setup);
3281
3282 /*
3283  * hugepagesz command line processing
3284  * A specific huge page size can only be specified once with hugepagesz.
3285  * hugepagesz is followed by hugepages on the command line.  The global
3286  * variable 'parsed_valid_hugepagesz' is used to determine if prior
3287  * hugepagesz argument was valid.
3288  */
3289 static int __init hugepagesz_setup(char *s)
3290 {
3291         unsigned long size;
3292         struct hstate *h;
3293
3294         parsed_valid_hugepagesz = false;
3295         size = (unsigned long)memparse(s, NULL);
3296
3297         if (!arch_hugetlb_valid_size(size)) {
3298                 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
3299                 return 0;
3300         }
3301
3302         h = size_to_hstate(size);
3303         if (h) {
3304                 /*
3305                  * hstate for this size already exists.  This is normally
3306                  * an error, but is allowed if the existing hstate is the
3307                  * default hstate.  More specifically, it is only allowed if
3308                  * the number of huge pages for the default hstate was not
3309                  * previously specified.
3310                  */
3311                 if (!parsed_default_hugepagesz ||  h != &default_hstate ||
3312                     default_hstate.max_huge_pages) {
3313                         pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
3314                         return 0;
3315                 }
3316
3317                 /*
3318                  * No need to call hugetlb_add_hstate() as hstate already
3319                  * exists.  But, do set parsed_hstate so that a following
3320                  * hugepages= parameter will be applied to this hstate.
3321                  */
3322                 parsed_hstate = h;
3323                 parsed_valid_hugepagesz = true;
3324                 return 1;
3325         }
3326
3327         hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3328         parsed_valid_hugepagesz = true;
3329         return 1;
3330 }
3331 __setup("hugepagesz=", hugepagesz_setup);
3332
3333 /*
3334  * default_hugepagesz command line input
3335  * Only one instance of default_hugepagesz allowed on command line.
3336  */
3337 static int __init default_hugepagesz_setup(char *s)
3338 {
3339         unsigned long size;
3340
3341         parsed_valid_hugepagesz = false;
3342         if (parsed_default_hugepagesz) {
3343                 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
3344                 return 0;
3345         }
3346
3347         size = (unsigned long)memparse(s, NULL);
3348
3349         if (!arch_hugetlb_valid_size(size)) {
3350                 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
3351                 return 0;
3352         }
3353
3354         hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3355         parsed_valid_hugepagesz = true;
3356         parsed_default_hugepagesz = true;
3357         default_hstate_idx = hstate_index(size_to_hstate(size));
3358
3359         /*
3360          * The number of default huge pages (for this size) could have been
3361          * specified as the first hugetlb parameter: hugepages=X.  If so,
3362          * then default_hstate_max_huge_pages is set.  If the default huge
3363          * page size is gigantic (>= MAX_ORDER), then the pages must be
3364          * allocated here from bootmem allocator.
3365          */
3366         if (default_hstate_max_huge_pages) {
3367                 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
3368                 if (hstate_is_gigantic(&default_hstate))
3369                         hugetlb_hstate_alloc_pages(&default_hstate);
3370                 default_hstate_max_huge_pages = 0;
3371         }
3372
3373         return 1;
3374 }
3375 __setup("default_hugepagesz=", default_hugepagesz_setup);
3376
3377 static unsigned int allowed_mems_nr(struct hstate *h)
3378 {
3379         int node;
3380         unsigned int nr = 0;
3381         nodemask_t *mpol_allowed;
3382         unsigned int *array = h->free_huge_pages_node;
3383         gfp_t gfp_mask = htlb_alloc_mask(h);
3384
3385         mpol_allowed = policy_nodemask_current(gfp_mask);
3386
3387         for_each_node_mask(node, cpuset_current_mems_allowed) {
3388                 if (!mpol_allowed || node_isset(node, *mpol_allowed))
3389                         nr += array[node];
3390         }
3391
3392         return nr;
3393 }
3394
3395 #ifdef CONFIG_SYSCTL
3396 static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
3397                                           void *buffer, size_t *length,
3398                                           loff_t *ppos, unsigned long *out)
3399 {
3400         struct ctl_table dup_table;
3401
3402         /*
3403          * In order to avoid races with __do_proc_doulongvec_minmax(), we
3404          * can duplicate the @table and alter the duplicate of it.
3405          */
3406         dup_table = *table;
3407         dup_table.data = out;
3408
3409         return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
3410 }
3411
3412 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
3413                          struct ctl_table *table, int write,
3414                          void *buffer, size_t *length, loff_t *ppos)
3415 {
3416         struct hstate *h = &default_hstate;
3417         unsigned long tmp = h->max_huge_pages;
3418         int ret;
3419
3420         if (!hugepages_supported())
3421                 return -EOPNOTSUPP;
3422
3423         ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3424                                              &tmp);
3425         if (ret)
3426                 goto out;
3427
3428         if (write)
3429                 ret = __nr_hugepages_store_common(obey_mempolicy, h,
3430                                                   NUMA_NO_NODE, tmp, *length);
3431 out:
3432         return ret;
3433 }
3434
3435 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
3436                           void *buffer, size_t *length, loff_t *ppos)
3437 {
3438
3439         return hugetlb_sysctl_handler_common(false, table, write,
3440                                                         buffer, length, ppos);
3441 }
3442
3443 #ifdef CONFIG_NUMA
3444 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
3445                           void *buffer, size_t *length, loff_t *ppos)
3446 {
3447         return hugetlb_sysctl_handler_common(true, table, write,
3448                                                         buffer, length, ppos);
3449 }
3450 #endif /* CONFIG_NUMA */
3451
3452 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
3453                 void *buffer, size_t *length, loff_t *ppos)
3454 {
3455         struct hstate *h = &default_hstate;
3456         unsigned long tmp;
3457         int ret;
3458
3459         if (!hugepages_supported())
3460                 return -EOPNOTSUPP;
3461
3462         tmp = h->nr_overcommit_huge_pages;
3463
3464         if (write && hstate_is_gigantic(h))
3465                 return -EINVAL;
3466
3467         ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3468                                              &tmp);
3469         if (ret)
3470                 goto out;
3471
3472         if (write) {
3473                 spin_lock(&hugetlb_lock);
3474                 h->nr_overcommit_huge_pages = tmp;
3475                 spin_unlock(&hugetlb_lock);
3476         }
3477 out:
3478         return ret;
3479 }
3480
3481 #endif /* CONFIG_SYSCTL */
3482
3483 void hugetlb_report_meminfo(struct seq_file *m)
3484 {
3485         struct hstate *h;
3486         unsigned long total = 0;
3487
3488         if (!hugepages_supported())
3489                 return;
3490
3491         for_each_hstate(h) {
3492                 unsigned long count = h->nr_huge_pages;
3493
3494                 total += huge_page_size(h) * count;
3495
3496                 if (h == &default_hstate)
3497                         seq_printf(m,
3498                                    "HugePages_Total:   %5lu\n"
3499                                    "HugePages_Free:    %5lu\n"
3500                                    "HugePages_Rsvd:    %5lu\n"
3501                                    "HugePages_Surp:    %5lu\n"
3502                                    "Hugepagesize:   %8lu kB\n",
3503                                    count,
3504                                    h->free_huge_pages,
3505                                    h->resv_huge_pages,
3506                                    h->surplus_huge_pages,
3507                                    huge_page_size(h) / SZ_1K);
3508         }
3509
3510         seq_printf(m, "Hugetlb:        %8lu kB\n", total / SZ_1K);
3511 }
3512
3513 int hugetlb_report_node_meminfo(char *buf, int len, int nid)
3514 {
3515         struct hstate *h = &default_hstate;
3516
3517         if (!hugepages_supported())
3518                 return 0;
3519
3520         return sysfs_emit_at(buf, len,
3521                              "Node %d HugePages_Total: %5u\n"
3522                              "Node %d HugePages_Free:  %5u\n"
3523                              "Node %d HugePages_Surp:  %5u\n",
3524                              nid, h->nr_huge_pages_node[nid],
3525                              nid, h->free_huge_pages_node[nid],
3526                              nid, h->surplus_huge_pages_node[nid]);
3527 }
3528
3529 void hugetlb_show_meminfo(void)
3530 {
3531         struct hstate *h;
3532         int nid;
3533
3534         if (!hugepages_supported())
3535                 return;
3536
3537         for_each_node_state(nid, N_MEMORY)
3538                 for_each_hstate(h)
3539                         pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3540                                 nid,
3541                                 h->nr_huge_pages_node[nid],
3542                                 h->free_huge_pages_node[nid],
3543                                 h->surplus_huge_pages_node[nid],
3544                                 huge_page_size(h) / SZ_1K);
3545 }
3546
3547 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3548 {
3549         seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3550                    atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3551 }
3552
3553 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3554 unsigned long hugetlb_total_pages(void)
3555 {
3556         struct hstate *h;
3557         unsigned long nr_total_pages = 0;
3558
3559         for_each_hstate(h)
3560                 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3561         return nr_total_pages;
3562 }
3563
3564 static int hugetlb_acct_memory(struct hstate *h, long delta)
3565 {
3566         int ret = -ENOMEM;
3567
3568         if (!delta)
3569                 return 0;
3570
3571         spin_lock(&hugetlb_lock);
3572         /*
3573          * When cpuset is configured, it breaks the strict hugetlb page
3574          * reservation as the accounting is done on a global variable. Such
3575          * reservation is completely rubbish in the presence of cpuset because
3576          * the reservation is not checked against page availability for the
3577          * current cpuset. Application can still potentially OOM'ed by kernel
3578          * with lack of free htlb page in cpuset that the task is in.
3579          * Attempt to enforce strict accounting with cpuset is almost
3580          * impossible (or too ugly) because cpuset is too fluid that
3581          * task or memory node can be dynamically moved between cpusets.
3582          *
3583          * The change of semantics for shared hugetlb mapping with cpuset is
3584          * undesirable. However, in order to preserve some of the semantics,
3585          * we fall back to check against current free page availability as
3586          * a best attempt and hopefully to minimize the impact of changing
3587          * semantics that cpuset has.
3588          *
3589          * Apart from cpuset, we also have memory policy mechanism that
3590          * also determines from which node the kernel will allocate memory
3591          * in a NUMA system. So similar to cpuset, we also should consider
3592          * the memory policy of the current task. Similar to the description
3593          * above.
3594          */
3595         if (delta > 0) {
3596                 if (gather_surplus_pages(h, delta) < 0)
3597                         goto out;
3598
3599                 if (delta > allowed_mems_nr(h)) {
3600                         return_unused_surplus_pages(h, delta);
3601                         goto out;
3602                 }
3603         }
3604
3605         ret = 0;
3606         if (delta < 0)
3607                 return_unused_surplus_pages(h, (unsigned long) -delta);
3608
3609 out:
3610         spin_unlock(&hugetlb_lock);
3611         return ret;
3612 }
3613
3614 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3615 {
3616         struct resv_map *resv = vma_resv_map(vma);
3617
3618         /*
3619          * This new VMA should share its siblings reservation map if present.
3620          * The VMA will only ever have a valid reservation map pointer where
3621          * it is being copied for another still existing VMA.  As that VMA
3622          * has a reference to the reservation map it cannot disappear until
3623          * after this open call completes.  It is therefore safe to take a
3624          * new reference here without additional locking.
3625          */
3626         if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3627                 kref_get(&resv->refs);
3628 }
3629
3630 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3631 {
3632         struct hstate *h = hstate_vma(vma);
3633         struct resv_map *resv = vma_resv_map(vma);
3634         struct hugepage_subpool *spool = subpool_vma(vma);
3635         unsigned long reserve, start, end;
3636         long gbl_reserve;
3637
3638         if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3639                 return;
3640
3641         start = vma_hugecache_offset(h, vma, vma->vm_start);
3642         end = vma_hugecache_offset(h, vma, vma->vm_end);
3643
3644         reserve = (end - start) - region_count(resv, start, end);
3645         hugetlb_cgroup_uncharge_counter(resv, start, end);
3646         if (reserve) {
3647                 /*
3648                  * Decrement reserve counts.  The global reserve count may be
3649                  * adjusted if the subpool has a minimum size.
3650                  */
3651                 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3652                 hugetlb_acct_memory(h, -gbl_reserve);
3653         }
3654
3655         kref_put(&resv->refs, resv_map_release);
3656 }
3657
3658 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3659 {
3660         if (addr & ~(huge_page_mask(hstate_vma(vma))))
3661                 return -EINVAL;
3662         return 0;
3663 }
3664
3665 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
3666 {
3667         return huge_page_size(hstate_vma(vma));
3668 }
3669
3670 /*
3671  * We cannot handle pagefaults against hugetlb pages at all.  They cause
3672  * handle_mm_fault() to try to instantiate regular-sized pages in the
3673  * hugepage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
3674  * this far.
3675  */
3676 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
3677 {
3678         BUG();
3679         return 0;
3680 }
3681
3682 /*
3683  * When a new function is introduced to vm_operations_struct and added
3684  * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3685  * This is because under System V memory model, mappings created via
3686  * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3687  * their original vm_ops are overwritten with shm_vm_ops.
3688  */
3689 const struct vm_operations_struct hugetlb_vm_ops = {
3690         .fault = hugetlb_vm_op_fault,
3691         .open = hugetlb_vm_op_open,
3692         .close = hugetlb_vm_op_close,
3693         .may_split = hugetlb_vm_op_split,
3694         .pagesize = hugetlb_vm_op_pagesize,
3695 };
3696
3697 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3698                                 int writable)
3699 {
3700         pte_t entry;
3701
3702         if (writable) {
3703                 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3704                                          vma->vm_page_prot)));
3705         } else {
3706                 entry = huge_pte_wrprotect(mk_huge_pte(page,
3707                                            vma->vm_page_prot));
3708         }
3709         entry = pte_mkyoung(entry);
3710         entry = pte_mkhuge(entry);
3711         entry = arch_make_huge_pte(entry, vma, page, writable);
3712
3713         return entry;
3714 }
3715
3716 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3717                                    unsigned long address, pte_t *ptep)
3718 {
3719         pte_t entry;
3720
3721         entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3722         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3723                 update_mmu_cache(vma, address, ptep);
3724 }
3725
3726 bool is_hugetlb_entry_migration(pte_t pte)
3727 {
3728         swp_entry_t swp;
3729
3730         if (huge_pte_none(pte) || pte_present(pte))
3731                 return false;
3732         swp = pte_to_swp_entry(pte);
3733         if (is_migration_entry(swp))
3734                 return true;
3735         else
3736                 return false;
3737 }
3738
3739 static bool is_hugetlb_entry_hwpoisoned(pte_t pte)
3740 {
3741         swp_entry_t swp;
3742
3743         if (huge_pte_none(pte) || pte_present(pte))
3744                 return false;
3745         swp = pte_to_swp_entry(pte);
3746         if (is_hwpoison_entry(swp))
3747                 return true;
3748         else
3749                 return false;
3750 }
3751
3752 static void
3753 hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr,
3754                      struct page *new_page)
3755 {
3756         __SetPageUptodate(new_page);
3757         set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1));
3758         hugepage_add_new_anon_rmap(new_page, vma, addr);
3759         hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm);
3760         ClearHPageRestoreReserve(new_page);
3761         SetHPageMigratable(new_page);
3762 }
3763
3764 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3765                             struct vm_area_struct *vma)
3766 {
3767         pte_t *src_pte, *dst_pte, entry, dst_entry;
3768         struct page *ptepage;
3769         unsigned long addr;
3770         bool cow = is_cow_mapping(vma->vm_flags);
3771         struct hstate *h = hstate_vma(vma);
3772         unsigned long sz = huge_page_size(h);
3773         unsigned long npages = pages_per_huge_page(h);
3774         struct address_space *mapping = vma->vm_file->f_mapping;
3775         struct mmu_notifier_range range;
3776         int ret = 0;
3777
3778         if (cow) {
3779                 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
3780                                         vma->vm_start,
3781                                         vma->vm_end);
3782                 mmu_notifier_invalidate_range_start(&range);
3783         } else {
3784                 /*
3785                  * For shared mappings i_mmap_rwsem must be held to call
3786                  * huge_pte_alloc, otherwise the returned ptep could go
3787                  * away if part of a shared pmd and another thread calls
3788                  * huge_pmd_unshare.
3789                  */
3790                 i_mmap_lock_read(mapping);
3791         }
3792
3793         for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3794                 spinlock_t *src_ptl, *dst_ptl;
3795                 src_pte = huge_pte_offset(src, addr, sz);
3796                 if (!src_pte)
3797                         continue;
3798                 dst_pte = huge_pte_alloc(dst, addr, sz);
3799                 if (!dst_pte) {
3800                         ret = -ENOMEM;
3801                         break;
3802                 }
3803
3804                 /*
3805                  * If the pagetables are shared don't copy or take references.
3806                  * dst_pte == src_pte is the common case of src/dest sharing.
3807                  *
3808                  * However, src could have 'unshared' and dst shares with
3809                  * another vma.  If dst_pte !none, this implies sharing.
3810                  * Check here before taking page table lock, and once again
3811                  * after taking the lock below.
3812                  */
3813                 dst_entry = huge_ptep_get(dst_pte);
3814                 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
3815                         continue;
3816
3817                 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3818                 src_ptl = huge_pte_lockptr(h, src, src_pte);
3819                 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3820                 entry = huge_ptep_get(src_pte);
3821                 dst_entry = huge_ptep_get(dst_pte);
3822 again:
3823                 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3824                         /*
3825                          * Skip if src entry none.  Also, skip in the
3826                          * unlikely case dst entry !none as this implies
3827                          * sharing with another vma.
3828                          */
3829                         ;
3830                 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3831                                     is_hugetlb_entry_hwpoisoned(entry))) {
3832                         swp_entry_t swp_entry = pte_to_swp_entry(entry);
3833
3834                         if (is_write_migration_entry(swp_entry) && cow) {
3835                                 /*
3836                                  * COW mappings require pages in both
3837                                  * parent and child to be set to read.
3838                                  */
3839                                 make_migration_entry_read(&swp_entry);
3840                                 entry = swp_entry_to_pte(swp_entry);
3841                                 set_huge_swap_pte_at(src, addr, src_pte,
3842                                                      entry, sz);
3843                         }
3844                         set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3845                 } else {
3846                         entry = huge_ptep_get(src_pte);
3847                         ptepage = pte_page(entry);
3848                         get_page(ptepage);
3849
3850                         /*
3851                          * This is a rare case where we see pinned hugetlb
3852                          * pages while they're prone to COW.  We need to do the
3853                          * COW earlier during fork.
3854                          *
3855                          * When pre-allocating the page or copying data, we
3856                          * need to be without the pgtable locks since we could
3857                          * sleep during the process.
3858                          */
3859                         if (unlikely(page_needs_cow_for_dma(vma, ptepage))) {
3860                                 pte_t src_pte_old = entry;
3861                                 struct page *new;
3862
3863                                 spin_unlock(src_ptl);
3864                                 spin_unlock(dst_ptl);
3865                                 /* Do not use reserve as it's private owned */
3866                                 new = alloc_huge_page(vma, addr, 1);
3867                                 if (IS_ERR(new)) {
3868                                         put_page(ptepage);
3869                                         ret = PTR_ERR(new);
3870                                         break;
3871                                 }
3872                                 copy_user_huge_page(new, ptepage, addr, vma,
3873                                                     npages);
3874                                 put_page(ptepage);
3875
3876                                 /* Install the new huge page if src pte stable */
3877                                 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3878                                 src_ptl = huge_pte_lockptr(h, src, src_pte);
3879                                 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3880                                 entry = huge_ptep_get(src_pte);
3881                                 if (!pte_same(src_pte_old, entry)) {
3882                                         put_page(new);
3883                                         /* dst_entry won't change as in child */
3884                                         goto again;
3885                                 }
3886                                 hugetlb_install_page(vma, dst_pte, addr, new);
3887                                 spin_unlock(src_ptl);
3888                                 spin_unlock(dst_ptl);
3889                                 continue;
3890                         }
3891
3892                         if (cow) {
3893                                 /*
3894                                  * No need to notify as we are downgrading page
3895                                  * table protection not changing it to point
3896                                  * to a new page.
3897                                  *
3898                                  * See Documentation/vm/mmu_notifier.rst
3899                                  */
3900                                 huge_ptep_set_wrprotect(src, addr, src_pte);
3901                         }
3902
3903                         page_dup_rmap(ptepage, true);
3904                         set_huge_pte_at(dst, addr, dst_pte, entry);
3905                         hugetlb_count_add(npages, dst);
3906                 }
3907                 spin_unlock(src_ptl);
3908                 spin_unlock(dst_ptl);
3909         }
3910
3911         if (cow)
3912                 mmu_notifier_invalidate_range_end(&range);
3913         else
3914                 i_mmap_unlock_read(mapping);
3915
3916         return ret;
3917 }
3918
3919 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3920                             unsigned long start, unsigned long end,
3921                             struct page *ref_page)
3922 {
3923         struct mm_struct *mm = vma->vm_mm;
3924         unsigned long address;
3925         pte_t *ptep;
3926         pte_t pte;
3927         spinlock_t *ptl;
3928         struct page *page;
3929         struct hstate *h = hstate_vma(vma);
3930         unsigned long sz = huge_page_size(h);
3931         struct mmu_notifier_range range;
3932
3933         WARN_ON(!is_vm_hugetlb_page(vma));
3934         BUG_ON(start & ~huge_page_mask(h));
3935         BUG_ON(end & ~huge_page_mask(h));
3936
3937         /*
3938          * This is a hugetlb vma, all the pte entries should point
3939          * to huge page.
3940          */
3941         tlb_change_page_size(tlb, sz);
3942         tlb_start_vma(tlb, vma);
3943
3944         /*
3945          * If sharing possible, alert mmu notifiers of worst case.
3946          */
3947         mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
3948                                 end);
3949         adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
3950         mmu_notifier_invalidate_range_start(&range);
3951         address = start;
3952         for (; address < end; address += sz) {
3953                 ptep = huge_pte_offset(mm, address, sz);
3954                 if (!ptep)
3955                         continue;
3956
3957                 ptl = huge_pte_lock(h, mm, ptep);
3958                 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
3959                         spin_unlock(ptl);
3960                         /*
3961                          * We just unmapped a page of PMDs by clearing a PUD.
3962                          * The caller's TLB flush range should cover this area.
3963                          */
3964                         continue;
3965                 }
3966
3967                 pte = huge_ptep_get(ptep);
3968                 if (huge_pte_none(pte)) {
3969                         spin_unlock(ptl);
3970                         continue;
3971                 }
3972
3973                 /*
3974                  * Migrating hugepage or HWPoisoned hugepage is already
3975                  * unmapped and its refcount is dropped, so just clear pte here.
3976                  */
3977                 if (unlikely(!pte_present(pte))) {
3978                         huge_pte_clear(mm, address, ptep, sz);
3979                         spin_unlock(ptl);
3980                         continue;
3981                 }
3982
3983                 page = pte_page(pte);
3984                 /*
3985                  * If a reference page is supplied, it is because a specific
3986                  * page is being unmapped, not a range. Ensure the page we
3987                  * are about to unmap is the actual page of interest.
3988                  */
3989                 if (ref_page) {
3990                         if (page != ref_page) {
3991                                 spin_unlock(ptl);
3992                                 continue;
3993                         }
3994                         /*
3995                          * Mark the VMA as having unmapped its page so that
3996                          * future faults in this VMA will fail rather than
3997                          * looking like data was lost
3998                          */
3999                         set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
4000                 }
4001
4002                 pte = huge_ptep_get_and_clear(mm, address, ptep);
4003                 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
4004                 if (huge_pte_dirty(pte))
4005                         set_page_dirty(page);
4006
4007                 hugetlb_count_sub(pages_per_huge_page(h), mm);
4008                 page_remove_rmap(page, true);
4009
4010                 spin_unlock(ptl);
4011                 tlb_remove_page_size(tlb, page, huge_page_size(h));
4012                 /*
4013                  * Bail out after unmapping reference page if supplied
4014                  */
4015                 if (ref_page)
4016                         break;
4017         }
4018         mmu_notifier_invalidate_range_end(&range);
4019         tlb_end_vma(tlb, vma);
4020 }
4021
4022 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
4023                           struct vm_area_struct *vma, unsigned long start,
4024                           unsigned long end, struct page *ref_page)
4025 {
4026         __unmap_hugepage_range(tlb, vma, start, end, ref_page);
4027
4028         /*
4029          * Clear this flag so that x86's huge_pmd_share page_table_shareable
4030          * test will fail on a vma being torn down, and not grab a page table
4031          * on its way out.  We're lucky that the flag has such an appropriate
4032          * name, and can in fact be safely cleared here. We could clear it
4033          * before the __unmap_hugepage_range above, but all that's necessary
4034          * is to clear it before releasing the i_mmap_rwsem. This works
4035          * because in the context this is called, the VMA is about to be
4036          * destroyed and the i_mmap_rwsem is held.
4037          */
4038         vma->vm_flags &= ~VM_MAYSHARE;
4039 }
4040
4041 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
4042                           unsigned long end, struct page *ref_page)
4043 {
4044         struct mmu_gather tlb;
4045
4046         tlb_gather_mmu(&tlb, vma->vm_mm);
4047         __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
4048         tlb_finish_mmu(&tlb);
4049 }
4050
4051 /*
4052  * This is called when the original mapper is failing to COW a MAP_PRIVATE
4053  * mapping it owns the reserve page for. The intention is to unmap the page
4054  * from other VMAs and let the children be SIGKILLed if they are faulting the
4055  * same region.
4056  */
4057 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
4058                               struct page *page, unsigned long address)
4059 {
4060         struct hstate *h = hstate_vma(vma);
4061         struct vm_area_struct *iter_vma;
4062         struct address_space *mapping;
4063         pgoff_t pgoff;
4064
4065         /*
4066          * vm_pgoff is in PAGE_SIZE units, hence the different calculation
4067          * from page cache lookup which is in HPAGE_SIZE units.
4068          */
4069         address = address & huge_page_mask(h);
4070         pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
4071                         vma->vm_pgoff;
4072         mapping = vma->vm_file->f_mapping;
4073
4074         /*
4075          * Take the mapping lock for the duration of the table walk. As
4076          * this mapping should be shared between all the VMAs,
4077          * __unmap_hugepage_range() is called as the lock is already held
4078          */
4079         i_mmap_lock_write(mapping);
4080         vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
4081                 /* Do not unmap the current VMA */
4082                 if (iter_vma == vma)
4083                         continue;
4084
4085                 /*
4086                  * Shared VMAs have their own reserves and do not affect
4087                  * MAP_PRIVATE accounting but it is possible that a shared
4088                  * VMA is using the same page so check and skip such VMAs.
4089                  */
4090                 if (iter_vma->vm_flags & VM_MAYSHARE)
4091                         continue;
4092
4093                 /*
4094                  * Unmap the page from other VMAs without their own reserves.
4095                  * They get marked to be SIGKILLed if they fault in these
4096                  * areas. This is because a future no-page fault on this VMA
4097                  * could insert a zeroed page instead of the data existing
4098                  * from the time of fork. This would look like data corruption
4099                  */
4100                 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
4101                         unmap_hugepage_range(iter_vma, address,
4102                                              address + huge_page_size(h), page);
4103         }
4104         i_mmap_unlock_write(mapping);
4105 }
4106
4107 /*
4108  * Hugetlb_cow() should be called with page lock of the original hugepage held.
4109  * Called with hugetlb_instantiation_mutex held and pte_page locked so we
4110  * cannot race with other handlers or page migration.
4111  * Keep the pte_same checks anyway to make transition from the mutex easier.
4112  */
4113 static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
4114                        unsigned long address, pte_t *ptep,
4115                        struct page *pagecache_page, spinlock_t *ptl)
4116 {
4117         pte_t pte;
4118         struct hstate *h = hstate_vma(vma);
4119         struct page *old_page, *new_page;
4120         int outside_reserve = 0;
4121         vm_fault_t ret = 0;
4122         unsigned long haddr = address & huge_page_mask(h);
4123         struct mmu_notifier_range range;
4124
4125         pte = huge_ptep_get(ptep);
4126         old_page = pte_page(pte);
4127
4128 retry_avoidcopy:
4129         /* If no-one else is actually using this page, avoid the copy
4130          * and just make the page writable */
4131         if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
4132                 page_move_anon_rmap(old_page, vma);
4133                 set_huge_ptep_writable(vma, haddr, ptep);
4134                 return 0;
4135         }
4136
4137         /*
4138          * If the process that created a MAP_PRIVATE mapping is about to
4139          * perform a COW due to a shared page count, attempt to satisfy
4140          * the allocation without using the existing reserves. The pagecache
4141          * page is used to determine if the reserve at this address was
4142          * consumed or not. If reserves were used, a partial faulted mapping
4143          * at the time of fork() could consume its reserves on COW instead
4144          * of the full address range.
4145          */
4146         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
4147                         old_page != pagecache_page)
4148                 outside_reserve = 1;
4149
4150         get_page(old_page);
4151
4152         /*
4153          * Drop page table lock as buddy allocator may be called. It will
4154          * be acquired again before returning to the caller, as expected.
4155          */
4156         spin_unlock(ptl);
4157         new_page = alloc_huge_page(vma, haddr, outside_reserve);
4158
4159         if (IS_ERR(new_page)) {
4160                 /*
4161                  * If a process owning a MAP_PRIVATE mapping fails to COW,
4162                  * it is due to references held by a child and an insufficient
4163                  * huge page pool. To guarantee the original mappers
4164                  * reliability, unmap the page from child processes. The child
4165                  * may get SIGKILLed if it later faults.
4166                  */
4167                 if (outside_reserve) {
4168                         struct address_space *mapping = vma->vm_file->f_mapping;
4169                         pgoff_t idx;
4170                         u32 hash;
4171
4172                         put_page(old_page);
4173                         BUG_ON(huge_pte_none(pte));
4174                         /*
4175                          * Drop hugetlb_fault_mutex and i_mmap_rwsem before
4176                          * unmapping.  unmapping needs to hold i_mmap_rwsem
4177                          * in write mode.  Dropping i_mmap_rwsem in read mode
4178                          * here is OK as COW mappings do not interact with
4179                          * PMD sharing.
4180                          *
4181                          * Reacquire both after unmap operation.
4182                          */
4183                         idx = vma_hugecache_offset(h, vma, haddr);
4184                         hash = hugetlb_fault_mutex_hash(mapping, idx);
4185                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4186                         i_mmap_unlock_read(mapping);
4187
4188                         unmap_ref_private(mm, vma, old_page, haddr);
4189
4190                         i_mmap_lock_read(mapping);
4191                         mutex_lock(&hugetlb_fault_mutex_table[hash]);
4192                         spin_lock(ptl);
4193                         ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4194                         if (likely(ptep &&
4195                                    pte_same(huge_ptep_get(ptep), pte)))
4196                                 goto retry_avoidcopy;
4197                         /*
4198                          * race occurs while re-acquiring page table
4199                          * lock, and our job is done.
4200                          */
4201                         return 0;
4202                 }
4203
4204                 ret = vmf_error(PTR_ERR(new_page));
4205                 goto out_release_old;
4206         }
4207
4208         /*
4209          * When the original hugepage is shared one, it does not have
4210          * anon_vma prepared.
4211          */
4212         if (unlikely(anon_vma_prepare(vma))) {
4213                 ret = VM_FAULT_OOM;
4214                 goto out_release_all;
4215         }
4216
4217         copy_user_huge_page(new_page, old_page, address, vma,
4218                             pages_per_huge_page(h));
4219         __SetPageUptodate(new_page);
4220
4221         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
4222                                 haddr + huge_page_size(h));
4223         mmu_notifier_invalidate_range_start(&range);
4224
4225         /*
4226          * Retake the page table lock to check for racing updates
4227          * before the page tables are altered
4228          */
4229         spin_lock(ptl);
4230         ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4231         if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
4232                 ClearHPageRestoreReserve(new_page);
4233
4234                 /* Break COW */
4235                 huge_ptep_clear_flush(vma, haddr, ptep);
4236                 mmu_notifier_invalidate_range(mm, range.start, range.end);
4237                 set_huge_pte_at(mm, haddr, ptep,
4238                                 make_huge_pte(vma, new_page, 1));
4239                 page_remove_rmap(old_page, true);
4240                 hugepage_add_new_anon_rmap(new_page, vma, haddr);
4241                 SetHPageMigratable(new_page);
4242                 /* Make the old page be freed below */
4243                 new_page = old_page;
4244         }
4245         spin_unlock(ptl);
4246         mmu_notifier_invalidate_range_end(&range);
4247 out_release_all:
4248         restore_reserve_on_error(h, vma, haddr, new_page);
4249         put_page(new_page);
4250 out_release_old:
4251         put_page(old_page);
4252
4253         spin_lock(ptl); /* Caller expects lock to be held */
4254         return ret;
4255 }
4256
4257 /* Return the pagecache page at a given address within a VMA */
4258 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
4259                         struct vm_area_struct *vma, unsigned long address)
4260 {
4261         struct address_space *mapping;
4262         pgoff_t idx;
4263
4264         mapping = vma->vm_file->f_mapping;
4265         idx = vma_hugecache_offset(h, vma, address);
4266
4267         return find_lock_page(mapping, idx);
4268 }
4269
4270 /*
4271  * Return whether there is a pagecache page to back given address within VMA.
4272  * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
4273  */
4274 static bool hugetlbfs_pagecache_present(struct hstate *h,
4275                         struct vm_area_struct *vma, unsigned long address)
4276 {
4277         struct address_space *mapping;
4278         pgoff_t idx;
4279         struct page *page;
4280
4281         mapping = vma->vm_file->f_mapping;
4282         idx = vma_hugecache_offset(h, vma, address);
4283
4284         page = find_get_page(mapping, idx);
4285         if (page)
4286                 put_page(page);
4287         return page != NULL;
4288 }
4289
4290 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
4291                            pgoff_t idx)
4292 {
4293         struct inode *inode = mapping->host;
4294         struct hstate *h = hstate_inode(inode);
4295         int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
4296
4297         if (err)
4298                 return err;
4299         ClearHPageRestoreReserve(page);
4300
4301         /*
4302          * set page dirty so that it will not be removed from cache/file
4303          * by non-hugetlbfs specific code paths.
4304          */
4305         set_page_dirty(page);
4306
4307         spin_lock(&inode->i_lock);
4308         inode->i_blocks += blocks_per_huge_page(h);
4309         spin_unlock(&inode->i_lock);
4310         return 0;
4311 }
4312
4313 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
4314                         struct vm_area_struct *vma,
4315                         struct address_space *mapping, pgoff_t idx,
4316                         unsigned long address, pte_t *ptep, unsigned int flags)
4317 {
4318         struct hstate *h = hstate_vma(vma);
4319         vm_fault_t ret = VM_FAULT_SIGBUS;
4320         int anon_rmap = 0;
4321         unsigned long size;
4322         struct page *page;
4323         pte_t new_pte;
4324         spinlock_t *ptl;
4325         unsigned long haddr = address & huge_page_mask(h);
4326         bool new_page = false;
4327
4328         /*
4329          * Currently, we are forced to kill the process in the event the
4330          * original mapper has unmapped pages from the child due to a failed
4331          * COW. Warn that such a situation has occurred as it may not be obvious
4332          */
4333         if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
4334                 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
4335                            current->pid);
4336                 return ret;
4337         }
4338
4339         /*
4340          * We can not race with truncation due to holding i_mmap_rwsem.
4341          * i_size is modified when holding i_mmap_rwsem, so check here
4342          * once for faults beyond end of file.
4343          */
4344         size = i_size_read(mapping->host) >> huge_page_shift(h);
4345         if (idx >= size)
4346                 goto out;
4347
4348 retry:
4349         page = find_lock_page(mapping, idx);
4350         if (!page) {
4351                 /*
4352                  * Check for page in userfault range
4353                  */
4354                 if (userfaultfd_missing(vma)) {
4355                         u32 hash;
4356                         struct vm_fault vmf = {
4357                                 .vma = vma,
4358                                 .address = haddr,
4359                                 .flags = flags,
4360                                 /*
4361                                  * Hard to debug if it ends up being
4362                                  * used by a callee that assumes
4363                                  * something about the other
4364                                  * uninitialized fields... same as in
4365                                  * memory.c
4366                                  */
4367                         };
4368
4369                         /*
4370                          * hugetlb_fault_mutex and i_mmap_rwsem must be
4371                          * dropped before handling userfault.  Reacquire
4372                          * after handling fault to make calling code simpler.
4373                          */
4374                         hash = hugetlb_fault_mutex_hash(mapping, idx);
4375                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4376                         i_mmap_unlock_read(mapping);
4377                         ret = handle_userfault(&vmf, VM_UFFD_MISSING);
4378                         i_mmap_lock_read(mapping);
4379                         mutex_lock(&hugetlb_fault_mutex_table[hash]);
4380                         goto out;
4381                 }
4382
4383                 page = alloc_huge_page(vma, haddr, 0);
4384                 if (IS_ERR(page)) {
4385                         /*
4386                          * Returning error will result in faulting task being
4387                          * sent SIGBUS.  The hugetlb fault mutex prevents two
4388                          * tasks from racing to fault in the same page which
4389                          * could result in false unable to allocate errors.
4390                          * Page migration does not take the fault mutex, but
4391                          * does a clear then write of pte's under page table
4392                          * lock.  Page fault code could race with migration,
4393                          * notice the clear pte and try to allocate a page
4394                          * here.  Before returning error, get ptl and make
4395                          * sure there really is no pte entry.
4396                          */
4397                         ptl = huge_pte_lock(h, mm, ptep);
4398                         if (!huge_pte_none(huge_ptep_get(ptep))) {
4399                                 ret = 0;
4400                                 spin_unlock(ptl);
4401                                 goto out;
4402                         }
4403                         spin_unlock(ptl);
4404                         ret = vmf_error(PTR_ERR(page));
4405                         goto out;
4406                 }
4407                 clear_huge_page(page, address, pages_per_huge_page(h));
4408                 __SetPageUptodate(page);
4409                 new_page = true;
4410
4411                 if (vma->vm_flags & VM_MAYSHARE) {
4412                         int err = huge_add_to_page_cache(page, mapping, idx);
4413                         if (err) {
4414                                 put_page(page);
4415                                 if (err == -EEXIST)
4416                                         goto retry;
4417                                 goto out;
4418                         }
4419                 } else {
4420                         lock_page(page);
4421                         if (unlikely(anon_vma_prepare(vma))) {
4422                                 ret = VM_FAULT_OOM;
4423                                 goto backout_unlocked;
4424                         }
4425                         anon_rmap = 1;
4426                 }
4427         } else {
4428                 /*
4429                  * If memory error occurs between mmap() and fault, some process
4430                  * don't have hwpoisoned swap entry for errored virtual address.
4431                  * So we need to block hugepage fault by PG_hwpoison bit check.
4432                  */
4433                 if (unlikely(PageHWPoison(page))) {
4434                         ret = VM_FAULT_HWPOISON_LARGE |
4435                                 VM_FAULT_SET_HINDEX(hstate_index(h));
4436                         goto backout_unlocked;
4437                 }
4438         }
4439
4440         /*
4441          * If we are going to COW a private mapping later, we examine the
4442          * pending reservations for this page now. This will ensure that
4443          * any allocations necessary to record that reservation occur outside
4444          * the spinlock.
4445          */
4446         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4447                 if (vma_needs_reservation(h, vma, haddr) < 0) {
4448                         ret = VM_FAULT_OOM;
4449                         goto backout_unlocked;
4450                 }
4451                 /* Just decrements count, does not deallocate */
4452                 vma_end_reservation(h, vma, haddr);
4453         }
4454
4455         ptl = huge_pte_lock(h, mm, ptep);
4456         ret = 0;
4457         if (!huge_pte_none(huge_ptep_get(ptep)))
4458                 goto backout;
4459
4460         if (anon_rmap) {
4461                 ClearHPageRestoreReserve(page);
4462                 hugepage_add_new_anon_rmap(page, vma, haddr);
4463         } else
4464                 page_dup_rmap(page, true);
4465         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
4466                                 && (vma->vm_flags & VM_SHARED)));
4467         set_huge_pte_at(mm, haddr, ptep, new_pte);
4468
4469         hugetlb_count_add(pages_per_huge_page(h), mm);
4470         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4471                 /* Optimization, do the COW without a second fault */
4472                 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
4473         }
4474
4475         spin_unlock(ptl);
4476
4477         /*
4478          * Only set HPageMigratable in newly allocated pages.  Existing pages
4479          * found in the pagecache may not have HPageMigratableset if they have
4480          * been isolated for migration.
4481          */
4482         if (new_page)
4483                 SetHPageMigratable(page);
4484
4485         unlock_page(page);
4486 out:
4487         return ret;
4488
4489 backout:
4490         spin_unlock(ptl);
4491 backout_unlocked:
4492         unlock_page(page);
4493         restore_reserve_on_error(h, vma, haddr, page);
4494         put_page(page);
4495         goto out;
4496 }
4497
4498 #ifdef CONFIG_SMP
4499 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4500 {
4501         unsigned long key[2];
4502         u32 hash;
4503
4504         key[0] = (unsigned long) mapping;
4505         key[1] = idx;
4506
4507         hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
4508
4509         return hash & (num_fault_mutexes - 1);
4510 }
4511 #else
4512 /*
4513  * For uniprocessor systems we always use a single mutex, so just
4514  * return 0 and avoid the hashing overhead.
4515  */
4516 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4517 {
4518         return 0;
4519 }
4520 #endif
4521
4522 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
4523                         unsigned long address, unsigned int flags)
4524 {
4525         pte_t *ptep, entry;
4526         spinlock_t *ptl;
4527         vm_fault_t ret;
4528         u32 hash;
4529         pgoff_t idx;
4530         struct page *page = NULL;
4531         struct page *pagecache_page = NULL;
4532         struct hstate *h = hstate_vma(vma);
4533         struct address_space *mapping;
4534         int need_wait_lock = 0;
4535         unsigned long haddr = address & huge_page_mask(h);
4536
4537         ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4538         if (ptep) {
4539                 /*
4540                  * Since we hold no locks, ptep could be stale.  That is
4541                  * OK as we are only making decisions based on content and
4542                  * not actually modifying content here.
4543                  */
4544                 entry = huge_ptep_get(ptep);
4545                 if (unlikely(is_hugetlb_entry_migration(entry))) {
4546                         migration_entry_wait_huge(vma, mm, ptep);
4547                         return 0;
4548                 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
4549                         return VM_FAULT_HWPOISON_LARGE |
4550                                 VM_FAULT_SET_HINDEX(hstate_index(h));
4551         }
4552
4553         /*
4554          * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold
4555          * until finished with ptep.  This serves two purposes:
4556          * 1) It prevents huge_pmd_unshare from being called elsewhere
4557          *    and making the ptep no longer valid.
4558          * 2) It synchronizes us with i_size modifications during truncation.
4559          *
4560          * ptep could have already be assigned via huge_pte_offset.  That
4561          * is OK, as huge_pte_alloc will return the same value unless
4562          * something has changed.
4563          */
4564         mapping = vma->vm_file->f_mapping;
4565         i_mmap_lock_read(mapping);
4566         ptep = huge_pte_alloc(mm, haddr, huge_page_size(h));
4567         if (!ptep) {
4568                 i_mmap_unlock_read(mapping);
4569                 return VM_FAULT_OOM;
4570         }
4571
4572         /*
4573          * Serialize hugepage allocation and instantiation, so that we don't
4574          * get spurious allocation failures if two CPUs race to instantiate
4575          * the same page in the page cache.
4576          */
4577         idx = vma_hugecache_offset(h, vma, haddr);
4578         hash = hugetlb_fault_mutex_hash(mapping, idx);
4579         mutex_lock(&hugetlb_fault_mutex_table[hash]);
4580
4581         entry = huge_ptep_get(ptep);
4582         if (huge_pte_none(entry)) {
4583                 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
4584                 goto out_mutex;
4585         }
4586
4587         ret = 0;
4588
4589         /*
4590          * entry could be a migration/hwpoison entry at this point, so this
4591          * check prevents the kernel from going below assuming that we have
4592          * an active hugepage in pagecache. This goto expects the 2nd page
4593          * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
4594          * properly handle it.
4595          */
4596         if (!pte_present(entry))
4597                 goto out_mutex;
4598
4599         /*
4600          * If we are going to COW the mapping later, we examine the pending
4601          * reservations for this page now. This will ensure that any
4602          * allocations necessary to record that reservation occur outside the
4603          * spinlock. For private mappings, we also lookup the pagecache
4604          * page now as it is used to determine if a reservation has been
4605          * consumed.
4606          */
4607         if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
4608                 if (vma_needs_reservation(h, vma, haddr) < 0) {
4609                         ret = VM_FAULT_OOM;
4610                         goto out_mutex;
4611                 }
4612                 /* Just decrements count, does not deallocate */
4613                 vma_end_reservation(h, vma, haddr);
4614
4615                 if (!(vma->vm_flags & VM_MAYSHARE))
4616                         pagecache_page = hugetlbfs_pagecache_page(h,
4617                                                                 vma, haddr);
4618         }
4619
4620         ptl = huge_pte_lock(h, mm, ptep);
4621
4622         /* Check for a racing update before calling hugetlb_cow */
4623         if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
4624                 goto out_ptl;
4625
4626         /*
4627          * hugetlb_cow() requires page locks of pte_page(entry) and
4628          * pagecache_page, so here we need take the former one
4629          * when page != pagecache_page or !pagecache_page.
4630          */
4631         page = pte_page(entry);
4632         if (page != pagecache_page)
4633                 if (!trylock_page(page)) {
4634                         need_wait_lock = 1;
4635                         goto out_ptl;
4636                 }
4637
4638         get_page(page);
4639
4640         if (flags & FAULT_FLAG_WRITE) {
4641                 if (!huge_pte_write(entry)) {
4642                         ret = hugetlb_cow(mm, vma, address, ptep,
4643                                           pagecache_page, ptl);
4644                         goto out_put_page;
4645                 }
4646                 entry = huge_pte_mkdirty(entry);
4647         }
4648         entry = pte_mkyoung(entry);
4649         if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
4650                                                 flags & FAULT_FLAG_WRITE))
4651                 update_mmu_cache(vma, haddr, ptep);
4652 out_put_page:
4653         if (page != pagecache_page)
4654                 unlock_page(page);
4655         put_page(page);
4656 out_ptl:
4657         spin_unlock(ptl);
4658
4659         if (pagecache_page) {
4660                 unlock_page(pagecache_page);
4661                 put_page(pagecache_page);
4662         }
4663 out_mutex:
4664         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4665         i_mmap_unlock_read(mapping);
4666         /*
4667          * Generally it's safe to hold refcount during waiting page lock. But
4668          * here we just wait to defer the next page fault to avoid busy loop and
4669          * the page is not used after unlocked before returning from the current
4670          * page fault. So we are safe from accessing freed page, even if we wait
4671          * here without taking refcount.
4672          */
4673         if (need_wait_lock)
4674                 wait_on_page_locked(page);
4675         return ret;
4676 }
4677
4678 /*
4679  * Used by userfaultfd UFFDIO_COPY.  Based on mcopy_atomic_pte with
4680  * modifications for huge pages.
4681  */
4682 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4683                             pte_t *dst_pte,
4684                             struct vm_area_struct *dst_vma,
4685                             unsigned long dst_addr,
4686                             unsigned long src_addr,
4687                             struct page **pagep)
4688 {
4689         struct address_space *mapping;
4690         pgoff_t idx;
4691         unsigned long size;
4692         int vm_shared = dst_vma->vm_flags & VM_SHARED;
4693         struct hstate *h = hstate_vma(dst_vma);
4694         pte_t _dst_pte;
4695         spinlock_t *ptl;
4696         int ret;
4697         struct page *page;
4698
4699         if (!*pagep) {
4700                 ret = -ENOMEM;
4701                 page = alloc_huge_page(dst_vma, dst_addr, 0);
4702                 if (IS_ERR(page))
4703                         goto out;
4704
4705                 ret = copy_huge_page_from_user(page,
4706                                                 (const void __user *) src_addr,
4707                                                 pages_per_huge_page(h), false);
4708
4709                 /* fallback to copy_from_user outside mmap_lock */
4710                 if (unlikely(ret)) {
4711                         ret = -ENOENT;
4712                         *pagep = page;
4713                         /* don't free the page */
4714                         goto out;
4715                 }
4716         } else {
4717                 page = *pagep;
4718                 *pagep = NULL;
4719         }
4720
4721         /*
4722          * The memory barrier inside __SetPageUptodate makes sure that
4723          * preceding stores to the page contents become visible before
4724          * the set_pte_at() write.
4725          */
4726         __SetPageUptodate(page);
4727
4728         mapping = dst_vma->vm_file->f_mapping;
4729         idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4730
4731         /*
4732          * If shared, add to page cache
4733          */
4734         if (vm_shared) {
4735                 size = i_size_read(mapping->host) >> huge_page_shift(h);
4736                 ret = -EFAULT;
4737                 if (idx >= size)
4738                         goto out_release_nounlock;
4739
4740                 /*
4741                  * Serialization between remove_inode_hugepages() and
4742                  * huge_add_to_page_cache() below happens through the
4743                  * hugetlb_fault_mutex_table that here must be hold by
4744                  * the caller.
4745                  */
4746                 ret = huge_add_to_page_cache(page, mapping, idx);
4747                 if (ret)
4748                         goto out_release_nounlock;
4749         }
4750
4751         ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4752         spin_lock(ptl);
4753
4754         /*
4755          * Recheck the i_size after holding PT lock to make sure not
4756          * to leave any page mapped (as page_mapped()) beyond the end
4757          * of the i_size (remove_inode_hugepages() is strict about
4758          * enforcing that). If we bail out here, we'll also leave a
4759          * page in the radix tree in the vm_shared case beyond the end
4760          * of the i_size, but remove_inode_hugepages() will take care
4761          * of it as soon as we drop the hugetlb_fault_mutex_table.
4762          */
4763         size = i_size_read(mapping->host) >> huge_page_shift(h);
4764         ret = -EFAULT;
4765         if (idx >= size)
4766                 goto out_release_unlock;
4767
4768         ret = -EEXIST;
4769         if (!huge_pte_none(huge_ptep_get(dst_pte)))
4770                 goto out_release_unlock;
4771
4772         if (vm_shared) {
4773                 page_dup_rmap(page, true);
4774         } else {
4775                 ClearHPageRestoreReserve(page);
4776                 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4777         }
4778
4779         _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4780         if (dst_vma->vm_flags & VM_WRITE)
4781                 _dst_pte = huge_pte_mkdirty(_dst_pte);
4782         _dst_pte = pte_mkyoung(_dst_pte);
4783
4784         set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4785
4786         (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4787                                         dst_vma->vm_flags & VM_WRITE);
4788         hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4789
4790         /* No need to invalidate - it was non-present before */
4791         update_mmu_cache(dst_vma, dst_addr, dst_pte);
4792
4793         spin_unlock(ptl);
4794         SetHPageMigratable(page);
4795         if (vm_shared)
4796                 unlock_page(page);
4797         ret = 0;
4798 out:
4799         return ret;
4800 out_release_unlock:
4801         spin_unlock(ptl);
4802         if (vm_shared)
4803                 unlock_page(page);
4804 out_release_nounlock:
4805         put_page(page);
4806         goto out;
4807 }
4808
4809 static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma,
4810                                  int refs, struct page **pages,
4811                                  struct vm_area_struct **vmas)
4812 {
4813         int nr;
4814
4815         for (nr = 0; nr < refs; nr++) {
4816                 if (likely(pages))
4817                         pages[nr] = mem_map_offset(page, nr);
4818                 if (vmas)
4819                         vmas[nr] = vma;
4820         }
4821 }
4822
4823 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4824                          struct page **pages, struct vm_area_struct **vmas,
4825                          unsigned long *position, unsigned long *nr_pages,
4826                          long i, unsigned int flags, int *locked)
4827 {
4828         unsigned long pfn_offset;
4829         unsigned long vaddr = *position;
4830         unsigned long remainder = *nr_pages;
4831         struct hstate *h = hstate_vma(vma);
4832         int err = -EFAULT, refs;
4833
4834         while (vaddr < vma->vm_end && remainder) {
4835                 pte_t *pte;
4836                 spinlock_t *ptl = NULL;
4837                 int absent;
4838                 struct page *page;
4839
4840                 /*
4841                  * If we have a pending SIGKILL, don't keep faulting pages and
4842                  * potentially allocating memory.
4843                  */
4844                 if (fatal_signal_pending(current)) {
4845                         remainder = 0;
4846                         break;
4847                 }
4848
4849                 /*
4850                  * Some archs (sparc64, sh*) have multiple pte_ts to
4851                  * each hugepage.  We have to make sure we get the
4852                  * first, for the page indexing below to work.
4853                  *
4854                  * Note that page table lock is not held when pte is null.
4855                  */
4856                 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4857                                       huge_page_size(h));
4858                 if (pte)
4859                         ptl = huge_pte_lock(h, mm, pte);
4860                 absent = !pte || huge_pte_none(huge_ptep_get(pte));
4861
4862                 /*
4863                  * When coredumping, it suits get_dump_page if we just return
4864                  * an error where there's an empty slot with no huge pagecache
4865                  * to back it.  This way, we avoid allocating a hugepage, and
4866                  * the sparse dumpfile avoids allocating disk blocks, but its
4867                  * huge holes still show up with zeroes where they need to be.
4868                  */
4869                 if (absent && (flags & FOLL_DUMP) &&
4870                     !hugetlbfs_pagecache_present(h, vma, vaddr)) {
4871                         if (pte)
4872                                 spin_unlock(ptl);
4873                         remainder = 0;
4874                         break;
4875                 }
4876
4877                 /*
4878                  * We need call hugetlb_fault for both hugepages under migration
4879                  * (in which case hugetlb_fault waits for the migration,) and
4880                  * hwpoisoned hugepages (in which case we need to prevent the
4881                  * caller from accessing to them.) In order to do this, we use
4882                  * here is_swap_pte instead of is_hugetlb_entry_migration and
4883                  * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4884                  * both cases, and because we can't follow correct pages
4885                  * directly from any kind of swap entries.
4886                  */
4887                 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4888                     ((flags & FOLL_WRITE) &&
4889                       !huge_pte_write(huge_ptep_get(pte)))) {
4890                         vm_fault_t ret;
4891                         unsigned int fault_flags = 0;
4892
4893                         if (pte)
4894                                 spin_unlock(ptl);
4895                         if (flags & FOLL_WRITE)
4896                                 fault_flags |= FAULT_FLAG_WRITE;
4897                         if (locked)
4898                                 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4899                                         FAULT_FLAG_KILLABLE;
4900                         if (flags & FOLL_NOWAIT)
4901                                 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4902                                         FAULT_FLAG_RETRY_NOWAIT;
4903                         if (flags & FOLL_TRIED) {
4904                                 /*
4905                                  * Note: FAULT_FLAG_ALLOW_RETRY and
4906                                  * FAULT_FLAG_TRIED can co-exist
4907                                  */
4908                                 fault_flags |= FAULT_FLAG_TRIED;
4909                         }
4910                         ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4911                         if (ret & VM_FAULT_ERROR) {
4912                                 err = vm_fault_to_errno(ret, flags);
4913                                 remainder = 0;
4914                                 break;
4915                         }
4916                         if (ret & VM_FAULT_RETRY) {
4917                                 if (locked &&
4918                                     !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
4919                                         *locked = 0;
4920                                 *nr_pages = 0;
4921                                 /*
4922                                  * VM_FAULT_RETRY must not return an
4923                                  * error, it will return zero
4924                                  * instead.
4925                                  *
4926                                  * No need to update "position" as the
4927                                  * caller will not check it after
4928                                  * *nr_pages is set to 0.
4929                                  */
4930                                 return i;
4931                         }
4932                         continue;
4933                 }
4934
4935                 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4936                 page = pte_page(huge_ptep_get(pte));
4937
4938                 /*
4939                  * If subpage information not requested, update counters
4940                  * and skip the same_page loop below.
4941                  */
4942                 if (!pages && !vmas && !pfn_offset &&
4943                     (vaddr + huge_page_size(h) < vma->vm_end) &&
4944                     (remainder >= pages_per_huge_page(h))) {
4945                         vaddr += huge_page_size(h);
4946                         remainder -= pages_per_huge_page(h);
4947                         i += pages_per_huge_page(h);
4948                         spin_unlock(ptl);
4949                         continue;
4950                 }
4951
4952                 refs = min3(pages_per_huge_page(h) - pfn_offset,
4953                             (vma->vm_end - vaddr) >> PAGE_SHIFT, remainder);
4954
4955                 if (pages || vmas)
4956                         record_subpages_vmas(mem_map_offset(page, pfn_offset),
4957                                              vma, refs,
4958                                              likely(pages) ? pages + i : NULL,
4959                                              vmas ? vmas + i : NULL);
4960
4961                 if (pages) {
4962                         /*
4963                          * try_grab_compound_head() should always succeed here,
4964                          * because: a) we hold the ptl lock, and b) we've just
4965                          * checked that the huge page is present in the page
4966                          * tables. If the huge page is present, then the tail
4967                          * pages must also be present. The ptl prevents the
4968                          * head page and tail pages from being rearranged in
4969                          * any way. So this page must be available at this
4970                          * point, unless the page refcount overflowed:
4971                          */
4972                         if (WARN_ON_ONCE(!try_grab_compound_head(pages[i],
4973                                                                  refs,
4974                                                                  flags))) {
4975                                 spin_unlock(ptl);
4976                                 remainder = 0;
4977                                 err = -ENOMEM;
4978                                 break;
4979                         }
4980                 }
4981
4982                 vaddr += (refs << PAGE_SHIFT);
4983                 remainder -= refs;
4984                 i += refs;
4985
4986                 spin_unlock(ptl);
4987         }
4988         *nr_pages = remainder;
4989         /*
4990          * setting position is actually required only if remainder is
4991          * not zero but it's faster not to add a "if (remainder)"
4992          * branch.
4993          */
4994         *position = vaddr;
4995
4996         return i ? i : err;
4997 }
4998
4999 #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
5000 /*
5001  * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
5002  * implement this.
5003  */
5004 #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
5005 #endif
5006
5007 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
5008                 unsigned long address, unsigned long end, pgprot_t newprot)
5009 {
5010         struct mm_struct *mm = vma->vm_mm;
5011         unsigned long start = address;
5012         pte_t *ptep;
5013         pte_t pte;
5014         struct hstate *h = hstate_vma(vma);
5015         unsigned long pages = 0;
5016         bool shared_pmd = false;
5017         struct mmu_notifier_range range;
5018
5019         /*
5020          * In the case of shared PMDs, the area to flush could be beyond
5021          * start/end.  Set range.start/range.end to cover the maximum possible
5022          * range if PMD sharing is possible.
5023          */
5024         mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
5025                                 0, vma, mm, start, end);
5026         adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5027
5028         BUG_ON(address >= end);
5029         flush_cache_range(vma, range.start, range.end);
5030
5031         mmu_notifier_invalidate_range_start(&range);
5032         i_mmap_lock_write(vma->vm_file->f_mapping);
5033         for (; address < end; address += huge_page_size(h)) {
5034                 spinlock_t *ptl;
5035                 ptep = huge_pte_offset(mm, address, huge_page_size(h));
5036                 if (!ptep)
5037                         continue;
5038                 ptl = huge_pte_lock(h, mm, ptep);
5039                 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
5040                         pages++;
5041                         spin_unlock(ptl);
5042                         shared_pmd = true;
5043                         continue;
5044                 }
5045                 pte = huge_ptep_get(ptep);
5046                 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
5047                         spin_unlock(ptl);
5048                         continue;
5049                 }
5050                 if (unlikely(is_hugetlb_entry_migration(pte))) {
5051                         swp_entry_t entry = pte_to_swp_entry(pte);
5052
5053                         if (is_write_migration_entry(entry)) {
5054                                 pte_t newpte;
5055
5056                                 make_migration_entry_read(&entry);
5057                                 newpte = swp_entry_to_pte(entry);
5058                                 set_huge_swap_pte_at(mm, address, ptep,
5059                                                      newpte, huge_page_size(h));
5060                                 pages++;
5061                         }
5062                         spin_unlock(ptl);
5063                         continue;
5064                 }
5065                 if (!huge_pte_none(pte)) {
5066                         pte_t old_pte;
5067
5068                         old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
5069                         pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
5070                         pte = arch_make_huge_pte(pte, vma, NULL, 0);
5071                         huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
5072                         pages++;
5073                 }
5074                 spin_unlock(ptl);
5075         }
5076         /*
5077          * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
5078          * may have cleared our pud entry and done put_page on the page table:
5079          * once we release i_mmap_rwsem, another task can do the final put_page
5080          * and that page table be reused and filled with junk.  If we actually
5081          * did unshare a page of pmds, flush the range corresponding to the pud.
5082          */
5083         if (shared_pmd)
5084                 flush_hugetlb_tlb_range(vma, range.start, range.end);
5085         else
5086                 flush_hugetlb_tlb_range(vma, start, end);
5087         /*
5088          * No need to call mmu_notifier_invalidate_range() we are downgrading
5089          * page table protection not changing it to point to a new page.
5090          *
5091          * See Documentation/vm/mmu_notifier.rst
5092          */
5093         i_mmap_unlock_write(vma->vm_file->f_mapping);
5094         mmu_notifier_invalidate_range_end(&range);
5095
5096         return pages << h->order;
5097 }
5098
5099 /* Return true if reservation was successful, false otherwise.  */
5100 bool hugetlb_reserve_pages(struct inode *inode,
5101                                         long from, long to,
5102                                         struct vm_area_struct *vma,
5103                                         vm_flags_t vm_flags)
5104 {
5105         long chg, add = -1;
5106         struct hstate *h = hstate_inode(inode);
5107         struct hugepage_subpool *spool = subpool_inode(inode);
5108         struct resv_map *resv_map;
5109         struct hugetlb_cgroup *h_cg = NULL;
5110         long gbl_reserve, regions_needed = 0;
5111
5112         /* This should never happen */
5113         if (from > to) {
5114                 VM_WARN(1, "%s called with a negative range\n", __func__);
5115                 return false;
5116         }
5117
5118         /*
5119          * Only apply hugepage reservation if asked. At fault time, an
5120          * attempt will be made for VM_NORESERVE to allocate a page
5121          * without using reserves
5122          */
5123         if (vm_flags & VM_NORESERVE)
5124                 return true;
5125
5126         /*
5127          * Shared mappings base their reservation on the number of pages that
5128          * are already allocated on behalf of the file. Private mappings need
5129          * to reserve the full area even if read-only as mprotect() may be
5130          * called to make the mapping read-write. Assume !vma is a shm mapping
5131          */
5132         if (!vma || vma->vm_flags & VM_MAYSHARE) {
5133                 /*
5134                  * resv_map can not be NULL as hugetlb_reserve_pages is only
5135                  * called for inodes for which resv_maps were created (see
5136                  * hugetlbfs_get_inode).
5137                  */
5138                 resv_map = inode_resv_map(inode);
5139
5140                 chg = region_chg(resv_map, from, to, &regions_needed);
5141
5142         } else {
5143                 /* Private mapping. */
5144                 resv_map = resv_map_alloc();
5145                 if (!resv_map)
5146                         return false;
5147
5148                 chg = to - from;
5149
5150                 set_vma_resv_map(vma, resv_map);
5151                 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
5152         }
5153
5154         if (chg < 0)
5155                 goto out_err;
5156
5157         if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h),
5158                                 chg * pages_per_huge_page(h), &h_cg) < 0)
5159                 goto out_err;
5160
5161         if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
5162                 /* For private mappings, the hugetlb_cgroup uncharge info hangs
5163                  * of the resv_map.
5164                  */
5165                 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
5166         }
5167
5168         /*
5169          * There must be enough pages in the subpool for the mapping. If
5170          * the subpool has a minimum size, there may be some global
5171          * reservations already in place (gbl_reserve).
5172          */
5173         gbl_reserve = hugepage_subpool_get_pages(spool, chg);
5174         if (gbl_reserve < 0)
5175                 goto out_uncharge_cgroup;
5176
5177         /*
5178          * Check enough hugepages are available for the reservation.
5179          * Hand the pages back to the subpool if there are not
5180          */
5181         if (hugetlb_acct_memory(h, gbl_reserve) < 0)
5182                 goto out_put_pages;
5183
5184         /*
5185          * Account for the reservations made. Shared mappings record regions
5186          * that have reservations as they are shared by multiple VMAs.
5187          * When the last VMA disappears, the region map says how much
5188          * the reservation was and the page cache tells how much of
5189          * the reservation was consumed. Private mappings are per-VMA and
5190          * only the consumed reservations are tracked. When the VMA
5191          * disappears, the original reservation is the VMA size and the
5192          * consumed reservations are stored in the map. Hence, nothing
5193          * else has to be done for private mappings here
5194          */
5195         if (!vma || vma->vm_flags & VM_MAYSHARE) {
5196                 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
5197
5198                 if (unlikely(add < 0)) {
5199                         hugetlb_acct_memory(h, -gbl_reserve);
5200                         goto out_put_pages;
5201                 } else if (unlikely(chg > add)) {
5202                         /*
5203                          * pages in this range were added to the reserve
5204                          * map between region_chg and region_add.  This
5205                          * indicates a race with alloc_huge_page.  Adjust
5206                          * the subpool and reserve counts modified above
5207                          * based on the difference.
5208                          */
5209                         long rsv_adjust;
5210
5211                         /*
5212                          * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the
5213                          * reference to h_cg->css. See comment below for detail.
5214                          */
5215                         hugetlb_cgroup_uncharge_cgroup_rsvd(
5216                                 hstate_index(h),
5217                                 (chg - add) * pages_per_huge_page(h), h_cg);
5218
5219                         rsv_adjust = hugepage_subpool_put_pages(spool,
5220                                                                 chg - add);
5221                         hugetlb_acct_memory(h, -rsv_adjust);
5222                 } else if (h_cg) {
5223                         /*
5224                          * The file_regions will hold their own reference to
5225                          * h_cg->css. So we should release the reference held
5226                          * via hugetlb_cgroup_charge_cgroup_rsvd() when we are
5227                          * done.
5228                          */
5229                         hugetlb_cgroup_put_rsvd_cgroup(h_cg);
5230                 }
5231         }
5232         return true;
5233
5234 out_put_pages:
5235         /* put back original number of pages, chg */
5236         (void)hugepage_subpool_put_pages(spool, chg);
5237 out_uncharge_cgroup:
5238         hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
5239                                             chg * pages_per_huge_page(h), h_cg);
5240 out_err:
5241         if (!vma || vma->vm_flags & VM_MAYSHARE)
5242                 /* Only call region_abort if the region_chg succeeded but the
5243                  * region_add failed or didn't run.
5244                  */
5245                 if (chg >= 0 && add < 0)
5246                         region_abort(resv_map, from, to, regions_needed);
5247         if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
5248                 kref_put(&resv_map->refs, resv_map_release);
5249         return false;
5250 }
5251
5252 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
5253                                                                 long freed)
5254 {
5255         struct hstate *h = hstate_inode(inode);
5256         struct resv_map *resv_map = inode_resv_map(inode);
5257         long chg = 0;
5258         struct hugepage_subpool *spool = subpool_inode(inode);
5259         long gbl_reserve;
5260
5261         /*
5262          * Since this routine can be called in the evict inode path for all
5263          * hugetlbfs inodes, resv_map could be NULL.
5264          */
5265         if (resv_map) {
5266                 chg = region_del(resv_map, start, end);
5267                 /*
5268                  * region_del() can fail in the rare case where a region
5269                  * must be split and another region descriptor can not be
5270                  * allocated.  If end == LONG_MAX, it will not fail.
5271                  */
5272                 if (chg < 0)
5273                         return chg;
5274         }
5275
5276         spin_lock(&inode->i_lock);
5277         inode->i_blocks -= (blocks_per_huge_page(h) * freed);
5278         spin_unlock(&inode->i_lock);
5279
5280         /*
5281          * If the subpool has a minimum size, the number of global
5282          * reservations to be released may be adjusted.
5283          */
5284         gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
5285         hugetlb_acct_memory(h, -gbl_reserve);
5286
5287         return 0;
5288 }
5289
5290 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
5291 static unsigned long page_table_shareable(struct vm_area_struct *svma,
5292                                 struct vm_area_struct *vma,
5293                                 unsigned long addr, pgoff_t idx)
5294 {
5295         unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
5296                                 svma->vm_start;
5297         unsigned long sbase = saddr & PUD_MASK;
5298         unsigned long s_end = sbase + PUD_SIZE;
5299
5300         /* Allow segments to share if only one is marked locked */
5301         unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
5302         unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
5303
5304         /*
5305          * match the virtual addresses, permission and the alignment of the
5306          * page table page.
5307          */
5308         if (pmd_index(addr) != pmd_index(saddr) ||
5309             vm_flags != svm_flags ||
5310             !range_in_vma(svma, sbase, s_end))
5311                 return 0;
5312
5313         return saddr;
5314 }
5315
5316 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
5317 {
5318         unsigned long base = addr & PUD_MASK;
5319         unsigned long end = base + PUD_SIZE;
5320
5321         /*
5322          * check on proper vm_flags and page table alignment
5323          */
5324         if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
5325                 return true;
5326         return false;
5327 }
5328
5329 /*
5330  * Determine if start,end range within vma could be mapped by shared pmd.
5331  * If yes, adjust start and end to cover range associated with possible
5332  * shared pmd mappings.
5333  */
5334 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5335                                 unsigned long *start, unsigned long *end)
5336 {
5337         unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
5338                 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
5339
5340         /*
5341          * vma need span at least one aligned PUD size and the start,end range
5342          * must at least partialy within it.
5343          */
5344         if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
5345                 (*end <= v_start) || (*start >= v_end))
5346                 return;
5347
5348         /* Extend the range to be PUD aligned for a worst case scenario */
5349         if (*start > v_start)
5350                 *start = ALIGN_DOWN(*start, PUD_SIZE);
5351
5352         if (*end < v_end)
5353                 *end = ALIGN(*end, PUD_SIZE);
5354 }
5355
5356 /*
5357  * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
5358  * and returns the corresponding pte. While this is not necessary for the
5359  * !shared pmd case because we can allocate the pmd later as well, it makes the
5360  * code much cleaner.
5361  *
5362  * This routine must be called with i_mmap_rwsem held in at least read mode if
5363  * sharing is possible.  For hugetlbfs, this prevents removal of any page
5364  * table entries associated with the address space.  This is important as we
5365  * are setting up sharing based on existing page table entries (mappings).
5366  *
5367  * NOTE: This routine is only called from huge_pte_alloc.  Some callers of
5368  * huge_pte_alloc know that sharing is not possible and do not take
5369  * i_mmap_rwsem as a performance optimization.  This is handled by the
5370  * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is
5371  * only required for subsequent processing.
5372  */
5373 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
5374 {
5375         struct vm_area_struct *vma = find_vma(mm, addr);
5376         struct address_space *mapping = vma->vm_file->f_mapping;
5377         pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
5378                         vma->vm_pgoff;
5379         struct vm_area_struct *svma;
5380         unsigned long saddr;
5381         pte_t *spte = NULL;
5382         pte_t *pte;
5383         spinlock_t *ptl;
5384
5385         if (!vma_shareable(vma, addr))
5386                 return (pte_t *)pmd_alloc(mm, pud, addr);
5387
5388         i_mmap_assert_locked(mapping);
5389         vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
5390                 if (svma == vma)
5391                         continue;
5392
5393                 saddr = page_table_shareable(svma, vma, addr, idx);
5394                 if (saddr) {
5395                         spte = huge_pte_offset(svma->vm_mm, saddr,
5396                                                vma_mmu_pagesize(svma));
5397                         if (spte) {
5398                                 get_page(virt_to_page(spte));
5399                                 break;
5400                         }
5401                 }
5402         }
5403
5404         if (!spte)
5405                 goto out;
5406
5407         ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
5408         if (pud_none(*pud)) {
5409                 pud_populate(mm, pud,
5410                                 (pmd_t *)((unsigned long)spte & PAGE_MASK));
5411                 mm_inc_nr_pmds(mm);
5412         } else {
5413                 put_page(virt_to_page(spte));
5414         }
5415         spin_unlock(ptl);
5416 out:
5417         pte = (pte_t *)pmd_alloc(mm, pud, addr);
5418         return pte;
5419 }
5420
5421 /*
5422  * unmap huge page backed by shared pte.
5423  *
5424  * Hugetlb pte page is ref counted at the time of mapping.  If pte is shared
5425  * indicated by page_count > 1, unmap is achieved by clearing pud and
5426  * decrementing the ref count. If count == 1, the pte page is not shared.
5427  *
5428  * Called with page table lock held and i_mmap_rwsem held in write mode.
5429  *
5430  * returns: 1 successfully unmapped a shared pte page
5431  *          0 the underlying pte page is not shared, or it is the last user
5432  */
5433 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
5434                                         unsigned long *addr, pte_t *ptep)
5435 {
5436         pgd_t *pgd = pgd_offset(mm, *addr);
5437         p4d_t *p4d = p4d_offset(pgd, *addr);
5438         pud_t *pud = pud_offset(p4d, *addr);
5439
5440         i_mmap_assert_write_locked(vma->vm_file->f_mapping);
5441         BUG_ON(page_count(virt_to_page(ptep)) == 0);
5442         if (page_count(virt_to_page(ptep)) == 1)
5443                 return 0;
5444
5445         pud_clear(pud);
5446         put_page(virt_to_page(ptep));
5447         mm_dec_nr_pmds(mm);
5448         *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
5449         return 1;
5450 }
5451 #define want_pmd_share()        (1)
5452 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5453 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
5454 {
5455         return NULL;
5456 }
5457
5458 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
5459                                 unsigned long *addr, pte_t *ptep)
5460 {
5461         return 0;
5462 }
5463
5464 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5465                                 unsigned long *start, unsigned long *end)
5466 {
5467 }
5468 #define want_pmd_share()        (0)
5469 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5470
5471 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
5472 pte_t *huge_pte_alloc(struct mm_struct *mm,
5473                         unsigned long addr, unsigned long sz)
5474 {
5475         pgd_t *pgd;
5476         p4d_t *p4d;
5477         pud_t *pud;
5478         pte_t *pte = NULL;
5479
5480         pgd = pgd_offset(mm, addr);
5481         p4d = p4d_alloc(mm, pgd, addr);
5482         if (!p4d)
5483                 return NULL;
5484         pud = pud_alloc(mm, p4d, addr);
5485         if (pud) {
5486                 if (sz == PUD_SIZE) {
5487                         pte = (pte_t *)pud;
5488                 } else {
5489                         BUG_ON(sz != PMD_SIZE);
5490                         if (want_pmd_share() && pud_none(*pud))
5491                                 pte = huge_pmd_share(mm, addr, pud);
5492                         else
5493                                 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5494                 }
5495         }
5496         BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
5497
5498         return pte;
5499 }
5500
5501 /*
5502  * huge_pte_offset() - Walk the page table to resolve the hugepage
5503  * entry at address @addr
5504  *
5505  * Return: Pointer to page table entry (PUD or PMD) for
5506  * address @addr, or NULL if a !p*d_present() entry is encountered and the
5507  * size @sz doesn't match the hugepage size at this level of the page
5508  * table.
5509  */
5510 pte_t *huge_pte_offset(struct mm_struct *mm,
5511                        unsigned long addr, unsigned long sz)
5512 {
5513         pgd_t *pgd;
5514         p4d_t *p4d;
5515         pud_t *pud;
5516         pmd_t *pmd;
5517
5518         pgd = pgd_offset(mm, addr);
5519         if (!pgd_present(*pgd))
5520                 return NULL;
5521         p4d = p4d_offset(pgd, addr);
5522         if (!p4d_present(*p4d))
5523                 return NULL;
5524
5525         pud = pud_offset(p4d, addr);
5526         if (sz == PUD_SIZE)
5527                 /* must be pud huge, non-present or none */
5528                 return (pte_t *)pud;
5529         if (!pud_present(*pud))
5530                 return NULL;
5531         /* must have a valid entry and size to go further */
5532
5533         pmd = pmd_offset(pud, addr);
5534         /* must be pmd huge, non-present or none */
5535         return (pte_t *)pmd;
5536 }
5537
5538 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
5539
5540 /*
5541  * These functions are overwritable if your architecture needs its own
5542  * behavior.
5543  */
5544 struct page * __weak
5545 follow_huge_addr(struct mm_struct *mm, unsigned long address,
5546                               int write)
5547 {
5548         return ERR_PTR(-EINVAL);
5549 }
5550
5551 struct page * __weak
5552 follow_huge_pd(struct vm_area_struct *vma,
5553                unsigned long address, hugepd_t hpd, int flags, int pdshift)
5554 {
5555         WARN(1, "hugepd follow called with no support for hugepage directory format\n");
5556         return NULL;
5557 }
5558
5559 struct page * __weak
5560 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
5561                 pmd_t *pmd, int flags)
5562 {
5563         struct page *page = NULL;
5564         spinlock_t *ptl;
5565         pte_t pte;
5566
5567         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
5568         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
5569                          (FOLL_PIN | FOLL_GET)))
5570                 return NULL;
5571
5572 retry:
5573         ptl = pmd_lockptr(mm, pmd);
5574         spin_lock(ptl);
5575         /*
5576          * make sure that the address range covered by this pmd is not
5577          * unmapped from other threads.
5578          */
5579         if (!pmd_huge(*pmd))
5580                 goto out;
5581         pte = huge_ptep_get((pte_t *)pmd);
5582         if (pte_present(pte)) {
5583                 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
5584                 /*
5585                  * try_grab_page() should always succeed here, because: a) we
5586                  * hold the pmd (ptl) lock, and b) we've just checked that the
5587                  * huge pmd (head) page is present in the page tables. The ptl
5588                  * prevents the head page and tail pages from being rearranged
5589                  * in any way. So this page must be available at this point,
5590                  * unless the page refcount overflowed:
5591                  */
5592                 if (WARN_ON_ONCE(!try_grab_page(page, flags))) {
5593                         page = NULL;
5594                         goto out;
5595                 }
5596         } else {
5597                 if (is_hugetlb_entry_migration(pte)) {
5598                         spin_unlock(ptl);
5599                         __migration_entry_wait(mm, (pte_t *)pmd, ptl);
5600                         goto retry;
5601                 }
5602                 /*
5603                  * hwpoisoned entry is treated as no_page_table in
5604                  * follow_page_mask().
5605                  */
5606         }
5607 out:
5608         spin_unlock(ptl);
5609         return page;
5610 }
5611
5612 struct page * __weak
5613 follow_huge_pud(struct mm_struct *mm, unsigned long address,
5614                 pud_t *pud, int flags)
5615 {
5616         if (flags & (FOLL_GET | FOLL_PIN))
5617                 return NULL;
5618
5619         return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
5620 }
5621
5622 struct page * __weak
5623 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
5624 {
5625         if (flags & (FOLL_GET | FOLL_PIN))
5626                 return NULL;
5627
5628         return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
5629 }
5630
5631 bool isolate_huge_page(struct page *page, struct list_head *list)
5632 {
5633         bool ret = true;
5634
5635         spin_lock(&hugetlb_lock);
5636         if (!PageHeadHuge(page) ||
5637             !HPageMigratable(page) ||
5638             !get_page_unless_zero(page)) {
5639                 ret = false;
5640                 goto unlock;
5641         }
5642         ClearHPageMigratable(page);
5643         list_move_tail(&page->lru, list);
5644 unlock:
5645         spin_unlock(&hugetlb_lock);
5646         return ret;
5647 }
5648
5649 void putback_active_hugepage(struct page *page)
5650 {
5651         spin_lock(&hugetlb_lock);
5652         SetHPageMigratable(page);
5653         list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
5654         spin_unlock(&hugetlb_lock);
5655         put_page(page);
5656 }
5657
5658 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
5659 {
5660         struct hstate *h = page_hstate(oldpage);
5661
5662         hugetlb_cgroup_migrate(oldpage, newpage);
5663         set_page_owner_migrate_reason(newpage, reason);
5664
5665         /*
5666          * transfer temporary state of the new huge page. This is
5667          * reverse to other transitions because the newpage is going to
5668          * be final while the old one will be freed so it takes over
5669          * the temporary status.
5670          *
5671          * Also note that we have to transfer the per-node surplus state
5672          * here as well otherwise the global surplus count will not match
5673          * the per-node's.
5674          */
5675         if (HPageTemporary(newpage)) {
5676                 int old_nid = page_to_nid(oldpage);
5677                 int new_nid = page_to_nid(newpage);
5678
5679                 SetHPageTemporary(oldpage);
5680                 ClearHPageTemporary(newpage);
5681
5682                 spin_lock(&hugetlb_lock);
5683                 if (h->surplus_huge_pages_node[old_nid]) {
5684                         h->surplus_huge_pages_node[old_nid]--;
5685                         h->surplus_huge_pages_node[new_nid]++;
5686                 }
5687                 spin_unlock(&hugetlb_lock);
5688         }
5689 }
5690
5691 #ifdef CONFIG_CMA
5692 static bool cma_reserve_called __initdata;
5693
5694 static int __init cmdline_parse_hugetlb_cma(char *p)
5695 {
5696         hugetlb_cma_size = memparse(p, &p);
5697         return 0;
5698 }
5699
5700 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
5701
5702 void __init hugetlb_cma_reserve(int order)
5703 {
5704         unsigned long size, reserved, per_node;
5705         int nid;
5706
5707         cma_reserve_called = true;
5708
5709         if (!hugetlb_cma_size)
5710                 return;
5711
5712         if (hugetlb_cma_size < (PAGE_SIZE << order)) {
5713                 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
5714                         (PAGE_SIZE << order) / SZ_1M);
5715                 return;
5716         }
5717
5718         /*
5719          * If 3 GB area is requested on a machine with 4 numa nodes,
5720          * let's allocate 1 GB on first three nodes and ignore the last one.
5721          */
5722         per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
5723         pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
5724                 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
5725
5726         reserved = 0;
5727         for_each_node_state(nid, N_ONLINE) {
5728                 int res;
5729                 char name[CMA_MAX_NAME];
5730
5731                 size = min(per_node, hugetlb_cma_size - reserved);
5732                 size = round_up(size, PAGE_SIZE << order);
5733
5734                 snprintf(name, sizeof(name), "hugetlb%d", nid);
5735                 res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order,
5736                                                  0, false, name,
5737                                                  &hugetlb_cma[nid], nid);
5738                 if (res) {
5739                         pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
5740                                 res, nid);
5741                         continue;
5742                 }
5743
5744                 reserved += size;
5745                 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
5746                         size / SZ_1M, nid);
5747
5748                 if (reserved >= hugetlb_cma_size)
5749                         break;
5750         }
5751 }
5752
5753 void __init hugetlb_cma_check(void)
5754 {
5755         if (!hugetlb_cma_size || cma_reserve_called)
5756                 return;
5757
5758         pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
5759 }
5760
5761 #endif /* CONFIG_CMA */