ASoC: samsung: i2s: Move registers cache to common driver data structure
[platform/kernel/linux-exynos.git] / mm / hugetlb.c
1 /*
2  * Generic hugetlb support.
3  * (C) Nadia Yvette Chambers, April 2004
4  */
5 #include <linux/list.h>
6 #include <linux/init.h>
7 #include <linux/mm.h>
8 #include <linux/seq_file.h>
9 #include <linux/sysctl.h>
10 #include <linux/highmem.h>
11 #include <linux/mmu_notifier.h>
12 #include <linux/nodemask.h>
13 #include <linux/pagemap.h>
14 #include <linux/mempolicy.h>
15 #include <linux/compiler.h>
16 #include <linux/cpuset.h>
17 #include <linux/mutex.h>
18 #include <linux/bootmem.h>
19 #include <linux/sysfs.h>
20 #include <linux/slab.h>
21 #include <linux/mmdebug.h>
22 #include <linux/sched/signal.h>
23 #include <linux/rmap.h>
24 #include <linux/string_helpers.h>
25 #include <linux/swap.h>
26 #include <linux/swapops.h>
27 #include <linux/jhash.h>
28
29 #include <asm/page.h>
30 #include <asm/pgtable.h>
31 #include <asm/tlb.h>
32
33 #include <linux/io.h>
34 #include <linux/hugetlb.h>
35 #include <linux/hugetlb_cgroup.h>
36 #include <linux/node.h>
37 #include <linux/userfaultfd_k.h>
38 #include "internal.h"
39
40 int hugepages_treat_as_movable;
41
42 int hugetlb_max_hstate __read_mostly;
43 unsigned int default_hstate_idx;
44 struct hstate hstates[HUGE_MAX_HSTATE];
45 /*
46  * Minimum page order among possible hugepage sizes, set to a proper value
47  * at boot time.
48  */
49 static unsigned int minimum_order __read_mostly = UINT_MAX;
50
51 __initdata LIST_HEAD(huge_boot_pages);
52
53 /* for command line parsing */
54 static struct hstate * __initdata parsed_hstate;
55 static unsigned long __initdata default_hstate_max_huge_pages;
56 static unsigned long __initdata default_hstate_size;
57 static bool __initdata parsed_valid_hugepagesz = true;
58
59 /*
60  * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
61  * free_huge_pages, and surplus_huge_pages.
62  */
63 DEFINE_SPINLOCK(hugetlb_lock);
64
65 /*
66  * Serializes faults on the same logical page.  This is used to
67  * prevent spurious OOMs when the hugepage pool is fully utilized.
68  */
69 static int num_fault_mutexes;
70 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
71
72 /* Forward declaration */
73 static int hugetlb_acct_memory(struct hstate *h, long delta);
74
75 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
76 {
77         bool free = (spool->count == 0) && (spool->used_hpages == 0);
78
79         spin_unlock(&spool->lock);
80
81         /* If no pages are used, and no other handles to the subpool
82          * remain, give up any reservations mased on minimum size and
83          * free the subpool */
84         if (free) {
85                 if (spool->min_hpages != -1)
86                         hugetlb_acct_memory(spool->hstate,
87                                                 -spool->min_hpages);
88                 kfree(spool);
89         }
90 }
91
92 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
93                                                 long min_hpages)
94 {
95         struct hugepage_subpool *spool;
96
97         spool = kzalloc(sizeof(*spool), GFP_KERNEL);
98         if (!spool)
99                 return NULL;
100
101         spin_lock_init(&spool->lock);
102         spool->count = 1;
103         spool->max_hpages = max_hpages;
104         spool->hstate = h;
105         spool->min_hpages = min_hpages;
106
107         if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
108                 kfree(spool);
109                 return NULL;
110         }
111         spool->rsv_hpages = min_hpages;
112
113         return spool;
114 }
115
116 void hugepage_put_subpool(struct hugepage_subpool *spool)
117 {
118         spin_lock(&spool->lock);
119         BUG_ON(!spool->count);
120         spool->count--;
121         unlock_or_release_subpool(spool);
122 }
123
124 /*
125  * Subpool accounting for allocating and reserving pages.
126  * Return -ENOMEM if there are not enough resources to satisfy the
127  * the request.  Otherwise, return the number of pages by which the
128  * global pools must be adjusted (upward).  The returned value may
129  * only be different than the passed value (delta) in the case where
130  * a subpool minimum size must be manitained.
131  */
132 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
133                                       long delta)
134 {
135         long ret = delta;
136
137         if (!spool)
138                 return ret;
139
140         spin_lock(&spool->lock);
141
142         if (spool->max_hpages != -1) {          /* maximum size accounting */
143                 if ((spool->used_hpages + delta) <= spool->max_hpages)
144                         spool->used_hpages += delta;
145                 else {
146                         ret = -ENOMEM;
147                         goto unlock_ret;
148                 }
149         }
150
151         /* minimum size accounting */
152         if (spool->min_hpages != -1 && spool->rsv_hpages) {
153                 if (delta > spool->rsv_hpages) {
154                         /*
155                          * Asking for more reserves than those already taken on
156                          * behalf of subpool.  Return difference.
157                          */
158                         ret = delta - spool->rsv_hpages;
159                         spool->rsv_hpages = 0;
160                 } else {
161                         ret = 0;        /* reserves already accounted for */
162                         spool->rsv_hpages -= delta;
163                 }
164         }
165
166 unlock_ret:
167         spin_unlock(&spool->lock);
168         return ret;
169 }
170
171 /*
172  * Subpool accounting for freeing and unreserving pages.
173  * Return the number of global page reservations that must be dropped.
174  * The return value may only be different than the passed value (delta)
175  * in the case where a subpool minimum size must be maintained.
176  */
177 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
178                                        long delta)
179 {
180         long ret = delta;
181
182         if (!spool)
183                 return delta;
184
185         spin_lock(&spool->lock);
186
187         if (spool->max_hpages != -1)            /* maximum size accounting */
188                 spool->used_hpages -= delta;
189
190          /* minimum size accounting */
191         if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
192                 if (spool->rsv_hpages + delta <= spool->min_hpages)
193                         ret = 0;
194                 else
195                         ret = spool->rsv_hpages + delta - spool->min_hpages;
196
197                 spool->rsv_hpages += delta;
198                 if (spool->rsv_hpages > spool->min_hpages)
199                         spool->rsv_hpages = spool->min_hpages;
200         }
201
202         /*
203          * If hugetlbfs_put_super couldn't free spool due to an outstanding
204          * quota reference, free it now.
205          */
206         unlock_or_release_subpool(spool);
207
208         return ret;
209 }
210
211 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
212 {
213         return HUGETLBFS_SB(inode->i_sb)->spool;
214 }
215
216 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
217 {
218         return subpool_inode(file_inode(vma->vm_file));
219 }
220
221 /*
222  * Region tracking -- allows tracking of reservations and instantiated pages
223  *                    across the pages in a mapping.
224  *
225  * The region data structures are embedded into a resv_map and protected
226  * by a resv_map's lock.  The set of regions within the resv_map represent
227  * reservations for huge pages, or huge pages that have already been
228  * instantiated within the map.  The from and to elements are huge page
229  * indicies into the associated mapping.  from indicates the starting index
230  * of the region.  to represents the first index past the end of  the region.
231  *
232  * For example, a file region structure with from == 0 and to == 4 represents
233  * four huge pages in a mapping.  It is important to note that the to element
234  * represents the first element past the end of the region. This is used in
235  * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
236  *
237  * Interval notation of the form [from, to) will be used to indicate that
238  * the endpoint from is inclusive and to is exclusive.
239  */
240 struct file_region {
241         struct list_head link;
242         long from;
243         long to;
244 };
245
246 /*
247  * Add the huge page range represented by [f, t) to the reserve
248  * map.  In the normal case, existing regions will be expanded
249  * to accommodate the specified range.  Sufficient regions should
250  * exist for expansion due to the previous call to region_chg
251  * with the same range.  However, it is possible that region_del
252  * could have been called after region_chg and modifed the map
253  * in such a way that no region exists to be expanded.  In this
254  * case, pull a region descriptor from the cache associated with
255  * the map and use that for the new range.
256  *
257  * Return the number of new huge pages added to the map.  This
258  * number is greater than or equal to zero.
259  */
260 static long region_add(struct resv_map *resv, long f, long t)
261 {
262         struct list_head *head = &resv->regions;
263         struct file_region *rg, *nrg, *trg;
264         long add = 0;
265
266         spin_lock(&resv->lock);
267         /* Locate the region we are either in or before. */
268         list_for_each_entry(rg, head, link)
269                 if (f <= rg->to)
270                         break;
271
272         /*
273          * If no region exists which can be expanded to include the
274          * specified range, the list must have been modified by an
275          * interleving call to region_del().  Pull a region descriptor
276          * from the cache and use it for this range.
277          */
278         if (&rg->link == head || t < rg->from) {
279                 VM_BUG_ON(resv->region_cache_count <= 0);
280
281                 resv->region_cache_count--;
282                 nrg = list_first_entry(&resv->region_cache, struct file_region,
283                                         link);
284                 list_del(&nrg->link);
285
286                 nrg->from = f;
287                 nrg->to = t;
288                 list_add(&nrg->link, rg->link.prev);
289
290                 add += t - f;
291                 goto out_locked;
292         }
293
294         /* Round our left edge to the current segment if it encloses us. */
295         if (f > rg->from)
296                 f = rg->from;
297
298         /* Check for and consume any regions we now overlap with. */
299         nrg = rg;
300         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
301                 if (&rg->link == head)
302                         break;
303                 if (rg->from > t)
304                         break;
305
306                 /* If this area reaches higher then extend our area to
307                  * include it completely.  If this is not the first area
308                  * which we intend to reuse, free it. */
309                 if (rg->to > t)
310                         t = rg->to;
311                 if (rg != nrg) {
312                         /* Decrement return value by the deleted range.
313                          * Another range will span this area so that by
314                          * end of routine add will be >= zero
315                          */
316                         add -= (rg->to - rg->from);
317                         list_del(&rg->link);
318                         kfree(rg);
319                 }
320         }
321
322         add += (nrg->from - f);         /* Added to beginning of region */
323         nrg->from = f;
324         add += t - nrg->to;             /* Added to end of region */
325         nrg->to = t;
326
327 out_locked:
328         resv->adds_in_progress--;
329         spin_unlock(&resv->lock);
330         VM_BUG_ON(add < 0);
331         return add;
332 }
333
334 /*
335  * Examine the existing reserve map and determine how many
336  * huge pages in the specified range [f, t) are NOT currently
337  * represented.  This routine is called before a subsequent
338  * call to region_add that will actually modify the reserve
339  * map to add the specified range [f, t).  region_chg does
340  * not change the number of huge pages represented by the
341  * map.  However, if the existing regions in the map can not
342  * be expanded to represent the new range, a new file_region
343  * structure is added to the map as a placeholder.  This is
344  * so that the subsequent region_add call will have all the
345  * regions it needs and will not fail.
346  *
347  * Upon entry, region_chg will also examine the cache of region descriptors
348  * associated with the map.  If there are not enough descriptors cached, one
349  * will be allocated for the in progress add operation.
350  *
351  * Returns the number of huge pages that need to be added to the existing
352  * reservation map for the range [f, t).  This number is greater or equal to
353  * zero.  -ENOMEM is returned if a new file_region structure or cache entry
354  * is needed and can not be allocated.
355  */
356 static long region_chg(struct resv_map *resv, long f, long t)
357 {
358         struct list_head *head = &resv->regions;
359         struct file_region *rg, *nrg = NULL;
360         long chg = 0;
361
362 retry:
363         spin_lock(&resv->lock);
364 retry_locked:
365         resv->adds_in_progress++;
366
367         /*
368          * Check for sufficient descriptors in the cache to accommodate
369          * the number of in progress add operations.
370          */
371         if (resv->adds_in_progress > resv->region_cache_count) {
372                 struct file_region *trg;
373
374                 VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1);
375                 /* Must drop lock to allocate a new descriptor. */
376                 resv->adds_in_progress--;
377                 spin_unlock(&resv->lock);
378
379                 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
380                 if (!trg) {
381                         kfree(nrg);
382                         return -ENOMEM;
383                 }
384
385                 spin_lock(&resv->lock);
386                 list_add(&trg->link, &resv->region_cache);
387                 resv->region_cache_count++;
388                 goto retry_locked;
389         }
390
391         /* Locate the region we are before or in. */
392         list_for_each_entry(rg, head, link)
393                 if (f <= rg->to)
394                         break;
395
396         /* If we are below the current region then a new region is required.
397          * Subtle, allocate a new region at the position but make it zero
398          * size such that we can guarantee to record the reservation. */
399         if (&rg->link == head || t < rg->from) {
400                 if (!nrg) {
401                         resv->adds_in_progress--;
402                         spin_unlock(&resv->lock);
403                         nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
404                         if (!nrg)
405                                 return -ENOMEM;
406
407                         nrg->from = f;
408                         nrg->to   = f;
409                         INIT_LIST_HEAD(&nrg->link);
410                         goto retry;
411                 }
412
413                 list_add(&nrg->link, rg->link.prev);
414                 chg = t - f;
415                 goto out_nrg;
416         }
417
418         /* Round our left edge to the current segment if it encloses us. */
419         if (f > rg->from)
420                 f = rg->from;
421         chg = t - f;
422
423         /* Check for and consume any regions we now overlap with. */
424         list_for_each_entry(rg, rg->link.prev, link) {
425                 if (&rg->link == head)
426                         break;
427                 if (rg->from > t)
428                         goto out;
429
430                 /* We overlap with this area, if it extends further than
431                  * us then we must extend ourselves.  Account for its
432                  * existing reservation. */
433                 if (rg->to > t) {
434                         chg += rg->to - t;
435                         t = rg->to;
436                 }
437                 chg -= rg->to - rg->from;
438         }
439
440 out:
441         spin_unlock(&resv->lock);
442         /*  We already know we raced and no longer need the new region */
443         kfree(nrg);
444         return chg;
445 out_nrg:
446         spin_unlock(&resv->lock);
447         return chg;
448 }
449
450 /*
451  * Abort the in progress add operation.  The adds_in_progress field
452  * of the resv_map keeps track of the operations in progress between
453  * calls to region_chg and region_add.  Operations are sometimes
454  * aborted after the call to region_chg.  In such cases, region_abort
455  * is called to decrement the adds_in_progress counter.
456  *
457  * NOTE: The range arguments [f, t) are not needed or used in this
458  * routine.  They are kept to make reading the calling code easier as
459  * arguments will match the associated region_chg call.
460  */
461 static void region_abort(struct resv_map *resv, long f, long t)
462 {
463         spin_lock(&resv->lock);
464         VM_BUG_ON(!resv->region_cache_count);
465         resv->adds_in_progress--;
466         spin_unlock(&resv->lock);
467 }
468
469 /*
470  * Delete the specified range [f, t) from the reserve map.  If the
471  * t parameter is LONG_MAX, this indicates that ALL regions after f
472  * should be deleted.  Locate the regions which intersect [f, t)
473  * and either trim, delete or split the existing regions.
474  *
475  * Returns the number of huge pages deleted from the reserve map.
476  * In the normal case, the return value is zero or more.  In the
477  * case where a region must be split, a new region descriptor must
478  * be allocated.  If the allocation fails, -ENOMEM will be returned.
479  * NOTE: If the parameter t == LONG_MAX, then we will never split
480  * a region and possibly return -ENOMEM.  Callers specifying
481  * t == LONG_MAX do not need to check for -ENOMEM error.
482  */
483 static long region_del(struct resv_map *resv, long f, long t)
484 {
485         struct list_head *head = &resv->regions;
486         struct file_region *rg, *trg;
487         struct file_region *nrg = NULL;
488         long del = 0;
489
490 retry:
491         spin_lock(&resv->lock);
492         list_for_each_entry_safe(rg, trg, head, link) {
493                 /*
494                  * Skip regions before the range to be deleted.  file_region
495                  * ranges are normally of the form [from, to).  However, there
496                  * may be a "placeholder" entry in the map which is of the form
497                  * (from, to) with from == to.  Check for placeholder entries
498                  * at the beginning of the range to be deleted.
499                  */
500                 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
501                         continue;
502
503                 if (rg->from >= t)
504                         break;
505
506                 if (f > rg->from && t < rg->to) { /* Must split region */
507                         /*
508                          * Check for an entry in the cache before dropping
509                          * lock and attempting allocation.
510                          */
511                         if (!nrg &&
512                             resv->region_cache_count > resv->adds_in_progress) {
513                                 nrg = list_first_entry(&resv->region_cache,
514                                                         struct file_region,
515                                                         link);
516                                 list_del(&nrg->link);
517                                 resv->region_cache_count--;
518                         }
519
520                         if (!nrg) {
521                                 spin_unlock(&resv->lock);
522                                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
523                                 if (!nrg)
524                                         return -ENOMEM;
525                                 goto retry;
526                         }
527
528                         del += t - f;
529
530                         /* New entry for end of split region */
531                         nrg->from = t;
532                         nrg->to = rg->to;
533                         INIT_LIST_HEAD(&nrg->link);
534
535                         /* Original entry is trimmed */
536                         rg->to = f;
537
538                         list_add(&nrg->link, &rg->link);
539                         nrg = NULL;
540                         break;
541                 }
542
543                 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
544                         del += rg->to - rg->from;
545                         list_del(&rg->link);
546                         kfree(rg);
547                         continue;
548                 }
549
550                 if (f <= rg->from) {    /* Trim beginning of region */
551                         del += t - rg->from;
552                         rg->from = t;
553                 } else {                /* Trim end of region */
554                         del += rg->to - f;
555                         rg->to = f;
556                 }
557         }
558
559         spin_unlock(&resv->lock);
560         kfree(nrg);
561         return del;
562 }
563
564 /*
565  * A rare out of memory error was encountered which prevented removal of
566  * the reserve map region for a page.  The huge page itself was free'ed
567  * and removed from the page cache.  This routine will adjust the subpool
568  * usage count, and the global reserve count if needed.  By incrementing
569  * these counts, the reserve map entry which could not be deleted will
570  * appear as a "reserved" entry instead of simply dangling with incorrect
571  * counts.
572  */
573 void hugetlb_fix_reserve_counts(struct inode *inode)
574 {
575         struct hugepage_subpool *spool = subpool_inode(inode);
576         long rsv_adjust;
577
578         rsv_adjust = hugepage_subpool_get_pages(spool, 1);
579         if (rsv_adjust) {
580                 struct hstate *h = hstate_inode(inode);
581
582                 hugetlb_acct_memory(h, 1);
583         }
584 }
585
586 /*
587  * Count and return the number of huge pages in the reserve map
588  * that intersect with the range [f, t).
589  */
590 static long region_count(struct resv_map *resv, long f, long t)
591 {
592         struct list_head *head = &resv->regions;
593         struct file_region *rg;
594         long chg = 0;
595
596         spin_lock(&resv->lock);
597         /* Locate each segment we overlap with, and count that overlap. */
598         list_for_each_entry(rg, head, link) {
599                 long seg_from;
600                 long seg_to;
601
602                 if (rg->to <= f)
603                         continue;
604                 if (rg->from >= t)
605                         break;
606
607                 seg_from = max(rg->from, f);
608                 seg_to = min(rg->to, t);
609
610                 chg += seg_to - seg_from;
611         }
612         spin_unlock(&resv->lock);
613
614         return chg;
615 }
616
617 /*
618  * Convert the address within this vma to the page offset within
619  * the mapping, in pagecache page units; huge pages here.
620  */
621 static pgoff_t vma_hugecache_offset(struct hstate *h,
622                         struct vm_area_struct *vma, unsigned long address)
623 {
624         return ((address - vma->vm_start) >> huge_page_shift(h)) +
625                         (vma->vm_pgoff >> huge_page_order(h));
626 }
627
628 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
629                                      unsigned long address)
630 {
631         return vma_hugecache_offset(hstate_vma(vma), vma, address);
632 }
633 EXPORT_SYMBOL_GPL(linear_hugepage_index);
634
635 /*
636  * Return the size of the pages allocated when backing a VMA. In the majority
637  * cases this will be same size as used by the page table entries.
638  */
639 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
640 {
641         struct hstate *hstate;
642
643         if (!is_vm_hugetlb_page(vma))
644                 return PAGE_SIZE;
645
646         hstate = hstate_vma(vma);
647
648         return 1UL << huge_page_shift(hstate);
649 }
650 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
651
652 /*
653  * Return the page size being used by the MMU to back a VMA. In the majority
654  * of cases, the page size used by the kernel matches the MMU size. On
655  * architectures where it differs, an architecture-specific version of this
656  * function is required.
657  */
658 #ifndef vma_mmu_pagesize
659 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
660 {
661         return vma_kernel_pagesize(vma);
662 }
663 #endif
664
665 /*
666  * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
667  * bits of the reservation map pointer, which are always clear due to
668  * alignment.
669  */
670 #define HPAGE_RESV_OWNER    (1UL << 0)
671 #define HPAGE_RESV_UNMAPPED (1UL << 1)
672 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
673
674 /*
675  * These helpers are used to track how many pages are reserved for
676  * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
677  * is guaranteed to have their future faults succeed.
678  *
679  * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
680  * the reserve counters are updated with the hugetlb_lock held. It is safe
681  * to reset the VMA at fork() time as it is not in use yet and there is no
682  * chance of the global counters getting corrupted as a result of the values.
683  *
684  * The private mapping reservation is represented in a subtly different
685  * manner to a shared mapping.  A shared mapping has a region map associated
686  * with the underlying file, this region map represents the backing file
687  * pages which have ever had a reservation assigned which this persists even
688  * after the page is instantiated.  A private mapping has a region map
689  * associated with the original mmap which is attached to all VMAs which
690  * reference it, this region map represents those offsets which have consumed
691  * reservation ie. where pages have been instantiated.
692  */
693 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
694 {
695         return (unsigned long)vma->vm_private_data;
696 }
697
698 static void set_vma_private_data(struct vm_area_struct *vma,
699                                                         unsigned long value)
700 {
701         vma->vm_private_data = (void *)value;
702 }
703
704 struct resv_map *resv_map_alloc(void)
705 {
706         struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
707         struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
708
709         if (!resv_map || !rg) {
710                 kfree(resv_map);
711                 kfree(rg);
712                 return NULL;
713         }
714
715         kref_init(&resv_map->refs);
716         spin_lock_init(&resv_map->lock);
717         INIT_LIST_HEAD(&resv_map->regions);
718
719         resv_map->adds_in_progress = 0;
720
721         INIT_LIST_HEAD(&resv_map->region_cache);
722         list_add(&rg->link, &resv_map->region_cache);
723         resv_map->region_cache_count = 1;
724
725         return resv_map;
726 }
727
728 void resv_map_release(struct kref *ref)
729 {
730         struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
731         struct list_head *head = &resv_map->region_cache;
732         struct file_region *rg, *trg;
733
734         /* Clear out any active regions before we release the map. */
735         region_del(resv_map, 0, LONG_MAX);
736
737         /* ... and any entries left in the cache */
738         list_for_each_entry_safe(rg, trg, head, link) {
739                 list_del(&rg->link);
740                 kfree(rg);
741         }
742
743         VM_BUG_ON(resv_map->adds_in_progress);
744
745         kfree(resv_map);
746 }
747
748 static inline struct resv_map *inode_resv_map(struct inode *inode)
749 {
750         return inode->i_mapping->private_data;
751 }
752
753 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
754 {
755         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
756         if (vma->vm_flags & VM_MAYSHARE) {
757                 struct address_space *mapping = vma->vm_file->f_mapping;
758                 struct inode *inode = mapping->host;
759
760                 return inode_resv_map(inode);
761
762         } else {
763                 return (struct resv_map *)(get_vma_private_data(vma) &
764                                                         ~HPAGE_RESV_MASK);
765         }
766 }
767
768 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
769 {
770         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
771         VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
772
773         set_vma_private_data(vma, (get_vma_private_data(vma) &
774                                 HPAGE_RESV_MASK) | (unsigned long)map);
775 }
776
777 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
778 {
779         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
780         VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
781
782         set_vma_private_data(vma, get_vma_private_data(vma) | flags);
783 }
784
785 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
786 {
787         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
788
789         return (get_vma_private_data(vma) & flag) != 0;
790 }
791
792 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
793 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
794 {
795         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
796         if (!(vma->vm_flags & VM_MAYSHARE))
797                 vma->vm_private_data = (void *)0;
798 }
799
800 /* Returns true if the VMA has associated reserve pages */
801 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
802 {
803         if (vma->vm_flags & VM_NORESERVE) {
804                 /*
805                  * This address is already reserved by other process(chg == 0),
806                  * so, we should decrement reserved count. Without decrementing,
807                  * reserve count remains after releasing inode, because this
808                  * allocated page will go into page cache and is regarded as
809                  * coming from reserved pool in releasing step.  Currently, we
810                  * don't have any other solution to deal with this situation
811                  * properly, so add work-around here.
812                  */
813                 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
814                         return true;
815                 else
816                         return false;
817         }
818
819         /* Shared mappings always use reserves */
820         if (vma->vm_flags & VM_MAYSHARE) {
821                 /*
822                  * We know VM_NORESERVE is not set.  Therefore, there SHOULD
823                  * be a region map for all pages.  The only situation where
824                  * there is no region map is if a hole was punched via
825                  * fallocate.  In this case, there really are no reverves to
826                  * use.  This situation is indicated if chg != 0.
827                  */
828                 if (chg)
829                         return false;
830                 else
831                         return true;
832         }
833
834         /*
835          * Only the process that called mmap() has reserves for
836          * private mappings.
837          */
838         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
839                 /*
840                  * Like the shared case above, a hole punch or truncate
841                  * could have been performed on the private mapping.
842                  * Examine the value of chg to determine if reserves
843                  * actually exist or were previously consumed.
844                  * Very Subtle - The value of chg comes from a previous
845                  * call to vma_needs_reserves().  The reserve map for
846                  * private mappings has different (opposite) semantics
847                  * than that of shared mappings.  vma_needs_reserves()
848                  * has already taken this difference in semantics into
849                  * account.  Therefore, the meaning of chg is the same
850                  * as in the shared case above.  Code could easily be
851                  * combined, but keeping it separate draws attention to
852                  * subtle differences.
853                  */
854                 if (chg)
855                         return false;
856                 else
857                         return true;
858         }
859
860         return false;
861 }
862
863 static void enqueue_huge_page(struct hstate *h, struct page *page)
864 {
865         int nid = page_to_nid(page);
866         list_move(&page->lru, &h->hugepage_freelists[nid]);
867         h->free_huge_pages++;
868         h->free_huge_pages_node[nid]++;
869 }
870
871 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
872 {
873         struct page *page;
874
875         list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
876                 if (!PageHWPoison(page))
877                         break;
878         /*
879          * if 'non-isolated free hugepage' not found on the list,
880          * the allocation fails.
881          */
882         if (&h->hugepage_freelists[nid] == &page->lru)
883                 return NULL;
884         list_move(&page->lru, &h->hugepage_activelist);
885         set_page_refcounted(page);
886         h->free_huge_pages--;
887         h->free_huge_pages_node[nid]--;
888         return page;
889 }
890
891 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
892                 nodemask_t *nmask)
893 {
894         unsigned int cpuset_mems_cookie;
895         struct zonelist *zonelist;
896         struct zone *zone;
897         struct zoneref *z;
898         int node = -1;
899
900         zonelist = node_zonelist(nid, gfp_mask);
901
902 retry_cpuset:
903         cpuset_mems_cookie = read_mems_allowed_begin();
904         for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
905                 struct page *page;
906
907                 if (!cpuset_zone_allowed(zone, gfp_mask))
908                         continue;
909                 /*
910                  * no need to ask again on the same node. Pool is node rather than
911                  * zone aware
912                  */
913                 if (zone_to_nid(zone) == node)
914                         continue;
915                 node = zone_to_nid(zone);
916
917                 page = dequeue_huge_page_node_exact(h, node);
918                 if (page)
919                         return page;
920         }
921         if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
922                 goto retry_cpuset;
923
924         return NULL;
925 }
926
927 /* Movability of hugepages depends on migration support. */
928 static inline gfp_t htlb_alloc_mask(struct hstate *h)
929 {
930         if (hugepages_treat_as_movable || hugepage_migration_supported(h))
931                 return GFP_HIGHUSER_MOVABLE;
932         else
933                 return GFP_HIGHUSER;
934 }
935
936 static struct page *dequeue_huge_page_vma(struct hstate *h,
937                                 struct vm_area_struct *vma,
938                                 unsigned long address, int avoid_reserve,
939                                 long chg)
940 {
941         struct page *page;
942         struct mempolicy *mpol;
943         gfp_t gfp_mask;
944         nodemask_t *nodemask;
945         int nid;
946
947         /*
948          * A child process with MAP_PRIVATE mappings created by their parent
949          * have no page reserves. This check ensures that reservations are
950          * not "stolen". The child may still get SIGKILLed
951          */
952         if (!vma_has_reserves(vma, chg) &&
953                         h->free_huge_pages - h->resv_huge_pages == 0)
954                 goto err;
955
956         /* If reserves cannot be used, ensure enough pages are in the pool */
957         if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
958                 goto err;
959
960         gfp_mask = htlb_alloc_mask(h);
961         nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
962         page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
963         if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
964                 SetPagePrivate(page);
965                 h->resv_huge_pages--;
966         }
967
968         mpol_cond_put(mpol);
969         return page;
970
971 err:
972         return NULL;
973 }
974
975 /*
976  * common helper functions for hstate_next_node_to_{alloc|free}.
977  * We may have allocated or freed a huge page based on a different
978  * nodes_allowed previously, so h->next_node_to_{alloc|free} might
979  * be outside of *nodes_allowed.  Ensure that we use an allowed
980  * node for alloc or free.
981  */
982 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
983 {
984         nid = next_node_in(nid, *nodes_allowed);
985         VM_BUG_ON(nid >= MAX_NUMNODES);
986
987         return nid;
988 }
989
990 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
991 {
992         if (!node_isset(nid, *nodes_allowed))
993                 nid = next_node_allowed(nid, nodes_allowed);
994         return nid;
995 }
996
997 /*
998  * returns the previously saved node ["this node"] from which to
999  * allocate a persistent huge page for the pool and advance the
1000  * next node from which to allocate, handling wrap at end of node
1001  * mask.
1002  */
1003 static int hstate_next_node_to_alloc(struct hstate *h,
1004                                         nodemask_t *nodes_allowed)
1005 {
1006         int nid;
1007
1008         VM_BUG_ON(!nodes_allowed);
1009
1010         nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1011         h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1012
1013         return nid;
1014 }
1015
1016 /*
1017  * helper for free_pool_huge_page() - return the previously saved
1018  * node ["this node"] from which to free a huge page.  Advance the
1019  * next node id whether or not we find a free huge page to free so
1020  * that the next attempt to free addresses the next node.
1021  */
1022 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1023 {
1024         int nid;
1025
1026         VM_BUG_ON(!nodes_allowed);
1027
1028         nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1029         h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1030
1031         return nid;
1032 }
1033
1034 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask)           \
1035         for (nr_nodes = nodes_weight(*mask);                            \
1036                 nr_nodes > 0 &&                                         \
1037                 ((node = hstate_next_node_to_alloc(hs, mask)) || 1);    \
1038                 nr_nodes--)
1039
1040 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask)            \
1041         for (nr_nodes = nodes_weight(*mask);                            \
1042                 nr_nodes > 0 &&                                         \
1043                 ((node = hstate_next_node_to_free(hs, mask)) || 1);     \
1044                 nr_nodes--)
1045
1046 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1047 static void destroy_compound_gigantic_page(struct page *page,
1048                                         unsigned int order)
1049 {
1050         int i;
1051         int nr_pages = 1 << order;
1052         struct page *p = page + 1;
1053
1054         atomic_set(compound_mapcount_ptr(page), 0);
1055         for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1056                 clear_compound_head(p);
1057                 set_page_refcounted(p);
1058         }
1059
1060         set_compound_order(page, 0);
1061         __ClearPageHead(page);
1062 }
1063
1064 static void free_gigantic_page(struct page *page, unsigned int order)
1065 {
1066         free_contig_range(page_to_pfn(page), 1 << order);
1067 }
1068
1069 static int __alloc_gigantic_page(unsigned long start_pfn,
1070                                 unsigned long nr_pages, gfp_t gfp_mask)
1071 {
1072         unsigned long end_pfn = start_pfn + nr_pages;
1073         return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
1074                                   gfp_mask);
1075 }
1076
1077 static bool pfn_range_valid_gigantic(struct zone *z,
1078                         unsigned long start_pfn, unsigned long nr_pages)
1079 {
1080         unsigned long i, end_pfn = start_pfn + nr_pages;
1081         struct page *page;
1082
1083         for (i = start_pfn; i < end_pfn; i++) {
1084                 if (!pfn_valid(i))
1085                         return false;
1086
1087                 page = pfn_to_page(i);
1088
1089                 if (page_zone(page) != z)
1090                         return false;
1091
1092                 if (PageReserved(page))
1093                         return false;
1094
1095                 if (page_count(page) > 0)
1096                         return false;
1097
1098                 if (PageHuge(page))
1099                         return false;
1100         }
1101
1102         return true;
1103 }
1104
1105 static bool zone_spans_last_pfn(const struct zone *zone,
1106                         unsigned long start_pfn, unsigned long nr_pages)
1107 {
1108         unsigned long last_pfn = start_pfn + nr_pages - 1;
1109         return zone_spans_pfn(zone, last_pfn);
1110 }
1111
1112 static struct page *alloc_gigantic_page(int nid, struct hstate *h)
1113 {
1114         unsigned int order = huge_page_order(h);
1115         unsigned long nr_pages = 1 << order;
1116         unsigned long ret, pfn, flags;
1117         struct zonelist *zonelist;
1118         struct zone *zone;
1119         struct zoneref *z;
1120         gfp_t gfp_mask;
1121
1122         gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1123         zonelist = node_zonelist(nid, gfp_mask);
1124         for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), NULL) {
1125                 spin_lock_irqsave(&zone->lock, flags);
1126
1127                 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
1128                 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
1129                         if (pfn_range_valid_gigantic(zone, pfn, nr_pages)) {
1130                                 /*
1131                                  * We release the zone lock here because
1132                                  * alloc_contig_range() will also lock the zone
1133                                  * at some point. If there's an allocation
1134                                  * spinning on this lock, it may win the race
1135                                  * and cause alloc_contig_range() to fail...
1136                                  */
1137                                 spin_unlock_irqrestore(&zone->lock, flags);
1138                                 ret = __alloc_gigantic_page(pfn, nr_pages, gfp_mask);
1139                                 if (!ret)
1140                                         return pfn_to_page(pfn);
1141                                 spin_lock_irqsave(&zone->lock, flags);
1142                         }
1143                         pfn += nr_pages;
1144                 }
1145
1146                 spin_unlock_irqrestore(&zone->lock, flags);
1147         }
1148
1149         return NULL;
1150 }
1151
1152 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1153 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1154
1155 static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid)
1156 {
1157         struct page *page;
1158
1159         page = alloc_gigantic_page(nid, h);
1160         if (page) {
1161                 prep_compound_gigantic_page(page, huge_page_order(h));
1162                 prep_new_huge_page(h, page, nid);
1163         }
1164
1165         return page;
1166 }
1167
1168 static int alloc_fresh_gigantic_page(struct hstate *h,
1169                                 nodemask_t *nodes_allowed)
1170 {
1171         struct page *page = NULL;
1172         int nr_nodes, node;
1173
1174         for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1175                 page = alloc_fresh_gigantic_page_node(h, node);
1176                 if (page)
1177                         return 1;
1178         }
1179
1180         return 0;
1181 }
1182
1183 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1184 static inline bool gigantic_page_supported(void) { return false; }
1185 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1186 static inline void destroy_compound_gigantic_page(struct page *page,
1187                                                 unsigned int order) { }
1188 static inline int alloc_fresh_gigantic_page(struct hstate *h,
1189                                         nodemask_t *nodes_allowed) { return 0; }
1190 #endif
1191
1192 static void update_and_free_page(struct hstate *h, struct page *page)
1193 {
1194         int i;
1195
1196         if (hstate_is_gigantic(h) && !gigantic_page_supported())
1197                 return;
1198
1199         h->nr_huge_pages--;
1200         h->nr_huge_pages_node[page_to_nid(page)]--;
1201         for (i = 0; i < pages_per_huge_page(h); i++) {
1202                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1203                                 1 << PG_referenced | 1 << PG_dirty |
1204                                 1 << PG_active | 1 << PG_private |
1205                                 1 << PG_writeback);
1206         }
1207         VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1208         set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1209         set_page_refcounted(page);
1210         if (hstate_is_gigantic(h)) {
1211                 destroy_compound_gigantic_page(page, huge_page_order(h));
1212                 free_gigantic_page(page, huge_page_order(h));
1213         } else {
1214                 __free_pages(page, huge_page_order(h));
1215         }
1216 }
1217
1218 struct hstate *size_to_hstate(unsigned long size)
1219 {
1220         struct hstate *h;
1221
1222         for_each_hstate(h) {
1223                 if (huge_page_size(h) == size)
1224                         return h;
1225         }
1226         return NULL;
1227 }
1228
1229 /*
1230  * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1231  * to hstate->hugepage_activelist.)
1232  *
1233  * This function can be called for tail pages, but never returns true for them.
1234  */
1235 bool page_huge_active(struct page *page)
1236 {
1237         VM_BUG_ON_PAGE(!PageHuge(page), page);
1238         return PageHead(page) && PagePrivate(&page[1]);
1239 }
1240
1241 /* never called for tail page */
1242 static void set_page_huge_active(struct page *page)
1243 {
1244         VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1245         SetPagePrivate(&page[1]);
1246 }
1247
1248 static void clear_page_huge_active(struct page *page)
1249 {
1250         VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1251         ClearPagePrivate(&page[1]);
1252 }
1253
1254 void free_huge_page(struct page *page)
1255 {
1256         /*
1257          * Can't pass hstate in here because it is called from the
1258          * compound page destructor.
1259          */
1260         struct hstate *h = page_hstate(page);
1261         int nid = page_to_nid(page);
1262         struct hugepage_subpool *spool =
1263                 (struct hugepage_subpool *)page_private(page);
1264         bool restore_reserve;
1265
1266         set_page_private(page, 0);
1267         page->mapping = NULL;
1268         VM_BUG_ON_PAGE(page_count(page), page);
1269         VM_BUG_ON_PAGE(page_mapcount(page), page);
1270         restore_reserve = PagePrivate(page);
1271         ClearPagePrivate(page);
1272
1273         /*
1274          * A return code of zero implies that the subpool will be under its
1275          * minimum size if the reservation is not restored after page is free.
1276          * Therefore, force restore_reserve operation.
1277          */
1278         if (hugepage_subpool_put_pages(spool, 1) == 0)
1279                 restore_reserve = true;
1280
1281         spin_lock(&hugetlb_lock);
1282         clear_page_huge_active(page);
1283         hugetlb_cgroup_uncharge_page(hstate_index(h),
1284                                      pages_per_huge_page(h), page);
1285         if (restore_reserve)
1286                 h->resv_huge_pages++;
1287
1288         if (h->surplus_huge_pages_node[nid]) {
1289                 /* remove the page from active list */
1290                 list_del(&page->lru);
1291                 update_and_free_page(h, page);
1292                 h->surplus_huge_pages--;
1293                 h->surplus_huge_pages_node[nid]--;
1294         } else {
1295                 arch_clear_hugepage_flags(page);
1296                 enqueue_huge_page(h, page);
1297         }
1298         spin_unlock(&hugetlb_lock);
1299 }
1300
1301 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1302 {
1303         INIT_LIST_HEAD(&page->lru);
1304         set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1305         spin_lock(&hugetlb_lock);
1306         set_hugetlb_cgroup(page, NULL);
1307         h->nr_huge_pages++;
1308         h->nr_huge_pages_node[nid]++;
1309         spin_unlock(&hugetlb_lock);
1310         put_page(page); /* free it into the hugepage allocator */
1311 }
1312
1313 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1314 {
1315         int i;
1316         int nr_pages = 1 << order;
1317         struct page *p = page + 1;
1318
1319         /* we rely on prep_new_huge_page to set the destructor */
1320         set_compound_order(page, order);
1321         __ClearPageReserved(page);
1322         __SetPageHead(page);
1323         for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1324                 /*
1325                  * For gigantic hugepages allocated through bootmem at
1326                  * boot, it's safer to be consistent with the not-gigantic
1327                  * hugepages and clear the PG_reserved bit from all tail pages
1328                  * too.  Otherwse drivers using get_user_pages() to access tail
1329                  * pages may get the reference counting wrong if they see
1330                  * PG_reserved set on a tail page (despite the head page not
1331                  * having PG_reserved set).  Enforcing this consistency between
1332                  * head and tail pages allows drivers to optimize away a check
1333                  * on the head page when they need know if put_page() is needed
1334                  * after get_user_pages().
1335                  */
1336                 __ClearPageReserved(p);
1337                 set_page_count(p, 0);
1338                 set_compound_head(p, page);
1339         }
1340         atomic_set(compound_mapcount_ptr(page), -1);
1341 }
1342
1343 /*
1344  * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1345  * transparent huge pages.  See the PageTransHuge() documentation for more
1346  * details.
1347  */
1348 int PageHuge(struct page *page)
1349 {
1350         if (!PageCompound(page))
1351                 return 0;
1352
1353         page = compound_head(page);
1354         return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1355 }
1356 EXPORT_SYMBOL_GPL(PageHuge);
1357
1358 /*
1359  * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1360  * normal or transparent huge pages.
1361  */
1362 int PageHeadHuge(struct page *page_head)
1363 {
1364         if (!PageHead(page_head))
1365                 return 0;
1366
1367         return get_compound_page_dtor(page_head) == free_huge_page;
1368 }
1369
1370 pgoff_t __basepage_index(struct page *page)
1371 {
1372         struct page *page_head = compound_head(page);
1373         pgoff_t index = page_index(page_head);
1374         unsigned long compound_idx;
1375
1376         if (!PageHuge(page_head))
1377                 return page_index(page);
1378
1379         if (compound_order(page_head) >= MAX_ORDER)
1380                 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1381         else
1382                 compound_idx = page - page_head;
1383
1384         return (index << compound_order(page_head)) + compound_idx;
1385 }
1386
1387 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
1388 {
1389         struct page *page;
1390
1391         page = __alloc_pages_node(nid,
1392                 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
1393                                                 __GFP_RETRY_MAYFAIL|__GFP_NOWARN,
1394                 huge_page_order(h));
1395         if (page) {
1396                 prep_new_huge_page(h, page, nid);
1397         }
1398
1399         return page;
1400 }
1401
1402 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1403 {
1404         struct page *page;
1405         int nr_nodes, node;
1406         int ret = 0;
1407
1408         for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1409                 page = alloc_fresh_huge_page_node(h, node);
1410                 if (page) {
1411                         ret = 1;
1412                         break;
1413                 }
1414         }
1415
1416         if (ret)
1417                 count_vm_event(HTLB_BUDDY_PGALLOC);
1418         else
1419                 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1420
1421         return ret;
1422 }
1423
1424 /*
1425  * Free huge page from pool from next node to free.
1426  * Attempt to keep persistent huge pages more or less
1427  * balanced over allowed nodes.
1428  * Called with hugetlb_lock locked.
1429  */
1430 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1431                                                          bool acct_surplus)
1432 {
1433         int nr_nodes, node;
1434         int ret = 0;
1435
1436         for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1437                 /*
1438                  * If we're returning unused surplus pages, only examine
1439                  * nodes with surplus pages.
1440                  */
1441                 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1442                     !list_empty(&h->hugepage_freelists[node])) {
1443                         struct page *page =
1444                                 list_entry(h->hugepage_freelists[node].next,
1445                                           struct page, lru);
1446                         list_del(&page->lru);
1447                         h->free_huge_pages--;
1448                         h->free_huge_pages_node[node]--;
1449                         if (acct_surplus) {
1450                                 h->surplus_huge_pages--;
1451                                 h->surplus_huge_pages_node[node]--;
1452                         }
1453                         update_and_free_page(h, page);
1454                         ret = 1;
1455                         break;
1456                 }
1457         }
1458
1459         return ret;
1460 }
1461
1462 /*
1463  * Dissolve a given free hugepage into free buddy pages. This function does
1464  * nothing for in-use (including surplus) hugepages. Returns -EBUSY if the
1465  * number of free hugepages would be reduced below the number of reserved
1466  * hugepages.
1467  */
1468 int dissolve_free_huge_page(struct page *page)
1469 {
1470         int rc = 0;
1471
1472         spin_lock(&hugetlb_lock);
1473         if (PageHuge(page) && !page_count(page)) {
1474                 struct page *head = compound_head(page);
1475                 struct hstate *h = page_hstate(head);
1476                 int nid = page_to_nid(head);
1477                 if (h->free_huge_pages - h->resv_huge_pages == 0) {
1478                         rc = -EBUSY;
1479                         goto out;
1480                 }
1481                 /*
1482                  * Move PageHWPoison flag from head page to the raw error page,
1483                  * which makes any subpages rather than the error page reusable.
1484                  */
1485                 if (PageHWPoison(head) && page != head) {
1486                         SetPageHWPoison(page);
1487                         ClearPageHWPoison(head);
1488                 }
1489                 list_del(&head->lru);
1490                 h->free_huge_pages--;
1491                 h->free_huge_pages_node[nid]--;
1492                 h->max_huge_pages--;
1493                 update_and_free_page(h, head);
1494         }
1495 out:
1496         spin_unlock(&hugetlb_lock);
1497         return rc;
1498 }
1499
1500 /*
1501  * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1502  * make specified memory blocks removable from the system.
1503  * Note that this will dissolve a free gigantic hugepage completely, if any
1504  * part of it lies within the given range.
1505  * Also note that if dissolve_free_huge_page() returns with an error, all
1506  * free hugepages that were dissolved before that error are lost.
1507  */
1508 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1509 {
1510         unsigned long pfn;
1511         struct page *page;
1512         int rc = 0;
1513
1514         if (!hugepages_supported())
1515                 return rc;
1516
1517         for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1518                 page = pfn_to_page(pfn);
1519                 if (PageHuge(page) && !page_count(page)) {
1520                         rc = dissolve_free_huge_page(page);
1521                         if (rc)
1522                                 break;
1523                 }
1524         }
1525
1526         return rc;
1527 }
1528
1529 static struct page *__hugetlb_alloc_buddy_huge_page(struct hstate *h,
1530                 gfp_t gfp_mask, int nid, nodemask_t *nmask)
1531 {
1532         int order = huge_page_order(h);
1533
1534         gfp_mask |= __GFP_COMP|__GFP_RETRY_MAYFAIL|__GFP_NOWARN;
1535         if (nid == NUMA_NO_NODE)
1536                 nid = numa_mem_id();
1537         return __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
1538 }
1539
1540 static struct page *__alloc_buddy_huge_page(struct hstate *h, gfp_t gfp_mask,
1541                 int nid, nodemask_t *nmask)
1542 {
1543         struct page *page;
1544         unsigned int r_nid;
1545
1546         if (hstate_is_gigantic(h))
1547                 return NULL;
1548
1549         /*
1550          * Assume we will successfully allocate the surplus page to
1551          * prevent racing processes from causing the surplus to exceed
1552          * overcommit
1553          *
1554          * This however introduces a different race, where a process B
1555          * tries to grow the static hugepage pool while alloc_pages() is
1556          * called by process A. B will only examine the per-node
1557          * counters in determining if surplus huge pages can be
1558          * converted to normal huge pages in adjust_pool_surplus(). A
1559          * won't be able to increment the per-node counter, until the
1560          * lock is dropped by B, but B doesn't drop hugetlb_lock until
1561          * no more huge pages can be converted from surplus to normal
1562          * state (and doesn't try to convert again). Thus, we have a
1563          * case where a surplus huge page exists, the pool is grown, and
1564          * the surplus huge page still exists after, even though it
1565          * should just have been converted to a normal huge page. This
1566          * does not leak memory, though, as the hugepage will be freed
1567          * once it is out of use. It also does not allow the counters to
1568          * go out of whack in adjust_pool_surplus() as we don't modify
1569          * the node values until we've gotten the hugepage and only the
1570          * per-node value is checked there.
1571          */
1572         spin_lock(&hugetlb_lock);
1573         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1574                 spin_unlock(&hugetlb_lock);
1575                 return NULL;
1576         } else {
1577                 h->nr_huge_pages++;
1578                 h->surplus_huge_pages++;
1579         }
1580         spin_unlock(&hugetlb_lock);
1581
1582         page = __hugetlb_alloc_buddy_huge_page(h, gfp_mask, nid, nmask);
1583
1584         spin_lock(&hugetlb_lock);
1585         if (page) {
1586                 INIT_LIST_HEAD(&page->lru);
1587                 r_nid = page_to_nid(page);
1588                 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1589                 set_hugetlb_cgroup(page, NULL);
1590                 /*
1591                  * We incremented the global counters already
1592                  */
1593                 h->nr_huge_pages_node[r_nid]++;
1594                 h->surplus_huge_pages_node[r_nid]++;
1595                 __count_vm_event(HTLB_BUDDY_PGALLOC);
1596         } else {
1597                 h->nr_huge_pages--;
1598                 h->surplus_huge_pages--;
1599                 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1600         }
1601         spin_unlock(&hugetlb_lock);
1602
1603         return page;
1604 }
1605
1606 /*
1607  * Use the VMA's mpolicy to allocate a huge page from the buddy.
1608  */
1609 static
1610 struct page *__alloc_buddy_huge_page_with_mpol(struct hstate *h,
1611                 struct vm_area_struct *vma, unsigned long addr)
1612 {
1613         struct page *page;
1614         struct mempolicy *mpol;
1615         gfp_t gfp_mask = htlb_alloc_mask(h);
1616         int nid;
1617         nodemask_t *nodemask;
1618
1619         nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1620         page = __alloc_buddy_huge_page(h, gfp_mask, nid, nodemask);
1621         mpol_cond_put(mpol);
1622
1623         return page;
1624 }
1625
1626 /*
1627  * This allocation function is useful in the context where vma is irrelevant.
1628  * E.g. soft-offlining uses this function because it only cares physical
1629  * address of error page.
1630  */
1631 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1632 {
1633         gfp_t gfp_mask = htlb_alloc_mask(h);
1634         struct page *page = NULL;
1635
1636         if (nid != NUMA_NO_NODE)
1637                 gfp_mask |= __GFP_THISNODE;
1638
1639         spin_lock(&hugetlb_lock);
1640         if (h->free_huge_pages - h->resv_huge_pages > 0)
1641                 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL);
1642         spin_unlock(&hugetlb_lock);
1643
1644         if (!page)
1645                 page = __alloc_buddy_huge_page(h, gfp_mask, nid, NULL);
1646
1647         return page;
1648 }
1649
1650
1651 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
1652                 nodemask_t *nmask)
1653 {
1654         gfp_t gfp_mask = htlb_alloc_mask(h);
1655
1656         spin_lock(&hugetlb_lock);
1657         if (h->free_huge_pages - h->resv_huge_pages > 0) {
1658                 struct page *page;
1659
1660                 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1661                 if (page) {
1662                         spin_unlock(&hugetlb_lock);
1663                         return page;
1664                 }
1665         }
1666         spin_unlock(&hugetlb_lock);
1667
1668         /* No reservations, try to overcommit */
1669
1670         return __alloc_buddy_huge_page(h, gfp_mask, preferred_nid, nmask);
1671 }
1672
1673 /*
1674  * Increase the hugetlb pool such that it can accommodate a reservation
1675  * of size 'delta'.
1676  */
1677 static int gather_surplus_pages(struct hstate *h, int delta)
1678 {
1679         struct list_head surplus_list;
1680         struct page *page, *tmp;
1681         int ret, i;
1682         int needed, allocated;
1683         bool alloc_ok = true;
1684
1685         needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1686         if (needed <= 0) {
1687                 h->resv_huge_pages += delta;
1688                 return 0;
1689         }
1690
1691         allocated = 0;
1692         INIT_LIST_HEAD(&surplus_list);
1693
1694         ret = -ENOMEM;
1695 retry:
1696         spin_unlock(&hugetlb_lock);
1697         for (i = 0; i < needed; i++) {
1698                 page = __alloc_buddy_huge_page(h, htlb_alloc_mask(h),
1699                                 NUMA_NO_NODE, NULL);
1700                 if (!page) {
1701                         alloc_ok = false;
1702                         break;
1703                 }
1704                 list_add(&page->lru, &surplus_list);
1705                 cond_resched();
1706         }
1707         allocated += i;
1708
1709         /*
1710          * After retaking hugetlb_lock, we need to recalculate 'needed'
1711          * because either resv_huge_pages or free_huge_pages may have changed.
1712          */
1713         spin_lock(&hugetlb_lock);
1714         needed = (h->resv_huge_pages + delta) -
1715                         (h->free_huge_pages + allocated);
1716         if (needed > 0) {
1717                 if (alloc_ok)
1718                         goto retry;
1719                 /*
1720                  * We were not able to allocate enough pages to
1721                  * satisfy the entire reservation so we free what
1722                  * we've allocated so far.
1723                  */
1724                 goto free;
1725         }
1726         /*
1727          * The surplus_list now contains _at_least_ the number of extra pages
1728          * needed to accommodate the reservation.  Add the appropriate number
1729          * of pages to the hugetlb pool and free the extras back to the buddy
1730          * allocator.  Commit the entire reservation here to prevent another
1731          * process from stealing the pages as they are added to the pool but
1732          * before they are reserved.
1733          */
1734         needed += allocated;
1735         h->resv_huge_pages += delta;
1736         ret = 0;
1737
1738         /* Free the needed pages to the hugetlb pool */
1739         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1740                 if ((--needed) < 0)
1741                         break;
1742                 /*
1743                  * This page is now managed by the hugetlb allocator and has
1744                  * no users -- drop the buddy allocator's reference.
1745                  */
1746                 put_page_testzero(page);
1747                 VM_BUG_ON_PAGE(page_count(page), page);
1748                 enqueue_huge_page(h, page);
1749         }
1750 free:
1751         spin_unlock(&hugetlb_lock);
1752
1753         /* Free unnecessary surplus pages to the buddy allocator */
1754         list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1755                 put_page(page);
1756         spin_lock(&hugetlb_lock);
1757
1758         return ret;
1759 }
1760
1761 /*
1762  * This routine has two main purposes:
1763  * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1764  *    in unused_resv_pages.  This corresponds to the prior adjustments made
1765  *    to the associated reservation map.
1766  * 2) Free any unused surplus pages that may have been allocated to satisfy
1767  *    the reservation.  As many as unused_resv_pages may be freed.
1768  *
1769  * Called with hugetlb_lock held.  However, the lock could be dropped (and
1770  * reacquired) during calls to cond_resched_lock.  Whenever dropping the lock,
1771  * we must make sure nobody else can claim pages we are in the process of
1772  * freeing.  Do this by ensuring resv_huge_page always is greater than the
1773  * number of huge pages we plan to free when dropping the lock.
1774  */
1775 static void return_unused_surplus_pages(struct hstate *h,
1776                                         unsigned long unused_resv_pages)
1777 {
1778         unsigned long nr_pages;
1779
1780         /* Cannot return gigantic pages currently */
1781         if (hstate_is_gigantic(h))
1782                 goto out;
1783
1784         /*
1785          * Part (or even all) of the reservation could have been backed
1786          * by pre-allocated pages. Only free surplus pages.
1787          */
1788         nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1789
1790         /*
1791          * We want to release as many surplus pages as possible, spread
1792          * evenly across all nodes with memory. Iterate across these nodes
1793          * until we can no longer free unreserved surplus pages. This occurs
1794          * when the nodes with surplus pages have no free pages.
1795          * free_pool_huge_page() will balance the the freed pages across the
1796          * on-line nodes with memory and will handle the hstate accounting.
1797          *
1798          * Note that we decrement resv_huge_pages as we free the pages.  If
1799          * we drop the lock, resv_huge_pages will still be sufficiently large
1800          * to cover subsequent pages we may free.
1801          */
1802         while (nr_pages--) {
1803                 h->resv_huge_pages--;
1804                 unused_resv_pages--;
1805                 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1806                         goto out;
1807                 cond_resched_lock(&hugetlb_lock);
1808         }
1809
1810 out:
1811         /* Fully uncommit the reservation */
1812         h->resv_huge_pages -= unused_resv_pages;
1813 }
1814
1815
1816 /*
1817  * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1818  * are used by the huge page allocation routines to manage reservations.
1819  *
1820  * vma_needs_reservation is called to determine if the huge page at addr
1821  * within the vma has an associated reservation.  If a reservation is
1822  * needed, the value 1 is returned.  The caller is then responsible for
1823  * managing the global reservation and subpool usage counts.  After
1824  * the huge page has been allocated, vma_commit_reservation is called
1825  * to add the page to the reservation map.  If the page allocation fails,
1826  * the reservation must be ended instead of committed.  vma_end_reservation
1827  * is called in such cases.
1828  *
1829  * In the normal case, vma_commit_reservation returns the same value
1830  * as the preceding vma_needs_reservation call.  The only time this
1831  * is not the case is if a reserve map was changed between calls.  It
1832  * is the responsibility of the caller to notice the difference and
1833  * take appropriate action.
1834  *
1835  * vma_add_reservation is used in error paths where a reservation must
1836  * be restored when a newly allocated huge page must be freed.  It is
1837  * to be called after calling vma_needs_reservation to determine if a
1838  * reservation exists.
1839  */
1840 enum vma_resv_mode {
1841         VMA_NEEDS_RESV,
1842         VMA_COMMIT_RESV,
1843         VMA_END_RESV,
1844         VMA_ADD_RESV,
1845 };
1846 static long __vma_reservation_common(struct hstate *h,
1847                                 struct vm_area_struct *vma, unsigned long addr,
1848                                 enum vma_resv_mode mode)
1849 {
1850         struct resv_map *resv;
1851         pgoff_t idx;
1852         long ret;
1853
1854         resv = vma_resv_map(vma);
1855         if (!resv)
1856                 return 1;
1857
1858         idx = vma_hugecache_offset(h, vma, addr);
1859         switch (mode) {
1860         case VMA_NEEDS_RESV:
1861                 ret = region_chg(resv, idx, idx + 1);
1862                 break;
1863         case VMA_COMMIT_RESV:
1864                 ret = region_add(resv, idx, idx + 1);
1865                 break;
1866         case VMA_END_RESV:
1867                 region_abort(resv, idx, idx + 1);
1868                 ret = 0;
1869                 break;
1870         case VMA_ADD_RESV:
1871                 if (vma->vm_flags & VM_MAYSHARE)
1872                         ret = region_add(resv, idx, idx + 1);
1873                 else {
1874                         region_abort(resv, idx, idx + 1);
1875                         ret = region_del(resv, idx, idx + 1);
1876                 }
1877                 break;
1878         default:
1879                 BUG();
1880         }
1881
1882         if (vma->vm_flags & VM_MAYSHARE)
1883                 return ret;
1884         else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
1885                 /*
1886                  * In most cases, reserves always exist for private mappings.
1887                  * However, a file associated with mapping could have been
1888                  * hole punched or truncated after reserves were consumed.
1889                  * As subsequent fault on such a range will not use reserves.
1890                  * Subtle - The reserve map for private mappings has the
1891                  * opposite meaning than that of shared mappings.  If NO
1892                  * entry is in the reserve map, it means a reservation exists.
1893                  * If an entry exists in the reserve map, it means the
1894                  * reservation has already been consumed.  As a result, the
1895                  * return value of this routine is the opposite of the
1896                  * value returned from reserve map manipulation routines above.
1897                  */
1898                 if (ret)
1899                         return 0;
1900                 else
1901                         return 1;
1902         }
1903         else
1904                 return ret < 0 ? ret : 0;
1905 }
1906
1907 static long vma_needs_reservation(struct hstate *h,
1908                         struct vm_area_struct *vma, unsigned long addr)
1909 {
1910         return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
1911 }
1912
1913 static long vma_commit_reservation(struct hstate *h,
1914                         struct vm_area_struct *vma, unsigned long addr)
1915 {
1916         return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1917 }
1918
1919 static void vma_end_reservation(struct hstate *h,
1920                         struct vm_area_struct *vma, unsigned long addr)
1921 {
1922         (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
1923 }
1924
1925 static long vma_add_reservation(struct hstate *h,
1926                         struct vm_area_struct *vma, unsigned long addr)
1927 {
1928         return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
1929 }
1930
1931 /*
1932  * This routine is called to restore a reservation on error paths.  In the
1933  * specific error paths, a huge page was allocated (via alloc_huge_page)
1934  * and is about to be freed.  If a reservation for the page existed,
1935  * alloc_huge_page would have consumed the reservation and set PagePrivate
1936  * in the newly allocated page.  When the page is freed via free_huge_page,
1937  * the global reservation count will be incremented if PagePrivate is set.
1938  * However, free_huge_page can not adjust the reserve map.  Adjust the
1939  * reserve map here to be consistent with global reserve count adjustments
1940  * to be made by free_huge_page.
1941  */
1942 static void restore_reserve_on_error(struct hstate *h,
1943                         struct vm_area_struct *vma, unsigned long address,
1944                         struct page *page)
1945 {
1946         if (unlikely(PagePrivate(page))) {
1947                 long rc = vma_needs_reservation(h, vma, address);
1948
1949                 if (unlikely(rc < 0)) {
1950                         /*
1951                          * Rare out of memory condition in reserve map
1952                          * manipulation.  Clear PagePrivate so that
1953                          * global reserve count will not be incremented
1954                          * by free_huge_page.  This will make it appear
1955                          * as though the reservation for this page was
1956                          * consumed.  This may prevent the task from
1957                          * faulting in the page at a later time.  This
1958                          * is better than inconsistent global huge page
1959                          * accounting of reserve counts.
1960                          */
1961                         ClearPagePrivate(page);
1962                 } else if (rc) {
1963                         rc = vma_add_reservation(h, vma, address);
1964                         if (unlikely(rc < 0))
1965                                 /*
1966                                  * See above comment about rare out of
1967                                  * memory condition.
1968                                  */
1969                                 ClearPagePrivate(page);
1970                 } else
1971                         vma_end_reservation(h, vma, address);
1972         }
1973 }
1974
1975 struct page *alloc_huge_page(struct vm_area_struct *vma,
1976                                     unsigned long addr, int avoid_reserve)
1977 {
1978         struct hugepage_subpool *spool = subpool_vma(vma);
1979         struct hstate *h = hstate_vma(vma);
1980         struct page *page;
1981         long map_chg, map_commit;
1982         long gbl_chg;
1983         int ret, idx;
1984         struct hugetlb_cgroup *h_cg;
1985
1986         idx = hstate_index(h);
1987         /*
1988          * Examine the region/reserve map to determine if the process
1989          * has a reservation for the page to be allocated.  A return
1990          * code of zero indicates a reservation exists (no change).
1991          */
1992         map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
1993         if (map_chg < 0)
1994                 return ERR_PTR(-ENOMEM);
1995
1996         /*
1997          * Processes that did not create the mapping will have no
1998          * reserves as indicated by the region/reserve map. Check
1999          * that the allocation will not exceed the subpool limit.
2000          * Allocations for MAP_NORESERVE mappings also need to be
2001          * checked against any subpool limit.
2002          */
2003         if (map_chg || avoid_reserve) {
2004                 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2005                 if (gbl_chg < 0) {
2006                         vma_end_reservation(h, vma, addr);
2007                         return ERR_PTR(-ENOSPC);
2008                 }
2009
2010                 /*
2011                  * Even though there was no reservation in the region/reserve
2012                  * map, there could be reservations associated with the
2013                  * subpool that can be used.  This would be indicated if the
2014                  * return value of hugepage_subpool_get_pages() is zero.
2015                  * However, if avoid_reserve is specified we still avoid even
2016                  * the subpool reservations.
2017                  */
2018                 if (avoid_reserve)
2019                         gbl_chg = 1;
2020         }
2021
2022         ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2023         if (ret)
2024                 goto out_subpool_put;
2025
2026         spin_lock(&hugetlb_lock);
2027         /*
2028          * glb_chg is passed to indicate whether or not a page must be taken
2029          * from the global free pool (global change).  gbl_chg == 0 indicates
2030          * a reservation exists for the allocation.
2031          */
2032         page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2033         if (!page) {
2034                 spin_unlock(&hugetlb_lock);
2035                 page = __alloc_buddy_huge_page_with_mpol(h, vma, addr);
2036                 if (!page)
2037                         goto out_uncharge_cgroup;
2038                 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2039                         SetPagePrivate(page);
2040                         h->resv_huge_pages--;
2041                 }
2042                 spin_lock(&hugetlb_lock);
2043                 list_move(&page->lru, &h->hugepage_activelist);
2044                 /* Fall through */
2045         }
2046         hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2047         spin_unlock(&hugetlb_lock);
2048
2049         set_page_private(page, (unsigned long)spool);
2050
2051         map_commit = vma_commit_reservation(h, vma, addr);
2052         if (unlikely(map_chg > map_commit)) {
2053                 /*
2054                  * The page was added to the reservation map between
2055                  * vma_needs_reservation and vma_commit_reservation.
2056                  * This indicates a race with hugetlb_reserve_pages.
2057                  * Adjust for the subpool count incremented above AND
2058                  * in hugetlb_reserve_pages for the same page.  Also,
2059                  * the reservation count added in hugetlb_reserve_pages
2060                  * no longer applies.
2061                  */
2062                 long rsv_adjust;
2063
2064                 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2065                 hugetlb_acct_memory(h, -rsv_adjust);
2066         }
2067         return page;
2068
2069 out_uncharge_cgroup:
2070         hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2071 out_subpool_put:
2072         if (map_chg || avoid_reserve)
2073                 hugepage_subpool_put_pages(spool, 1);
2074         vma_end_reservation(h, vma, addr);
2075         return ERR_PTR(-ENOSPC);
2076 }
2077
2078 /*
2079  * alloc_huge_page()'s wrapper which simply returns the page if allocation
2080  * succeeds, otherwise NULL. This function is called from new_vma_page(),
2081  * where no ERR_VALUE is expected to be returned.
2082  */
2083 struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
2084                                 unsigned long addr, int avoid_reserve)
2085 {
2086         struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
2087         if (IS_ERR(page))
2088                 page = NULL;
2089         return page;
2090 }
2091
2092 int alloc_bootmem_huge_page(struct hstate *h)
2093         __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2094 int __alloc_bootmem_huge_page(struct hstate *h)
2095 {
2096         struct huge_bootmem_page *m;
2097         int nr_nodes, node;
2098
2099         for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2100                 void *addr;
2101
2102                 addr = memblock_virt_alloc_try_nid_nopanic(
2103                                 huge_page_size(h), huge_page_size(h),
2104                                 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
2105                 if (addr) {
2106                         /*
2107                          * Use the beginning of the huge page to store the
2108                          * huge_bootmem_page struct (until gather_bootmem
2109                          * puts them into the mem_map).
2110                          */
2111                         m = addr;
2112                         goto found;
2113                 }
2114         }
2115         return 0;
2116
2117 found:
2118         BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2119         /* Put them into a private list first because mem_map is not up yet */
2120         list_add(&m->list, &huge_boot_pages);
2121         m->hstate = h;
2122         return 1;
2123 }
2124
2125 static void __init prep_compound_huge_page(struct page *page,
2126                 unsigned int order)
2127 {
2128         if (unlikely(order > (MAX_ORDER - 1)))
2129                 prep_compound_gigantic_page(page, order);
2130         else
2131                 prep_compound_page(page, order);
2132 }
2133
2134 /* Put bootmem huge pages into the standard lists after mem_map is up */
2135 static void __init gather_bootmem_prealloc(void)
2136 {
2137         struct huge_bootmem_page *m;
2138
2139         list_for_each_entry(m, &huge_boot_pages, list) {
2140                 struct hstate *h = m->hstate;
2141                 struct page *page;
2142
2143 #ifdef CONFIG_HIGHMEM
2144                 page = pfn_to_page(m->phys >> PAGE_SHIFT);
2145                 memblock_free_late(__pa(m),
2146                                    sizeof(struct huge_bootmem_page));
2147 #else
2148                 page = virt_to_page(m);
2149 #endif
2150                 WARN_ON(page_count(page) != 1);
2151                 prep_compound_huge_page(page, h->order);
2152                 WARN_ON(PageReserved(page));
2153                 prep_new_huge_page(h, page, page_to_nid(page));
2154                 /*
2155                  * If we had gigantic hugepages allocated at boot time, we need
2156                  * to restore the 'stolen' pages to totalram_pages in order to
2157                  * fix confusing memory reports from free(1) and another
2158                  * side-effects, like CommitLimit going negative.
2159                  */
2160                 if (hstate_is_gigantic(h))
2161                         adjust_managed_page_count(page, 1 << h->order);
2162                 cond_resched();
2163         }
2164 }
2165
2166 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2167 {
2168         unsigned long i;
2169
2170         for (i = 0; i < h->max_huge_pages; ++i) {
2171                 if (hstate_is_gigantic(h)) {
2172                         if (!alloc_bootmem_huge_page(h))
2173                                 break;
2174                 } else if (!alloc_fresh_huge_page(h,
2175                                          &node_states[N_MEMORY]))
2176                         break;
2177                 cond_resched();
2178         }
2179         if (i < h->max_huge_pages) {
2180                 char buf[32];
2181
2182                 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2183                 pr_warn("HugeTLB: allocating %lu of page size %s failed.  Only allocated %lu hugepages.\n",
2184                         h->max_huge_pages, buf, i);
2185                 h->max_huge_pages = i;
2186         }
2187 }
2188
2189 static void __init hugetlb_init_hstates(void)
2190 {
2191         struct hstate *h;
2192
2193         for_each_hstate(h) {
2194                 if (minimum_order > huge_page_order(h))
2195                         minimum_order = huge_page_order(h);
2196
2197                 /* oversize hugepages were init'ed in early boot */
2198                 if (!hstate_is_gigantic(h))
2199                         hugetlb_hstate_alloc_pages(h);
2200         }
2201         VM_BUG_ON(minimum_order == UINT_MAX);
2202 }
2203
2204 static void __init report_hugepages(void)
2205 {
2206         struct hstate *h;
2207
2208         for_each_hstate(h) {
2209                 char buf[32];
2210
2211                 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2212                 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2213                         buf, h->free_huge_pages);
2214         }
2215 }
2216
2217 #ifdef CONFIG_HIGHMEM
2218 static void try_to_free_low(struct hstate *h, unsigned long count,
2219                                                 nodemask_t *nodes_allowed)
2220 {
2221         int i;
2222
2223         if (hstate_is_gigantic(h))
2224                 return;
2225
2226         for_each_node_mask(i, *nodes_allowed) {
2227                 struct page *page, *next;
2228                 struct list_head *freel = &h->hugepage_freelists[i];
2229                 list_for_each_entry_safe(page, next, freel, lru) {
2230                         if (count >= h->nr_huge_pages)
2231                                 return;
2232                         if (PageHighMem(page))
2233                                 continue;
2234                         list_del(&page->lru);
2235                         update_and_free_page(h, page);
2236                         h->free_huge_pages--;
2237                         h->free_huge_pages_node[page_to_nid(page)]--;
2238                 }
2239         }
2240 }
2241 #else
2242 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2243                                                 nodemask_t *nodes_allowed)
2244 {
2245 }
2246 #endif
2247
2248 /*
2249  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
2250  * balanced by operating on them in a round-robin fashion.
2251  * Returns 1 if an adjustment was made.
2252  */
2253 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2254                                 int delta)
2255 {
2256         int nr_nodes, node;
2257
2258         VM_BUG_ON(delta != -1 && delta != 1);
2259
2260         if (delta < 0) {
2261                 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2262                         if (h->surplus_huge_pages_node[node])
2263                                 goto found;
2264                 }
2265         } else {
2266                 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2267                         if (h->surplus_huge_pages_node[node] <
2268                                         h->nr_huge_pages_node[node])
2269                                 goto found;
2270                 }
2271         }
2272         return 0;
2273
2274 found:
2275         h->surplus_huge_pages += delta;
2276         h->surplus_huge_pages_node[node] += delta;
2277         return 1;
2278 }
2279
2280 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2281 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
2282                                                 nodemask_t *nodes_allowed)
2283 {
2284         unsigned long min_count, ret;
2285
2286         if (hstate_is_gigantic(h) && !gigantic_page_supported())
2287                 return h->max_huge_pages;
2288
2289         /*
2290          * Increase the pool size
2291          * First take pages out of surplus state.  Then make up the
2292          * remaining difference by allocating fresh huge pages.
2293          *
2294          * We might race with __alloc_buddy_huge_page() here and be unable
2295          * to convert a surplus huge page to a normal huge page. That is
2296          * not critical, though, it just means the overall size of the
2297          * pool might be one hugepage larger than it needs to be, but
2298          * within all the constraints specified by the sysctls.
2299          */
2300         spin_lock(&hugetlb_lock);
2301         while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2302                 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2303                         break;
2304         }
2305
2306         while (count > persistent_huge_pages(h)) {
2307                 /*
2308                  * If this allocation races such that we no longer need the
2309                  * page, free_huge_page will handle it by freeing the page
2310                  * and reducing the surplus.
2311                  */
2312                 spin_unlock(&hugetlb_lock);
2313
2314                 /* yield cpu to avoid soft lockup */
2315                 cond_resched();
2316
2317                 if (hstate_is_gigantic(h))
2318                         ret = alloc_fresh_gigantic_page(h, nodes_allowed);
2319                 else
2320                         ret = alloc_fresh_huge_page(h, nodes_allowed);
2321                 spin_lock(&hugetlb_lock);
2322                 if (!ret)
2323                         goto out;
2324
2325                 /* Bail for signals. Probably ctrl-c from user */
2326                 if (signal_pending(current))
2327                         goto out;
2328         }
2329
2330         /*
2331          * Decrease the pool size
2332          * First return free pages to the buddy allocator (being careful
2333          * to keep enough around to satisfy reservations).  Then place
2334          * pages into surplus state as needed so the pool will shrink
2335          * to the desired size as pages become free.
2336          *
2337          * By placing pages into the surplus state independent of the
2338          * overcommit value, we are allowing the surplus pool size to
2339          * exceed overcommit. There are few sane options here. Since
2340          * __alloc_buddy_huge_page() is checking the global counter,
2341          * though, we'll note that we're not allowed to exceed surplus
2342          * and won't grow the pool anywhere else. Not until one of the
2343          * sysctls are changed, or the surplus pages go out of use.
2344          */
2345         min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2346         min_count = max(count, min_count);
2347         try_to_free_low(h, min_count, nodes_allowed);
2348         while (min_count < persistent_huge_pages(h)) {
2349                 if (!free_pool_huge_page(h, nodes_allowed, 0))
2350                         break;
2351                 cond_resched_lock(&hugetlb_lock);
2352         }
2353         while (count < persistent_huge_pages(h)) {
2354                 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2355                         break;
2356         }
2357 out:
2358         ret = persistent_huge_pages(h);
2359         spin_unlock(&hugetlb_lock);
2360         return ret;
2361 }
2362
2363 #define HSTATE_ATTR_RO(_name) \
2364         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2365
2366 #define HSTATE_ATTR(_name) \
2367         static struct kobj_attribute _name##_attr = \
2368                 __ATTR(_name, 0644, _name##_show, _name##_store)
2369
2370 static struct kobject *hugepages_kobj;
2371 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2372
2373 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2374
2375 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2376 {
2377         int i;
2378
2379         for (i = 0; i < HUGE_MAX_HSTATE; i++)
2380                 if (hstate_kobjs[i] == kobj) {
2381                         if (nidp)
2382                                 *nidp = NUMA_NO_NODE;
2383                         return &hstates[i];
2384                 }
2385
2386         return kobj_to_node_hstate(kobj, nidp);
2387 }
2388
2389 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2390                                         struct kobj_attribute *attr, char *buf)
2391 {
2392         struct hstate *h;
2393         unsigned long nr_huge_pages;
2394         int nid;
2395
2396         h = kobj_to_hstate(kobj, &nid);
2397         if (nid == NUMA_NO_NODE)
2398                 nr_huge_pages = h->nr_huge_pages;
2399         else
2400                 nr_huge_pages = h->nr_huge_pages_node[nid];
2401
2402         return sprintf(buf, "%lu\n", nr_huge_pages);
2403 }
2404
2405 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2406                                            struct hstate *h, int nid,
2407                                            unsigned long count, size_t len)
2408 {
2409         int err;
2410         NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
2411
2412         if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
2413                 err = -EINVAL;
2414                 goto out;
2415         }
2416
2417         if (nid == NUMA_NO_NODE) {
2418                 /*
2419                  * global hstate attribute
2420                  */
2421                 if (!(obey_mempolicy &&
2422                                 init_nodemask_of_mempolicy(nodes_allowed))) {
2423                         NODEMASK_FREE(nodes_allowed);
2424                         nodes_allowed = &node_states[N_MEMORY];
2425                 }
2426         } else if (nodes_allowed) {
2427                 /*
2428                  * per node hstate attribute: adjust count to global,
2429                  * but restrict alloc/free to the specified node.
2430                  */
2431                 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2432                 init_nodemask_of_node(nodes_allowed, nid);
2433         } else
2434                 nodes_allowed = &node_states[N_MEMORY];
2435
2436         h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
2437
2438         if (nodes_allowed != &node_states[N_MEMORY])
2439                 NODEMASK_FREE(nodes_allowed);
2440
2441         return len;
2442 out:
2443         NODEMASK_FREE(nodes_allowed);
2444         return err;
2445 }
2446
2447 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2448                                          struct kobject *kobj, const char *buf,
2449                                          size_t len)
2450 {
2451         struct hstate *h;
2452         unsigned long count;
2453         int nid;
2454         int err;
2455
2456         err = kstrtoul(buf, 10, &count);
2457         if (err)
2458                 return err;
2459
2460         h = kobj_to_hstate(kobj, &nid);
2461         return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2462 }
2463
2464 static ssize_t nr_hugepages_show(struct kobject *kobj,
2465                                        struct kobj_attribute *attr, char *buf)
2466 {
2467         return nr_hugepages_show_common(kobj, attr, buf);
2468 }
2469
2470 static ssize_t nr_hugepages_store(struct kobject *kobj,
2471                struct kobj_attribute *attr, const char *buf, size_t len)
2472 {
2473         return nr_hugepages_store_common(false, kobj, buf, len);
2474 }
2475 HSTATE_ATTR(nr_hugepages);
2476
2477 #ifdef CONFIG_NUMA
2478
2479 /*
2480  * hstate attribute for optionally mempolicy-based constraint on persistent
2481  * huge page alloc/free.
2482  */
2483 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2484                                        struct kobj_attribute *attr, char *buf)
2485 {
2486         return nr_hugepages_show_common(kobj, attr, buf);
2487 }
2488
2489 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2490                struct kobj_attribute *attr, const char *buf, size_t len)
2491 {
2492         return nr_hugepages_store_common(true, kobj, buf, len);
2493 }
2494 HSTATE_ATTR(nr_hugepages_mempolicy);
2495 #endif
2496
2497
2498 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2499                                         struct kobj_attribute *attr, char *buf)
2500 {
2501         struct hstate *h = kobj_to_hstate(kobj, NULL);
2502         return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2503 }
2504
2505 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2506                 struct kobj_attribute *attr, const char *buf, size_t count)
2507 {
2508         int err;
2509         unsigned long input;
2510         struct hstate *h = kobj_to_hstate(kobj, NULL);
2511
2512         if (hstate_is_gigantic(h))
2513                 return -EINVAL;
2514
2515         err = kstrtoul(buf, 10, &input);
2516         if (err)
2517                 return err;
2518
2519         spin_lock(&hugetlb_lock);
2520         h->nr_overcommit_huge_pages = input;
2521         spin_unlock(&hugetlb_lock);
2522
2523         return count;
2524 }
2525 HSTATE_ATTR(nr_overcommit_hugepages);
2526
2527 static ssize_t free_hugepages_show(struct kobject *kobj,
2528                                         struct kobj_attribute *attr, char *buf)
2529 {
2530         struct hstate *h;
2531         unsigned long free_huge_pages;
2532         int nid;
2533
2534         h = kobj_to_hstate(kobj, &nid);
2535         if (nid == NUMA_NO_NODE)
2536                 free_huge_pages = h->free_huge_pages;
2537         else
2538                 free_huge_pages = h->free_huge_pages_node[nid];
2539
2540         return sprintf(buf, "%lu\n", free_huge_pages);
2541 }
2542 HSTATE_ATTR_RO(free_hugepages);
2543
2544 static ssize_t resv_hugepages_show(struct kobject *kobj,
2545                                         struct kobj_attribute *attr, char *buf)
2546 {
2547         struct hstate *h = kobj_to_hstate(kobj, NULL);
2548         return sprintf(buf, "%lu\n", h->resv_huge_pages);
2549 }
2550 HSTATE_ATTR_RO(resv_hugepages);
2551
2552 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2553                                         struct kobj_attribute *attr, char *buf)
2554 {
2555         struct hstate *h;
2556         unsigned long surplus_huge_pages;
2557         int nid;
2558
2559         h = kobj_to_hstate(kobj, &nid);
2560         if (nid == NUMA_NO_NODE)
2561                 surplus_huge_pages = h->surplus_huge_pages;
2562         else
2563                 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2564
2565         return sprintf(buf, "%lu\n", surplus_huge_pages);
2566 }
2567 HSTATE_ATTR_RO(surplus_hugepages);
2568
2569 static struct attribute *hstate_attrs[] = {
2570         &nr_hugepages_attr.attr,
2571         &nr_overcommit_hugepages_attr.attr,
2572         &free_hugepages_attr.attr,
2573         &resv_hugepages_attr.attr,
2574         &surplus_hugepages_attr.attr,
2575 #ifdef CONFIG_NUMA
2576         &nr_hugepages_mempolicy_attr.attr,
2577 #endif
2578         NULL,
2579 };
2580
2581 static const struct attribute_group hstate_attr_group = {
2582         .attrs = hstate_attrs,
2583 };
2584
2585 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2586                                     struct kobject **hstate_kobjs,
2587                                     const struct attribute_group *hstate_attr_group)
2588 {
2589         int retval;
2590         int hi = hstate_index(h);
2591
2592         hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2593         if (!hstate_kobjs[hi])
2594                 return -ENOMEM;
2595
2596         retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2597         if (retval)
2598                 kobject_put(hstate_kobjs[hi]);
2599
2600         return retval;
2601 }
2602
2603 static void __init hugetlb_sysfs_init(void)
2604 {
2605         struct hstate *h;
2606         int err;
2607
2608         hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2609         if (!hugepages_kobj)
2610                 return;
2611
2612         for_each_hstate(h) {
2613                 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2614                                          hstate_kobjs, &hstate_attr_group);
2615                 if (err)
2616                         pr_err("Hugetlb: Unable to add hstate %s", h->name);
2617         }
2618 }
2619
2620 #ifdef CONFIG_NUMA
2621
2622 /*
2623  * node_hstate/s - associate per node hstate attributes, via their kobjects,
2624  * with node devices in node_devices[] using a parallel array.  The array
2625  * index of a node device or _hstate == node id.
2626  * This is here to avoid any static dependency of the node device driver, in
2627  * the base kernel, on the hugetlb module.
2628  */
2629 struct node_hstate {
2630         struct kobject          *hugepages_kobj;
2631         struct kobject          *hstate_kobjs[HUGE_MAX_HSTATE];
2632 };
2633 static struct node_hstate node_hstates[MAX_NUMNODES];
2634
2635 /*
2636  * A subset of global hstate attributes for node devices
2637  */
2638 static struct attribute *per_node_hstate_attrs[] = {
2639         &nr_hugepages_attr.attr,
2640         &free_hugepages_attr.attr,
2641         &surplus_hugepages_attr.attr,
2642         NULL,
2643 };
2644
2645 static const struct attribute_group per_node_hstate_attr_group = {
2646         .attrs = per_node_hstate_attrs,
2647 };
2648
2649 /*
2650  * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2651  * Returns node id via non-NULL nidp.
2652  */
2653 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2654 {
2655         int nid;
2656
2657         for (nid = 0; nid < nr_node_ids; nid++) {
2658                 struct node_hstate *nhs = &node_hstates[nid];
2659                 int i;
2660                 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2661                         if (nhs->hstate_kobjs[i] == kobj) {
2662                                 if (nidp)
2663                                         *nidp = nid;
2664                                 return &hstates[i];
2665                         }
2666         }
2667
2668         BUG();
2669         return NULL;
2670 }
2671
2672 /*
2673  * Unregister hstate attributes from a single node device.
2674  * No-op if no hstate attributes attached.
2675  */
2676 static void hugetlb_unregister_node(struct node *node)
2677 {
2678         struct hstate *h;
2679         struct node_hstate *nhs = &node_hstates[node->dev.id];
2680
2681         if (!nhs->hugepages_kobj)
2682                 return;         /* no hstate attributes */
2683
2684         for_each_hstate(h) {
2685                 int idx = hstate_index(h);
2686                 if (nhs->hstate_kobjs[idx]) {
2687                         kobject_put(nhs->hstate_kobjs[idx]);
2688                         nhs->hstate_kobjs[idx] = NULL;
2689                 }
2690         }
2691
2692         kobject_put(nhs->hugepages_kobj);
2693         nhs->hugepages_kobj = NULL;
2694 }
2695
2696
2697 /*
2698  * Register hstate attributes for a single node device.
2699  * No-op if attributes already registered.
2700  */
2701 static void hugetlb_register_node(struct node *node)
2702 {
2703         struct hstate *h;
2704         struct node_hstate *nhs = &node_hstates[node->dev.id];
2705         int err;
2706
2707         if (nhs->hugepages_kobj)
2708                 return;         /* already allocated */
2709
2710         nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2711                                                         &node->dev.kobj);
2712         if (!nhs->hugepages_kobj)
2713                 return;
2714
2715         for_each_hstate(h) {
2716                 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2717                                                 nhs->hstate_kobjs,
2718                                                 &per_node_hstate_attr_group);
2719                 if (err) {
2720                         pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2721                                 h->name, node->dev.id);
2722                         hugetlb_unregister_node(node);
2723                         break;
2724                 }
2725         }
2726 }
2727
2728 /*
2729  * hugetlb init time:  register hstate attributes for all registered node
2730  * devices of nodes that have memory.  All on-line nodes should have
2731  * registered their associated device by this time.
2732  */
2733 static void __init hugetlb_register_all_nodes(void)
2734 {
2735         int nid;
2736
2737         for_each_node_state(nid, N_MEMORY) {
2738                 struct node *node = node_devices[nid];
2739                 if (node->dev.id == nid)
2740                         hugetlb_register_node(node);
2741         }
2742
2743         /*
2744          * Let the node device driver know we're here so it can
2745          * [un]register hstate attributes on node hotplug.
2746          */
2747         register_hugetlbfs_with_node(hugetlb_register_node,
2748                                      hugetlb_unregister_node);
2749 }
2750 #else   /* !CONFIG_NUMA */
2751
2752 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2753 {
2754         BUG();
2755         if (nidp)
2756                 *nidp = -1;
2757         return NULL;
2758 }
2759
2760 static void hugetlb_register_all_nodes(void) { }
2761
2762 #endif
2763
2764 static int __init hugetlb_init(void)
2765 {
2766         int i;
2767
2768         if (!hugepages_supported())
2769                 return 0;
2770
2771         if (!size_to_hstate(default_hstate_size)) {
2772                 if (default_hstate_size != 0) {
2773                         pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n",
2774                                default_hstate_size, HPAGE_SIZE);
2775                 }
2776
2777                 default_hstate_size = HPAGE_SIZE;
2778                 if (!size_to_hstate(default_hstate_size))
2779                         hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2780         }
2781         default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2782         if (default_hstate_max_huge_pages) {
2783                 if (!default_hstate.max_huge_pages)
2784                         default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2785         }
2786
2787         hugetlb_init_hstates();
2788         gather_bootmem_prealloc();
2789         report_hugepages();
2790
2791         hugetlb_sysfs_init();
2792         hugetlb_register_all_nodes();
2793         hugetlb_cgroup_file_init();
2794
2795 #ifdef CONFIG_SMP
2796         num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2797 #else
2798         num_fault_mutexes = 1;
2799 #endif
2800         hugetlb_fault_mutex_table =
2801                 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
2802         BUG_ON(!hugetlb_fault_mutex_table);
2803
2804         for (i = 0; i < num_fault_mutexes; i++)
2805                 mutex_init(&hugetlb_fault_mutex_table[i]);
2806         return 0;
2807 }
2808 subsys_initcall(hugetlb_init);
2809
2810 /* Should be called on processing a hugepagesz=... option */
2811 void __init hugetlb_bad_size(void)
2812 {
2813         parsed_valid_hugepagesz = false;
2814 }
2815
2816 void __init hugetlb_add_hstate(unsigned int order)
2817 {
2818         struct hstate *h;
2819         unsigned long i;
2820
2821         if (size_to_hstate(PAGE_SIZE << order)) {
2822                 pr_warn("hugepagesz= specified twice, ignoring\n");
2823                 return;
2824         }
2825         BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2826         BUG_ON(order == 0);
2827         h = &hstates[hugetlb_max_hstate++];
2828         h->order = order;
2829         h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2830         h->nr_huge_pages = 0;
2831         h->free_huge_pages = 0;
2832         for (i = 0; i < MAX_NUMNODES; ++i)
2833                 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2834         INIT_LIST_HEAD(&h->hugepage_activelist);
2835         h->next_nid_to_alloc = first_memory_node;
2836         h->next_nid_to_free = first_memory_node;
2837         snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2838                                         huge_page_size(h)/1024);
2839
2840         parsed_hstate = h;
2841 }
2842
2843 static int __init hugetlb_nrpages_setup(char *s)
2844 {
2845         unsigned long *mhp;
2846         static unsigned long *last_mhp;
2847
2848         if (!parsed_valid_hugepagesz) {
2849                 pr_warn("hugepages = %s preceded by "
2850                         "an unsupported hugepagesz, ignoring\n", s);
2851                 parsed_valid_hugepagesz = true;
2852                 return 1;
2853         }
2854         /*
2855          * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2856          * so this hugepages= parameter goes to the "default hstate".
2857          */
2858         else if (!hugetlb_max_hstate)
2859                 mhp = &default_hstate_max_huge_pages;
2860         else
2861                 mhp = &parsed_hstate->max_huge_pages;
2862
2863         if (mhp == last_mhp) {
2864                 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
2865                 return 1;
2866         }
2867
2868         if (sscanf(s, "%lu", mhp) <= 0)
2869                 *mhp = 0;
2870
2871         /*
2872          * Global state is always initialized later in hugetlb_init.
2873          * But we need to allocate >= MAX_ORDER hstates here early to still
2874          * use the bootmem allocator.
2875          */
2876         if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2877                 hugetlb_hstate_alloc_pages(parsed_hstate);
2878
2879         last_mhp = mhp;
2880
2881         return 1;
2882 }
2883 __setup("hugepages=", hugetlb_nrpages_setup);
2884
2885 static int __init hugetlb_default_setup(char *s)
2886 {
2887         default_hstate_size = memparse(s, &s);
2888         return 1;
2889 }
2890 __setup("default_hugepagesz=", hugetlb_default_setup);
2891
2892 static unsigned int cpuset_mems_nr(unsigned int *array)
2893 {
2894         int node;
2895         unsigned int nr = 0;
2896
2897         for_each_node_mask(node, cpuset_current_mems_allowed)
2898                 nr += array[node];
2899
2900         return nr;
2901 }
2902
2903 #ifdef CONFIG_SYSCTL
2904 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2905                          struct ctl_table *table, int write,
2906                          void __user *buffer, size_t *length, loff_t *ppos)
2907 {
2908         struct hstate *h = &default_hstate;
2909         unsigned long tmp = h->max_huge_pages;
2910         int ret;
2911
2912         if (!hugepages_supported())
2913                 return -EOPNOTSUPP;
2914
2915         table->data = &tmp;
2916         table->maxlen = sizeof(unsigned long);
2917         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2918         if (ret)
2919                 goto out;
2920
2921         if (write)
2922                 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2923                                                   NUMA_NO_NODE, tmp, *length);
2924 out:
2925         return ret;
2926 }
2927
2928 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2929                           void __user *buffer, size_t *length, loff_t *ppos)
2930 {
2931
2932         return hugetlb_sysctl_handler_common(false, table, write,
2933                                                         buffer, length, ppos);
2934 }
2935
2936 #ifdef CONFIG_NUMA
2937 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2938                           void __user *buffer, size_t *length, loff_t *ppos)
2939 {
2940         return hugetlb_sysctl_handler_common(true, table, write,
2941                                                         buffer, length, ppos);
2942 }
2943 #endif /* CONFIG_NUMA */
2944
2945 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2946                         void __user *buffer,
2947                         size_t *length, loff_t *ppos)
2948 {
2949         struct hstate *h = &default_hstate;
2950         unsigned long tmp;
2951         int ret;
2952
2953         if (!hugepages_supported())
2954                 return -EOPNOTSUPP;
2955
2956         tmp = h->nr_overcommit_huge_pages;
2957
2958         if (write && hstate_is_gigantic(h))
2959                 return -EINVAL;
2960
2961         table->data = &tmp;
2962         table->maxlen = sizeof(unsigned long);
2963         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2964         if (ret)
2965                 goto out;
2966
2967         if (write) {
2968                 spin_lock(&hugetlb_lock);
2969                 h->nr_overcommit_huge_pages = tmp;
2970                 spin_unlock(&hugetlb_lock);
2971         }
2972 out:
2973         return ret;
2974 }
2975
2976 #endif /* CONFIG_SYSCTL */
2977
2978 void hugetlb_report_meminfo(struct seq_file *m)
2979 {
2980         struct hstate *h = &default_hstate;
2981         if (!hugepages_supported())
2982                 return;
2983         seq_printf(m,
2984                         "HugePages_Total:   %5lu\n"
2985                         "HugePages_Free:    %5lu\n"
2986                         "HugePages_Rsvd:    %5lu\n"
2987                         "HugePages_Surp:    %5lu\n"
2988                         "Hugepagesize:   %8lu kB\n",
2989                         h->nr_huge_pages,
2990                         h->free_huge_pages,
2991                         h->resv_huge_pages,
2992                         h->surplus_huge_pages,
2993                         1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2994 }
2995
2996 int hugetlb_report_node_meminfo(int nid, char *buf)
2997 {
2998         struct hstate *h = &default_hstate;
2999         if (!hugepages_supported())
3000                 return 0;
3001         return sprintf(buf,
3002                 "Node %d HugePages_Total: %5u\n"
3003                 "Node %d HugePages_Free:  %5u\n"
3004                 "Node %d HugePages_Surp:  %5u\n",
3005                 nid, h->nr_huge_pages_node[nid],
3006                 nid, h->free_huge_pages_node[nid],
3007                 nid, h->surplus_huge_pages_node[nid]);
3008 }
3009
3010 void hugetlb_show_meminfo(void)
3011 {
3012         struct hstate *h;
3013         int nid;
3014
3015         if (!hugepages_supported())
3016                 return;
3017
3018         for_each_node_state(nid, N_MEMORY)
3019                 for_each_hstate(h)
3020                         pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3021                                 nid,
3022                                 h->nr_huge_pages_node[nid],
3023                                 h->free_huge_pages_node[nid],
3024                                 h->surplus_huge_pages_node[nid],
3025                                 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3026 }
3027
3028 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3029 {
3030         seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3031                    atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3032 }
3033
3034 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3035 unsigned long hugetlb_total_pages(void)
3036 {
3037         struct hstate *h;
3038         unsigned long nr_total_pages = 0;
3039
3040         for_each_hstate(h)
3041                 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3042         return nr_total_pages;
3043 }
3044
3045 static int hugetlb_acct_memory(struct hstate *h, long delta)
3046 {
3047         int ret = -ENOMEM;
3048
3049         spin_lock(&hugetlb_lock);
3050         /*
3051          * When cpuset is configured, it breaks the strict hugetlb page
3052          * reservation as the accounting is done on a global variable. Such
3053          * reservation is completely rubbish in the presence of cpuset because
3054          * the reservation is not checked against page availability for the
3055          * current cpuset. Application can still potentially OOM'ed by kernel
3056          * with lack of free htlb page in cpuset that the task is in.
3057          * Attempt to enforce strict accounting with cpuset is almost
3058          * impossible (or too ugly) because cpuset is too fluid that
3059          * task or memory node can be dynamically moved between cpusets.
3060          *
3061          * The change of semantics for shared hugetlb mapping with cpuset is
3062          * undesirable. However, in order to preserve some of the semantics,
3063          * we fall back to check against current free page availability as
3064          * a best attempt and hopefully to minimize the impact of changing
3065          * semantics that cpuset has.
3066          */
3067         if (delta > 0) {
3068                 if (gather_surplus_pages(h, delta) < 0)
3069                         goto out;
3070
3071                 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
3072                         return_unused_surplus_pages(h, delta);
3073                         goto out;
3074                 }
3075         }
3076
3077         ret = 0;
3078         if (delta < 0)
3079                 return_unused_surplus_pages(h, (unsigned long) -delta);
3080
3081 out:
3082         spin_unlock(&hugetlb_lock);
3083         return ret;
3084 }
3085
3086 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3087 {
3088         struct resv_map *resv = vma_resv_map(vma);
3089
3090         /*
3091          * This new VMA should share its siblings reservation map if present.
3092          * The VMA will only ever have a valid reservation map pointer where
3093          * it is being copied for another still existing VMA.  As that VMA
3094          * has a reference to the reservation map it cannot disappear until
3095          * after this open call completes.  It is therefore safe to take a
3096          * new reference here without additional locking.
3097          */
3098         if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3099                 kref_get(&resv->refs);
3100 }
3101
3102 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3103 {
3104         struct hstate *h = hstate_vma(vma);
3105         struct resv_map *resv = vma_resv_map(vma);
3106         struct hugepage_subpool *spool = subpool_vma(vma);
3107         unsigned long reserve, start, end;
3108         long gbl_reserve;
3109
3110         if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3111                 return;
3112
3113         start = vma_hugecache_offset(h, vma, vma->vm_start);
3114         end = vma_hugecache_offset(h, vma, vma->vm_end);
3115
3116         reserve = (end - start) - region_count(resv, start, end);
3117
3118         kref_put(&resv->refs, resv_map_release);
3119
3120         if (reserve) {
3121                 /*
3122                  * Decrement reserve counts.  The global reserve count may be
3123                  * adjusted if the subpool has a minimum size.
3124                  */
3125                 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3126                 hugetlb_acct_memory(h, -gbl_reserve);
3127         }
3128 }
3129
3130 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3131 {
3132         if (addr & ~(huge_page_mask(hstate_vma(vma))))
3133                 return -EINVAL;
3134         return 0;
3135 }
3136
3137 /*
3138  * We cannot handle pagefaults against hugetlb pages at all.  They cause
3139  * handle_mm_fault() to try to instantiate regular-sized pages in the
3140  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
3141  * this far.
3142  */
3143 static int hugetlb_vm_op_fault(struct vm_fault *vmf)
3144 {
3145         BUG();
3146         return 0;
3147 }
3148
3149 const struct vm_operations_struct hugetlb_vm_ops = {
3150         .fault = hugetlb_vm_op_fault,
3151         .open = hugetlb_vm_op_open,
3152         .close = hugetlb_vm_op_close,
3153         .split = hugetlb_vm_op_split,
3154 };
3155
3156 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3157                                 int writable)
3158 {
3159         pte_t entry;
3160
3161         if (writable) {
3162                 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3163                                          vma->vm_page_prot)));
3164         } else {
3165                 entry = huge_pte_wrprotect(mk_huge_pte(page,
3166                                            vma->vm_page_prot));
3167         }
3168         entry = pte_mkyoung(entry);
3169         entry = pte_mkhuge(entry);
3170         entry = arch_make_huge_pte(entry, vma, page, writable);
3171
3172         return entry;
3173 }
3174
3175 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3176                                    unsigned long address, pte_t *ptep)
3177 {
3178         pte_t entry;
3179
3180         entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3181         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3182                 update_mmu_cache(vma, address, ptep);
3183 }
3184
3185 bool is_hugetlb_entry_migration(pte_t pte)
3186 {
3187         swp_entry_t swp;
3188
3189         if (huge_pte_none(pte) || pte_present(pte))
3190                 return false;
3191         swp = pte_to_swp_entry(pte);
3192         if (non_swap_entry(swp) && is_migration_entry(swp))
3193                 return true;
3194         else
3195                 return false;
3196 }
3197
3198 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3199 {
3200         swp_entry_t swp;
3201
3202         if (huge_pte_none(pte) || pte_present(pte))
3203                 return 0;
3204         swp = pte_to_swp_entry(pte);
3205         if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3206                 return 1;
3207         else
3208                 return 0;
3209 }
3210
3211 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3212                             struct vm_area_struct *vma)
3213 {
3214         pte_t *src_pte, *dst_pte, entry, dst_entry;
3215         struct page *ptepage;
3216         unsigned long addr;
3217         int cow;
3218         struct hstate *h = hstate_vma(vma);
3219         unsigned long sz = huge_page_size(h);
3220         unsigned long mmun_start;       /* For mmu_notifiers */
3221         unsigned long mmun_end;         /* For mmu_notifiers */
3222         int ret = 0;
3223
3224         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3225
3226         mmun_start = vma->vm_start;
3227         mmun_end = vma->vm_end;
3228         if (cow)
3229                 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
3230
3231         for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3232                 spinlock_t *src_ptl, *dst_ptl;
3233                 src_pte = huge_pte_offset(src, addr, sz);
3234                 if (!src_pte)
3235                         continue;
3236                 dst_pte = huge_pte_alloc(dst, addr, sz);
3237                 if (!dst_pte) {
3238                         ret = -ENOMEM;
3239                         break;
3240                 }
3241
3242                 /*
3243                  * If the pagetables are shared don't copy or take references.
3244                  * dst_pte == src_pte is the common case of src/dest sharing.
3245                  *
3246                  * However, src could have 'unshared' and dst shares with
3247                  * another vma.  If dst_pte !none, this implies sharing.
3248                  * Check here before taking page table lock, and once again
3249                  * after taking the lock below.
3250                  */
3251                 dst_entry = huge_ptep_get(dst_pte);
3252                 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
3253                         continue;
3254
3255                 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3256                 src_ptl = huge_pte_lockptr(h, src, src_pte);
3257                 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3258                 entry = huge_ptep_get(src_pte);
3259                 dst_entry = huge_ptep_get(dst_pte);
3260                 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3261                         /*
3262                          * Skip if src entry none.  Also, skip in the
3263                          * unlikely case dst entry !none as this implies
3264                          * sharing with another vma.
3265                          */
3266                         ;
3267                 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3268                                     is_hugetlb_entry_hwpoisoned(entry))) {
3269                         swp_entry_t swp_entry = pte_to_swp_entry(entry);
3270
3271                         if (is_write_migration_entry(swp_entry) && cow) {
3272                                 /*
3273                                  * COW mappings require pages in both
3274                                  * parent and child to be set to read.
3275                                  */
3276                                 make_migration_entry_read(&swp_entry);
3277                                 entry = swp_entry_to_pte(swp_entry);
3278                                 set_huge_swap_pte_at(src, addr, src_pte,
3279                                                      entry, sz);
3280                         }
3281                         set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3282                 } else {
3283                         if (cow) {
3284                                 huge_ptep_set_wrprotect(src, addr, src_pte);
3285                                 mmu_notifier_invalidate_range(src, mmun_start,
3286                                                                    mmun_end);
3287                         }
3288                         entry = huge_ptep_get(src_pte);
3289                         ptepage = pte_page(entry);
3290                         get_page(ptepage);
3291                         page_dup_rmap(ptepage, true);
3292                         set_huge_pte_at(dst, addr, dst_pte, entry);
3293                         hugetlb_count_add(pages_per_huge_page(h), dst);
3294                 }
3295                 spin_unlock(src_ptl);
3296                 spin_unlock(dst_ptl);
3297         }
3298
3299         if (cow)
3300                 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
3301
3302         return ret;
3303 }
3304
3305 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3306                             unsigned long start, unsigned long end,
3307                             struct page *ref_page)
3308 {
3309         struct mm_struct *mm = vma->vm_mm;
3310         unsigned long address;
3311         pte_t *ptep;
3312         pte_t pte;
3313         spinlock_t *ptl;
3314         struct page *page;
3315         struct hstate *h = hstate_vma(vma);
3316         unsigned long sz = huge_page_size(h);
3317         const unsigned long mmun_start = start; /* For mmu_notifiers */
3318         const unsigned long mmun_end   = end;   /* For mmu_notifiers */
3319
3320         WARN_ON(!is_vm_hugetlb_page(vma));
3321         BUG_ON(start & ~huge_page_mask(h));
3322         BUG_ON(end & ~huge_page_mask(h));
3323
3324         /*
3325          * This is a hugetlb vma, all the pte entries should point
3326          * to huge page.
3327          */
3328         tlb_remove_check_page_size_change(tlb, sz);
3329         tlb_start_vma(tlb, vma);
3330         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3331         address = start;
3332         for (; address < end; address += sz) {
3333                 ptep = huge_pte_offset(mm, address, sz);
3334                 if (!ptep)
3335                         continue;
3336
3337                 ptl = huge_pte_lock(h, mm, ptep);
3338                 if (huge_pmd_unshare(mm, &address, ptep)) {
3339                         spin_unlock(ptl);
3340                         continue;
3341                 }
3342
3343                 pte = huge_ptep_get(ptep);
3344                 if (huge_pte_none(pte)) {
3345                         spin_unlock(ptl);
3346                         continue;
3347                 }
3348
3349                 /*
3350                  * Migrating hugepage or HWPoisoned hugepage is already
3351                  * unmapped and its refcount is dropped, so just clear pte here.
3352                  */
3353                 if (unlikely(!pte_present(pte))) {
3354                         huge_pte_clear(mm, address, ptep, sz);
3355                         spin_unlock(ptl);
3356                         continue;
3357                 }
3358
3359                 page = pte_page(pte);
3360                 /*
3361                  * If a reference page is supplied, it is because a specific
3362                  * page is being unmapped, not a range. Ensure the page we
3363                  * are about to unmap is the actual page of interest.
3364                  */
3365                 if (ref_page) {
3366                         if (page != ref_page) {
3367                                 spin_unlock(ptl);
3368                                 continue;
3369                         }
3370                         /*
3371                          * Mark the VMA as having unmapped its page so that
3372                          * future faults in this VMA will fail rather than
3373                          * looking like data was lost
3374                          */
3375                         set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3376                 }
3377
3378                 pte = huge_ptep_get_and_clear(mm, address, ptep);
3379                 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
3380                 if (huge_pte_dirty(pte))
3381                         set_page_dirty(page);
3382
3383                 hugetlb_count_sub(pages_per_huge_page(h), mm);
3384                 page_remove_rmap(page, true);
3385
3386                 spin_unlock(ptl);
3387                 tlb_remove_page_size(tlb, page, huge_page_size(h));
3388                 /*
3389                  * Bail out after unmapping reference page if supplied
3390                  */
3391                 if (ref_page)
3392                         break;
3393         }
3394         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3395         tlb_end_vma(tlb, vma);
3396 }
3397
3398 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3399                           struct vm_area_struct *vma, unsigned long start,
3400                           unsigned long end, struct page *ref_page)
3401 {
3402         __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3403
3404         /*
3405          * Clear this flag so that x86's huge_pmd_share page_table_shareable
3406          * test will fail on a vma being torn down, and not grab a page table
3407          * on its way out.  We're lucky that the flag has such an appropriate
3408          * name, and can in fact be safely cleared here. We could clear it
3409          * before the __unmap_hugepage_range above, but all that's necessary
3410          * is to clear it before releasing the i_mmap_rwsem. This works
3411          * because in the context this is called, the VMA is about to be
3412          * destroyed and the i_mmap_rwsem is held.
3413          */
3414         vma->vm_flags &= ~VM_MAYSHARE;
3415 }
3416
3417 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3418                           unsigned long end, struct page *ref_page)
3419 {
3420         struct mm_struct *mm;
3421         struct mmu_gather tlb;
3422
3423         mm = vma->vm_mm;
3424
3425         tlb_gather_mmu(&tlb, mm, start, end);
3426         __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3427         tlb_finish_mmu(&tlb, start, end);
3428 }
3429
3430 /*
3431  * This is called when the original mapper is failing to COW a MAP_PRIVATE
3432  * mappping it owns the reserve page for. The intention is to unmap the page
3433  * from other VMAs and let the children be SIGKILLed if they are faulting the
3434  * same region.
3435  */
3436 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3437                               struct page *page, unsigned long address)
3438 {
3439         struct hstate *h = hstate_vma(vma);
3440         struct vm_area_struct *iter_vma;
3441         struct address_space *mapping;
3442         pgoff_t pgoff;
3443
3444         /*
3445          * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3446          * from page cache lookup which is in HPAGE_SIZE units.
3447          */
3448         address = address & huge_page_mask(h);
3449         pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3450                         vma->vm_pgoff;
3451         mapping = vma->vm_file->f_mapping;
3452
3453         /*
3454          * Take the mapping lock for the duration of the table walk. As
3455          * this mapping should be shared between all the VMAs,
3456          * __unmap_hugepage_range() is called as the lock is already held
3457          */
3458         i_mmap_lock_write(mapping);
3459         vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3460                 /* Do not unmap the current VMA */
3461                 if (iter_vma == vma)
3462                         continue;
3463
3464                 /*
3465                  * Shared VMAs have their own reserves and do not affect
3466                  * MAP_PRIVATE accounting but it is possible that a shared
3467                  * VMA is using the same page so check and skip such VMAs.
3468                  */
3469                 if (iter_vma->vm_flags & VM_MAYSHARE)
3470                         continue;
3471
3472                 /*
3473                  * Unmap the page from other VMAs without their own reserves.
3474                  * They get marked to be SIGKILLed if they fault in these
3475                  * areas. This is because a future no-page fault on this VMA
3476                  * could insert a zeroed page instead of the data existing
3477                  * from the time of fork. This would look like data corruption
3478                  */
3479                 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3480                         unmap_hugepage_range(iter_vma, address,
3481                                              address + huge_page_size(h), page);
3482         }
3483         i_mmap_unlock_write(mapping);
3484 }
3485
3486 /*
3487  * Hugetlb_cow() should be called with page lock of the original hugepage held.
3488  * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3489  * cannot race with other handlers or page migration.
3490  * Keep the pte_same checks anyway to make transition from the mutex easier.
3491  */
3492 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3493                        unsigned long address, pte_t *ptep,
3494                        struct page *pagecache_page, spinlock_t *ptl)
3495 {
3496         pte_t pte;
3497         struct hstate *h = hstate_vma(vma);
3498         struct page *old_page, *new_page;
3499         int ret = 0, outside_reserve = 0;
3500         unsigned long mmun_start;       /* For mmu_notifiers */
3501         unsigned long mmun_end;         /* For mmu_notifiers */
3502
3503         pte = huge_ptep_get(ptep);
3504         old_page = pte_page(pte);
3505
3506 retry_avoidcopy:
3507         /* If no-one else is actually using this page, avoid the copy
3508          * and just make the page writable */
3509         if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3510                 page_move_anon_rmap(old_page, vma);
3511                 set_huge_ptep_writable(vma, address, ptep);
3512                 return 0;
3513         }
3514
3515         /*
3516          * If the process that created a MAP_PRIVATE mapping is about to
3517          * perform a COW due to a shared page count, attempt to satisfy
3518          * the allocation without using the existing reserves. The pagecache
3519          * page is used to determine if the reserve at this address was
3520          * consumed or not. If reserves were used, a partial faulted mapping
3521          * at the time of fork() could consume its reserves on COW instead
3522          * of the full address range.
3523          */
3524         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3525                         old_page != pagecache_page)
3526                 outside_reserve = 1;
3527
3528         get_page(old_page);
3529
3530         /*
3531          * Drop page table lock as buddy allocator may be called. It will
3532          * be acquired again before returning to the caller, as expected.
3533          */
3534         spin_unlock(ptl);
3535         new_page = alloc_huge_page(vma, address, outside_reserve);
3536
3537         if (IS_ERR(new_page)) {
3538                 /*
3539                  * If a process owning a MAP_PRIVATE mapping fails to COW,
3540                  * it is due to references held by a child and an insufficient
3541                  * huge page pool. To guarantee the original mappers
3542                  * reliability, unmap the page from child processes. The child
3543                  * may get SIGKILLed if it later faults.
3544                  */
3545                 if (outside_reserve) {
3546                         put_page(old_page);
3547                         BUG_ON(huge_pte_none(pte));
3548                         unmap_ref_private(mm, vma, old_page, address);
3549                         BUG_ON(huge_pte_none(pte));
3550                         spin_lock(ptl);
3551                         ptep = huge_pte_offset(mm, address & huge_page_mask(h),
3552                                                huge_page_size(h));
3553                         if (likely(ptep &&
3554                                    pte_same(huge_ptep_get(ptep), pte)))
3555                                 goto retry_avoidcopy;
3556                         /*
3557                          * race occurs while re-acquiring page table
3558                          * lock, and our job is done.
3559                          */
3560                         return 0;
3561                 }
3562
3563                 ret = (PTR_ERR(new_page) == -ENOMEM) ?
3564                         VM_FAULT_OOM : VM_FAULT_SIGBUS;
3565                 goto out_release_old;
3566         }
3567
3568         /*
3569          * When the original hugepage is shared one, it does not have
3570          * anon_vma prepared.
3571          */
3572         if (unlikely(anon_vma_prepare(vma))) {
3573                 ret = VM_FAULT_OOM;
3574                 goto out_release_all;
3575         }
3576
3577         copy_user_huge_page(new_page, old_page, address, vma,
3578                             pages_per_huge_page(h));
3579         __SetPageUptodate(new_page);
3580         set_page_huge_active(new_page);
3581
3582         mmun_start = address & huge_page_mask(h);
3583         mmun_end = mmun_start + huge_page_size(h);
3584         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3585
3586         /*
3587          * Retake the page table lock to check for racing updates
3588          * before the page tables are altered
3589          */
3590         spin_lock(ptl);
3591         ptep = huge_pte_offset(mm, address & huge_page_mask(h),
3592                                huge_page_size(h));
3593         if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3594                 ClearPagePrivate(new_page);
3595
3596                 /* Break COW */
3597                 huge_ptep_clear_flush(vma, address, ptep);
3598                 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
3599                 set_huge_pte_at(mm, address, ptep,
3600                                 make_huge_pte(vma, new_page, 1));
3601                 page_remove_rmap(old_page, true);
3602                 hugepage_add_new_anon_rmap(new_page, vma, address);
3603                 /* Make the old page be freed below */
3604                 new_page = old_page;
3605         }
3606         spin_unlock(ptl);
3607         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3608 out_release_all:
3609         restore_reserve_on_error(h, vma, address, new_page);
3610         put_page(new_page);
3611 out_release_old:
3612         put_page(old_page);
3613
3614         spin_lock(ptl); /* Caller expects lock to be held */
3615         return ret;
3616 }
3617
3618 /* Return the pagecache page at a given address within a VMA */
3619 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3620                         struct vm_area_struct *vma, unsigned long address)
3621 {
3622         struct address_space *mapping;
3623         pgoff_t idx;
3624
3625         mapping = vma->vm_file->f_mapping;
3626         idx = vma_hugecache_offset(h, vma, address);
3627
3628         return find_lock_page(mapping, idx);
3629 }
3630
3631 /*
3632  * Return whether there is a pagecache page to back given address within VMA.
3633  * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3634  */
3635 static bool hugetlbfs_pagecache_present(struct hstate *h,
3636                         struct vm_area_struct *vma, unsigned long address)
3637 {
3638         struct address_space *mapping;
3639         pgoff_t idx;
3640         struct page *page;
3641
3642         mapping = vma->vm_file->f_mapping;
3643         idx = vma_hugecache_offset(h, vma, address);
3644
3645         page = find_get_page(mapping, idx);
3646         if (page)
3647                 put_page(page);
3648         return page != NULL;
3649 }
3650
3651 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3652                            pgoff_t idx)
3653 {
3654         struct inode *inode = mapping->host;
3655         struct hstate *h = hstate_inode(inode);
3656         int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3657
3658         if (err)
3659                 return err;
3660         ClearPagePrivate(page);
3661
3662         /*
3663          * set page dirty so that it will not be removed from cache/file
3664          * by non-hugetlbfs specific code paths.
3665          */
3666         set_page_dirty(page);
3667
3668         spin_lock(&inode->i_lock);
3669         inode->i_blocks += blocks_per_huge_page(h);
3670         spin_unlock(&inode->i_lock);
3671         return 0;
3672 }
3673
3674 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
3675                            struct address_space *mapping, pgoff_t idx,
3676                            unsigned long address, pte_t *ptep, unsigned int flags)
3677 {
3678         struct hstate *h = hstate_vma(vma);
3679         int ret = VM_FAULT_SIGBUS;
3680         int anon_rmap = 0;
3681         unsigned long size;
3682         struct page *page;
3683         pte_t new_pte;
3684         spinlock_t *ptl;
3685
3686         /*
3687          * Currently, we are forced to kill the process in the event the
3688          * original mapper has unmapped pages from the child due to a failed
3689          * COW. Warn that such a situation has occurred as it may not be obvious
3690          */
3691         if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3692                 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
3693                            current->pid);
3694                 return ret;
3695         }
3696
3697         /*
3698          * Use page lock to guard against racing truncation
3699          * before we get page_table_lock.
3700          */
3701 retry:
3702         page = find_lock_page(mapping, idx);
3703         if (!page) {
3704                 size = i_size_read(mapping->host) >> huge_page_shift(h);
3705                 if (idx >= size)
3706                         goto out;
3707
3708                 /*
3709                  * Check for page in userfault range
3710                  */
3711                 if (userfaultfd_missing(vma)) {
3712                         u32 hash;
3713                         struct vm_fault vmf = {
3714                                 .vma = vma,
3715                                 .address = address,
3716                                 .flags = flags,
3717                                 /*
3718                                  * Hard to debug if it ends up being
3719                                  * used by a callee that assumes
3720                                  * something about the other
3721                                  * uninitialized fields... same as in
3722                                  * memory.c
3723                                  */
3724                         };
3725
3726                         /*
3727                          * hugetlb_fault_mutex must be dropped before
3728                          * handling userfault.  Reacquire after handling
3729                          * fault to make calling code simpler.
3730                          */
3731                         hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping,
3732                                                         idx, address);
3733                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3734                         ret = handle_userfault(&vmf, VM_UFFD_MISSING);
3735                         mutex_lock(&hugetlb_fault_mutex_table[hash]);
3736                         goto out;
3737                 }
3738
3739                 page = alloc_huge_page(vma, address, 0);
3740                 if (IS_ERR(page)) {
3741                         ret = PTR_ERR(page);
3742                         if (ret == -ENOMEM)
3743                                 ret = VM_FAULT_OOM;
3744                         else
3745                                 ret = VM_FAULT_SIGBUS;
3746                         goto out;
3747                 }
3748                 clear_huge_page(page, address, pages_per_huge_page(h));
3749                 __SetPageUptodate(page);
3750                 set_page_huge_active(page);
3751
3752                 if (vma->vm_flags & VM_MAYSHARE) {
3753                         int err = huge_add_to_page_cache(page, mapping, idx);
3754                         if (err) {
3755                                 put_page(page);
3756                                 if (err == -EEXIST)
3757                                         goto retry;
3758                                 goto out;
3759                         }
3760                 } else {
3761                         lock_page(page);
3762                         if (unlikely(anon_vma_prepare(vma))) {
3763                                 ret = VM_FAULT_OOM;
3764                                 goto backout_unlocked;
3765                         }
3766                         anon_rmap = 1;
3767                 }
3768         } else {
3769                 /*
3770                  * If memory error occurs between mmap() and fault, some process
3771                  * don't have hwpoisoned swap entry for errored virtual address.
3772                  * So we need to block hugepage fault by PG_hwpoison bit check.
3773                  */
3774                 if (unlikely(PageHWPoison(page))) {
3775                         ret = VM_FAULT_HWPOISON |
3776                                 VM_FAULT_SET_HINDEX(hstate_index(h));
3777                         goto backout_unlocked;
3778                 }
3779         }
3780
3781         /*
3782          * If we are going to COW a private mapping later, we examine the
3783          * pending reservations for this page now. This will ensure that
3784          * any allocations necessary to record that reservation occur outside
3785          * the spinlock.
3786          */
3787         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3788                 if (vma_needs_reservation(h, vma, address) < 0) {
3789                         ret = VM_FAULT_OOM;
3790                         goto backout_unlocked;
3791                 }
3792                 /* Just decrements count, does not deallocate */
3793                 vma_end_reservation(h, vma, address);
3794         }
3795
3796         ptl = huge_pte_lock(h, mm, ptep);
3797         size = i_size_read(mapping->host) >> huge_page_shift(h);
3798         if (idx >= size)
3799                 goto backout;
3800
3801         ret = 0;
3802         if (!huge_pte_none(huge_ptep_get(ptep)))
3803                 goto backout;
3804
3805         if (anon_rmap) {
3806                 ClearPagePrivate(page);
3807                 hugepage_add_new_anon_rmap(page, vma, address);
3808         } else
3809                 page_dup_rmap(page, true);
3810         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3811                                 && (vma->vm_flags & VM_SHARED)));
3812         set_huge_pte_at(mm, address, ptep, new_pte);
3813
3814         hugetlb_count_add(pages_per_huge_page(h), mm);
3815         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3816                 /* Optimization, do the COW without a second fault */
3817                 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
3818         }
3819
3820         spin_unlock(ptl);
3821         unlock_page(page);
3822 out:
3823         return ret;
3824
3825 backout:
3826         spin_unlock(ptl);
3827 backout_unlocked:
3828         unlock_page(page);
3829         restore_reserve_on_error(h, vma, address, page);
3830         put_page(page);
3831         goto out;
3832 }
3833
3834 #ifdef CONFIG_SMP
3835 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3836                             struct vm_area_struct *vma,
3837                             struct address_space *mapping,
3838                             pgoff_t idx, unsigned long address)
3839 {
3840         unsigned long key[2];
3841         u32 hash;
3842
3843         if (vma->vm_flags & VM_SHARED) {
3844                 key[0] = (unsigned long) mapping;
3845                 key[1] = idx;
3846         } else {
3847                 key[0] = (unsigned long) mm;
3848                 key[1] = address >> huge_page_shift(h);
3849         }
3850
3851         hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3852
3853         return hash & (num_fault_mutexes - 1);
3854 }
3855 #else
3856 /*
3857  * For uniprocesor systems we always use a single mutex, so just
3858  * return 0 and avoid the hashing overhead.
3859  */
3860 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3861                             struct vm_area_struct *vma,
3862                             struct address_space *mapping,
3863                             pgoff_t idx, unsigned long address)
3864 {
3865         return 0;
3866 }
3867 #endif
3868
3869 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3870                         unsigned long address, unsigned int flags)
3871 {
3872         pte_t *ptep, entry;
3873         spinlock_t *ptl;
3874         int ret;
3875         u32 hash;
3876         pgoff_t idx;
3877         struct page *page = NULL;
3878         struct page *pagecache_page = NULL;
3879         struct hstate *h = hstate_vma(vma);
3880         struct address_space *mapping;
3881         int need_wait_lock = 0;
3882
3883         address &= huge_page_mask(h);
3884
3885         ptep = huge_pte_offset(mm, address, huge_page_size(h));
3886         if (ptep) {
3887                 entry = huge_ptep_get(ptep);
3888                 if (unlikely(is_hugetlb_entry_migration(entry))) {
3889                         migration_entry_wait_huge(vma, mm, ptep);
3890                         return 0;
3891                 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
3892                         return VM_FAULT_HWPOISON_LARGE |
3893                                 VM_FAULT_SET_HINDEX(hstate_index(h));
3894         } else {
3895                 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
3896                 if (!ptep)
3897                         return VM_FAULT_OOM;
3898         }
3899
3900         mapping = vma->vm_file->f_mapping;
3901         idx = vma_hugecache_offset(h, vma, address);
3902
3903         /*
3904          * Serialize hugepage allocation and instantiation, so that we don't
3905          * get spurious allocation failures if two CPUs race to instantiate
3906          * the same page in the page cache.
3907          */
3908         hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address);
3909         mutex_lock(&hugetlb_fault_mutex_table[hash]);
3910
3911         entry = huge_ptep_get(ptep);
3912         if (huge_pte_none(entry)) {
3913                 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
3914                 goto out_mutex;
3915         }
3916
3917         ret = 0;
3918
3919         /*
3920          * entry could be a migration/hwpoison entry at this point, so this
3921          * check prevents the kernel from going below assuming that we have
3922          * a active hugepage in pagecache. This goto expects the 2nd page fault,
3923          * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3924          * handle it.
3925          */
3926         if (!pte_present(entry))
3927                 goto out_mutex;
3928
3929         /*
3930          * If we are going to COW the mapping later, we examine the pending
3931          * reservations for this page now. This will ensure that any
3932          * allocations necessary to record that reservation occur outside the
3933          * spinlock. For private mappings, we also lookup the pagecache
3934          * page now as it is used to determine if a reservation has been
3935          * consumed.
3936          */
3937         if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
3938                 if (vma_needs_reservation(h, vma, address) < 0) {
3939                         ret = VM_FAULT_OOM;
3940                         goto out_mutex;
3941                 }
3942                 /* Just decrements count, does not deallocate */
3943                 vma_end_reservation(h, vma, address);
3944
3945                 if (!(vma->vm_flags & VM_MAYSHARE))
3946                         pagecache_page = hugetlbfs_pagecache_page(h,
3947                                                                 vma, address);
3948         }
3949
3950         ptl = huge_pte_lock(h, mm, ptep);
3951
3952         /* Check for a racing update before calling hugetlb_cow */
3953         if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
3954                 goto out_ptl;
3955
3956         /*
3957          * hugetlb_cow() requires page locks of pte_page(entry) and
3958          * pagecache_page, so here we need take the former one
3959          * when page != pagecache_page or !pagecache_page.
3960          */
3961         page = pte_page(entry);
3962         if (page != pagecache_page)
3963                 if (!trylock_page(page)) {
3964                         need_wait_lock = 1;
3965                         goto out_ptl;
3966                 }
3967
3968         get_page(page);
3969
3970         if (flags & FAULT_FLAG_WRITE) {
3971                 if (!huge_pte_write(entry)) {
3972                         ret = hugetlb_cow(mm, vma, address, ptep,
3973                                           pagecache_page, ptl);
3974                         goto out_put_page;
3975                 }
3976                 entry = huge_pte_mkdirty(entry);
3977         }
3978         entry = pte_mkyoung(entry);
3979         if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3980                                                 flags & FAULT_FLAG_WRITE))
3981                 update_mmu_cache(vma, address, ptep);
3982 out_put_page:
3983         if (page != pagecache_page)
3984                 unlock_page(page);
3985         put_page(page);
3986 out_ptl:
3987         spin_unlock(ptl);
3988
3989         if (pagecache_page) {
3990                 unlock_page(pagecache_page);
3991                 put_page(pagecache_page);
3992         }
3993 out_mutex:
3994         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3995         /*
3996          * Generally it's safe to hold refcount during waiting page lock. But
3997          * here we just wait to defer the next page fault to avoid busy loop and
3998          * the page is not used after unlocked before returning from the current
3999          * page fault. So we are safe from accessing freed page, even if we wait
4000          * here without taking refcount.
4001          */
4002         if (need_wait_lock)
4003                 wait_on_page_locked(page);
4004         return ret;
4005 }
4006
4007 /*
4008  * Used by userfaultfd UFFDIO_COPY.  Based on mcopy_atomic_pte with
4009  * modifications for huge pages.
4010  */
4011 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4012                             pte_t *dst_pte,
4013                             struct vm_area_struct *dst_vma,
4014                             unsigned long dst_addr,
4015                             unsigned long src_addr,
4016                             struct page **pagep)
4017 {
4018         struct address_space *mapping;
4019         pgoff_t idx;
4020         unsigned long size;
4021         int vm_shared = dst_vma->vm_flags & VM_SHARED;
4022         struct hstate *h = hstate_vma(dst_vma);
4023         pte_t _dst_pte;
4024         spinlock_t *ptl;
4025         int ret;
4026         struct page *page;
4027
4028         if (!*pagep) {
4029                 ret = -ENOMEM;
4030                 page = alloc_huge_page(dst_vma, dst_addr, 0);
4031                 if (IS_ERR(page))
4032                         goto out;
4033
4034                 ret = copy_huge_page_from_user(page,
4035                                                 (const void __user *) src_addr,
4036                                                 pages_per_huge_page(h), false);
4037
4038                 /* fallback to copy_from_user outside mmap_sem */
4039                 if (unlikely(ret)) {
4040                         ret = -ENOENT;
4041                         *pagep = page;
4042                         /* don't free the page */
4043                         goto out;
4044                 }
4045         } else {
4046                 page = *pagep;
4047                 *pagep = NULL;
4048         }
4049
4050         /*
4051          * The memory barrier inside __SetPageUptodate makes sure that
4052          * preceding stores to the page contents become visible before
4053          * the set_pte_at() write.
4054          */
4055         __SetPageUptodate(page);
4056         set_page_huge_active(page);
4057
4058         mapping = dst_vma->vm_file->f_mapping;
4059         idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4060
4061         /*
4062          * If shared, add to page cache
4063          */
4064         if (vm_shared) {
4065                 size = i_size_read(mapping->host) >> huge_page_shift(h);
4066                 ret = -EFAULT;
4067                 if (idx >= size)
4068                         goto out_release_nounlock;
4069
4070                 /*
4071                  * Serialization between remove_inode_hugepages() and
4072                  * huge_add_to_page_cache() below happens through the
4073                  * hugetlb_fault_mutex_table that here must be hold by
4074                  * the caller.
4075                  */
4076                 ret = huge_add_to_page_cache(page, mapping, idx);
4077                 if (ret)
4078                         goto out_release_nounlock;
4079         }
4080
4081         ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4082         spin_lock(ptl);
4083
4084         /*
4085          * Recheck the i_size after holding PT lock to make sure not
4086          * to leave any page mapped (as page_mapped()) beyond the end
4087          * of the i_size (remove_inode_hugepages() is strict about
4088          * enforcing that). If we bail out here, we'll also leave a
4089          * page in the radix tree in the vm_shared case beyond the end
4090          * of the i_size, but remove_inode_hugepages() will take care
4091          * of it as soon as we drop the hugetlb_fault_mutex_table.
4092          */
4093         size = i_size_read(mapping->host) >> huge_page_shift(h);
4094         ret = -EFAULT;
4095         if (idx >= size)
4096                 goto out_release_unlock;
4097
4098         ret = -EEXIST;
4099         if (!huge_pte_none(huge_ptep_get(dst_pte)))
4100                 goto out_release_unlock;
4101
4102         if (vm_shared) {
4103                 page_dup_rmap(page, true);
4104         } else {
4105                 ClearPagePrivate(page);
4106                 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4107         }
4108
4109         _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4110         if (dst_vma->vm_flags & VM_WRITE)
4111                 _dst_pte = huge_pte_mkdirty(_dst_pte);
4112         _dst_pte = pte_mkyoung(_dst_pte);
4113
4114         set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4115
4116         (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4117                                         dst_vma->vm_flags & VM_WRITE);
4118         hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4119
4120         /* No need to invalidate - it was non-present before */
4121         update_mmu_cache(dst_vma, dst_addr, dst_pte);
4122
4123         spin_unlock(ptl);
4124         if (vm_shared)
4125                 unlock_page(page);
4126         ret = 0;
4127 out:
4128         return ret;
4129 out_release_unlock:
4130         spin_unlock(ptl);
4131         if (vm_shared)
4132                 unlock_page(page);
4133 out_release_nounlock:
4134         put_page(page);
4135         goto out;
4136 }
4137
4138 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4139                          struct page **pages, struct vm_area_struct **vmas,
4140                          unsigned long *position, unsigned long *nr_pages,
4141                          long i, unsigned int flags, int *nonblocking)
4142 {
4143         unsigned long pfn_offset;
4144         unsigned long vaddr = *position;
4145         unsigned long remainder = *nr_pages;
4146         struct hstate *h = hstate_vma(vma);
4147         int err = -EFAULT;
4148
4149         while (vaddr < vma->vm_end && remainder) {
4150                 pte_t *pte;
4151                 spinlock_t *ptl = NULL;
4152                 int absent;
4153                 struct page *page;
4154
4155                 /*
4156                  * If we have a pending SIGKILL, don't keep faulting pages and
4157                  * potentially allocating memory.
4158                  */
4159                 if (unlikely(fatal_signal_pending(current))) {
4160                         remainder = 0;
4161                         break;
4162                 }
4163
4164                 /*
4165                  * Some archs (sparc64, sh*) have multiple pte_ts to
4166                  * each hugepage.  We have to make sure we get the
4167                  * first, for the page indexing below to work.
4168                  *
4169                  * Note that page table lock is not held when pte is null.
4170                  */
4171                 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4172                                       huge_page_size(h));
4173                 if (pte)
4174                         ptl = huge_pte_lock(h, mm, pte);
4175                 absent = !pte || huge_pte_none(huge_ptep_get(pte));
4176
4177                 /*
4178                  * When coredumping, it suits get_dump_page if we just return
4179                  * an error where there's an empty slot with no huge pagecache
4180                  * to back it.  This way, we avoid allocating a hugepage, and
4181                  * the sparse dumpfile avoids allocating disk blocks, but its
4182                  * huge holes still show up with zeroes where they need to be.
4183                  */
4184                 if (absent && (flags & FOLL_DUMP) &&
4185                     !hugetlbfs_pagecache_present(h, vma, vaddr)) {
4186                         if (pte)
4187                                 spin_unlock(ptl);
4188                         remainder = 0;
4189                         break;
4190                 }
4191
4192                 /*
4193                  * We need call hugetlb_fault for both hugepages under migration
4194                  * (in which case hugetlb_fault waits for the migration,) and
4195                  * hwpoisoned hugepages (in which case we need to prevent the
4196                  * caller from accessing to them.) In order to do this, we use
4197                  * here is_swap_pte instead of is_hugetlb_entry_migration and
4198                  * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4199                  * both cases, and because we can't follow correct pages
4200                  * directly from any kind of swap entries.
4201                  */
4202                 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4203                     ((flags & FOLL_WRITE) &&
4204                       !huge_pte_write(huge_ptep_get(pte)))) {
4205                         int ret;
4206                         unsigned int fault_flags = 0;
4207
4208                         if (pte)
4209                                 spin_unlock(ptl);
4210                         if (flags & FOLL_WRITE)
4211                                 fault_flags |= FAULT_FLAG_WRITE;
4212                         if (nonblocking)
4213                                 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
4214                         if (flags & FOLL_NOWAIT)
4215                                 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4216                                         FAULT_FLAG_RETRY_NOWAIT;
4217                         if (flags & FOLL_TRIED) {
4218                                 VM_WARN_ON_ONCE(fault_flags &
4219                                                 FAULT_FLAG_ALLOW_RETRY);
4220                                 fault_flags |= FAULT_FLAG_TRIED;
4221                         }
4222                         ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4223                         if (ret & VM_FAULT_ERROR) {
4224                                 err = vm_fault_to_errno(ret, flags);
4225                                 remainder = 0;
4226                                 break;
4227                         }
4228                         if (ret & VM_FAULT_RETRY) {
4229                                 if (nonblocking)
4230                                         *nonblocking = 0;
4231                                 *nr_pages = 0;
4232                                 /*
4233                                  * VM_FAULT_RETRY must not return an
4234                                  * error, it will return zero
4235                                  * instead.
4236                                  *
4237                                  * No need to update "position" as the
4238                                  * caller will not check it after
4239                                  * *nr_pages is set to 0.
4240                                  */
4241                                 return i;
4242                         }
4243                         continue;
4244                 }
4245
4246                 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4247                 page = pte_page(huge_ptep_get(pte));
4248 same_page:
4249                 if (pages) {
4250                         pages[i] = mem_map_offset(page, pfn_offset);
4251                         get_page(pages[i]);
4252                 }
4253
4254                 if (vmas)
4255                         vmas[i] = vma;
4256
4257                 vaddr += PAGE_SIZE;
4258                 ++pfn_offset;
4259                 --remainder;
4260                 ++i;
4261                 if (vaddr < vma->vm_end && remainder &&
4262                                 pfn_offset < pages_per_huge_page(h)) {
4263                         /*
4264                          * We use pfn_offset to avoid touching the pageframes
4265                          * of this compound page.
4266                          */
4267                         goto same_page;
4268                 }
4269                 spin_unlock(ptl);
4270         }
4271         *nr_pages = remainder;
4272         /*
4273          * setting position is actually required only if remainder is
4274          * not zero but it's faster not to add a "if (remainder)"
4275          * branch.
4276          */
4277         *position = vaddr;
4278
4279         return i ? i : err;
4280 }
4281
4282 #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4283 /*
4284  * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4285  * implement this.
4286  */
4287 #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4288 #endif
4289
4290 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
4291                 unsigned long address, unsigned long end, pgprot_t newprot)
4292 {
4293         struct mm_struct *mm = vma->vm_mm;
4294         unsigned long start = address;
4295         pte_t *ptep;
4296         pte_t pte;
4297         struct hstate *h = hstate_vma(vma);
4298         unsigned long pages = 0;
4299
4300         BUG_ON(address >= end);
4301         flush_cache_range(vma, address, end);
4302
4303         mmu_notifier_invalidate_range_start(mm, start, end);
4304         i_mmap_lock_write(vma->vm_file->f_mapping);
4305         for (; address < end; address += huge_page_size(h)) {
4306                 spinlock_t *ptl;
4307                 ptep = huge_pte_offset(mm, address, huge_page_size(h));
4308                 if (!ptep)
4309                         continue;
4310                 ptl = huge_pte_lock(h, mm, ptep);
4311                 if (huge_pmd_unshare(mm, &address, ptep)) {
4312                         pages++;
4313                         spin_unlock(ptl);
4314                         continue;
4315                 }
4316                 pte = huge_ptep_get(ptep);
4317                 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
4318                         spin_unlock(ptl);
4319                         continue;
4320                 }
4321                 if (unlikely(is_hugetlb_entry_migration(pte))) {
4322                         swp_entry_t entry = pte_to_swp_entry(pte);
4323
4324                         if (is_write_migration_entry(entry)) {
4325                                 pte_t newpte;
4326
4327                                 make_migration_entry_read(&entry);
4328                                 newpte = swp_entry_to_pte(entry);
4329                                 set_huge_swap_pte_at(mm, address, ptep,
4330                                                      newpte, huge_page_size(h));
4331                                 pages++;
4332                         }
4333                         spin_unlock(ptl);
4334                         continue;
4335                 }
4336                 if (!huge_pte_none(pte)) {
4337                         pte = huge_ptep_get_and_clear(mm, address, ptep);
4338                         pte = pte_mkhuge(huge_pte_modify(pte, newprot));
4339                         pte = arch_make_huge_pte(pte, vma, NULL, 0);
4340                         set_huge_pte_at(mm, address, ptep, pte);
4341                         pages++;
4342                 }
4343                 spin_unlock(ptl);
4344         }
4345         /*
4346          * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
4347          * may have cleared our pud entry and done put_page on the page table:
4348          * once we release i_mmap_rwsem, another task can do the final put_page
4349          * and that page table be reused and filled with junk.
4350          */
4351         flush_hugetlb_tlb_range(vma, start, end);
4352         mmu_notifier_invalidate_range(mm, start, end);
4353         i_mmap_unlock_write(vma->vm_file->f_mapping);
4354         mmu_notifier_invalidate_range_end(mm, start, end);
4355
4356         return pages << h->order;
4357 }
4358
4359 int hugetlb_reserve_pages(struct inode *inode,
4360                                         long from, long to,
4361                                         struct vm_area_struct *vma,
4362                                         vm_flags_t vm_flags)
4363 {
4364         long ret, chg;
4365         struct hstate *h = hstate_inode(inode);
4366         struct hugepage_subpool *spool = subpool_inode(inode);
4367         struct resv_map *resv_map;
4368         long gbl_reserve;
4369
4370         /* This should never happen */
4371         if (from > to) {
4372                 VM_WARN(1, "%s called with a negative range\n", __func__);
4373                 return -EINVAL;
4374         }
4375
4376         /*
4377          * Only apply hugepage reservation if asked. At fault time, an
4378          * attempt will be made for VM_NORESERVE to allocate a page
4379          * without using reserves
4380          */
4381         if (vm_flags & VM_NORESERVE)
4382                 return 0;
4383
4384         /*
4385          * Shared mappings base their reservation on the number of pages that
4386          * are already allocated on behalf of the file. Private mappings need
4387          * to reserve the full area even if read-only as mprotect() may be
4388          * called to make the mapping read-write. Assume !vma is a shm mapping
4389          */
4390         if (!vma || vma->vm_flags & VM_MAYSHARE) {
4391                 resv_map = inode_resv_map(inode);
4392
4393                 chg = region_chg(resv_map, from, to);
4394
4395         } else {
4396                 resv_map = resv_map_alloc();
4397                 if (!resv_map)
4398                         return -ENOMEM;
4399
4400                 chg = to - from;
4401
4402                 set_vma_resv_map(vma, resv_map);
4403                 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4404         }
4405
4406         if (chg < 0) {
4407                 ret = chg;
4408                 goto out_err;
4409         }
4410
4411         /*
4412          * There must be enough pages in the subpool for the mapping. If
4413          * the subpool has a minimum size, there may be some global
4414          * reservations already in place (gbl_reserve).
4415          */
4416         gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4417         if (gbl_reserve < 0) {
4418                 ret = -ENOSPC;
4419                 goto out_err;
4420         }
4421
4422         /*
4423          * Check enough hugepages are available for the reservation.
4424          * Hand the pages back to the subpool if there are not
4425          */
4426         ret = hugetlb_acct_memory(h, gbl_reserve);
4427         if (ret < 0) {
4428                 /* put back original number of pages, chg */
4429                 (void)hugepage_subpool_put_pages(spool, chg);
4430                 goto out_err;
4431         }
4432
4433         /*
4434          * Account for the reservations made. Shared mappings record regions
4435          * that have reservations as they are shared by multiple VMAs.
4436          * When the last VMA disappears, the region map says how much
4437          * the reservation was and the page cache tells how much of
4438          * the reservation was consumed. Private mappings are per-VMA and
4439          * only the consumed reservations are tracked. When the VMA
4440          * disappears, the original reservation is the VMA size and the
4441          * consumed reservations are stored in the map. Hence, nothing
4442          * else has to be done for private mappings here
4443          */
4444         if (!vma || vma->vm_flags & VM_MAYSHARE) {
4445                 long add = region_add(resv_map, from, to);
4446
4447                 if (unlikely(chg > add)) {
4448                         /*
4449                          * pages in this range were added to the reserve
4450                          * map between region_chg and region_add.  This
4451                          * indicates a race with alloc_huge_page.  Adjust
4452                          * the subpool and reserve counts modified above
4453                          * based on the difference.
4454                          */
4455                         long rsv_adjust;
4456
4457                         rsv_adjust = hugepage_subpool_put_pages(spool,
4458                                                                 chg - add);
4459                         hugetlb_acct_memory(h, -rsv_adjust);
4460                 }
4461         }
4462         return 0;
4463 out_err:
4464         if (!vma || vma->vm_flags & VM_MAYSHARE)
4465                 /* Don't call region_abort if region_chg failed */
4466                 if (chg >= 0)
4467                         region_abort(resv_map, from, to);
4468         if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4469                 kref_put(&resv_map->refs, resv_map_release);
4470         return ret;
4471 }
4472
4473 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4474                                                                 long freed)
4475 {
4476         struct hstate *h = hstate_inode(inode);
4477         struct resv_map *resv_map = inode_resv_map(inode);
4478         long chg = 0;
4479         struct hugepage_subpool *spool = subpool_inode(inode);
4480         long gbl_reserve;
4481
4482         if (resv_map) {
4483                 chg = region_del(resv_map, start, end);
4484                 /*
4485                  * region_del() can fail in the rare case where a region
4486                  * must be split and another region descriptor can not be
4487                  * allocated.  If end == LONG_MAX, it will not fail.
4488                  */
4489                 if (chg < 0)
4490                         return chg;
4491         }
4492
4493         spin_lock(&inode->i_lock);
4494         inode->i_blocks -= (blocks_per_huge_page(h) * freed);
4495         spin_unlock(&inode->i_lock);
4496
4497         /*
4498          * If the subpool has a minimum size, the number of global
4499          * reservations to be released may be adjusted.
4500          */
4501         gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4502         hugetlb_acct_memory(h, -gbl_reserve);
4503
4504         return 0;
4505 }
4506
4507 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4508 static unsigned long page_table_shareable(struct vm_area_struct *svma,
4509                                 struct vm_area_struct *vma,
4510                                 unsigned long addr, pgoff_t idx)
4511 {
4512         unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4513                                 svma->vm_start;
4514         unsigned long sbase = saddr & PUD_MASK;
4515         unsigned long s_end = sbase + PUD_SIZE;
4516
4517         /* Allow segments to share if only one is marked locked */
4518         unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4519         unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
4520
4521         /*
4522          * match the virtual addresses, permission and the alignment of the
4523          * page table page.
4524          */
4525         if (pmd_index(addr) != pmd_index(saddr) ||
4526             vm_flags != svm_flags ||
4527             sbase < svma->vm_start || svma->vm_end < s_end)
4528                 return 0;
4529
4530         return saddr;
4531 }
4532
4533 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
4534 {
4535         unsigned long base = addr & PUD_MASK;
4536         unsigned long end = base + PUD_SIZE;
4537
4538         /*
4539          * check on proper vm_flags and page table alignment
4540          */
4541         if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
4542                 return true;
4543         return false;
4544 }
4545
4546 /*
4547  * Determine if start,end range within vma could be mapped by shared pmd.
4548  * If yes, adjust start and end to cover range associated with possible
4549  * shared pmd mappings.
4550  */
4551 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
4552                                 unsigned long *start, unsigned long *end)
4553 {
4554         unsigned long check_addr = *start;
4555
4556         if (!(vma->vm_flags & VM_MAYSHARE))
4557                 return;
4558
4559         for (check_addr = *start; check_addr < *end; check_addr += PUD_SIZE) {
4560                 unsigned long a_start = check_addr & PUD_MASK;
4561                 unsigned long a_end = a_start + PUD_SIZE;
4562
4563                 /*
4564                  * If sharing is possible, adjust start/end if necessary.
4565                  */
4566                 if (range_in_vma(vma, a_start, a_end)) {
4567                         if (a_start < *start)
4568                                 *start = a_start;
4569                         if (a_end > *end)
4570                                 *end = a_end;
4571                 }
4572         }
4573 }
4574
4575 /*
4576  * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4577  * and returns the corresponding pte. While this is not necessary for the
4578  * !shared pmd case because we can allocate the pmd later as well, it makes the
4579  * code much cleaner. pmd allocation is essential for the shared case because
4580  * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4581  * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4582  * bad pmd for sharing.
4583  */
4584 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4585 {
4586         struct vm_area_struct *vma = find_vma(mm, addr);
4587         struct address_space *mapping = vma->vm_file->f_mapping;
4588         pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4589                         vma->vm_pgoff;
4590         struct vm_area_struct *svma;
4591         unsigned long saddr;
4592         pte_t *spte = NULL;
4593         pte_t *pte;
4594         spinlock_t *ptl;
4595
4596         if (!vma_shareable(vma, addr))
4597                 return (pte_t *)pmd_alloc(mm, pud, addr);
4598
4599         i_mmap_lock_write(mapping);
4600         vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4601                 if (svma == vma)
4602                         continue;
4603
4604                 saddr = page_table_shareable(svma, vma, addr, idx);
4605                 if (saddr) {
4606                         spte = huge_pte_offset(svma->vm_mm, saddr,
4607                                                vma_mmu_pagesize(svma));
4608                         if (spte) {
4609                                 get_page(virt_to_page(spte));
4610                                 break;
4611                         }
4612                 }
4613         }
4614
4615         if (!spte)
4616                 goto out;
4617
4618         ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
4619         if (pud_none(*pud)) {
4620                 pud_populate(mm, pud,
4621                                 (pmd_t *)((unsigned long)spte & PAGE_MASK));
4622                 mm_inc_nr_pmds(mm);
4623         } else {
4624                 put_page(virt_to_page(spte));
4625         }
4626         spin_unlock(ptl);
4627 out:
4628         pte = (pte_t *)pmd_alloc(mm, pud, addr);
4629         i_mmap_unlock_write(mapping);
4630         return pte;
4631 }
4632
4633 /*
4634  * unmap huge page backed by shared pte.
4635  *
4636  * Hugetlb pte page is ref counted at the time of mapping.  If pte is shared
4637  * indicated by page_count > 1, unmap is achieved by clearing pud and
4638  * decrementing the ref count. If count == 1, the pte page is not shared.
4639  *
4640  * called with page table lock held.
4641  *
4642  * returns: 1 successfully unmapped a shared pte page
4643  *          0 the underlying pte page is not shared, or it is the last user
4644  */
4645 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4646 {
4647         pgd_t *pgd = pgd_offset(mm, *addr);
4648         p4d_t *p4d = p4d_offset(pgd, *addr);
4649         pud_t *pud = pud_offset(p4d, *addr);
4650
4651         BUG_ON(page_count(virt_to_page(ptep)) == 0);
4652         if (page_count(virt_to_page(ptep)) == 1)
4653                 return 0;
4654
4655         pud_clear(pud);
4656         put_page(virt_to_page(ptep));
4657         mm_dec_nr_pmds(mm);
4658         *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4659         return 1;
4660 }
4661 #define want_pmd_share()        (1)
4662 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4663 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4664 {
4665         return NULL;
4666 }
4667
4668 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4669 {
4670         return 0;
4671 }
4672
4673 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
4674                                 unsigned long *start, unsigned long *end)
4675 {
4676 }
4677 #define want_pmd_share()        (0)
4678 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4679
4680 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4681 pte_t *huge_pte_alloc(struct mm_struct *mm,
4682                         unsigned long addr, unsigned long sz)
4683 {
4684         pgd_t *pgd;
4685         p4d_t *p4d;
4686         pud_t *pud;
4687         pte_t *pte = NULL;
4688
4689         pgd = pgd_offset(mm, addr);
4690         p4d = p4d_alloc(mm, pgd, addr);
4691         if (!p4d)
4692                 return NULL;
4693         pud = pud_alloc(mm, p4d, addr);
4694         if (pud) {
4695                 if (sz == PUD_SIZE) {
4696                         pte = (pte_t *)pud;
4697                 } else {
4698                         BUG_ON(sz != PMD_SIZE);
4699                         if (want_pmd_share() && pud_none(*pud))
4700                                 pte = huge_pmd_share(mm, addr, pud);
4701                         else
4702                                 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4703                 }
4704         }
4705         BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
4706
4707         return pte;
4708 }
4709
4710 /*
4711  * huge_pte_offset() - Walk the page table to resolve the hugepage
4712  * entry at address @addr
4713  *
4714  * Return: Pointer to page table or swap entry (PUD or PMD) for
4715  * address @addr, or NULL if a p*d_none() entry is encountered and the
4716  * size @sz doesn't match the hugepage size at this level of the page
4717  * table.
4718  */
4719 pte_t *huge_pte_offset(struct mm_struct *mm,
4720                        unsigned long addr, unsigned long sz)
4721 {
4722         pgd_t *pgd;
4723         p4d_t *p4d;
4724         pud_t *pud;
4725         pmd_t *pmd;
4726
4727         pgd = pgd_offset(mm, addr);
4728         if (!pgd_present(*pgd))
4729                 return NULL;
4730         p4d = p4d_offset(pgd, addr);
4731         if (!p4d_present(*p4d))
4732                 return NULL;
4733
4734         pud = pud_offset(p4d, addr);
4735         if (sz != PUD_SIZE && pud_none(*pud))
4736                 return NULL;
4737         /* hugepage or swap? */
4738         if (pud_huge(*pud) || !pud_present(*pud))
4739                 return (pte_t *)pud;
4740
4741         pmd = pmd_offset(pud, addr);
4742         if (sz != PMD_SIZE && pmd_none(*pmd))
4743                 return NULL;
4744         /* hugepage or swap? */
4745         if (pmd_huge(*pmd) || !pmd_present(*pmd))
4746                 return (pte_t *)pmd;
4747
4748         return NULL;
4749 }
4750
4751 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4752
4753 /*
4754  * These functions are overwritable if your architecture needs its own
4755  * behavior.
4756  */
4757 struct page * __weak
4758 follow_huge_addr(struct mm_struct *mm, unsigned long address,
4759                               int write)
4760 {
4761         return ERR_PTR(-EINVAL);
4762 }
4763
4764 struct page * __weak
4765 follow_huge_pd(struct vm_area_struct *vma,
4766                unsigned long address, hugepd_t hpd, int flags, int pdshift)
4767 {
4768         WARN(1, "hugepd follow called with no support for hugepage directory format\n");
4769         return NULL;
4770 }
4771
4772 struct page * __weak
4773 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
4774                 pmd_t *pmd, int flags)
4775 {
4776         struct page *page = NULL;
4777         spinlock_t *ptl;
4778         pte_t pte;
4779 retry:
4780         ptl = pmd_lockptr(mm, pmd);
4781         spin_lock(ptl);
4782         /*
4783          * make sure that the address range covered by this pmd is not
4784          * unmapped from other threads.
4785          */
4786         if (!pmd_huge(*pmd))
4787                 goto out;
4788         pte = huge_ptep_get((pte_t *)pmd);
4789         if (pte_present(pte)) {
4790                 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
4791                 if (flags & FOLL_GET)
4792                         get_page(page);
4793         } else {
4794                 if (is_hugetlb_entry_migration(pte)) {
4795                         spin_unlock(ptl);
4796                         __migration_entry_wait(mm, (pte_t *)pmd, ptl);
4797                         goto retry;
4798                 }
4799                 /*
4800                  * hwpoisoned entry is treated as no_page_table in
4801                  * follow_page_mask().
4802                  */
4803         }
4804 out:
4805         spin_unlock(ptl);
4806         return page;
4807 }
4808
4809 struct page * __weak
4810 follow_huge_pud(struct mm_struct *mm, unsigned long address,
4811                 pud_t *pud, int flags)
4812 {
4813         if (flags & FOLL_GET)
4814                 return NULL;
4815
4816         return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
4817 }
4818
4819 struct page * __weak
4820 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
4821 {
4822         if (flags & FOLL_GET)
4823                 return NULL;
4824
4825         return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
4826 }
4827
4828 bool isolate_huge_page(struct page *page, struct list_head *list)
4829 {
4830         bool ret = true;
4831
4832         VM_BUG_ON_PAGE(!PageHead(page), page);
4833         spin_lock(&hugetlb_lock);
4834         if (!page_huge_active(page) || !get_page_unless_zero(page)) {
4835                 ret = false;
4836                 goto unlock;
4837         }
4838         clear_page_huge_active(page);
4839         list_move_tail(&page->lru, list);
4840 unlock:
4841         spin_unlock(&hugetlb_lock);
4842         return ret;
4843 }
4844
4845 void putback_active_hugepage(struct page *page)
4846 {
4847         VM_BUG_ON_PAGE(!PageHead(page), page);
4848         spin_lock(&hugetlb_lock);
4849         set_page_huge_active(page);
4850         list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
4851         spin_unlock(&hugetlb_lock);
4852         put_page(page);
4853 }