51d31352a5bfd4f23d7ed3cfdbf8933956b286ff
[muen/linux.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/sched/signal.h>
22 #include <linux/rmap.h>
23 #include <linux/swap.h>
24 #include <linux/swapops.h>
25 #include <linux/page-isolation.h>
26 #include <linux/jhash.h>
27
28 #include <asm/page.h>
29 #include <asm/pgtable.h>
30 #include <asm/tlb.h>
31
32 #include <linux/io.h>
33 #include <linux/hugetlb.h>
34 #include <linux/hugetlb_cgroup.h>
35 #include <linux/node.h>
36 #include <linux/userfaultfd_k.h>
37 #include "internal.h"
38
39 int hugepages_treat_as_movable;
40
41 int hugetlb_max_hstate __read_mostly;
42 unsigned int default_hstate_idx;
43 struct hstate hstates[HUGE_MAX_HSTATE];
44 /*
45  * Minimum page order among possible hugepage sizes, set to a proper value
46  * at boot time.
47  */
48 static unsigned int minimum_order __read_mostly = UINT_MAX;
49
50 __initdata LIST_HEAD(huge_boot_pages);
51
52 /* for command line parsing */
53 static struct hstate * __initdata parsed_hstate;
54 static unsigned long __initdata default_hstate_max_huge_pages;
55 static unsigned long __initdata default_hstate_size;
56 static bool __initdata parsed_valid_hugepagesz = true;
57
58 /*
59  * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
60  * free_huge_pages, and surplus_huge_pages.
61  */
62 DEFINE_SPINLOCK(hugetlb_lock);
63
64 /*
65  * Serializes faults on the same logical page.  This is used to
66  * prevent spurious OOMs when the hugepage pool is fully utilized.
67  */
68 static int num_fault_mutexes;
69 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
70
71 /* Forward declaration */
72 static int hugetlb_acct_memory(struct hstate *h, long delta);
73
74 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
75 {
76         bool free = (spool->count == 0) && (spool->used_hpages == 0);
77
78         spin_unlock(&spool->lock);
79
80         /* If no pages are used, and no other handles to the subpool
81          * remain, give up any reservations mased on minimum size and
82          * free the subpool */
83         if (free) {
84                 if (spool->min_hpages != -1)
85                         hugetlb_acct_memory(spool->hstate,
86                                                 -spool->min_hpages);
87                 kfree(spool);
88         }
89 }
90
91 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
92                                                 long min_hpages)
93 {
94         struct hugepage_subpool *spool;
95
96         spool = kzalloc(sizeof(*spool), GFP_KERNEL);
97         if (!spool)
98                 return NULL;
99
100         spin_lock_init(&spool->lock);
101         spool->count = 1;
102         spool->max_hpages = max_hpages;
103         spool->hstate = h;
104         spool->min_hpages = min_hpages;
105
106         if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
107                 kfree(spool);
108                 return NULL;
109         }
110         spool->rsv_hpages = min_hpages;
111
112         return spool;
113 }
114
115 void hugepage_put_subpool(struct hugepage_subpool *spool)
116 {
117         spin_lock(&spool->lock);
118         BUG_ON(!spool->count);
119         spool->count--;
120         unlock_or_release_subpool(spool);
121 }
122
123 /*
124  * Subpool accounting for allocating and reserving pages.
125  * Return -ENOMEM if there are not enough resources to satisfy the
126  * the request.  Otherwise, return the number of pages by which the
127  * global pools must be adjusted (upward).  The returned value may
128  * only be different than the passed value (delta) in the case where
129  * a subpool minimum size must be manitained.
130  */
131 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
132                                       long delta)
133 {
134         long ret = delta;
135
136         if (!spool)
137                 return ret;
138
139         spin_lock(&spool->lock);
140
141         if (spool->max_hpages != -1) {          /* maximum size accounting */
142                 if ((spool->used_hpages + delta) <= spool->max_hpages)
143                         spool->used_hpages += delta;
144                 else {
145                         ret = -ENOMEM;
146                         goto unlock_ret;
147                 }
148         }
149
150         /* minimum size accounting */
151         if (spool->min_hpages != -1 && spool->rsv_hpages) {
152                 if (delta > spool->rsv_hpages) {
153                         /*
154                          * Asking for more reserves than those already taken on
155                          * behalf of subpool.  Return difference.
156                          */
157                         ret = delta - spool->rsv_hpages;
158                         spool->rsv_hpages = 0;
159                 } else {
160                         ret = 0;        /* reserves already accounted for */
161                         spool->rsv_hpages -= delta;
162                 }
163         }
164
165 unlock_ret:
166         spin_unlock(&spool->lock);
167         return ret;
168 }
169
170 /*
171  * Subpool accounting for freeing and unreserving pages.
172  * Return the number of global page reservations that must be dropped.
173  * The return value may only be different than the passed value (delta)
174  * in the case where a subpool minimum size must be maintained.
175  */
176 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
177                                        long delta)
178 {
179         long ret = delta;
180
181         if (!spool)
182                 return delta;
183
184         spin_lock(&spool->lock);
185
186         if (spool->max_hpages != -1)            /* maximum size accounting */
187                 spool->used_hpages -= delta;
188
189          /* minimum size accounting */
190         if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
191                 if (spool->rsv_hpages + delta <= spool->min_hpages)
192                         ret = 0;
193                 else
194                         ret = spool->rsv_hpages + delta - spool->min_hpages;
195
196                 spool->rsv_hpages += delta;
197                 if (spool->rsv_hpages > spool->min_hpages)
198                         spool->rsv_hpages = spool->min_hpages;
199         }
200
201         /*
202          * If hugetlbfs_put_super couldn't free spool due to an outstanding
203          * quota reference, free it now.
204          */
205         unlock_or_release_subpool(spool);
206
207         return ret;
208 }
209
210 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
211 {
212         return HUGETLBFS_SB(inode->i_sb)->spool;
213 }
214
215 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
216 {
217         return subpool_inode(file_inode(vma->vm_file));
218 }
219
220 /*
221  * Region tracking -- allows tracking of reservations and instantiated pages
222  *                    across the pages in a mapping.
223  *
224  * The region data structures are embedded into a resv_map and protected
225  * by a resv_map's lock.  The set of regions within the resv_map represent
226  * reservations for huge pages, or huge pages that have already been
227  * instantiated within the map.  The from and to elements are huge page
228  * indicies into the associated mapping.  from indicates the starting index
229  * of the region.  to represents the first index past the end of  the region.
230  *
231  * For example, a file region structure with from == 0 and to == 4 represents
232  * four huge pages in a mapping.  It is important to note that the to element
233  * represents the first element past the end of the region. This is used in
234  * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
235  *
236  * Interval notation of the form [from, to) will be used to indicate that
237  * the endpoint from is inclusive and to is exclusive.
238  */
239 struct file_region {
240         struct list_head link;
241         long from;
242         long to;
243 };
244
245 /*
246  * Add the huge page range represented by [f, t) to the reserve
247  * map.  In the normal case, existing regions will be expanded
248  * to accommodate the specified range.  Sufficient regions should
249  * exist for expansion due to the previous call to region_chg
250  * with the same range.  However, it is possible that region_del
251  * could have been called after region_chg and modifed the map
252  * in such a way that no region exists to be expanded.  In this
253  * case, pull a region descriptor from the cache associated with
254  * the map and use that for the new range.
255  *
256  * Return the number of new huge pages added to the map.  This
257  * number is greater than or equal to zero.
258  */
259 static long region_add(struct resv_map *resv, long f, long t)
260 {
261         struct list_head *head = &resv->regions;
262         struct file_region *rg, *nrg, *trg;
263         long add = 0;
264
265         spin_lock(&resv->lock);
266         /* Locate the region we are either in or before. */
267         list_for_each_entry(rg, head, link)
268                 if (f <= rg->to)
269                         break;
270
271         /*
272          * If no region exists which can be expanded to include the
273          * specified range, the list must have been modified by an
274          * interleving call to region_del().  Pull a region descriptor
275          * from the cache and use it for this range.
276          */
277         if (&rg->link == head || t < rg->from) {
278                 VM_BUG_ON(resv->region_cache_count <= 0);
279
280                 resv->region_cache_count--;
281                 nrg = list_first_entry(&resv->region_cache, struct file_region,
282                                         link);
283                 list_del(&nrg->link);
284
285                 nrg->from = f;
286                 nrg->to = t;
287                 list_add(&nrg->link, rg->link.prev);
288
289                 add += t - f;
290                 goto out_locked;
291         }
292
293         /* Round our left edge to the current segment if it encloses us. */
294         if (f > rg->from)
295                 f = rg->from;
296
297         /* Check for and consume any regions we now overlap with. */
298         nrg = rg;
299         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
300                 if (&rg->link == head)
301                         break;
302                 if (rg->from > t)
303                         break;
304
305                 /* If this area reaches higher then extend our area to
306                  * include it completely.  If this is not the first area
307                  * which we intend to reuse, free it. */
308                 if (rg->to > t)
309                         t = rg->to;
310                 if (rg != nrg) {
311                         /* Decrement return value by the deleted range.
312                          * Another range will span this area so that by
313                          * end of routine add will be >= zero
314                          */
315                         add -= (rg->to - rg->from);
316                         list_del(&rg->link);
317                         kfree(rg);
318                 }
319         }
320
321         add += (nrg->from - f);         /* Added to beginning of region */
322         nrg->from = f;
323         add += t - nrg->to;             /* Added to end of region */
324         nrg->to = t;
325
326 out_locked:
327         resv->adds_in_progress--;
328         spin_unlock(&resv->lock);
329         VM_BUG_ON(add < 0);
330         return add;
331 }
332
333 /*
334  * Examine the existing reserve map and determine how many
335  * huge pages in the specified range [f, t) are NOT currently
336  * represented.  This routine is called before a subsequent
337  * call to region_add that will actually modify the reserve
338  * map to add the specified range [f, t).  region_chg does
339  * not change the number of huge pages represented by the
340  * map.  However, if the existing regions in the map can not
341  * be expanded to represent the new range, a new file_region
342  * structure is added to the map as a placeholder.  This is
343  * so that the subsequent region_add call will have all the
344  * regions it needs and will not fail.
345  *
346  * Upon entry, region_chg will also examine the cache of region descriptors
347  * associated with the map.  If there are not enough descriptors cached, one
348  * will be allocated for the in progress add operation.
349  *
350  * Returns the number of huge pages that need to be added to the existing
351  * reservation map for the range [f, t).  This number is greater or equal to
352  * zero.  -ENOMEM is returned if a new file_region structure or cache entry
353  * is needed and can not be allocated.
354  */
355 static long region_chg(struct resv_map *resv, long f, long t)
356 {
357         struct list_head *head = &resv->regions;
358         struct file_region *rg, *nrg = NULL;
359         long chg = 0;
360
361 retry:
362         spin_lock(&resv->lock);
363 retry_locked:
364         resv->adds_in_progress++;
365
366         /*
367          * Check for sufficient descriptors in the cache to accommodate
368          * the number of in progress add operations.
369          */
370         if (resv->adds_in_progress > resv->region_cache_count) {
371                 struct file_region *trg;
372
373                 VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1);
374                 /* Must drop lock to allocate a new descriptor. */
375                 resv->adds_in_progress--;
376                 spin_unlock(&resv->lock);
377
378                 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
379                 if (!trg) {
380                         kfree(nrg);
381                         return -ENOMEM;
382                 }
383
384                 spin_lock(&resv->lock);
385                 list_add(&trg->link, &resv->region_cache);
386                 resv->region_cache_count++;
387                 goto retry_locked;
388         }
389
390         /* Locate the region we are before or in. */
391         list_for_each_entry(rg, head, link)
392                 if (f <= rg->to)
393                         break;
394
395         /* If we are below the current region then a new region is required.
396          * Subtle, allocate a new region at the position but make it zero
397          * size such that we can guarantee to record the reservation. */
398         if (&rg->link == head || t < rg->from) {
399                 if (!nrg) {
400                         resv->adds_in_progress--;
401                         spin_unlock(&resv->lock);
402                         nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
403                         if (!nrg)
404                                 return -ENOMEM;
405
406                         nrg->from = f;
407                         nrg->to   = f;
408                         INIT_LIST_HEAD(&nrg->link);
409                         goto retry;
410                 }
411
412                 list_add(&nrg->link, rg->link.prev);
413                 chg = t - f;
414                 goto out_nrg;
415         }
416
417         /* Round our left edge to the current segment if it encloses us. */
418         if (f > rg->from)
419                 f = rg->from;
420         chg = t - f;
421
422         /* Check for and consume any regions we now overlap with. */
423         list_for_each_entry(rg, rg->link.prev, link) {
424                 if (&rg->link == head)
425                         break;
426                 if (rg->from > t)
427                         goto out;
428
429                 /* We overlap with this area, if it extends further than
430                  * us then we must extend ourselves.  Account for its
431                  * existing reservation. */
432                 if (rg->to > t) {
433                         chg += rg->to - t;
434                         t = rg->to;
435                 }
436                 chg -= rg->to - rg->from;
437         }
438
439 out:
440         spin_unlock(&resv->lock);
441         /*  We already know we raced and no longer need the new region */
442         kfree(nrg);
443         return chg;
444 out_nrg:
445         spin_unlock(&resv->lock);
446         return chg;
447 }
448
449 /*
450  * Abort the in progress add operation.  The adds_in_progress field
451  * of the resv_map keeps track of the operations in progress between
452  * calls to region_chg and region_add.  Operations are sometimes
453  * aborted after the call to region_chg.  In such cases, region_abort
454  * is called to decrement the adds_in_progress counter.
455  *
456  * NOTE: The range arguments [f, t) are not needed or used in this
457  * routine.  They are kept to make reading the calling code easier as
458  * arguments will match the associated region_chg call.
459  */
460 static void region_abort(struct resv_map *resv, long f, long t)
461 {
462         spin_lock(&resv->lock);
463         VM_BUG_ON(!resv->region_cache_count);
464         resv->adds_in_progress--;
465         spin_unlock(&resv->lock);
466 }
467
468 /*
469  * Delete the specified range [f, t) from the reserve map.  If the
470  * t parameter is LONG_MAX, this indicates that ALL regions after f
471  * should be deleted.  Locate the regions which intersect [f, t)
472  * and either trim, delete or split the existing regions.
473  *
474  * Returns the number of huge pages deleted from the reserve map.
475  * In the normal case, the return value is zero or more.  In the
476  * case where a region must be split, a new region descriptor must
477  * be allocated.  If the allocation fails, -ENOMEM will be returned.
478  * NOTE: If the parameter t == LONG_MAX, then we will never split
479  * a region and possibly return -ENOMEM.  Callers specifying
480  * t == LONG_MAX do not need to check for -ENOMEM error.
481  */
482 static long region_del(struct resv_map *resv, long f, long t)
483 {
484         struct list_head *head = &resv->regions;
485         struct file_region *rg, *trg;
486         struct file_region *nrg = NULL;
487         long del = 0;
488
489 retry:
490         spin_lock(&resv->lock);
491         list_for_each_entry_safe(rg, trg, head, link) {
492                 /*
493                  * Skip regions before the range to be deleted.  file_region
494                  * ranges are normally of the form [from, to).  However, there
495                  * may be a "placeholder" entry in the map which is of the form
496                  * (from, to) with from == to.  Check for placeholder entries
497                  * at the beginning of the range to be deleted.
498                  */
499                 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
500                         continue;
501
502                 if (rg->from >= t)
503                         break;
504
505                 if (f > rg->from && t < rg->to) { /* Must split region */
506                         /*
507                          * Check for an entry in the cache before dropping
508                          * lock and attempting allocation.
509                          */
510                         if (!nrg &&
511                             resv->region_cache_count > resv->adds_in_progress) {
512                                 nrg = list_first_entry(&resv->region_cache,
513                                                         struct file_region,
514                                                         link);
515                                 list_del(&nrg->link);
516                                 resv->region_cache_count--;
517                         }
518
519                         if (!nrg) {
520                                 spin_unlock(&resv->lock);
521                                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
522                                 if (!nrg)
523                                         return -ENOMEM;
524                                 goto retry;
525                         }
526
527                         del += t - f;
528
529                         /* New entry for end of split region */
530                         nrg->from = t;
531                         nrg->to = rg->to;
532                         INIT_LIST_HEAD(&nrg->link);
533
534                         /* Original entry is trimmed */
535                         rg->to = f;
536
537                         list_add(&nrg->link, &rg->link);
538                         nrg = NULL;
539                         break;
540                 }
541
542                 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
543                         del += rg->to - rg->from;
544                         list_del(&rg->link);
545                         kfree(rg);
546                         continue;
547                 }
548
549                 if (f <= rg->from) {    /* Trim beginning of region */
550                         del += t - rg->from;
551                         rg->from = t;
552                 } else {                /* Trim end of region */
553                         del += rg->to - f;
554                         rg->to = f;
555                 }
556         }
557
558         spin_unlock(&resv->lock);
559         kfree(nrg);
560         return del;
561 }
562
563 /*
564  * A rare out of memory error was encountered which prevented removal of
565  * the reserve map region for a page.  The huge page itself was free'ed
566  * and removed from the page cache.  This routine will adjust the subpool
567  * usage count, and the global reserve count if needed.  By incrementing
568  * these counts, the reserve map entry which could not be deleted will
569  * appear as a "reserved" entry instead of simply dangling with incorrect
570  * counts.
571  */
572 void hugetlb_fix_reserve_counts(struct inode *inode)
573 {
574         struct hugepage_subpool *spool = subpool_inode(inode);
575         long rsv_adjust;
576
577         rsv_adjust = hugepage_subpool_get_pages(spool, 1);
578         if (rsv_adjust) {
579                 struct hstate *h = hstate_inode(inode);
580
581                 hugetlb_acct_memory(h, 1);
582         }
583 }
584
585 /*
586  * Count and return the number of huge pages in the reserve map
587  * that intersect with the range [f, t).
588  */
589 static long region_count(struct resv_map *resv, long f, long t)
590 {
591         struct list_head *head = &resv->regions;
592         struct file_region *rg;
593         long chg = 0;
594
595         spin_lock(&resv->lock);
596         /* Locate each segment we overlap with, and count that overlap. */
597         list_for_each_entry(rg, head, link) {
598                 long seg_from;
599                 long seg_to;
600
601                 if (rg->to <= f)
602                         continue;
603                 if (rg->from >= t)
604                         break;
605
606                 seg_from = max(rg->from, f);
607                 seg_to = min(rg->to, t);
608
609                 chg += seg_to - seg_from;
610         }
611         spin_unlock(&resv->lock);
612
613         return chg;
614 }
615
616 /*
617  * Convert the address within this vma to the page offset within
618  * the mapping, in pagecache page units; huge pages here.
619  */
620 static pgoff_t vma_hugecache_offset(struct hstate *h,
621                         struct vm_area_struct *vma, unsigned long address)
622 {
623         return ((address - vma->vm_start) >> huge_page_shift(h)) +
624                         (vma->vm_pgoff >> huge_page_order(h));
625 }
626
627 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
628                                      unsigned long address)
629 {
630         return vma_hugecache_offset(hstate_vma(vma), vma, address);
631 }
632 EXPORT_SYMBOL_GPL(linear_hugepage_index);
633
634 /*
635  * Return the size of the pages allocated when backing a VMA. In the majority
636  * cases this will be same size as used by the page table entries.
637  */
638 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
639 {
640         struct hstate *hstate;
641
642         if (!is_vm_hugetlb_page(vma))
643                 return PAGE_SIZE;
644
645         hstate = hstate_vma(vma);
646
647         return 1UL << huge_page_shift(hstate);
648 }
649 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
650
651 /*
652  * Return the page size being used by the MMU to back a VMA. In the majority
653  * of cases, the page size used by the kernel matches the MMU size. On
654  * architectures where it differs, an architecture-specific version of this
655  * function is required.
656  */
657 #ifndef vma_mmu_pagesize
658 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
659 {
660         return vma_kernel_pagesize(vma);
661 }
662 #endif
663
664 /*
665  * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
666  * bits of the reservation map pointer, which are always clear due to
667  * alignment.
668  */
669 #define HPAGE_RESV_OWNER    (1UL << 0)
670 #define HPAGE_RESV_UNMAPPED (1UL << 1)
671 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
672
673 /*
674  * These helpers are used to track how many pages are reserved for
675  * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
676  * is guaranteed to have their future faults succeed.
677  *
678  * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
679  * the reserve counters are updated with the hugetlb_lock held. It is safe
680  * to reset the VMA at fork() time as it is not in use yet and there is no
681  * chance of the global counters getting corrupted as a result of the values.
682  *
683  * The private mapping reservation is represented in a subtly different
684  * manner to a shared mapping.  A shared mapping has a region map associated
685  * with the underlying file, this region map represents the backing file
686  * pages which have ever had a reservation assigned which this persists even
687  * after the page is instantiated.  A private mapping has a region map
688  * associated with the original mmap which is attached to all VMAs which
689  * reference it, this region map represents those offsets which have consumed
690  * reservation ie. where pages have been instantiated.
691  */
692 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
693 {
694         return (unsigned long)vma->vm_private_data;
695 }
696
697 static void set_vma_private_data(struct vm_area_struct *vma,
698                                                         unsigned long value)
699 {
700         vma->vm_private_data = (void *)value;
701 }
702
703 struct resv_map *resv_map_alloc(void)
704 {
705         struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
706         struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
707
708         if (!resv_map || !rg) {
709                 kfree(resv_map);
710                 kfree(rg);
711                 return NULL;
712         }
713
714         kref_init(&resv_map->refs);
715         spin_lock_init(&resv_map->lock);
716         INIT_LIST_HEAD(&resv_map->regions);
717
718         resv_map->adds_in_progress = 0;
719
720         INIT_LIST_HEAD(&resv_map->region_cache);
721         list_add(&rg->link, &resv_map->region_cache);
722         resv_map->region_cache_count = 1;
723
724         return resv_map;
725 }
726
727 void resv_map_release(struct kref *ref)
728 {
729         struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
730         struct list_head *head = &resv_map->region_cache;
731         struct file_region *rg, *trg;
732
733         /* Clear out any active regions before we release the map. */
734         region_del(resv_map, 0, LONG_MAX);
735
736         /* ... and any entries left in the cache */
737         list_for_each_entry_safe(rg, trg, head, link) {
738                 list_del(&rg->link);
739                 kfree(rg);
740         }
741
742         VM_BUG_ON(resv_map->adds_in_progress);
743
744         kfree(resv_map);
745 }
746
747 static inline struct resv_map *inode_resv_map(struct inode *inode)
748 {
749         return inode->i_mapping->private_data;
750 }
751
752 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
753 {
754         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
755         if (vma->vm_flags & VM_MAYSHARE) {
756                 struct address_space *mapping = vma->vm_file->f_mapping;
757                 struct inode *inode = mapping->host;
758
759                 return inode_resv_map(inode);
760
761         } else {
762                 return (struct resv_map *)(get_vma_private_data(vma) &
763                                                         ~HPAGE_RESV_MASK);
764         }
765 }
766
767 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
768 {
769         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
770         VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
771
772         set_vma_private_data(vma, (get_vma_private_data(vma) &
773                                 HPAGE_RESV_MASK) | (unsigned long)map);
774 }
775
776 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
777 {
778         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
779         VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
780
781         set_vma_private_data(vma, get_vma_private_data(vma) | flags);
782 }
783
784 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
785 {
786         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
787
788         return (get_vma_private_data(vma) & flag) != 0;
789 }
790
791 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
792 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
793 {
794         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
795         if (!(vma->vm_flags & VM_MAYSHARE))
796                 vma->vm_private_data = (void *)0;
797 }
798
799 /* Returns true if the VMA has associated reserve pages */
800 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
801 {
802         if (vma->vm_flags & VM_NORESERVE) {
803                 /*
804                  * This address is already reserved by other process(chg == 0),
805                  * so, we should decrement reserved count. Without decrementing,
806                  * reserve count remains after releasing inode, because this
807                  * allocated page will go into page cache and is regarded as
808                  * coming from reserved pool in releasing step.  Currently, we
809                  * don't have any other solution to deal with this situation
810                  * properly, so add work-around here.
811                  */
812                 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
813                         return true;
814                 else
815                         return false;
816         }
817
818         /* Shared mappings always use reserves */
819         if (vma->vm_flags & VM_MAYSHARE) {
820                 /*
821                  * We know VM_NORESERVE is not set.  Therefore, there SHOULD
822                  * be a region map for all pages.  The only situation where
823                  * there is no region map is if a hole was punched via
824                  * fallocate.  In this case, there really are no reverves to
825                  * use.  This situation is indicated if chg != 0.
826                  */
827                 if (chg)
828                         return false;
829                 else
830                         return true;
831         }
832
833         /*
834          * Only the process that called mmap() has reserves for
835          * private mappings.
836          */
837         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
838                 /*
839                  * Like the shared case above, a hole punch or truncate
840                  * could have been performed on the private mapping.
841                  * Examine the value of chg to determine if reserves
842                  * actually exist or were previously consumed.
843                  * Very Subtle - The value of chg comes from a previous
844                  * call to vma_needs_reserves().  The reserve map for
845                  * private mappings has different (opposite) semantics
846                  * than that of shared mappings.  vma_needs_reserves()
847                  * has already taken this difference in semantics into
848                  * account.  Therefore, the meaning of chg is the same
849                  * as in the shared case above.  Code could easily be
850                  * combined, but keeping it separate draws attention to
851                  * subtle differences.
852                  */
853                 if (chg)
854                         return false;
855                 else
856                         return true;
857         }
858
859         return false;
860 }
861
862 static void enqueue_huge_page(struct hstate *h, struct page *page)
863 {
864         int nid = page_to_nid(page);
865         list_move(&page->lru, &h->hugepage_freelists[nid]);
866         h->free_huge_pages++;
867         h->free_huge_pages_node[nid]++;
868 }
869
870 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
871 {
872         struct page *page;
873
874         list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
875                 if (!is_migrate_isolate_page(page))
876                         break;
877         /*
878          * if 'non-isolated free hugepage' not found on the list,
879          * the allocation fails.
880          */
881         if (&h->hugepage_freelists[nid] == &page->lru)
882                 return NULL;
883         list_move(&page->lru, &h->hugepage_activelist);
884         set_page_refcounted(page);
885         h->free_huge_pages--;
886         h->free_huge_pages_node[nid]--;
887         return page;
888 }
889
890 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
891 {
892         struct page *page;
893         int node;
894
895         if (nid != NUMA_NO_NODE)
896                 return dequeue_huge_page_node_exact(h, nid);
897
898         for_each_online_node(node) {
899                 page = dequeue_huge_page_node_exact(h, node);
900                 if (page)
901                         return page;
902         }
903         return NULL;
904 }
905
906 /* Movability of hugepages depends on migration support. */
907 static inline gfp_t htlb_alloc_mask(struct hstate *h)
908 {
909         if (hugepages_treat_as_movable || hugepage_migration_supported(h))
910                 return GFP_HIGHUSER_MOVABLE;
911         else
912                 return GFP_HIGHUSER;
913 }
914
915 static struct page *dequeue_huge_page_vma(struct hstate *h,
916                                 struct vm_area_struct *vma,
917                                 unsigned long address, int avoid_reserve,
918                                 long chg)
919 {
920         struct page *page = NULL;
921         struct mempolicy *mpol;
922         nodemask_t *nodemask;
923         struct zonelist *zonelist;
924         struct zone *zone;
925         struct zoneref *z;
926         unsigned int cpuset_mems_cookie;
927
928         /*
929          * A child process with MAP_PRIVATE mappings created by their parent
930          * have no page reserves. This check ensures that reservations are
931          * not "stolen". The child may still get SIGKILLed
932          */
933         if (!vma_has_reserves(vma, chg) &&
934                         h->free_huge_pages - h->resv_huge_pages == 0)
935                 goto err;
936
937         /* If reserves cannot be used, ensure enough pages are in the pool */
938         if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
939                 goto err;
940
941 retry_cpuset:
942         cpuset_mems_cookie = read_mems_allowed_begin();
943         zonelist = huge_zonelist(vma, address,
944                                         htlb_alloc_mask(h), &mpol, &nodemask);
945
946         for_each_zone_zonelist_nodemask(zone, z, zonelist,
947                                                 MAX_NR_ZONES - 1, nodemask) {
948                 if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) {
949                         page = dequeue_huge_page_node(h, zone_to_nid(zone));
950                         if (page) {
951                                 if (avoid_reserve)
952                                         break;
953                                 if (!vma_has_reserves(vma, chg))
954                                         break;
955
956                                 SetPagePrivate(page);
957                                 h->resv_huge_pages--;
958                                 break;
959                         }
960                 }
961         }
962
963         mpol_cond_put(mpol);
964         if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
965                 goto retry_cpuset;
966         return page;
967
968 err:
969         return NULL;
970 }
971
972 /*
973  * common helper functions for hstate_next_node_to_{alloc|free}.
974  * We may have allocated or freed a huge page based on a different
975  * nodes_allowed previously, so h->next_node_to_{alloc|free} might
976  * be outside of *nodes_allowed.  Ensure that we use an allowed
977  * node for alloc or free.
978  */
979 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
980 {
981         nid = next_node_in(nid, *nodes_allowed);
982         VM_BUG_ON(nid >= MAX_NUMNODES);
983
984         return nid;
985 }
986
987 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
988 {
989         if (!node_isset(nid, *nodes_allowed))
990                 nid = next_node_allowed(nid, nodes_allowed);
991         return nid;
992 }
993
994 /*
995  * returns the previously saved node ["this node"] from which to
996  * allocate a persistent huge page for the pool and advance the
997  * next node from which to allocate, handling wrap at end of node
998  * mask.
999  */
1000 static int hstate_next_node_to_alloc(struct hstate *h,
1001                                         nodemask_t *nodes_allowed)
1002 {
1003         int nid;
1004
1005         VM_BUG_ON(!nodes_allowed);
1006
1007         nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1008         h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1009
1010         return nid;
1011 }
1012
1013 /*
1014  * helper for free_pool_huge_page() - return the previously saved
1015  * node ["this node"] from which to free a huge page.  Advance the
1016  * next node id whether or not we find a free huge page to free so
1017  * that the next attempt to free addresses the next node.
1018  */
1019 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1020 {
1021         int nid;
1022
1023         VM_BUG_ON(!nodes_allowed);
1024
1025         nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1026         h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1027
1028         return nid;
1029 }
1030
1031 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask)           \
1032         for (nr_nodes = nodes_weight(*mask);                            \
1033                 nr_nodes > 0 &&                                         \
1034                 ((node = hstate_next_node_to_alloc(hs, mask)) || 1);    \
1035                 nr_nodes--)
1036
1037 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask)            \
1038         for (nr_nodes = nodes_weight(*mask);                            \
1039                 nr_nodes > 0 &&                                         \
1040                 ((node = hstate_next_node_to_free(hs, mask)) || 1);     \
1041                 nr_nodes--)
1042
1043 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1044 static void destroy_compound_gigantic_page(struct page *page,
1045                                         unsigned int order)
1046 {
1047         int i;
1048         int nr_pages = 1 << order;
1049         struct page *p = page + 1;
1050
1051         atomic_set(compound_mapcount_ptr(page), 0);
1052         for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1053                 clear_compound_head(p);
1054                 set_page_refcounted(p);
1055         }
1056
1057         set_compound_order(page, 0);
1058         __ClearPageHead(page);
1059 }
1060
1061 static void free_gigantic_page(struct page *page, unsigned int order)
1062 {
1063         free_contig_range(page_to_pfn(page), 1 << order);
1064 }
1065
1066 static int __alloc_gigantic_page(unsigned long start_pfn,
1067                                 unsigned long nr_pages)
1068 {
1069         unsigned long end_pfn = start_pfn + nr_pages;
1070         return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
1071                                   GFP_KERNEL);
1072 }
1073
1074 static bool pfn_range_valid_gigantic(struct zone *z,
1075                         unsigned long start_pfn, unsigned long nr_pages)
1076 {
1077         unsigned long i, end_pfn = start_pfn + nr_pages;
1078         struct page *page;
1079
1080         for (i = start_pfn; i < end_pfn; i++) {
1081                 if (!pfn_valid(i))
1082                         return false;
1083
1084                 page = pfn_to_page(i);
1085
1086                 if (page_zone(page) != z)
1087                         return false;
1088
1089                 if (PageReserved(page))
1090                         return false;
1091
1092                 if (page_count(page) > 0)
1093                         return false;
1094
1095                 if (PageHuge(page))
1096                         return false;
1097         }
1098
1099         return true;
1100 }
1101
1102 static bool zone_spans_last_pfn(const struct zone *zone,
1103                         unsigned long start_pfn, unsigned long nr_pages)
1104 {
1105         unsigned long last_pfn = start_pfn + nr_pages - 1;
1106         return zone_spans_pfn(zone, last_pfn);
1107 }
1108
1109 static struct page *alloc_gigantic_page(int nid, unsigned int order)
1110 {
1111         unsigned long nr_pages = 1 << order;
1112         unsigned long ret, pfn, flags;
1113         struct zone *z;
1114
1115         z = NODE_DATA(nid)->node_zones;
1116         for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) {
1117                 spin_lock_irqsave(&z->lock, flags);
1118
1119                 pfn = ALIGN(z->zone_start_pfn, nr_pages);
1120                 while (zone_spans_last_pfn(z, pfn, nr_pages)) {
1121                         if (pfn_range_valid_gigantic(z, pfn, nr_pages)) {
1122                                 /*
1123                                  * We release the zone lock here because
1124                                  * alloc_contig_range() will also lock the zone
1125                                  * at some point. If there's an allocation
1126                                  * spinning on this lock, it may win the race
1127                                  * and cause alloc_contig_range() to fail...
1128                                  */
1129                                 spin_unlock_irqrestore(&z->lock, flags);
1130                                 ret = __alloc_gigantic_page(pfn, nr_pages);
1131                                 if (!ret)
1132                                         return pfn_to_page(pfn);
1133                                 spin_lock_irqsave(&z->lock, flags);
1134                         }
1135                         pfn += nr_pages;
1136                 }
1137
1138                 spin_unlock_irqrestore(&z->lock, flags);
1139         }
1140
1141         return NULL;
1142 }
1143
1144 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1145 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1146
1147 static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid)
1148 {
1149         struct page *page;
1150
1151         page = alloc_gigantic_page(nid, huge_page_order(h));
1152         if (page) {
1153                 prep_compound_gigantic_page(page, huge_page_order(h));
1154                 prep_new_huge_page(h, page, nid);
1155         }
1156
1157         return page;
1158 }
1159
1160 static int alloc_fresh_gigantic_page(struct hstate *h,
1161                                 nodemask_t *nodes_allowed)
1162 {
1163         struct page *page = NULL;
1164         int nr_nodes, node;
1165
1166         for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1167                 page = alloc_fresh_gigantic_page_node(h, node);
1168                 if (page)
1169                         return 1;
1170         }
1171
1172         return 0;
1173 }
1174
1175 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1176 static inline bool gigantic_page_supported(void) { return false; }
1177 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1178 static inline void destroy_compound_gigantic_page(struct page *page,
1179                                                 unsigned int order) { }
1180 static inline int alloc_fresh_gigantic_page(struct hstate *h,
1181                                         nodemask_t *nodes_allowed) { return 0; }
1182 #endif
1183
1184 static void update_and_free_page(struct hstate *h, struct page *page)
1185 {
1186         int i;
1187
1188         if (hstate_is_gigantic(h) && !gigantic_page_supported())
1189                 return;
1190
1191         h->nr_huge_pages--;
1192         h->nr_huge_pages_node[page_to_nid(page)]--;
1193         for (i = 0; i < pages_per_huge_page(h); i++) {
1194                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1195                                 1 << PG_referenced | 1 << PG_dirty |
1196                                 1 << PG_active | 1 << PG_private |
1197                                 1 << PG_writeback);
1198         }
1199         VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1200         set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1201         set_page_refcounted(page);
1202         if (hstate_is_gigantic(h)) {
1203                 destroy_compound_gigantic_page(page, huge_page_order(h));
1204                 free_gigantic_page(page, huge_page_order(h));
1205         } else {
1206                 __free_pages(page, huge_page_order(h));
1207         }
1208 }
1209
1210 struct hstate *size_to_hstate(unsigned long size)
1211 {
1212         struct hstate *h;
1213
1214         for_each_hstate(h) {
1215                 if (huge_page_size(h) == size)
1216                         return h;
1217         }
1218         return NULL;
1219 }
1220
1221 /*
1222  * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1223  * to hstate->hugepage_activelist.)
1224  *
1225  * This function can be called for tail pages, but never returns true for them.
1226  */
1227 bool page_huge_active(struct page *page)
1228 {
1229         VM_BUG_ON_PAGE(!PageHuge(page), page);
1230         return PageHead(page) && PagePrivate(&page[1]);
1231 }
1232
1233 /* never called for tail page */
1234 static void set_page_huge_active(struct page *page)
1235 {
1236         VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1237         SetPagePrivate(&page[1]);
1238 }
1239
1240 static void clear_page_huge_active(struct page *page)
1241 {
1242         VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1243         ClearPagePrivate(&page[1]);
1244 }
1245
1246 void free_huge_page(struct page *page)
1247 {
1248         /*
1249          * Can't pass hstate in here because it is called from the
1250          * compound page destructor.
1251          */
1252         struct hstate *h = page_hstate(page);
1253         int nid = page_to_nid(page);
1254         struct hugepage_subpool *spool =
1255                 (struct hugepage_subpool *)page_private(page);
1256         bool restore_reserve;
1257
1258         set_page_private(page, 0);
1259         page->mapping = NULL;
1260         VM_BUG_ON_PAGE(page_count(page), page);
1261         VM_BUG_ON_PAGE(page_mapcount(page), page);
1262         restore_reserve = PagePrivate(page);
1263         ClearPagePrivate(page);
1264
1265         /*
1266          * A return code of zero implies that the subpool will be under its
1267          * minimum size if the reservation is not restored after page is free.
1268          * Therefore, force restore_reserve operation.
1269          */
1270         if (hugepage_subpool_put_pages(spool, 1) == 0)
1271                 restore_reserve = true;
1272
1273         spin_lock(&hugetlb_lock);
1274         clear_page_huge_active(page);
1275         hugetlb_cgroup_uncharge_page(hstate_index(h),
1276                                      pages_per_huge_page(h), page);
1277         if (restore_reserve)
1278                 h->resv_huge_pages++;
1279
1280         if (h->surplus_huge_pages_node[nid]) {
1281                 /* remove the page from active list */
1282                 list_del(&page->lru);
1283                 update_and_free_page(h, page);
1284                 h->surplus_huge_pages--;
1285                 h->surplus_huge_pages_node[nid]--;
1286         } else {
1287                 arch_clear_hugepage_flags(page);
1288                 enqueue_huge_page(h, page);
1289         }
1290         spin_unlock(&hugetlb_lock);
1291 }
1292
1293 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1294 {
1295         INIT_LIST_HEAD(&page->lru);
1296         set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1297         spin_lock(&hugetlb_lock);
1298         set_hugetlb_cgroup(page, NULL);
1299         h->nr_huge_pages++;
1300         h->nr_huge_pages_node[nid]++;
1301         spin_unlock(&hugetlb_lock);
1302         put_page(page); /* free it into the hugepage allocator */
1303 }
1304
1305 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1306 {
1307         int i;
1308         int nr_pages = 1 << order;
1309         struct page *p = page + 1;
1310
1311         /* we rely on prep_new_huge_page to set the destructor */
1312         set_compound_order(page, order);
1313         __ClearPageReserved(page);
1314         __SetPageHead(page);
1315         for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1316                 /*
1317                  * For gigantic hugepages allocated through bootmem at
1318                  * boot, it's safer to be consistent with the not-gigantic
1319                  * hugepages and clear the PG_reserved bit from all tail pages
1320                  * too.  Otherwse drivers using get_user_pages() to access tail
1321                  * pages may get the reference counting wrong if they see
1322                  * PG_reserved set on a tail page (despite the head page not
1323                  * having PG_reserved set).  Enforcing this consistency between
1324                  * head and tail pages allows drivers to optimize away a check
1325                  * on the head page when they need know if put_page() is needed
1326                  * after get_user_pages().
1327                  */
1328                 __ClearPageReserved(p);
1329                 set_page_count(p, 0);
1330                 set_compound_head(p, page);
1331         }
1332         atomic_set(compound_mapcount_ptr(page), -1);
1333 }
1334
1335 /*
1336  * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1337  * transparent huge pages.  See the PageTransHuge() documentation for more
1338  * details.
1339  */
1340 int PageHuge(struct page *page)
1341 {
1342         if (!PageCompound(page))
1343                 return 0;
1344
1345         page = compound_head(page);
1346         return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1347 }
1348 EXPORT_SYMBOL_GPL(PageHuge);
1349
1350 /*
1351  * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1352  * normal or transparent huge pages.
1353  */
1354 int PageHeadHuge(struct page *page_head)
1355 {
1356         if (!PageHead(page_head))
1357                 return 0;
1358
1359         return get_compound_page_dtor(page_head) == free_huge_page;
1360 }
1361
1362 pgoff_t __basepage_index(struct page *page)
1363 {
1364         struct page *page_head = compound_head(page);
1365         pgoff_t index = page_index(page_head);
1366         unsigned long compound_idx;
1367
1368         if (!PageHuge(page_head))
1369                 return page_index(page);
1370
1371         if (compound_order(page_head) >= MAX_ORDER)
1372                 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1373         else
1374                 compound_idx = page - page_head;
1375
1376         return (index << compound_order(page_head)) + compound_idx;
1377 }
1378
1379 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
1380 {
1381         struct page *page;
1382
1383         page = __alloc_pages_node(nid,
1384                 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
1385                                                 __GFP_REPEAT|__GFP_NOWARN,
1386                 huge_page_order(h));
1387         if (page) {
1388                 prep_new_huge_page(h, page, nid);
1389         }
1390
1391         return page;
1392 }
1393
1394 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1395 {
1396         struct page *page;
1397         int nr_nodes, node;
1398         int ret = 0;
1399
1400         for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1401                 page = alloc_fresh_huge_page_node(h, node);
1402                 if (page) {
1403                         ret = 1;
1404                         break;
1405                 }
1406         }
1407
1408         if (ret)
1409                 count_vm_event(HTLB_BUDDY_PGALLOC);
1410         else
1411                 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1412
1413         return ret;
1414 }
1415
1416 /*
1417  * Free huge page from pool from next node to free.
1418  * Attempt to keep persistent huge pages more or less
1419  * balanced over allowed nodes.
1420  * Called with hugetlb_lock locked.
1421  */
1422 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1423                                                          bool acct_surplus)
1424 {
1425         int nr_nodes, node;
1426         int ret = 0;
1427
1428         for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1429                 /*
1430                  * If we're returning unused surplus pages, only examine
1431                  * nodes with surplus pages.
1432                  */
1433                 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1434                     !list_empty(&h->hugepage_freelists[node])) {
1435                         struct page *page =
1436                                 list_entry(h->hugepage_freelists[node].next,
1437                                           struct page, lru);
1438                         list_del(&page->lru);
1439                         h->free_huge_pages--;
1440                         h->free_huge_pages_node[node]--;
1441                         if (acct_surplus) {
1442                                 h->surplus_huge_pages--;
1443                                 h->surplus_huge_pages_node[node]--;
1444                         }
1445                         update_and_free_page(h, page);
1446                         ret = 1;
1447                         break;
1448                 }
1449         }
1450
1451         return ret;
1452 }
1453
1454 /*
1455  * Dissolve a given free hugepage into free buddy pages. This function does
1456  * nothing for in-use (including surplus) hugepages. Returns -EBUSY if the
1457  * number of free hugepages would be reduced below the number of reserved
1458  * hugepages.
1459  */
1460 static int dissolve_free_huge_page(struct page *page)
1461 {
1462         int rc = 0;
1463
1464         spin_lock(&hugetlb_lock);
1465         if (PageHuge(page) && !page_count(page)) {
1466                 struct page *head = compound_head(page);
1467                 struct hstate *h = page_hstate(head);
1468                 int nid = page_to_nid(head);
1469                 if (h->free_huge_pages - h->resv_huge_pages == 0) {
1470                         rc = -EBUSY;
1471                         goto out;
1472                 }
1473                 list_del(&head->lru);
1474                 h->free_huge_pages--;
1475                 h->free_huge_pages_node[nid]--;
1476                 h->max_huge_pages--;
1477                 update_and_free_page(h, head);
1478         }
1479 out:
1480         spin_unlock(&hugetlb_lock);
1481         return rc;
1482 }
1483
1484 /*
1485  * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1486  * make specified memory blocks removable from the system.
1487  * Note that this will dissolve a free gigantic hugepage completely, if any
1488  * part of it lies within the given range.
1489  * Also note that if dissolve_free_huge_page() returns with an error, all
1490  * free hugepages that were dissolved before that error are lost.
1491  */
1492 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1493 {
1494         unsigned long pfn;
1495         struct page *page;
1496         int rc = 0;
1497
1498         if (!hugepages_supported())
1499                 return rc;
1500
1501         for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1502                 page = pfn_to_page(pfn);
1503                 if (PageHuge(page) && !page_count(page)) {
1504                         rc = dissolve_free_huge_page(page);
1505                         if (rc)
1506                                 break;
1507                 }
1508         }
1509
1510         return rc;
1511 }
1512
1513 /*
1514  * There are 3 ways this can get called:
1515  * 1. With vma+addr: we use the VMA's memory policy
1516  * 2. With !vma, but nid=NUMA_NO_NODE:  We try to allocate a huge
1517  *    page from any node, and let the buddy allocator itself figure
1518  *    it out.
1519  * 3. With !vma, but nid!=NUMA_NO_NODE.  We allocate a huge page
1520  *    strictly from 'nid'
1521  */
1522 static struct page *__hugetlb_alloc_buddy_huge_page(struct hstate *h,
1523                 struct vm_area_struct *vma, unsigned long addr, int nid)
1524 {
1525         int order = huge_page_order(h);
1526         gfp_t gfp = htlb_alloc_mask(h)|__GFP_COMP|__GFP_REPEAT|__GFP_NOWARN;
1527         unsigned int cpuset_mems_cookie;
1528
1529         /*
1530          * We need a VMA to get a memory policy.  If we do not
1531          * have one, we use the 'nid' argument.
1532          *
1533          * The mempolicy stuff below has some non-inlined bits
1534          * and calls ->vm_ops.  That makes it hard to optimize at
1535          * compile-time, even when NUMA is off and it does
1536          * nothing.  This helps the compiler optimize it out.
1537          */
1538         if (!IS_ENABLED(CONFIG_NUMA) || !vma) {
1539                 /*
1540                  * If a specific node is requested, make sure to
1541                  * get memory from there, but only when a node
1542                  * is explicitly specified.
1543                  */
1544                 if (nid != NUMA_NO_NODE)
1545                         gfp |= __GFP_THISNODE;
1546                 /*
1547                  * Make sure to call something that can handle
1548                  * nid=NUMA_NO_NODE
1549                  */
1550                 return alloc_pages_node(nid, gfp, order);
1551         }
1552
1553         /*
1554          * OK, so we have a VMA.  Fetch the mempolicy and try to
1555          * allocate a huge page with it.  We will only reach this
1556          * when CONFIG_NUMA=y.
1557          */
1558         do {
1559                 struct page *page;
1560                 struct mempolicy *mpol;
1561                 struct zonelist *zl;
1562                 nodemask_t *nodemask;
1563
1564                 cpuset_mems_cookie = read_mems_allowed_begin();
1565                 zl = huge_zonelist(vma, addr, gfp, &mpol, &nodemask);
1566                 mpol_cond_put(mpol);
1567                 page = __alloc_pages_nodemask(gfp, order, zl, nodemask);
1568                 if (page)
1569                         return page;
1570         } while (read_mems_allowed_retry(cpuset_mems_cookie));
1571
1572         return NULL;
1573 }
1574
1575 /*
1576  * There are two ways to allocate a huge page:
1577  * 1. When you have a VMA and an address (like a fault)
1578  * 2. When you have no VMA (like when setting /proc/.../nr_hugepages)
1579  *
1580  * 'vma' and 'addr' are only for (1).  'nid' is always NUMA_NO_NODE in
1581  * this case which signifies that the allocation should be done with
1582  * respect for the VMA's memory policy.
1583  *
1584  * For (2), we ignore 'vma' and 'addr' and use 'nid' exclusively. This
1585  * implies that memory policies will not be taken in to account.
1586  */
1587 static struct page *__alloc_buddy_huge_page(struct hstate *h,
1588                 struct vm_area_struct *vma, unsigned long addr, int nid)
1589 {
1590         struct page *page;
1591         unsigned int r_nid;
1592
1593         if (hstate_is_gigantic(h))
1594                 return NULL;
1595
1596         /*
1597          * Make sure that anyone specifying 'nid' is not also specifying a VMA.
1598          * This makes sure the caller is picking _one_ of the modes with which
1599          * we can call this function, not both.
1600          */
1601         if (vma || (addr != -1)) {
1602                 VM_WARN_ON_ONCE(addr == -1);
1603                 VM_WARN_ON_ONCE(nid != NUMA_NO_NODE);
1604         }
1605         /*
1606          * Assume we will successfully allocate the surplus page to
1607          * prevent racing processes from causing the surplus to exceed
1608          * overcommit
1609          *
1610          * This however introduces a different race, where a process B
1611          * tries to grow the static hugepage pool while alloc_pages() is
1612          * called by process A. B will only examine the per-node
1613          * counters in determining if surplus huge pages can be
1614          * converted to normal huge pages in adjust_pool_surplus(). A
1615          * won't be able to increment the per-node counter, until the
1616          * lock is dropped by B, but B doesn't drop hugetlb_lock until
1617          * no more huge pages can be converted from surplus to normal
1618          * state (and doesn't try to convert again). Thus, we have a
1619          * case where a surplus huge page exists, the pool is grown, and
1620          * the surplus huge page still exists after, even though it
1621          * should just have been converted to a normal huge page. This
1622          * does not leak memory, though, as the hugepage will be freed
1623          * once it is out of use. It also does not allow the counters to
1624          * go out of whack in adjust_pool_surplus() as we don't modify
1625          * the node values until we've gotten the hugepage and only the
1626          * per-node value is checked there.
1627          */
1628         spin_lock(&hugetlb_lock);
1629         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1630                 spin_unlock(&hugetlb_lock);
1631                 return NULL;
1632         } else {
1633                 h->nr_huge_pages++;
1634                 h->surplus_huge_pages++;
1635         }
1636         spin_unlock(&hugetlb_lock);
1637
1638         page = __hugetlb_alloc_buddy_huge_page(h, vma, addr, nid);
1639
1640         spin_lock(&hugetlb_lock);
1641         if (page) {
1642                 INIT_LIST_HEAD(&page->lru);
1643                 r_nid = page_to_nid(page);
1644                 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1645                 set_hugetlb_cgroup(page, NULL);
1646                 /*
1647                  * We incremented the global counters already
1648                  */
1649                 h->nr_huge_pages_node[r_nid]++;
1650                 h->surplus_huge_pages_node[r_nid]++;
1651                 __count_vm_event(HTLB_BUDDY_PGALLOC);
1652         } else {
1653                 h->nr_huge_pages--;
1654                 h->surplus_huge_pages--;
1655                 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1656         }
1657         spin_unlock(&hugetlb_lock);
1658
1659         return page;
1660 }
1661
1662 /*
1663  * Allocate a huge page from 'nid'.  Note, 'nid' may be
1664  * NUMA_NO_NODE, which means that it may be allocated
1665  * anywhere.
1666  */
1667 static
1668 struct page *__alloc_buddy_huge_page_no_mpol(struct hstate *h, int nid)
1669 {
1670         unsigned long addr = -1;
1671
1672         return __alloc_buddy_huge_page(h, NULL, addr, nid);
1673 }
1674
1675 /*
1676  * Use the VMA's mpolicy to allocate a huge page from the buddy.
1677  */
1678 static
1679 struct page *__alloc_buddy_huge_page_with_mpol(struct hstate *h,
1680                 struct vm_area_struct *vma, unsigned long addr)
1681 {
1682         return __alloc_buddy_huge_page(h, vma, addr, NUMA_NO_NODE);
1683 }
1684
1685 /*
1686  * This allocation function is useful in the context where vma is irrelevant.
1687  * E.g. soft-offlining uses this function because it only cares physical
1688  * address of error page.
1689  */
1690 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1691 {
1692         struct page *page = NULL;
1693
1694         spin_lock(&hugetlb_lock);
1695         if (h->free_huge_pages - h->resv_huge_pages > 0)
1696                 page = dequeue_huge_page_node(h, nid);
1697         spin_unlock(&hugetlb_lock);
1698
1699         if (!page)
1700                 page = __alloc_buddy_huge_page_no_mpol(h, nid);
1701
1702         return page;
1703 }
1704
1705 /*
1706  * Increase the hugetlb pool such that it can accommodate a reservation
1707  * of size 'delta'.
1708  */
1709 static int gather_surplus_pages(struct hstate *h, int delta)
1710 {
1711         struct list_head surplus_list;
1712         struct page *page, *tmp;
1713         int ret, i;
1714         int needed, allocated;
1715         bool alloc_ok = true;
1716
1717         needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1718         if (needed <= 0) {
1719                 h->resv_huge_pages += delta;
1720                 return 0;
1721         }
1722
1723         allocated = 0;
1724         INIT_LIST_HEAD(&surplus_list);
1725
1726         ret = -ENOMEM;
1727 retry:
1728         spin_unlock(&hugetlb_lock);
1729         for (i = 0; i < needed; i++) {
1730                 page = __alloc_buddy_huge_page_no_mpol(h, NUMA_NO_NODE);
1731                 if (!page) {
1732                         alloc_ok = false;
1733                         break;
1734                 }
1735                 list_add(&page->lru, &surplus_list);
1736         }
1737         allocated += i;
1738
1739         /*
1740          * After retaking hugetlb_lock, we need to recalculate 'needed'
1741          * because either resv_huge_pages or free_huge_pages may have changed.
1742          */
1743         spin_lock(&hugetlb_lock);
1744         needed = (h->resv_huge_pages + delta) -
1745                         (h->free_huge_pages + allocated);
1746         if (needed > 0) {
1747                 if (alloc_ok)
1748                         goto retry;
1749                 /*
1750                  * We were not able to allocate enough pages to
1751                  * satisfy the entire reservation so we free what
1752                  * we've allocated so far.
1753                  */
1754                 goto free;
1755         }
1756         /*
1757          * The surplus_list now contains _at_least_ the number of extra pages
1758          * needed to accommodate the reservation.  Add the appropriate number
1759          * of pages to the hugetlb pool and free the extras back to the buddy
1760          * allocator.  Commit the entire reservation here to prevent another
1761          * process from stealing the pages as they are added to the pool but
1762          * before they are reserved.
1763          */
1764         needed += allocated;
1765         h->resv_huge_pages += delta;
1766         ret = 0;
1767
1768         /* Free the needed pages to the hugetlb pool */
1769         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1770                 if ((--needed) < 0)
1771                         break;
1772                 /*
1773                  * This page is now managed by the hugetlb allocator and has
1774                  * no users -- drop the buddy allocator's reference.
1775                  */
1776                 put_page_testzero(page);
1777                 VM_BUG_ON_PAGE(page_count(page), page);
1778                 enqueue_huge_page(h, page);
1779         }
1780 free:
1781         spin_unlock(&hugetlb_lock);
1782
1783         /* Free unnecessary surplus pages to the buddy allocator */
1784         list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1785                 put_page(page);
1786         spin_lock(&hugetlb_lock);
1787
1788         return ret;
1789 }
1790
1791 /*
1792  * This routine has two main purposes:
1793  * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1794  *    in unused_resv_pages.  This corresponds to the prior adjustments made
1795  *    to the associated reservation map.
1796  * 2) Free any unused surplus pages that may have been allocated to satisfy
1797  *    the reservation.  As many as unused_resv_pages may be freed.
1798  *
1799  * Called with hugetlb_lock held.  However, the lock could be dropped (and
1800  * reacquired) during calls to cond_resched_lock.  Whenever dropping the lock,
1801  * we must make sure nobody else can claim pages we are in the process of
1802  * freeing.  Do this by ensuring resv_huge_page always is greater than the
1803  * number of huge pages we plan to free when dropping the lock.
1804  */
1805 static void return_unused_surplus_pages(struct hstate *h,
1806                                         unsigned long unused_resv_pages)
1807 {
1808         unsigned long nr_pages;
1809
1810         /* Cannot return gigantic pages currently */
1811         if (hstate_is_gigantic(h))
1812                 goto out;
1813
1814         /*
1815          * Part (or even all) of the reservation could have been backed
1816          * by pre-allocated pages. Only free surplus pages.
1817          */
1818         nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1819
1820         /*
1821          * We want to release as many surplus pages as possible, spread
1822          * evenly across all nodes with memory. Iterate across these nodes
1823          * until we can no longer free unreserved surplus pages. This occurs
1824          * when the nodes with surplus pages have no free pages.
1825          * free_pool_huge_page() will balance the the freed pages across the
1826          * on-line nodes with memory and will handle the hstate accounting.
1827          *
1828          * Note that we decrement resv_huge_pages as we free the pages.  If
1829          * we drop the lock, resv_huge_pages will still be sufficiently large
1830          * to cover subsequent pages we may free.
1831          */
1832         while (nr_pages--) {
1833                 h->resv_huge_pages--;
1834                 unused_resv_pages--;
1835                 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1836                         goto out;
1837                 cond_resched_lock(&hugetlb_lock);
1838         }
1839
1840 out:
1841         /* Fully uncommit the reservation */
1842         h->resv_huge_pages -= unused_resv_pages;
1843 }
1844
1845
1846 /*
1847  * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1848  * are used by the huge page allocation routines to manage reservations.
1849  *
1850  * vma_needs_reservation is called to determine if the huge page at addr
1851  * within the vma has an associated reservation.  If a reservation is
1852  * needed, the value 1 is returned.  The caller is then responsible for
1853  * managing the global reservation and subpool usage counts.  After
1854  * the huge page has been allocated, vma_commit_reservation is called
1855  * to add the page to the reservation map.  If the page allocation fails,
1856  * the reservation must be ended instead of committed.  vma_end_reservation
1857  * is called in such cases.
1858  *
1859  * In the normal case, vma_commit_reservation returns the same value
1860  * as the preceding vma_needs_reservation call.  The only time this
1861  * is not the case is if a reserve map was changed between calls.  It
1862  * is the responsibility of the caller to notice the difference and
1863  * take appropriate action.
1864  *
1865  * vma_add_reservation is used in error paths where a reservation must
1866  * be restored when a newly allocated huge page must be freed.  It is
1867  * to be called after calling vma_needs_reservation to determine if a
1868  * reservation exists.
1869  */
1870 enum vma_resv_mode {
1871         VMA_NEEDS_RESV,
1872         VMA_COMMIT_RESV,
1873         VMA_END_RESV,
1874         VMA_ADD_RESV,
1875 };
1876 static long __vma_reservation_common(struct hstate *h,
1877                                 struct vm_area_struct *vma, unsigned long addr,
1878                                 enum vma_resv_mode mode)
1879 {
1880         struct resv_map *resv;
1881         pgoff_t idx;
1882         long ret;
1883
1884         resv = vma_resv_map(vma);
1885         if (!resv)
1886                 return 1;
1887
1888         idx = vma_hugecache_offset(h, vma, addr);
1889         switch (mode) {
1890         case VMA_NEEDS_RESV:
1891                 ret = region_chg(resv, idx, idx + 1);
1892                 break;
1893         case VMA_COMMIT_RESV:
1894                 ret = region_add(resv, idx, idx + 1);
1895                 break;
1896         case VMA_END_RESV:
1897                 region_abort(resv, idx, idx + 1);
1898                 ret = 0;
1899                 break;
1900         case VMA_ADD_RESV:
1901                 if (vma->vm_flags & VM_MAYSHARE)
1902                         ret = region_add(resv, idx, idx + 1);
1903                 else {
1904                         region_abort(resv, idx, idx + 1);
1905                         ret = region_del(resv, idx, idx + 1);
1906                 }
1907                 break;
1908         default:
1909                 BUG();
1910         }
1911
1912         if (vma->vm_flags & VM_MAYSHARE)
1913                 return ret;
1914         else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
1915                 /*
1916                  * In most cases, reserves always exist for private mappings.
1917                  * However, a file associated with mapping could have been
1918                  * hole punched or truncated after reserves were consumed.
1919                  * As subsequent fault on such a range will not use reserves.
1920                  * Subtle - The reserve map for private mappings has the
1921                  * opposite meaning than that of shared mappings.  If NO
1922                  * entry is in the reserve map, it means a reservation exists.
1923                  * If an entry exists in the reserve map, it means the
1924                  * reservation has already been consumed.  As a result, the
1925                  * return value of this routine is the opposite of the
1926                  * value returned from reserve map manipulation routines above.
1927                  */
1928                 if (ret)
1929                         return 0;
1930                 else
1931                         return 1;
1932         }
1933         else
1934                 return ret < 0 ? ret : 0;
1935 }
1936
1937 static long vma_needs_reservation(struct hstate *h,
1938                         struct vm_area_struct *vma, unsigned long addr)
1939 {
1940         return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
1941 }
1942
1943 static long vma_commit_reservation(struct hstate *h,
1944                         struct vm_area_struct *vma, unsigned long addr)
1945 {
1946         return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1947 }
1948
1949 static void vma_end_reservation(struct hstate *h,
1950                         struct vm_area_struct *vma, unsigned long addr)
1951 {
1952         (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
1953 }
1954
1955 static long vma_add_reservation(struct hstate *h,
1956                         struct vm_area_struct *vma, unsigned long addr)
1957 {
1958         return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
1959 }
1960
1961 /*
1962  * This routine is called to restore a reservation on error paths.  In the
1963  * specific error paths, a huge page was allocated (via alloc_huge_page)
1964  * and is about to be freed.  If a reservation for the page existed,
1965  * alloc_huge_page would have consumed the reservation and set PagePrivate
1966  * in the newly allocated page.  When the page is freed via free_huge_page,
1967  * the global reservation count will be incremented if PagePrivate is set.
1968  * However, free_huge_page can not adjust the reserve map.  Adjust the
1969  * reserve map here to be consistent with global reserve count adjustments
1970  * to be made by free_huge_page.
1971  */
1972 static void restore_reserve_on_error(struct hstate *h,
1973                         struct vm_area_struct *vma, unsigned long address,
1974                         struct page *page)
1975 {
1976         if (unlikely(PagePrivate(page))) {
1977                 long rc = vma_needs_reservation(h, vma, address);
1978
1979                 if (unlikely(rc < 0)) {
1980                         /*
1981                          * Rare out of memory condition in reserve map
1982                          * manipulation.  Clear PagePrivate so that
1983                          * global reserve count will not be incremented
1984                          * by free_huge_page.  This will make it appear
1985                          * as though the reservation for this page was
1986                          * consumed.  This may prevent the task from
1987                          * faulting in the page at a later time.  This
1988                          * is better than inconsistent global huge page
1989                          * accounting of reserve counts.
1990                          */
1991                         ClearPagePrivate(page);
1992                 } else if (rc) {
1993                         rc = vma_add_reservation(h, vma, address);
1994                         if (unlikely(rc < 0))
1995                                 /*
1996                                  * See above comment about rare out of
1997                                  * memory condition.
1998                                  */
1999                                 ClearPagePrivate(page);
2000                 } else
2001                         vma_end_reservation(h, vma, address);
2002         }
2003 }
2004
2005 struct page *alloc_huge_page(struct vm_area_struct *vma,
2006                                     unsigned long addr, int avoid_reserve)
2007 {
2008         struct hugepage_subpool *spool = subpool_vma(vma);
2009         struct hstate *h = hstate_vma(vma);
2010         struct page *page;
2011         long map_chg, map_commit;
2012         long gbl_chg;
2013         int ret, idx;
2014         struct hugetlb_cgroup *h_cg;
2015
2016         idx = hstate_index(h);
2017         /*
2018          * Examine the region/reserve map to determine if the process
2019          * has a reservation for the page to be allocated.  A return
2020          * code of zero indicates a reservation exists (no change).
2021          */
2022         map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2023         if (map_chg < 0)
2024                 return ERR_PTR(-ENOMEM);
2025
2026         /*
2027          * Processes that did not create the mapping will have no
2028          * reserves as indicated by the region/reserve map. Check
2029          * that the allocation will not exceed the subpool limit.
2030          * Allocations for MAP_NORESERVE mappings also need to be
2031          * checked against any subpool limit.
2032          */
2033         if (map_chg || avoid_reserve) {
2034                 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2035                 if (gbl_chg < 0) {
2036                         vma_end_reservation(h, vma, addr);
2037                         return ERR_PTR(-ENOSPC);
2038                 }
2039
2040                 /*
2041                  * Even though there was no reservation in the region/reserve
2042                  * map, there could be reservations associated with the
2043                  * subpool that can be used.  This would be indicated if the
2044                  * return value of hugepage_subpool_get_pages() is zero.
2045                  * However, if avoid_reserve is specified we still avoid even
2046                  * the subpool reservations.
2047                  */
2048                 if (avoid_reserve)
2049                         gbl_chg = 1;
2050         }
2051
2052         ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2053         if (ret)
2054                 goto out_subpool_put;
2055
2056         spin_lock(&hugetlb_lock);
2057         /*
2058          * glb_chg is passed to indicate whether or not a page must be taken
2059          * from the global free pool (global change).  gbl_chg == 0 indicates
2060          * a reservation exists for the allocation.
2061          */
2062         page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2063         if (!page) {
2064                 spin_unlock(&hugetlb_lock);
2065                 page = __alloc_buddy_huge_page_with_mpol(h, vma, addr);
2066                 if (!page)
2067                         goto out_uncharge_cgroup;
2068                 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2069                         SetPagePrivate(page);
2070                         h->resv_huge_pages--;
2071                 }
2072                 spin_lock(&hugetlb_lock);
2073                 list_move(&page->lru, &h->hugepage_activelist);
2074                 /* Fall through */
2075         }
2076         hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2077         spin_unlock(&hugetlb_lock);
2078
2079         set_page_private(page, (unsigned long)spool);
2080
2081         map_commit = vma_commit_reservation(h, vma, addr);
2082         if (unlikely(map_chg > map_commit)) {
2083                 /*
2084                  * The page was added to the reservation map between
2085                  * vma_needs_reservation and vma_commit_reservation.
2086                  * This indicates a race with hugetlb_reserve_pages.
2087                  * Adjust for the subpool count incremented above AND
2088                  * in hugetlb_reserve_pages for the same page.  Also,
2089                  * the reservation count added in hugetlb_reserve_pages
2090                  * no longer applies.
2091                  */
2092                 long rsv_adjust;
2093
2094                 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2095                 hugetlb_acct_memory(h, -rsv_adjust);
2096         }
2097         return page;
2098
2099 out_uncharge_cgroup:
2100         hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2101 out_subpool_put:
2102         if (map_chg || avoid_reserve)
2103                 hugepage_subpool_put_pages(spool, 1);
2104         vma_end_reservation(h, vma, addr);
2105         return ERR_PTR(-ENOSPC);
2106 }
2107
2108 /*
2109  * alloc_huge_page()'s wrapper which simply returns the page if allocation
2110  * succeeds, otherwise NULL. This function is called from new_vma_page(),
2111  * where no ERR_VALUE is expected to be returned.
2112  */
2113 struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
2114                                 unsigned long addr, int avoid_reserve)
2115 {
2116         struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
2117         if (IS_ERR(page))
2118                 page = NULL;
2119         return page;
2120 }
2121
2122 int __weak alloc_bootmem_huge_page(struct hstate *h)
2123 {
2124         struct huge_bootmem_page *m;
2125         int nr_nodes, node;
2126
2127         for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2128                 void *addr;
2129
2130                 addr = memblock_virt_alloc_try_nid_nopanic(
2131                                 huge_page_size(h), huge_page_size(h),
2132                                 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
2133                 if (addr) {
2134                         /*
2135                          * Use the beginning of the huge page to store the
2136                          * huge_bootmem_page struct (until gather_bootmem
2137                          * puts them into the mem_map).
2138                          */
2139                         m = addr;
2140                         goto found;
2141                 }
2142         }
2143         return 0;
2144
2145 found:
2146         BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2147         /* Put them into a private list first because mem_map is not up yet */
2148         list_add(&m->list, &huge_boot_pages);
2149         m->hstate = h;
2150         return 1;
2151 }
2152
2153 static void __init prep_compound_huge_page(struct page *page,
2154                 unsigned int order)
2155 {
2156         if (unlikely(order > (MAX_ORDER - 1)))
2157                 prep_compound_gigantic_page(page, order);
2158         else
2159                 prep_compound_page(page, order);
2160 }
2161
2162 /* Put bootmem huge pages into the standard lists after mem_map is up */
2163 static void __init gather_bootmem_prealloc(void)
2164 {
2165         struct huge_bootmem_page *m;
2166
2167         list_for_each_entry(m, &huge_boot_pages, list) {
2168                 struct hstate *h = m->hstate;
2169                 struct page *page;
2170
2171 #ifdef CONFIG_HIGHMEM
2172                 page = pfn_to_page(m->phys >> PAGE_SHIFT);
2173                 memblock_free_late(__pa(m),
2174                                    sizeof(struct huge_bootmem_page));
2175 #else
2176                 page = virt_to_page(m);
2177 #endif
2178                 WARN_ON(page_count(page) != 1);
2179                 prep_compound_huge_page(page, h->order);
2180                 WARN_ON(PageReserved(page));
2181                 prep_new_huge_page(h, page, page_to_nid(page));
2182                 /*
2183                  * If we had gigantic hugepages allocated at boot time, we need
2184                  * to restore the 'stolen' pages to totalram_pages in order to
2185                  * fix confusing memory reports from free(1) and another
2186                  * side-effects, like CommitLimit going negative.
2187                  */
2188                 if (hstate_is_gigantic(h))
2189                         adjust_managed_page_count(page, 1 << h->order);
2190         }
2191 }
2192
2193 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2194 {
2195         unsigned long i;
2196
2197         for (i = 0; i < h->max_huge_pages; ++i) {
2198                 if (hstate_is_gigantic(h)) {
2199                         if (!alloc_bootmem_huge_page(h))
2200                                 break;
2201                 } else if (!alloc_fresh_huge_page(h,
2202                                          &node_states[N_MEMORY]))
2203                         break;
2204         }
2205         h->max_huge_pages = i;
2206 }
2207
2208 static void __init hugetlb_init_hstates(void)
2209 {
2210         struct hstate *h;
2211
2212         for_each_hstate(h) {
2213                 if (minimum_order > huge_page_order(h))
2214                         minimum_order = huge_page_order(h);
2215
2216                 /* oversize hugepages were init'ed in early boot */
2217                 if (!hstate_is_gigantic(h))
2218                         hugetlb_hstate_alloc_pages(h);
2219         }
2220         VM_BUG_ON(minimum_order == UINT_MAX);
2221 }
2222
2223 static char * __init memfmt(char *buf, unsigned long n)
2224 {
2225         if (n >= (1UL << 30))
2226                 sprintf(buf, "%lu GB", n >> 30);
2227         else if (n >= (1UL << 20))
2228                 sprintf(buf, "%lu MB", n >> 20);
2229         else
2230                 sprintf(buf, "%lu KB", n >> 10);
2231         return buf;
2232 }
2233
2234 static void __init report_hugepages(void)
2235 {
2236         struct hstate *h;
2237
2238         for_each_hstate(h) {
2239                 char buf[32];
2240                 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2241                         memfmt(buf, huge_page_size(h)),
2242                         h->free_huge_pages);
2243         }
2244 }
2245
2246 #ifdef CONFIG_HIGHMEM
2247 static void try_to_free_low(struct hstate *h, unsigned long count,
2248                                                 nodemask_t *nodes_allowed)
2249 {
2250         int i;
2251
2252         if (hstate_is_gigantic(h))
2253                 return;
2254
2255         for_each_node_mask(i, *nodes_allowed) {
2256                 struct page *page, *next;
2257                 struct list_head *freel = &h->hugepage_freelists[i];
2258                 list_for_each_entry_safe(page, next, freel, lru) {
2259                         if (count >= h->nr_huge_pages)
2260                                 return;
2261                         if (PageHighMem(page))
2262                                 continue;
2263                         list_del(&page->lru);
2264                         update_and_free_page(h, page);
2265                         h->free_huge_pages--;
2266                         h->free_huge_pages_node[page_to_nid(page)]--;
2267                 }
2268         }
2269 }
2270 #else
2271 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2272                                                 nodemask_t *nodes_allowed)
2273 {
2274 }
2275 #endif
2276
2277 /*
2278  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
2279  * balanced by operating on them in a round-robin fashion.
2280  * Returns 1 if an adjustment was made.
2281  */
2282 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2283                                 int delta)
2284 {
2285         int nr_nodes, node;
2286
2287         VM_BUG_ON(delta != -1 && delta != 1);
2288
2289         if (delta < 0) {
2290                 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2291                         if (h->surplus_huge_pages_node[node])
2292                                 goto found;
2293                 }
2294         } else {
2295                 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2296                         if (h->surplus_huge_pages_node[node] <
2297                                         h->nr_huge_pages_node[node])
2298                                 goto found;
2299                 }
2300         }
2301         return 0;
2302
2303 found:
2304         h->surplus_huge_pages += delta;
2305         h->surplus_huge_pages_node[node] += delta;
2306         return 1;
2307 }
2308
2309 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2310 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
2311                                                 nodemask_t *nodes_allowed)
2312 {
2313         unsigned long min_count, ret;
2314
2315         if (hstate_is_gigantic(h) && !gigantic_page_supported())
2316                 return h->max_huge_pages;
2317
2318         /*
2319          * Increase the pool size
2320          * First take pages out of surplus state.  Then make up the
2321          * remaining difference by allocating fresh huge pages.
2322          *
2323          * We might race with __alloc_buddy_huge_page() here and be unable
2324          * to convert a surplus huge page to a normal huge page. That is
2325          * not critical, though, it just means the overall size of the
2326          * pool might be one hugepage larger than it needs to be, but
2327          * within all the constraints specified by the sysctls.
2328          */
2329         spin_lock(&hugetlb_lock);
2330         while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2331                 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2332                         break;
2333         }
2334
2335         while (count > persistent_huge_pages(h)) {
2336                 /*
2337                  * If this allocation races such that we no longer need the
2338                  * page, free_huge_page will handle it by freeing the page
2339                  * and reducing the surplus.
2340                  */
2341                 spin_unlock(&hugetlb_lock);
2342
2343                 /* yield cpu to avoid soft lockup */
2344                 cond_resched();
2345
2346                 if (hstate_is_gigantic(h))
2347                         ret = alloc_fresh_gigantic_page(h, nodes_allowed);
2348                 else
2349                         ret = alloc_fresh_huge_page(h, nodes_allowed);
2350                 spin_lock(&hugetlb_lock);
2351                 if (!ret)
2352                         goto out;
2353
2354                 /* Bail for signals. Probably ctrl-c from user */
2355                 if (signal_pending(current))
2356                         goto out;
2357         }
2358
2359         /*
2360          * Decrease the pool size
2361          * First return free pages to the buddy allocator (being careful
2362          * to keep enough around to satisfy reservations).  Then place
2363          * pages into surplus state as needed so the pool will shrink
2364          * to the desired size as pages become free.
2365          *
2366          * By placing pages into the surplus state independent of the
2367          * overcommit value, we are allowing the surplus pool size to
2368          * exceed overcommit. There are few sane options here. Since
2369          * __alloc_buddy_huge_page() is checking the global counter,
2370          * though, we'll note that we're not allowed to exceed surplus
2371          * and won't grow the pool anywhere else. Not until one of the
2372          * sysctls are changed, or the surplus pages go out of use.
2373          */
2374         min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2375         min_count = max(count, min_count);
2376         try_to_free_low(h, min_count, nodes_allowed);
2377         while (min_count < persistent_huge_pages(h)) {
2378                 if (!free_pool_huge_page(h, nodes_allowed, 0))
2379                         break;
2380                 cond_resched_lock(&hugetlb_lock);
2381         }
2382         while (count < persistent_huge_pages(h)) {
2383                 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2384                         break;
2385         }
2386 out:
2387         ret = persistent_huge_pages(h);
2388         spin_unlock(&hugetlb_lock);
2389         return ret;
2390 }
2391
2392 #define HSTATE_ATTR_RO(_name) \
2393         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2394
2395 #define HSTATE_ATTR(_name) \
2396         static struct kobj_attribute _name##_attr = \
2397                 __ATTR(_name, 0644, _name##_show, _name##_store)
2398
2399 static struct kobject *hugepages_kobj;
2400 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2401
2402 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2403
2404 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2405 {
2406         int i;
2407
2408         for (i = 0; i < HUGE_MAX_HSTATE; i++)
2409                 if (hstate_kobjs[i] == kobj) {
2410                         if (nidp)
2411                                 *nidp = NUMA_NO_NODE;
2412                         return &hstates[i];
2413                 }
2414
2415         return kobj_to_node_hstate(kobj, nidp);
2416 }
2417
2418 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2419                                         struct kobj_attribute *attr, char *buf)
2420 {
2421         struct hstate *h;
2422         unsigned long nr_huge_pages;
2423         int nid;
2424
2425         h = kobj_to_hstate(kobj, &nid);
2426         if (nid == NUMA_NO_NODE)
2427                 nr_huge_pages = h->nr_huge_pages;
2428         else
2429                 nr_huge_pages = h->nr_huge_pages_node[nid];
2430
2431         return sprintf(buf, "%lu\n", nr_huge_pages);
2432 }
2433
2434 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2435                                            struct hstate *h, int nid,
2436                                            unsigned long count, size_t len)
2437 {
2438         int err;
2439         NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
2440
2441         if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
2442                 err = -EINVAL;
2443                 goto out;
2444         }
2445
2446         if (nid == NUMA_NO_NODE) {
2447                 /*
2448                  * global hstate attribute
2449                  */
2450                 if (!(obey_mempolicy &&
2451                                 init_nodemask_of_mempolicy(nodes_allowed))) {
2452                         NODEMASK_FREE(nodes_allowed);
2453                         nodes_allowed = &node_states[N_MEMORY];
2454                 }
2455         } else if (nodes_allowed) {
2456                 /*
2457                  * per node hstate attribute: adjust count to global,
2458                  * but restrict alloc/free to the specified node.
2459                  */
2460                 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2461                 init_nodemask_of_node(nodes_allowed, nid);
2462         } else
2463                 nodes_allowed = &node_states[N_MEMORY];
2464
2465         h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
2466
2467         if (nodes_allowed != &node_states[N_MEMORY])
2468                 NODEMASK_FREE(nodes_allowed);
2469
2470         return len;
2471 out:
2472         NODEMASK_FREE(nodes_allowed);
2473         return err;
2474 }
2475
2476 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2477                                          struct kobject *kobj, const char *buf,
2478                                          size_t len)
2479 {
2480         struct hstate *h;
2481         unsigned long count;
2482         int nid;
2483         int err;
2484
2485         err = kstrtoul(buf, 10, &count);
2486         if (err)
2487                 return err;
2488
2489         h = kobj_to_hstate(kobj, &nid);
2490         return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2491 }
2492
2493 static ssize_t nr_hugepages_show(struct kobject *kobj,
2494                                        struct kobj_attribute *attr, char *buf)
2495 {
2496         return nr_hugepages_show_common(kobj, attr, buf);
2497 }
2498
2499 static ssize_t nr_hugepages_store(struct kobject *kobj,
2500                struct kobj_attribute *attr, const char *buf, size_t len)
2501 {
2502         return nr_hugepages_store_common(false, kobj, buf, len);
2503 }
2504 HSTATE_ATTR(nr_hugepages);
2505
2506 #ifdef CONFIG_NUMA
2507
2508 /*
2509  * hstate attribute for optionally mempolicy-based constraint on persistent
2510  * huge page alloc/free.
2511  */
2512 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2513                                        struct kobj_attribute *attr, char *buf)
2514 {
2515         return nr_hugepages_show_common(kobj, attr, buf);
2516 }
2517
2518 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2519                struct kobj_attribute *attr, const char *buf, size_t len)
2520 {
2521         return nr_hugepages_store_common(true, kobj, buf, len);
2522 }
2523 HSTATE_ATTR(nr_hugepages_mempolicy);
2524 #endif
2525
2526
2527 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2528                                         struct kobj_attribute *attr, char *buf)
2529 {
2530         struct hstate *h = kobj_to_hstate(kobj, NULL);
2531         return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2532 }
2533
2534 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2535                 struct kobj_attribute *attr, const char *buf, size_t count)
2536 {
2537         int err;
2538         unsigned long input;
2539         struct hstate *h = kobj_to_hstate(kobj, NULL);
2540
2541         if (hstate_is_gigantic(h))
2542                 return -EINVAL;
2543
2544         err = kstrtoul(buf, 10, &input);
2545         if (err)
2546                 return err;
2547
2548         spin_lock(&hugetlb_lock);
2549         h->nr_overcommit_huge_pages = input;
2550         spin_unlock(&hugetlb_lock);
2551
2552         return count;
2553 }
2554 HSTATE_ATTR(nr_overcommit_hugepages);
2555
2556 static ssize_t free_hugepages_show(struct kobject *kobj,
2557                                         struct kobj_attribute *attr, char *buf)
2558 {
2559         struct hstate *h;
2560         unsigned long free_huge_pages;
2561         int nid;
2562
2563         h = kobj_to_hstate(kobj, &nid);
2564         if (nid == NUMA_NO_NODE)
2565                 free_huge_pages = h->free_huge_pages;
2566         else
2567                 free_huge_pages = h->free_huge_pages_node[nid];
2568
2569         return sprintf(buf, "%lu\n", free_huge_pages);
2570 }
2571 HSTATE_ATTR_RO(free_hugepages);
2572
2573 static ssize_t resv_hugepages_show(struct kobject *kobj,
2574                                         struct kobj_attribute *attr, char *buf)
2575 {
2576         struct hstate *h = kobj_to_hstate(kobj, NULL);
2577         return sprintf(buf, "%lu\n", h->resv_huge_pages);
2578 }
2579 HSTATE_ATTR_RO(resv_hugepages);
2580
2581 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2582                                         struct kobj_attribute *attr, char *buf)
2583 {
2584         struct hstate *h;
2585         unsigned long surplus_huge_pages;
2586         int nid;
2587
2588         h = kobj_to_hstate(kobj, &nid);
2589         if (nid == NUMA_NO_NODE)
2590                 surplus_huge_pages = h->surplus_huge_pages;
2591         else
2592                 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2593
2594         return sprintf(buf, "%lu\n", surplus_huge_pages);
2595 }
2596 HSTATE_ATTR_RO(surplus_hugepages);
2597
2598 static struct attribute *hstate_attrs[] = {
2599         &nr_hugepages_attr.attr,
2600         &nr_overcommit_hugepages_attr.attr,
2601         &free_hugepages_attr.attr,
2602         &resv_hugepages_attr.attr,
2603         &surplus_hugepages_attr.attr,
2604 #ifdef CONFIG_NUMA
2605         &nr_hugepages_mempolicy_attr.attr,
2606 #endif
2607         NULL,
2608 };
2609
2610 static struct attribute_group hstate_attr_group = {
2611         .attrs = hstate_attrs,
2612 };
2613
2614 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2615                                     struct kobject **hstate_kobjs,
2616                                     struct attribute_group *hstate_attr_group)
2617 {
2618         int retval;
2619         int hi = hstate_index(h);
2620
2621         hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2622         if (!hstate_kobjs[hi])
2623                 return -ENOMEM;
2624
2625         retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2626         if (retval)
2627                 kobject_put(hstate_kobjs[hi]);
2628
2629         return retval;
2630 }
2631
2632 static void __init hugetlb_sysfs_init(void)
2633 {
2634         struct hstate *h;
2635         int err;
2636
2637         hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2638         if (!hugepages_kobj)
2639                 return;
2640
2641         for_each_hstate(h) {
2642                 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2643                                          hstate_kobjs, &hstate_attr_group);
2644                 if (err)
2645                         pr_err("Hugetlb: Unable to add hstate %s", h->name);
2646         }
2647 }
2648
2649 #ifdef CONFIG_NUMA
2650
2651 /*
2652  * node_hstate/s - associate per node hstate attributes, via their kobjects,
2653  * with node devices in node_devices[] using a parallel array.  The array
2654  * index of a node device or _hstate == node id.
2655  * This is here to avoid any static dependency of the node device driver, in
2656  * the base kernel, on the hugetlb module.
2657  */
2658 struct node_hstate {
2659         struct kobject          *hugepages_kobj;
2660         struct kobject          *hstate_kobjs[HUGE_MAX_HSTATE];
2661 };
2662 static struct node_hstate node_hstates[MAX_NUMNODES];
2663
2664 /*
2665  * A subset of global hstate attributes for node devices
2666  */
2667 static struct attribute *per_node_hstate_attrs[] = {
2668         &nr_hugepages_attr.attr,
2669         &free_hugepages_attr.attr,
2670         &surplus_hugepages_attr.attr,
2671         NULL,
2672 };
2673
2674 static struct attribute_group per_node_hstate_attr_group = {
2675         .attrs = per_node_hstate_attrs,
2676 };
2677
2678 /*
2679  * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2680  * Returns node id via non-NULL nidp.
2681  */
2682 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2683 {
2684         int nid;
2685
2686         for (nid = 0; nid < nr_node_ids; nid++) {
2687                 struct node_hstate *nhs = &node_hstates[nid];
2688                 int i;
2689                 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2690                         if (nhs->hstate_kobjs[i] == kobj) {
2691                                 if (nidp)
2692                                         *nidp = nid;
2693                                 return &hstates[i];
2694                         }
2695         }
2696
2697         BUG();
2698         return NULL;
2699 }
2700
2701 /*
2702  * Unregister hstate attributes from a single node device.
2703  * No-op if no hstate attributes attached.
2704  */
2705 static void hugetlb_unregister_node(struct node *node)
2706 {
2707         struct hstate *h;
2708         struct node_hstate *nhs = &node_hstates[node->dev.id];
2709
2710         if (!nhs->hugepages_kobj)
2711                 return;         /* no hstate attributes */
2712
2713         for_each_hstate(h) {
2714                 int idx = hstate_index(h);
2715                 if (nhs->hstate_kobjs[idx]) {
2716                         kobject_put(nhs->hstate_kobjs[idx]);
2717                         nhs->hstate_kobjs[idx] = NULL;
2718                 }
2719         }
2720
2721         kobject_put(nhs->hugepages_kobj);
2722         nhs->hugepages_kobj = NULL;
2723 }
2724
2725
2726 /*
2727  * Register hstate attributes for a single node device.
2728  * No-op if attributes already registered.
2729  */
2730 static void hugetlb_register_node(struct node *node)
2731 {
2732         struct hstate *h;
2733         struct node_hstate *nhs = &node_hstates[node->dev.id];
2734         int err;
2735
2736         if (nhs->hugepages_kobj)
2737                 return;         /* already allocated */
2738
2739         nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2740                                                         &node->dev.kobj);
2741         if (!nhs->hugepages_kobj)
2742                 return;
2743
2744         for_each_hstate(h) {
2745                 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2746                                                 nhs->hstate_kobjs,
2747                                                 &per_node_hstate_attr_group);
2748                 if (err) {
2749                         pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2750                                 h->name, node->dev.id);
2751                         hugetlb_unregister_node(node);
2752                         break;
2753                 }
2754         }
2755 }
2756
2757 /*
2758  * hugetlb init time:  register hstate attributes for all registered node
2759  * devices of nodes that have memory.  All on-line nodes should have
2760  * registered their associated device by this time.
2761  */
2762 static void __init hugetlb_register_all_nodes(void)
2763 {
2764         int nid;
2765
2766         for_each_node_state(nid, N_MEMORY) {
2767                 struct node *node = node_devices[nid];
2768                 if (node->dev.id == nid)
2769                         hugetlb_register_node(node);
2770         }
2771
2772         /*
2773          * Let the node device driver know we're here so it can
2774          * [un]register hstate attributes on node hotplug.
2775          */
2776         register_hugetlbfs_with_node(hugetlb_register_node,
2777                                      hugetlb_unregister_node);
2778 }
2779 #else   /* !CONFIG_NUMA */
2780
2781 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2782 {
2783         BUG();
2784         if (nidp)
2785                 *nidp = -1;
2786         return NULL;
2787 }
2788
2789 static void hugetlb_register_all_nodes(void) { }
2790
2791 #endif
2792
2793 static int __init hugetlb_init(void)
2794 {
2795         int i;
2796
2797         if (!hugepages_supported())
2798                 return 0;
2799
2800         if (!size_to_hstate(default_hstate_size)) {
2801                 default_hstate_size = HPAGE_SIZE;
2802                 if (!size_to_hstate(default_hstate_size))
2803                         hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2804         }
2805         default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2806         if (default_hstate_max_huge_pages) {
2807                 if (!default_hstate.max_huge_pages)
2808                         default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2809         }
2810
2811         hugetlb_init_hstates();
2812         gather_bootmem_prealloc();
2813         report_hugepages();
2814
2815         hugetlb_sysfs_init();
2816         hugetlb_register_all_nodes();
2817         hugetlb_cgroup_file_init();
2818
2819 #ifdef CONFIG_SMP
2820         num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2821 #else
2822         num_fault_mutexes = 1;
2823 #endif
2824         hugetlb_fault_mutex_table =
2825                 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
2826         BUG_ON(!hugetlb_fault_mutex_table);
2827
2828         for (i = 0; i < num_fault_mutexes; i++)
2829                 mutex_init(&hugetlb_fault_mutex_table[i]);
2830         return 0;
2831 }
2832 subsys_initcall(hugetlb_init);
2833
2834 /* Should be called on processing a hugepagesz=... option */
2835 void __init hugetlb_bad_size(void)
2836 {
2837         parsed_valid_hugepagesz = false;
2838 }
2839
2840 void __init hugetlb_add_hstate(unsigned int order)
2841 {
2842         struct hstate *h;
2843         unsigned long i;
2844
2845         if (size_to_hstate(PAGE_SIZE << order)) {
2846                 pr_warn("hugepagesz= specified twice, ignoring\n");
2847                 return;
2848         }
2849         BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2850         BUG_ON(order == 0);
2851         h = &hstates[hugetlb_max_hstate++];
2852         h->order = order;
2853         h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2854         h->nr_huge_pages = 0;
2855         h->free_huge_pages = 0;
2856         for (i = 0; i < MAX_NUMNODES; ++i)
2857                 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2858         INIT_LIST_HEAD(&h->hugepage_activelist);
2859         h->next_nid_to_alloc = first_memory_node;
2860         h->next_nid_to_free = first_memory_node;
2861         snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2862                                         huge_page_size(h)/1024);
2863
2864         parsed_hstate = h;
2865 }
2866
2867 static int __init hugetlb_nrpages_setup(char *s)
2868 {
2869         unsigned long *mhp;
2870         static unsigned long *last_mhp;
2871
2872         if (!parsed_valid_hugepagesz) {
2873                 pr_warn("hugepages = %s preceded by "
2874                         "an unsupported hugepagesz, ignoring\n", s);
2875                 parsed_valid_hugepagesz = true;
2876                 return 1;
2877         }
2878         /*
2879          * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2880          * so this hugepages= parameter goes to the "default hstate".
2881          */
2882         else if (!hugetlb_max_hstate)
2883                 mhp = &default_hstate_max_huge_pages;
2884         else
2885                 mhp = &parsed_hstate->max_huge_pages;
2886
2887         if (mhp == last_mhp) {
2888                 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
2889                 return 1;
2890         }
2891
2892         if (sscanf(s, "%lu", mhp) <= 0)
2893                 *mhp = 0;
2894
2895         /*
2896          * Global state is always initialized later in hugetlb_init.
2897          * But we need to allocate >= MAX_ORDER hstates here early to still
2898          * use the bootmem allocator.
2899          */
2900         if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2901                 hugetlb_hstate_alloc_pages(parsed_hstate);
2902
2903         last_mhp = mhp;
2904
2905         return 1;
2906 }
2907 __setup("hugepages=", hugetlb_nrpages_setup);
2908
2909 static int __init hugetlb_default_setup(char *s)
2910 {
2911         default_hstate_size = memparse(s, &s);
2912         return 1;
2913 }
2914 __setup("default_hugepagesz=", hugetlb_default_setup);
2915
2916 static unsigned int cpuset_mems_nr(unsigned int *array)
2917 {
2918         int node;
2919         unsigned int nr = 0;
2920
2921         for_each_node_mask(node, cpuset_current_mems_allowed)
2922                 nr += array[node];
2923
2924         return nr;
2925 }
2926
2927 #ifdef CONFIG_SYSCTL
2928 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2929                          struct ctl_table *table, int write,
2930                          void __user *buffer, size_t *length, loff_t *ppos)
2931 {
2932         struct hstate *h = &default_hstate;
2933         unsigned long tmp = h->max_huge_pages;
2934         int ret;
2935
2936         if (!hugepages_supported())
2937                 return -EOPNOTSUPP;
2938
2939         table->data = &tmp;
2940         table->maxlen = sizeof(unsigned long);
2941         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2942         if (ret)
2943                 goto out;
2944
2945         if (write)
2946                 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2947                                                   NUMA_NO_NODE, tmp, *length);
2948 out:
2949         return ret;
2950 }
2951
2952 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2953                           void __user *buffer, size_t *length, loff_t *ppos)
2954 {
2955
2956         return hugetlb_sysctl_handler_common(false, table, write,
2957                                                         buffer, length, ppos);
2958 }
2959
2960 #ifdef CONFIG_NUMA
2961 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2962                           void __user *buffer, size_t *length, loff_t *ppos)
2963 {
2964         return hugetlb_sysctl_handler_common(true, table, write,
2965                                                         buffer, length, ppos);
2966 }
2967 #endif /* CONFIG_NUMA */
2968
2969 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2970                         void __user *buffer,
2971                         size_t *length, loff_t *ppos)
2972 {
2973         struct hstate *h = &default_hstate;
2974         unsigned long tmp;
2975         int ret;
2976
2977         if (!hugepages_supported())
2978                 return -EOPNOTSUPP;
2979
2980         tmp = h->nr_overcommit_huge_pages;
2981
2982         if (write && hstate_is_gigantic(h))
2983                 return -EINVAL;
2984
2985         table->data = &tmp;
2986         table->maxlen = sizeof(unsigned long);
2987         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2988         if (ret)
2989                 goto out;
2990
2991         if (write) {
2992                 spin_lock(&hugetlb_lock);
2993                 h->nr_overcommit_huge_pages = tmp;
2994                 spin_unlock(&hugetlb_lock);
2995         }
2996 out:
2997         return ret;
2998 }
2999
3000 #endif /* CONFIG_SYSCTL */
3001
3002 void hugetlb_report_meminfo(struct seq_file *m)
3003 {
3004         struct hstate *h = &default_hstate;
3005         if (!hugepages_supported())
3006                 return;
3007         seq_printf(m,
3008                         "HugePages_Total:   %5lu\n"
3009                         "HugePages_Free:    %5lu\n"
3010                         "HugePages_Rsvd:    %5lu\n"
3011                         "HugePages_Surp:    %5lu\n"
3012                         "Hugepagesize:   %8lu kB\n",
3013                         h->nr_huge_pages,
3014                         h->free_huge_pages,
3015                         h->resv_huge_pages,
3016                         h->surplus_huge_pages,
3017                         1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3018 }
3019
3020 int hugetlb_report_node_meminfo(int nid, char *buf)
3021 {
3022         struct hstate *h = &default_hstate;
3023         if (!hugepages_supported())
3024                 return 0;
3025         return sprintf(buf,
3026                 "Node %d HugePages_Total: %5u\n"
3027                 "Node %d HugePages_Free:  %5u\n"
3028                 "Node %d HugePages_Surp:  %5u\n",
3029                 nid, h->nr_huge_pages_node[nid],
3030                 nid, h->free_huge_pages_node[nid],
3031                 nid, h->surplus_huge_pages_node[nid]);
3032 }
3033
3034 void hugetlb_show_meminfo(void)
3035 {
3036         struct hstate *h;
3037         int nid;
3038
3039         if (!hugepages_supported())
3040                 return;
3041
3042         for_each_node_state(nid, N_MEMORY)
3043                 for_each_hstate(h)
3044                         pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3045                                 nid,
3046                                 h->nr_huge_pages_node[nid],
3047                                 h->free_huge_pages_node[nid],
3048                                 h->surplus_huge_pages_node[nid],
3049                                 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3050 }
3051
3052 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3053 {
3054         seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3055                    atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3056 }
3057
3058 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3059 unsigned long hugetlb_total_pages(void)
3060 {
3061         struct hstate *h;
3062         unsigned long nr_total_pages = 0;
3063
3064         for_each_hstate(h)
3065                 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3066         return nr_total_pages;
3067 }
3068
3069 static int hugetlb_acct_memory(struct hstate *h, long delta)
3070 {
3071         int ret = -ENOMEM;
3072
3073         spin_lock(&hugetlb_lock);
3074         /*
3075          * When cpuset is configured, it breaks the strict hugetlb page
3076          * reservation as the accounting is done on a global variable. Such
3077          * reservation is completely rubbish in the presence of cpuset because
3078          * the reservation is not checked against page availability for the
3079          * current cpuset. Application can still potentially OOM'ed by kernel
3080          * with lack of free htlb page in cpuset that the task is in.
3081          * Attempt to enforce strict accounting with cpuset is almost
3082          * impossible (or too ugly) because cpuset is too fluid that
3083          * task or memory node can be dynamically moved between cpusets.
3084          *
3085          * The change of semantics for shared hugetlb mapping with cpuset is
3086          * undesirable. However, in order to preserve some of the semantics,
3087          * we fall back to check against current free page availability as
3088          * a best attempt and hopefully to minimize the impact of changing
3089          * semantics that cpuset has.
3090          */
3091         if (delta > 0) {
3092                 if (gather_surplus_pages(h, delta) < 0)
3093                         goto out;
3094
3095                 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
3096                         return_unused_surplus_pages(h, delta);
3097                         goto out;
3098                 }
3099         }
3100
3101         ret = 0;
3102         if (delta < 0)
3103                 return_unused_surplus_pages(h, (unsigned long) -delta);
3104
3105 out:
3106         spin_unlock(&hugetlb_lock);
3107         return ret;
3108 }
3109
3110 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3111 {
3112         struct resv_map *resv = vma_resv_map(vma);
3113
3114         /*
3115          * This new VMA should share its siblings reservation map if present.
3116          * The VMA will only ever have a valid reservation map pointer where
3117          * it is being copied for another still existing VMA.  As that VMA
3118          * has a reference to the reservation map it cannot disappear until
3119          * after this open call completes.  It is therefore safe to take a
3120          * new reference here without additional locking.
3121          */
3122         if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3123                 kref_get(&resv->refs);
3124 }
3125
3126 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3127 {
3128         struct hstate *h = hstate_vma(vma);
3129         struct resv_map *resv = vma_resv_map(vma);
3130         struct hugepage_subpool *spool = subpool_vma(vma);
3131         unsigned long reserve, start, end;
3132         long gbl_reserve;
3133
3134         if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3135                 return;
3136
3137         start = vma_hugecache_offset(h, vma, vma->vm_start);
3138         end = vma_hugecache_offset(h, vma, vma->vm_end);
3139
3140         reserve = (end - start) - region_count(resv, start, end);
3141
3142         kref_put(&resv->refs, resv_map_release);
3143
3144         if (reserve) {
3145                 /*
3146                  * Decrement reserve counts.  The global reserve count may be
3147                  * adjusted if the subpool has a minimum size.
3148                  */
3149                 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3150                 hugetlb_acct_memory(h, -gbl_reserve);
3151         }
3152 }
3153
3154 /*
3155  * We cannot handle pagefaults against hugetlb pages at all.  They cause
3156  * handle_mm_fault() to try to instantiate regular-sized pages in the
3157  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
3158  * this far.
3159  */
3160 static int hugetlb_vm_op_fault(struct vm_fault *vmf)
3161 {
3162         BUG();
3163         return 0;
3164 }
3165
3166 const struct vm_operations_struct hugetlb_vm_ops = {
3167         .fault = hugetlb_vm_op_fault,
3168         .open = hugetlb_vm_op_open,
3169         .close = hugetlb_vm_op_close,
3170 };
3171
3172 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3173                                 int writable)
3174 {
3175         pte_t entry;
3176
3177         if (writable) {
3178                 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3179                                          vma->vm_page_prot)));
3180         } else {
3181                 entry = huge_pte_wrprotect(mk_huge_pte(page,
3182                                            vma->vm_page_prot));
3183         }
3184         entry = pte_mkyoung(entry);
3185         entry = pte_mkhuge(entry);
3186         entry = arch_make_huge_pte(entry, vma, page, writable);
3187
3188         return entry;
3189 }
3190
3191 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3192                                    unsigned long address, pte_t *ptep)
3193 {
3194         pte_t entry;
3195
3196         entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3197         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3198                 update_mmu_cache(vma, address, ptep);
3199 }
3200
3201 bool is_hugetlb_entry_migration(pte_t pte)
3202 {
3203         swp_entry_t swp;
3204
3205         if (huge_pte_none(pte) || pte_present(pte))
3206                 return false;
3207         swp = pte_to_swp_entry(pte);
3208         if (non_swap_entry(swp) && is_migration_entry(swp))
3209                 return true;
3210         else
3211                 return false;
3212 }
3213
3214 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3215 {
3216         swp_entry_t swp;
3217
3218         if (huge_pte_none(pte) || pte_present(pte))
3219                 return 0;
3220         swp = pte_to_swp_entry(pte);
3221         if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3222                 return 1;
3223         else
3224                 return 0;
3225 }
3226
3227 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3228                             struct vm_area_struct *vma)
3229 {
3230         pte_t *src_pte, *dst_pte, entry;
3231         struct page *ptepage;
3232         unsigned long addr;
3233         int cow;
3234         struct hstate *h = hstate_vma(vma);
3235         unsigned long sz = huge_page_size(h);
3236         unsigned long mmun_start;       /* For mmu_notifiers */
3237         unsigned long mmun_end;         /* For mmu_notifiers */
3238         int ret = 0;
3239
3240         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3241
3242         mmun_start = vma->vm_start;
3243         mmun_end = vma->vm_end;
3244         if (cow)
3245                 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
3246
3247         for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3248                 spinlock_t *src_ptl, *dst_ptl;
3249                 src_pte = huge_pte_offset(src, addr, sz);
3250                 if (!src_pte)
3251                         continue;
3252                 dst_pte = huge_pte_alloc(dst, addr, sz);
3253                 if (!dst_pte) {
3254                         ret = -ENOMEM;
3255                         break;
3256                 }
3257
3258                 /* If the pagetables are shared don't copy or take references */
3259                 if (dst_pte == src_pte)
3260                         continue;
3261
3262                 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3263                 src_ptl = huge_pte_lockptr(h, src, src_pte);
3264                 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3265                 entry = huge_ptep_get(src_pte);
3266                 if (huge_pte_none(entry)) { /* skip none entry */
3267                         ;
3268                 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3269                                     is_hugetlb_entry_hwpoisoned(entry))) {
3270                         swp_entry_t swp_entry = pte_to_swp_entry(entry);
3271
3272                         if (is_write_migration_entry(swp_entry) && cow) {
3273                                 /*
3274                                  * COW mappings require pages in both
3275                                  * parent and child to be set to read.
3276                                  */
3277                                 make_migration_entry_read(&swp_entry);
3278                                 entry = swp_entry_to_pte(swp_entry);
3279                                 set_huge_swap_pte_at(src, addr, src_pte,
3280                                                      entry, sz);
3281                         }
3282                         set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3283                 } else {
3284                         if (cow) {
3285                                 huge_ptep_set_wrprotect(src, addr, src_pte);
3286                                 mmu_notifier_invalidate_range(src, mmun_start,
3287                                                                    mmun_end);
3288                         }
3289                         entry = huge_ptep_get(src_pte);
3290                         ptepage = pte_page(entry);
3291                         get_page(ptepage);
3292                         page_dup_rmap(ptepage, true);
3293                         set_huge_pte_at(dst, addr, dst_pte, entry);
3294                         hugetlb_count_add(pages_per_huge_page(h), dst);
3295                 }
3296                 spin_unlock(src_ptl);
3297                 spin_unlock(dst_ptl);
3298         }
3299
3300         if (cow)
3301                 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
3302
3303         return ret;
3304 }
3305
3306 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3307                             unsigned long start, unsigned long end,
3308                             struct page *ref_page)
3309 {
3310         struct mm_struct *mm = vma->vm_mm;
3311         unsigned long address;
3312         pte_t *ptep;
3313         pte_t pte;
3314         spinlock_t *ptl;
3315         struct page *page;
3316         struct hstate *h = hstate_vma(vma);
3317         unsigned long sz = huge_page_size(h);
3318         const unsigned long mmun_start = start; /* For mmu_notifiers */
3319         const unsigned long mmun_end   = end;   /* For mmu_notifiers */
3320
3321         WARN_ON(!is_vm_hugetlb_page(vma));
3322         BUG_ON(start & ~huge_page_mask(h));
3323         BUG_ON(end & ~huge_page_mask(h));
3324
3325         /*
3326          * This is a hugetlb vma, all the pte entries should point
3327          * to huge page.
3328          */
3329         tlb_remove_check_page_size_change(tlb, sz);
3330         tlb_start_vma(tlb, vma);
3331         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3332         address = start;
3333         for (; address < end; address += sz) {
3334                 ptep = huge_pte_offset(mm, address, sz);
3335                 if (!ptep)
3336                         continue;
3337
3338                 ptl = huge_pte_lock(h, mm, ptep);
3339                 if (huge_pmd_unshare(mm, &address, ptep)) {
3340                         spin_unlock(ptl);
3341                         continue;
3342                 }
3343
3344                 pte = huge_ptep_get(ptep);
3345                 if (huge_pte_none(pte)) {
3346                         spin_unlock(ptl);
3347                         continue;
3348                 }
3349
3350                 /*
3351                  * Migrating hugepage or HWPoisoned hugepage is already
3352                  * unmapped and its refcount is dropped, so just clear pte here.
3353                  */
3354                 if (unlikely(!pte_present(pte))) {
3355                         huge_pte_clear(mm, address, ptep, sz);
3356                         spin_unlock(ptl);
3357                         continue;
3358                 }
3359
3360                 page = pte_page(pte);
3361                 /*
3362                  * If a reference page is supplied, it is because a specific
3363                  * page is being unmapped, not a range. Ensure the page we
3364                  * are about to unmap is the actual page of interest.
3365                  */
3366                 if (ref_page) {
3367                         if (page != ref_page) {
3368                                 spin_unlock(ptl);
3369                                 continue;
3370                         }
3371                         /*
3372                          * Mark the VMA as having unmapped its page so that
3373                          * future faults in this VMA will fail rather than
3374                          * looking like data was lost
3375                          */
3376                         set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3377                 }
3378
3379                 pte = huge_ptep_get_and_clear(mm, address, ptep);
3380                 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
3381                 if (huge_pte_dirty(pte))
3382                         set_page_dirty(page);
3383
3384                 hugetlb_count_sub(pages_per_huge_page(h), mm);
3385                 page_remove_rmap(page, true);
3386
3387                 spin_unlock(ptl);
3388                 tlb_remove_page_size(tlb, page, huge_page_size(h));
3389                 /*
3390                  * Bail out after unmapping reference page if supplied
3391                  */
3392                 if (ref_page)
3393                         break;
3394         }
3395         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3396         tlb_end_vma(tlb, vma);
3397 }
3398
3399 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3400                           struct vm_area_struct *vma, unsigned long start,
3401                           unsigned long end, struct page *ref_page)
3402 {
3403         __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3404
3405         /*
3406          * Clear this flag so that x86's huge_pmd_share page_table_shareable
3407          * test will fail on a vma being torn down, and not grab a page table
3408          * on its way out.  We're lucky that the flag has such an appropriate
3409          * name, and can in fact be safely cleared here. We could clear it
3410          * before the __unmap_hugepage_range above, but all that's necessary
3411          * is to clear it before releasing the i_mmap_rwsem. This works
3412          * because in the context this is called, the VMA is about to be
3413          * destroyed and the i_mmap_rwsem is held.
3414          */
3415         vma->vm_flags &= ~VM_MAYSHARE;
3416 }
3417
3418 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3419                           unsigned long end, struct page *ref_page)
3420 {
3421         struct mm_struct *mm;
3422         struct mmu_gather tlb;
3423
3424         mm = vma->vm_mm;
3425
3426         tlb_gather_mmu(&tlb, mm, start, end);
3427         __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3428         tlb_finish_mmu(&tlb, start, end);
3429 }
3430
3431 /*
3432  * This is called when the original mapper is failing to COW a MAP_PRIVATE
3433  * mappping it owns the reserve page for. The intention is to unmap the page
3434  * from other VMAs and let the children be SIGKILLed if they are faulting the
3435  * same region.
3436  */
3437 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3438                               struct page *page, unsigned long address)
3439 {
3440         struct hstate *h = hstate_vma(vma);
3441         struct vm_area_struct *iter_vma;
3442         struct address_space *mapping;
3443         pgoff_t pgoff;
3444
3445         /*
3446          * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3447          * from page cache lookup which is in HPAGE_SIZE units.
3448          */
3449         address = address & huge_page_mask(h);
3450         pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3451                         vma->vm_pgoff;
3452         mapping = vma->vm_file->f_mapping;
3453
3454         /*
3455          * Take the mapping lock for the duration of the table walk. As
3456          * this mapping should be shared between all the VMAs,
3457          * __unmap_hugepage_range() is called as the lock is already held
3458          */
3459         i_mmap_lock_write(mapping);
3460         vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3461                 /* Do not unmap the current VMA */
3462                 if (iter_vma == vma)
3463                         continue;
3464
3465                 /*
3466                  * Shared VMAs have their own reserves and do not affect
3467                  * MAP_PRIVATE accounting but it is possible that a shared
3468                  * VMA is using the same page so check and skip such VMAs.
3469                  */
3470                 if (iter_vma->vm_flags & VM_MAYSHARE)
3471                         continue;
3472
3473                 /*
3474                  * Unmap the page from other VMAs without their own reserves.
3475                  * They get marked to be SIGKILLed if they fault in these
3476                  * areas. This is because a future no-page fault on this VMA
3477                  * could insert a zeroed page instead of the data existing
3478                  * from the time of fork. This would look like data corruption
3479                  */
3480                 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3481                         unmap_hugepage_range(iter_vma, address,
3482                                              address + huge_page_size(h), page);
3483         }
3484         i_mmap_unlock_write(mapping);
3485 }
3486
3487 /*
3488  * Hugetlb_cow() should be called with page lock of the original hugepage held.
3489  * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3490  * cannot race with other handlers or page migration.
3491  * Keep the pte_same checks anyway to make transition from the mutex easier.
3492  */
3493 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3494                        unsigned long address, pte_t *ptep,
3495                        struct page *pagecache_page, spinlock_t *ptl)
3496 {
3497         pte_t pte;
3498         struct hstate *h = hstate_vma(vma);
3499         struct page *old_page, *new_page;
3500         int ret = 0, outside_reserve = 0;
3501         unsigned long mmun_start;       /* For mmu_notifiers */
3502         unsigned long mmun_end;         /* For mmu_notifiers */
3503
3504         pte = huge_ptep_get(ptep);
3505         old_page = pte_page(pte);
3506
3507 retry_avoidcopy:
3508         /* If no-one else is actually using this page, avoid the copy
3509          * and just make the page writable */
3510         if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3511                 page_move_anon_rmap(old_page, vma);
3512                 set_huge_ptep_writable(vma, address, ptep);
3513                 return 0;
3514         }
3515
3516         /*
3517          * If the process that created a MAP_PRIVATE mapping is about to
3518          * perform a COW due to a shared page count, attempt to satisfy
3519          * the allocation without using the existing reserves. The pagecache
3520          * page is used to determine if the reserve at this address was
3521          * consumed or not. If reserves were used, a partial faulted mapping
3522          * at the time of fork() could consume its reserves on COW instead
3523          * of the full address range.
3524          */
3525         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3526                         old_page != pagecache_page)
3527                 outside_reserve = 1;
3528
3529         get_page(old_page);
3530
3531         /*
3532          * Drop page table lock as buddy allocator may be called. It will
3533          * be acquired again before returning to the caller, as expected.
3534          */
3535         spin_unlock(ptl);
3536         new_page = alloc_huge_page(vma, address, outside_reserve);
3537
3538         if (IS_ERR(new_page)) {
3539                 /*
3540                  * If a process owning a MAP_PRIVATE mapping fails to COW,
3541                  * it is due to references held by a child and an insufficient
3542                  * huge page pool. To guarantee the original mappers
3543                  * reliability, unmap the page from child processes. The child
3544                  * may get SIGKILLed if it later faults.
3545                  */
3546                 if (outside_reserve) {
3547                         put_page(old_page);
3548                         BUG_ON(huge_pte_none(pte));
3549                         unmap_ref_private(mm, vma, old_page, address);
3550                         BUG_ON(huge_pte_none(pte));
3551                         spin_lock(ptl);
3552                         ptep = huge_pte_offset(mm, address & huge_page_mask(h),
3553                                                huge_page_size(h));
3554                         if (likely(ptep &&
3555                                    pte_same(huge_ptep_get(ptep), pte)))
3556                                 goto retry_avoidcopy;
3557                         /*
3558                          * race occurs while re-acquiring page table
3559                          * lock, and our job is done.
3560                          */
3561                         return 0;
3562                 }
3563
3564                 ret = (PTR_ERR(new_page) == -ENOMEM) ?
3565                         VM_FAULT_OOM : VM_FAULT_SIGBUS;
3566                 goto out_release_old;
3567         }
3568
3569         /*
3570          * When the original hugepage is shared one, it does not have
3571          * anon_vma prepared.
3572          */
3573         if (unlikely(anon_vma_prepare(vma))) {
3574                 ret = VM_FAULT_OOM;
3575                 goto out_release_all;
3576         }
3577
3578         copy_user_huge_page(new_page, old_page, address, vma,
3579                             pages_per_huge_page(h));
3580         __SetPageUptodate(new_page);
3581         set_page_huge_active(new_page);
3582
3583         mmun_start = address & huge_page_mask(h);
3584         mmun_end = mmun_start + huge_page_size(h);
3585         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3586
3587         /*
3588          * Retake the page table lock to check for racing updates
3589          * before the page tables are altered
3590          */
3591         spin_lock(ptl);
3592         ptep = huge_pte_offset(mm, address & huge_page_mask(h),
3593                                huge_page_size(h));
3594         if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3595                 ClearPagePrivate(new_page);
3596
3597                 /* Break COW */
3598                 huge_ptep_clear_flush(vma, address, ptep);
3599                 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
3600                 set_huge_pte_at(mm, address, ptep,
3601                                 make_huge_pte(vma, new_page, 1));
3602                 page_remove_rmap(old_page, true);
3603                 hugepage_add_new_anon_rmap(new_page, vma, address);
3604                 /* Make the old page be freed below */
3605                 new_page = old_page;
3606         }
3607         spin_unlock(ptl);
3608         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3609 out_release_all:
3610         restore_reserve_on_error(h, vma, address, new_page);
3611         put_page(new_page);
3612 out_release_old:
3613         put_page(old_page);
3614
3615         spin_lock(ptl); /* Caller expects lock to be held */
3616         return ret;
3617 }
3618
3619 /* Return the pagecache page at a given address within a VMA */
3620 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3621                         struct vm_area_struct *vma, unsigned long address)
3622 {
3623         struct address_space *mapping;
3624         pgoff_t idx;
3625
3626         mapping = vma->vm_file->f_mapping;
3627         idx = vma_hugecache_offset(h, vma, address);
3628
3629         return find_lock_page(mapping, idx);
3630 }
3631
3632 /*
3633  * Return whether there is a pagecache page to back given address within VMA.
3634  * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3635  */
3636 static bool hugetlbfs_pagecache_present(struct hstate *h,
3637                         struct vm_area_struct *vma, unsigned long address)
3638 {
3639         struct address_space *mapping;
3640         pgoff_t idx;
3641         struct page *page;
3642
3643         mapping = vma->vm_file->f_mapping;
3644         idx = vma_hugecache_offset(h, vma, address);
3645
3646         page = find_get_page(mapping, idx);
3647         if (page)
3648                 put_page(page);
3649         return page != NULL;
3650 }
3651
3652 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3653                            pgoff_t idx)
3654 {
3655         struct inode *inode = mapping->host;
3656         struct hstate *h = hstate_inode(inode);
3657         int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3658
3659         if (err)
3660                 return err;
3661         ClearPagePrivate(page);
3662
3663         spin_lock(&inode->i_lock);
3664         inode->i_blocks += blocks_per_huge_page(h);
3665         spin_unlock(&inode->i_lock);
3666         return 0;
3667 }
3668
3669 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
3670                            struct address_space *mapping, pgoff_t idx,
3671                            unsigned long address, pte_t *ptep, unsigned int flags)
3672 {
3673         struct hstate *h = hstate_vma(vma);
3674         int ret = VM_FAULT_SIGBUS;
3675         int anon_rmap = 0;
3676         unsigned long size;
3677         struct page *page;
3678         pte_t new_pte;
3679         spinlock_t *ptl;
3680
3681         /*
3682          * Currently, we are forced to kill the process in the event the
3683          * original mapper has unmapped pages from the child due to a failed
3684          * COW. Warn that such a situation has occurred as it may not be obvious
3685          */
3686         if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3687                 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
3688                            current->pid);
3689                 return ret;
3690         }
3691
3692         /*
3693          * Use page lock to guard against racing truncation
3694          * before we get page_table_lock.
3695          */
3696 retry:
3697         page = find_lock_page(mapping, idx);
3698         if (!page) {
3699                 size = i_size_read(mapping->host) >> huge_page_shift(h);
3700                 if (idx >= size)
3701                         goto out;
3702
3703                 /*
3704                  * Check for page in userfault range
3705                  */
3706                 if (userfaultfd_missing(vma)) {
3707                         u32 hash;
3708                         struct vm_fault vmf = {
3709                                 .vma = vma,
3710                                 .address = address,
3711                                 .flags = flags,
3712                                 /*
3713                                  * Hard to debug if it ends up being
3714                                  * used by a callee that assumes
3715                                  * something about the other
3716                                  * uninitialized fields... same as in
3717                                  * memory.c
3718                                  */
3719                         };
3720
3721                         /*
3722                          * hugetlb_fault_mutex must be dropped before
3723                          * handling userfault.  Reacquire after handling
3724                          * fault to make calling code simpler.
3725                          */
3726                         hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping,
3727                                                         idx, address);
3728                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3729                         ret = handle_userfault(&vmf, VM_UFFD_MISSING);
3730                         mutex_lock(&hugetlb_fault_mutex_table[hash]);
3731                         goto out;
3732                 }
3733
3734                 page = alloc_huge_page(vma, address, 0);
3735                 if (IS_ERR(page)) {
3736                         ret = PTR_ERR(page);
3737                         if (ret == -ENOMEM)
3738                                 ret = VM_FAULT_OOM;
3739                         else
3740                                 ret = VM_FAULT_SIGBUS;
3741                         goto out;
3742                 }
3743                 clear_huge_page(page, address, pages_per_huge_page(h));
3744                 __SetPageUptodate(page);
3745                 set_page_huge_active(page);
3746
3747                 if (vma->vm_flags & VM_MAYSHARE) {
3748                         int err = huge_add_to_page_cache(page, mapping, idx);
3749                         if (err) {
3750                                 put_page(page);
3751                                 if (err == -EEXIST)
3752                                         goto retry;
3753                                 goto out;
3754                         }
3755                 } else {
3756                         lock_page(page);
3757                         if (unlikely(anon_vma_prepare(vma))) {
3758                                 ret = VM_FAULT_OOM;
3759                                 goto backout_unlocked;
3760                         }
3761                         anon_rmap = 1;
3762                 }
3763         } else {
3764                 /*
3765                  * If memory error occurs between mmap() and fault, some process
3766                  * don't have hwpoisoned swap entry for errored virtual address.
3767                  * So we need to block hugepage fault by PG_hwpoison bit check.
3768                  */
3769                 if (unlikely(PageHWPoison(page))) {
3770                         ret = VM_FAULT_HWPOISON |
3771                                 VM_FAULT_SET_HINDEX(hstate_index(h));
3772                         goto backout_unlocked;
3773                 }
3774         }
3775
3776         /*
3777          * If we are going to COW a private mapping later, we examine the
3778          * pending reservations for this page now. This will ensure that
3779          * any allocations necessary to record that reservation occur outside
3780          * the spinlock.
3781          */
3782         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3783                 if (vma_needs_reservation(h, vma, address) < 0) {
3784                         ret = VM_FAULT_OOM;
3785                         goto backout_unlocked;
3786                 }
3787                 /* Just decrements count, does not deallocate */
3788                 vma_end_reservation(h, vma, address);
3789         }
3790
3791         ptl = huge_pte_lock(h, mm, ptep);
3792         size = i_size_read(mapping->host) >> huge_page_shift(h);
3793         if (idx >= size)
3794                 goto backout;
3795
3796         ret = 0;
3797         if (!huge_pte_none(huge_ptep_get(ptep)))
3798                 goto backout;
3799
3800         if (anon_rmap) {
3801                 ClearPagePrivate(page);
3802                 hugepage_add_new_anon_rmap(page, vma, address);
3803         } else
3804                 page_dup_rmap(page, true);
3805         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3806                                 && (vma->vm_flags & VM_SHARED)));
3807         set_huge_pte_at(mm, address, ptep, new_pte);
3808
3809         hugetlb_count_add(pages_per_huge_page(h), mm);
3810         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3811                 /* Optimization, do the COW without a second fault */
3812                 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
3813         }
3814
3815         spin_unlock(ptl);
3816         unlock_page(page);
3817 out:
3818         return ret;
3819
3820 backout:
3821         spin_unlock(ptl);
3822 backout_unlocked:
3823         unlock_page(page);
3824         restore_reserve_on_error(h, vma, address, page);
3825         put_page(page);
3826         goto out;
3827 }
3828
3829 #ifdef CONFIG_SMP
3830 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3831                             struct vm_area_struct *vma,
3832                             struct address_space *mapping,
3833                             pgoff_t idx, unsigned long address)
3834 {
3835         unsigned long key[2];
3836         u32 hash;
3837
3838         if (vma->vm_flags & VM_SHARED) {
3839                 key[0] = (unsigned long) mapping;
3840                 key[1] = idx;
3841         } else {
3842                 key[0] = (unsigned long) mm;
3843                 key[1] = address >> huge_page_shift(h);
3844         }
3845
3846         hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3847
3848         return hash & (num_fault_mutexes - 1);
3849 }
3850 #else
3851 /*
3852  * For uniprocesor systems we always use a single mutex, so just
3853  * return 0 and avoid the hashing overhead.
3854  */
3855 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3856                             struct vm_area_struct *vma,
3857                             struct address_space *mapping,
3858                             pgoff_t idx, unsigned long address)
3859 {
3860         return 0;
3861 }
3862 #endif
3863
3864 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3865                         unsigned long address, unsigned int flags)
3866 {
3867         pte_t *ptep, entry;
3868         spinlock_t *ptl;
3869         int ret;
3870         u32 hash;
3871         pgoff_t idx;
3872         struct page *page = NULL;
3873         struct page *pagecache_page = NULL;
3874         struct hstate *h = hstate_vma(vma);
3875         struct address_space *mapping;
3876         int need_wait_lock = 0;
3877
3878         address &= huge_page_mask(h);
3879
3880         ptep = huge_pte_offset(mm, address, huge_page_size(h));
3881         if (ptep) {
3882                 entry = huge_ptep_get(ptep);
3883                 if (unlikely(is_hugetlb_entry_migration(entry))) {
3884                         migration_entry_wait_huge(vma, mm, ptep);
3885                         return 0;
3886                 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
3887                         return VM_FAULT_HWPOISON_LARGE |
3888                                 VM_FAULT_SET_HINDEX(hstate_index(h));
3889         } else {
3890                 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
3891                 if (!ptep)
3892                         return VM_FAULT_OOM;
3893         }
3894
3895         mapping = vma->vm_file->f_mapping;
3896         idx = vma_hugecache_offset(h, vma, address);
3897
3898         /*
3899          * Serialize hugepage allocation and instantiation, so that we don't
3900          * get spurious allocation failures if two CPUs race to instantiate
3901          * the same page in the page cache.
3902          */
3903         hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address);
3904         mutex_lock(&hugetlb_fault_mutex_table[hash]);
3905
3906         entry = huge_ptep_get(ptep);
3907         if (huge_pte_none(entry)) {
3908                 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
3909                 goto out_mutex;
3910         }
3911
3912         ret = 0;
3913
3914         /*
3915          * entry could be a migration/hwpoison entry at this point, so this
3916          * check prevents the kernel from going below assuming that we have
3917          * a active hugepage in pagecache. This goto expects the 2nd page fault,
3918          * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3919          * handle it.
3920          */
3921         if (!pte_present(entry))
3922                 goto out_mutex;
3923
3924         /*
3925          * If we are going to COW the mapping later, we examine the pending
3926          * reservations for this page now. This will ensure that any
3927          * allocations necessary to record that reservation occur outside the
3928          * spinlock. For private mappings, we also lookup the pagecache
3929          * page now as it is used to determine if a reservation has been
3930          * consumed.
3931          */
3932         if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
3933                 if (vma_needs_reservation(h, vma, address) < 0) {
3934                         ret = VM_FAULT_OOM;
3935                         goto out_mutex;
3936                 }
3937                 /* Just decrements count, does not deallocate */
3938                 vma_end_reservation(h, vma, address);
3939
3940                 if (!(vma->vm_flags & VM_MAYSHARE))
3941                         pagecache_page = hugetlbfs_pagecache_page(h,
3942                                                                 vma, address);
3943         }
3944
3945         ptl = huge_pte_lock(h, mm, ptep);
3946
3947         /* Check for a racing update before calling hugetlb_cow */
3948         if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
3949                 goto out_ptl;
3950
3951         /*
3952          * hugetlb_cow() requires page locks of pte_page(entry) and
3953          * pagecache_page, so here we need take the former one
3954          * when page != pagecache_page or !pagecache_page.
3955          */
3956         page = pte_page(entry);
3957         if (page != pagecache_page)
3958                 if (!trylock_page(page)) {
3959                         need_wait_lock = 1;
3960                         goto out_ptl;
3961                 }
3962
3963         get_page(page);
3964
3965         if (flags & FAULT_FLAG_WRITE) {
3966                 if (!huge_pte_write(entry)) {
3967                         ret = hugetlb_cow(mm, vma, address, ptep,
3968                                           pagecache_page, ptl);
3969                         goto out_put_page;
3970                 }
3971                 entry = huge_pte_mkdirty(entry);
3972         }
3973         entry = pte_mkyoung(entry);
3974         if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3975                                                 flags & FAULT_FLAG_WRITE))
3976                 update_mmu_cache(vma, address, ptep);
3977 out_put_page:
3978         if (page != pagecache_page)
3979                 unlock_page(page);
3980         put_page(page);
3981 out_ptl:
3982         spin_unlock(ptl);
3983
3984         if (pagecache_page) {
3985                 unlock_page(pagecache_page);
3986                 put_page(pagecache_page);
3987         }
3988 out_mutex:
3989         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3990         /*
3991          * Generally it's safe to hold refcount during waiting page lock. But
3992          * here we just wait to defer the next page fault to avoid busy loop and
3993          * the page is not used after unlocked before returning from the current
3994          * page fault. So we are safe from accessing freed page, even if we wait
3995          * here without taking refcount.
3996          */
3997         if (need_wait_lock)
3998                 wait_on_page_locked(page);
3999         return ret;
4000 }
4001
4002 /*
4003  * Used by userfaultfd UFFDIO_COPY.  Based on mcopy_atomic_pte with
4004  * modifications for huge pages.
4005  */
4006 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4007                             pte_t *dst_pte,
4008                             struct vm_area_struct *dst_vma,
4009                             unsigned long dst_addr,
4010                             unsigned long src_addr,
4011                             struct page **pagep)
4012 {
4013         int vm_shared = dst_vma->vm_flags & VM_SHARED;
4014         struct hstate *h = hstate_vma(dst_vma);
4015         pte_t _dst_pte;
4016         spinlock_t *ptl;
4017         int ret;
4018         struct page *page;
4019
4020         if (!*pagep) {
4021                 ret = -ENOMEM;
4022                 page = alloc_huge_page(dst_vma, dst_addr, 0);
4023                 if (IS_ERR(page))
4024                         goto out;
4025
4026                 ret = copy_huge_page_from_user(page,
4027                                                 (const void __user *) src_addr,
4028                                                 pages_per_huge_page(h), false);
4029
4030                 /* fallback to copy_from_user outside mmap_sem */
4031                 if (unlikely(ret)) {
4032                         ret = -EFAULT;
4033                         *pagep = page;
4034                         /* don't free the page */
4035                         goto out;
4036                 }
4037         } else {
4038                 page = *pagep;
4039                 *pagep = NULL;
4040         }
4041
4042         /*
4043          * The memory barrier inside __SetPageUptodate makes sure that
4044          * preceding stores to the page contents become visible before
4045          * the set_pte_at() write.
4046          */
4047         __SetPageUptodate(page);
4048         set_page_huge_active(page);
4049
4050         /*
4051          * If shared, add to page cache
4052          */
4053         if (vm_shared) {
4054                 struct address_space *mapping = dst_vma->vm_file->f_mapping;
4055                 pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4056
4057                 ret = huge_add_to_page_cache(page, mapping, idx);
4058                 if (ret)
4059                         goto out_release_nounlock;
4060         }
4061
4062         ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4063         spin_lock(ptl);
4064
4065         ret = -EEXIST;
4066         if (!huge_pte_none(huge_ptep_get(dst_pte)))
4067                 goto out_release_unlock;
4068
4069         if (vm_shared) {
4070                 page_dup_rmap(page, true);
4071         } else {
4072                 ClearPagePrivate(page);
4073                 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4074         }
4075
4076         _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4077         if (dst_vma->vm_flags & VM_WRITE)
4078                 _dst_pte = huge_pte_mkdirty(_dst_pte);
4079         _dst_pte = pte_mkyoung(_dst_pte);
4080
4081         set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4082
4083         (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4084                                         dst_vma->vm_flags & VM_WRITE);
4085         hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4086
4087         /* No need to invalidate - it was non-present before */
4088         update_mmu_cache(dst_vma, dst_addr, dst_pte);
4089
4090         spin_unlock(ptl);
4091         if (vm_shared)
4092                 unlock_page(page);
4093         ret = 0;
4094 out:
4095         return ret;
4096 out_release_unlock:
4097         spin_unlock(ptl);
4098 out_release_nounlock:
4099         if (vm_shared)
4100                 unlock_page(page);
4101         put_page(page);
4102         goto out;
4103 }
4104
4105 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4106                          struct page **pages, struct vm_area_struct **vmas,
4107                          unsigned long *position, unsigned long *nr_pages,
4108                          long i, unsigned int flags, int *nonblocking)
4109 {
4110         unsigned long pfn_offset;
4111         unsigned long vaddr = *position;
4112         unsigned long remainder = *nr_pages;
4113         struct hstate *h = hstate_vma(vma);
4114
4115         while (vaddr < vma->vm_end && remainder) {
4116                 pte_t *pt