Merge tag 'reset-fixes-for-4.18' of git://git.pengutronix.de/git/pza/linux into next...
[platform/kernel/linux-rpi.git] / mm / hugetlb.c
1 /*
2  * Generic hugetlb support.
3  * (C) Nadia Yvette Chambers, April 2004
4  */
5 #include <linux/list.h>
6 #include <linux/init.h>
7 #include <linux/mm.h>
8 #include <linux/seq_file.h>
9 #include <linux/sysctl.h>
10 #include <linux/highmem.h>
11 #include <linux/mmu_notifier.h>
12 #include <linux/nodemask.h>
13 #include <linux/pagemap.h>
14 #include <linux/mempolicy.h>
15 #include <linux/compiler.h>
16 #include <linux/cpuset.h>
17 #include <linux/mutex.h>
18 #include <linux/bootmem.h>
19 #include <linux/sysfs.h>
20 #include <linux/slab.h>
21 #include <linux/mmdebug.h>
22 #include <linux/sched/signal.h>
23 #include <linux/rmap.h>
24 #include <linux/string_helpers.h>
25 #include <linux/swap.h>
26 #include <linux/swapops.h>
27 #include <linux/jhash.h>
28
29 #include <asm/page.h>
30 #include <asm/pgtable.h>
31 #include <asm/tlb.h>
32
33 #include <linux/io.h>
34 #include <linux/hugetlb.h>
35 #include <linux/hugetlb_cgroup.h>
36 #include <linux/node.h>
37 #include <linux/userfaultfd_k.h>
38 #include <linux/page_owner.h>
39 #include "internal.h"
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         if (vma->vm_ops && vma->vm_ops->pagesize)
641                 return vma->vm_ops->pagesize(vma);
642         return PAGE_SIZE;
643 }
644 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
645
646 /*
647  * Return the page size being used by the MMU to back a VMA. In the majority
648  * of cases, the page size used by the kernel matches the MMU size. On
649  * architectures where it differs, an architecture-specific 'strong'
650  * version of this symbol is required.
651  */
652 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
653 {
654         return vma_kernel_pagesize(vma);
655 }
656
657 /*
658  * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
659  * bits of the reservation map pointer, which are always clear due to
660  * alignment.
661  */
662 #define HPAGE_RESV_OWNER    (1UL << 0)
663 #define HPAGE_RESV_UNMAPPED (1UL << 1)
664 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
665
666 /*
667  * These helpers are used to track how many pages are reserved for
668  * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
669  * is guaranteed to have their future faults succeed.
670  *
671  * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
672  * the reserve counters are updated with the hugetlb_lock held. It is safe
673  * to reset the VMA at fork() time as it is not in use yet and there is no
674  * chance of the global counters getting corrupted as a result of the values.
675  *
676  * The private mapping reservation is represented in a subtly different
677  * manner to a shared mapping.  A shared mapping has a region map associated
678  * with the underlying file, this region map represents the backing file
679  * pages which have ever had a reservation assigned which this persists even
680  * after the page is instantiated.  A private mapping has a region map
681  * associated with the original mmap which is attached to all VMAs which
682  * reference it, this region map represents those offsets which have consumed
683  * reservation ie. where pages have been instantiated.
684  */
685 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
686 {
687         return (unsigned long)vma->vm_private_data;
688 }
689
690 static void set_vma_private_data(struct vm_area_struct *vma,
691                                                         unsigned long value)
692 {
693         vma->vm_private_data = (void *)value;
694 }
695
696 struct resv_map *resv_map_alloc(void)
697 {
698         struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
699         struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
700
701         if (!resv_map || !rg) {
702                 kfree(resv_map);
703                 kfree(rg);
704                 return NULL;
705         }
706
707         kref_init(&resv_map->refs);
708         spin_lock_init(&resv_map->lock);
709         INIT_LIST_HEAD(&resv_map->regions);
710
711         resv_map->adds_in_progress = 0;
712
713         INIT_LIST_HEAD(&resv_map->region_cache);
714         list_add(&rg->link, &resv_map->region_cache);
715         resv_map->region_cache_count = 1;
716
717         return resv_map;
718 }
719
720 void resv_map_release(struct kref *ref)
721 {
722         struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
723         struct list_head *head = &resv_map->region_cache;
724         struct file_region *rg, *trg;
725
726         /* Clear out any active regions before we release the map. */
727         region_del(resv_map, 0, LONG_MAX);
728
729         /* ... and any entries left in the cache */
730         list_for_each_entry_safe(rg, trg, head, link) {
731                 list_del(&rg->link);
732                 kfree(rg);
733         }
734
735         VM_BUG_ON(resv_map->adds_in_progress);
736
737         kfree(resv_map);
738 }
739
740 static inline struct resv_map *inode_resv_map(struct inode *inode)
741 {
742         return inode->i_mapping->private_data;
743 }
744
745 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
746 {
747         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
748         if (vma->vm_flags & VM_MAYSHARE) {
749                 struct address_space *mapping = vma->vm_file->f_mapping;
750                 struct inode *inode = mapping->host;
751
752                 return inode_resv_map(inode);
753
754         } else {
755                 return (struct resv_map *)(get_vma_private_data(vma) &
756                                                         ~HPAGE_RESV_MASK);
757         }
758 }
759
760 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
761 {
762         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
763         VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
764
765         set_vma_private_data(vma, (get_vma_private_data(vma) &
766                                 HPAGE_RESV_MASK) | (unsigned long)map);
767 }
768
769 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
770 {
771         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
772         VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
773
774         set_vma_private_data(vma, get_vma_private_data(vma) | flags);
775 }
776
777 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
778 {
779         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
780
781         return (get_vma_private_data(vma) & flag) != 0;
782 }
783
784 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
785 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
786 {
787         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
788         if (!(vma->vm_flags & VM_MAYSHARE))
789                 vma->vm_private_data = (void *)0;
790 }
791
792 /* Returns true if the VMA has associated reserve pages */
793 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
794 {
795         if (vma->vm_flags & VM_NORESERVE) {
796                 /*
797                  * This address is already reserved by other process(chg == 0),
798                  * so, we should decrement reserved count. Without decrementing,
799                  * reserve count remains after releasing inode, because this
800                  * allocated page will go into page cache and is regarded as
801                  * coming from reserved pool in releasing step.  Currently, we
802                  * don't have any other solution to deal with this situation
803                  * properly, so add work-around here.
804                  */
805                 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
806                         return true;
807                 else
808                         return false;
809         }
810
811         /* Shared mappings always use reserves */
812         if (vma->vm_flags & VM_MAYSHARE) {
813                 /*
814                  * We know VM_NORESERVE is not set.  Therefore, there SHOULD
815                  * be a region map for all pages.  The only situation where
816                  * there is no region map is if a hole was punched via
817                  * fallocate.  In this case, there really are no reverves to
818                  * use.  This situation is indicated if chg != 0.
819                  */
820                 if (chg)
821                         return false;
822                 else
823                         return true;
824         }
825
826         /*
827          * Only the process that called mmap() has reserves for
828          * private mappings.
829          */
830         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
831                 /*
832                  * Like the shared case above, a hole punch or truncate
833                  * could have been performed on the private mapping.
834                  * Examine the value of chg to determine if reserves
835                  * actually exist or were previously consumed.
836                  * Very Subtle - The value of chg comes from a previous
837                  * call to vma_needs_reserves().  The reserve map for
838                  * private mappings has different (opposite) semantics
839                  * than that of shared mappings.  vma_needs_reserves()
840                  * has already taken this difference in semantics into
841                  * account.  Therefore, the meaning of chg is the same
842                  * as in the shared case above.  Code could easily be
843                  * combined, but keeping it separate draws attention to
844                  * subtle differences.
845                  */
846                 if (chg)
847                         return false;
848                 else
849                         return true;
850         }
851
852         return false;
853 }
854
855 static void enqueue_huge_page(struct hstate *h, struct page *page)
856 {
857         int nid = page_to_nid(page);
858         list_move(&page->lru, &h->hugepage_freelists[nid]);
859         h->free_huge_pages++;
860         h->free_huge_pages_node[nid]++;
861 }
862
863 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
864 {
865         struct page *page;
866
867         list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
868                 if (!PageHWPoison(page))
869                         break;
870         /*
871          * if 'non-isolated free hugepage' not found on the list,
872          * the allocation fails.
873          */
874         if (&h->hugepage_freelists[nid] == &page->lru)
875                 return NULL;
876         list_move(&page->lru, &h->hugepage_activelist);
877         set_page_refcounted(page);
878         h->free_huge_pages--;
879         h->free_huge_pages_node[nid]--;
880         return page;
881 }
882
883 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
884                 nodemask_t *nmask)
885 {
886         unsigned int cpuset_mems_cookie;
887         struct zonelist *zonelist;
888         struct zone *zone;
889         struct zoneref *z;
890         int node = -1;
891
892         zonelist = node_zonelist(nid, gfp_mask);
893
894 retry_cpuset:
895         cpuset_mems_cookie = read_mems_allowed_begin();
896         for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
897                 struct page *page;
898
899                 if (!cpuset_zone_allowed(zone, gfp_mask))
900                         continue;
901                 /*
902                  * no need to ask again on the same node. Pool is node rather than
903                  * zone aware
904                  */
905                 if (zone_to_nid(zone) == node)
906                         continue;
907                 node = zone_to_nid(zone);
908
909                 page = dequeue_huge_page_node_exact(h, node);
910                 if (page)
911                         return page;
912         }
913         if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
914                 goto retry_cpuset;
915
916         return NULL;
917 }
918
919 /* Movability of hugepages depends on migration support. */
920 static inline gfp_t htlb_alloc_mask(struct hstate *h)
921 {
922         if (hugepage_migration_supported(h))
923                 return GFP_HIGHUSER_MOVABLE;
924         else
925                 return GFP_HIGHUSER;
926 }
927
928 static struct page *dequeue_huge_page_vma(struct hstate *h,
929                                 struct vm_area_struct *vma,
930                                 unsigned long address, int avoid_reserve,
931                                 long chg)
932 {
933         struct page *page;
934         struct mempolicy *mpol;
935         gfp_t gfp_mask;
936         nodemask_t *nodemask;
937         int nid;
938
939         /*
940          * A child process with MAP_PRIVATE mappings created by their parent
941          * have no page reserves. This check ensures that reservations are
942          * not "stolen". The child may still get SIGKILLed
943          */
944         if (!vma_has_reserves(vma, chg) &&
945                         h->free_huge_pages - h->resv_huge_pages == 0)
946                 goto err;
947
948         /* If reserves cannot be used, ensure enough pages are in the pool */
949         if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
950                 goto err;
951
952         gfp_mask = htlb_alloc_mask(h);
953         nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
954         page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
955         if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
956                 SetPagePrivate(page);
957                 h->resv_huge_pages--;
958         }
959
960         mpol_cond_put(mpol);
961         return page;
962
963 err:
964         return NULL;
965 }
966
967 /*
968  * common helper functions for hstate_next_node_to_{alloc|free}.
969  * We may have allocated or freed a huge page based on a different
970  * nodes_allowed previously, so h->next_node_to_{alloc|free} might
971  * be outside of *nodes_allowed.  Ensure that we use an allowed
972  * node for alloc or free.
973  */
974 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
975 {
976         nid = next_node_in(nid, *nodes_allowed);
977         VM_BUG_ON(nid >= MAX_NUMNODES);
978
979         return nid;
980 }
981
982 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
983 {
984         if (!node_isset(nid, *nodes_allowed))
985                 nid = next_node_allowed(nid, nodes_allowed);
986         return nid;
987 }
988
989 /*
990  * returns the previously saved node ["this node"] from which to
991  * allocate a persistent huge page for the pool and advance the
992  * next node from which to allocate, handling wrap at end of node
993  * mask.
994  */
995 static int hstate_next_node_to_alloc(struct hstate *h,
996                                         nodemask_t *nodes_allowed)
997 {
998         int nid;
999
1000         VM_BUG_ON(!nodes_allowed);
1001
1002         nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1003         h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1004
1005         return nid;
1006 }
1007
1008 /*
1009  * helper for free_pool_huge_page() - return the previously saved
1010  * node ["this node"] from which to free a huge page.  Advance the
1011  * next node id whether or not we find a free huge page to free so
1012  * that the next attempt to free addresses the next node.
1013  */
1014 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1015 {
1016         int nid;
1017
1018         VM_BUG_ON(!nodes_allowed);
1019
1020         nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1021         h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1022
1023         return nid;
1024 }
1025
1026 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask)           \
1027         for (nr_nodes = nodes_weight(*mask);                            \
1028                 nr_nodes > 0 &&                                         \
1029                 ((node = hstate_next_node_to_alloc(hs, mask)) || 1);    \
1030                 nr_nodes--)
1031
1032 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask)            \
1033         for (nr_nodes = nodes_weight(*mask);                            \
1034                 nr_nodes > 0 &&                                         \
1035                 ((node = hstate_next_node_to_free(hs, mask)) || 1);     \
1036                 nr_nodes--)
1037
1038 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1039 static void destroy_compound_gigantic_page(struct page *page,
1040                                         unsigned int order)
1041 {
1042         int i;
1043         int nr_pages = 1 << order;
1044         struct page *p = page + 1;
1045
1046         atomic_set(compound_mapcount_ptr(page), 0);
1047         for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1048                 clear_compound_head(p);
1049                 set_page_refcounted(p);
1050         }
1051
1052         set_compound_order(page, 0);
1053         __ClearPageHead(page);
1054 }
1055
1056 static void free_gigantic_page(struct page *page, unsigned int order)
1057 {
1058         free_contig_range(page_to_pfn(page), 1 << order);
1059 }
1060
1061 static int __alloc_gigantic_page(unsigned long start_pfn,
1062                                 unsigned long nr_pages, gfp_t gfp_mask)
1063 {
1064         unsigned long end_pfn = start_pfn + nr_pages;
1065         return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
1066                                   gfp_mask);
1067 }
1068
1069 static bool pfn_range_valid_gigantic(struct zone *z,
1070                         unsigned long start_pfn, unsigned long nr_pages)
1071 {
1072         unsigned long i, end_pfn = start_pfn + nr_pages;
1073         struct page *page;
1074
1075         for (i = start_pfn; i < end_pfn; i++) {
1076                 if (!pfn_valid(i))
1077                         return false;
1078
1079                 page = pfn_to_page(i);
1080
1081                 if (page_zone(page) != z)
1082                         return false;
1083
1084                 if (PageReserved(page))
1085                         return false;
1086
1087                 if (page_count(page) > 0)
1088                         return false;
1089
1090                 if (PageHuge(page))
1091                         return false;
1092         }
1093
1094         return true;
1095 }
1096
1097 static bool zone_spans_last_pfn(const struct zone *zone,
1098                         unsigned long start_pfn, unsigned long nr_pages)
1099 {
1100         unsigned long last_pfn = start_pfn + nr_pages - 1;
1101         return zone_spans_pfn(zone, last_pfn);
1102 }
1103
1104 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1105                 int nid, nodemask_t *nodemask)
1106 {
1107         unsigned int order = huge_page_order(h);
1108         unsigned long nr_pages = 1 << order;
1109         unsigned long ret, pfn, flags;
1110         struct zonelist *zonelist;
1111         struct zone *zone;
1112         struct zoneref *z;
1113
1114         zonelist = node_zonelist(nid, gfp_mask);
1115         for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nodemask) {
1116                 spin_lock_irqsave(&zone->lock, flags);
1117
1118                 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
1119                 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
1120                         if (pfn_range_valid_gigantic(zone, pfn, nr_pages)) {
1121                                 /*
1122                                  * We release the zone lock here because
1123                                  * alloc_contig_range() will also lock the zone
1124                                  * at some point. If there's an allocation
1125                                  * spinning on this lock, it may win the race
1126                                  * and cause alloc_contig_range() to fail...
1127                                  */
1128                                 spin_unlock_irqrestore(&zone->lock, flags);
1129                                 ret = __alloc_gigantic_page(pfn, nr_pages, gfp_mask);
1130                                 if (!ret)
1131                                         return pfn_to_page(pfn);
1132                                 spin_lock_irqsave(&zone->lock, flags);
1133                         }
1134                         pfn += nr_pages;
1135                 }
1136
1137                 spin_unlock_irqrestore(&zone->lock, flags);
1138         }
1139
1140         return NULL;
1141 }
1142
1143 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1144 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1145
1146 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1147 static inline bool gigantic_page_supported(void) { return false; }
1148 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1149                 int nid, nodemask_t *nodemask) { return NULL; }
1150 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1151 static inline void destroy_compound_gigantic_page(struct page *page,
1152                                                 unsigned int order) { }
1153 #endif
1154
1155 static void update_and_free_page(struct hstate *h, struct page *page)
1156 {
1157         int i;
1158
1159         if (hstate_is_gigantic(h) && !gigantic_page_supported())
1160                 return;
1161
1162         h->nr_huge_pages--;
1163         h->nr_huge_pages_node[page_to_nid(page)]--;
1164         for (i = 0; i < pages_per_huge_page(h); i++) {
1165                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1166                                 1 << PG_referenced | 1 << PG_dirty |
1167                                 1 << PG_active | 1 << PG_private |
1168                                 1 << PG_writeback);
1169         }
1170         VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1171         set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1172         set_page_refcounted(page);
1173         if (hstate_is_gigantic(h)) {
1174                 destroy_compound_gigantic_page(page, huge_page_order(h));
1175                 free_gigantic_page(page, huge_page_order(h));
1176         } else {
1177                 __free_pages(page, huge_page_order(h));
1178         }
1179 }
1180
1181 struct hstate *size_to_hstate(unsigned long size)
1182 {
1183         struct hstate *h;
1184
1185         for_each_hstate(h) {
1186                 if (huge_page_size(h) == size)
1187                         return h;
1188         }
1189         return NULL;
1190 }
1191
1192 /*
1193  * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1194  * to hstate->hugepage_activelist.)
1195  *
1196  * This function can be called for tail pages, but never returns true for them.
1197  */
1198 bool page_huge_active(struct page *page)
1199 {
1200         VM_BUG_ON_PAGE(!PageHuge(page), page);
1201         return PageHead(page) && PagePrivate(&page[1]);
1202 }
1203
1204 /* never called for tail page */
1205 static void set_page_huge_active(struct page *page)
1206 {
1207         VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1208         SetPagePrivate(&page[1]);
1209 }
1210
1211 static void clear_page_huge_active(struct page *page)
1212 {
1213         VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1214         ClearPagePrivate(&page[1]);
1215 }
1216
1217 /*
1218  * Internal hugetlb specific page flag. Do not use outside of the hugetlb
1219  * code
1220  */
1221 static inline bool PageHugeTemporary(struct page *page)
1222 {
1223         if (!PageHuge(page))
1224                 return false;
1225
1226         return (unsigned long)page[2].mapping == -1U;
1227 }
1228
1229 static inline void SetPageHugeTemporary(struct page *page)
1230 {
1231         page[2].mapping = (void *)-1U;
1232 }
1233
1234 static inline void ClearPageHugeTemporary(struct page *page)
1235 {
1236         page[2].mapping = NULL;
1237 }
1238
1239 void free_huge_page(struct page *page)
1240 {
1241         /*
1242          * Can't pass hstate in here because it is called from the
1243          * compound page destructor.
1244          */
1245         struct hstate *h = page_hstate(page);
1246         int nid = page_to_nid(page);
1247         struct hugepage_subpool *spool =
1248                 (struct hugepage_subpool *)page_private(page);
1249         bool restore_reserve;
1250
1251         set_page_private(page, 0);
1252         page->mapping = NULL;
1253         VM_BUG_ON_PAGE(page_count(page), page);
1254         VM_BUG_ON_PAGE(page_mapcount(page), page);
1255         restore_reserve = PagePrivate(page);
1256         ClearPagePrivate(page);
1257
1258         /*
1259          * A return code of zero implies that the subpool will be under its
1260          * minimum size if the reservation is not restored after page is free.
1261          * Therefore, force restore_reserve operation.
1262          */
1263         if (hugepage_subpool_put_pages(spool, 1) == 0)
1264                 restore_reserve = true;
1265
1266         spin_lock(&hugetlb_lock);
1267         clear_page_huge_active(page);
1268         hugetlb_cgroup_uncharge_page(hstate_index(h),
1269                                      pages_per_huge_page(h), page);
1270         if (restore_reserve)
1271                 h->resv_huge_pages++;
1272
1273         if (PageHugeTemporary(page)) {
1274                 list_del(&page->lru);
1275                 ClearPageHugeTemporary(page);
1276                 update_and_free_page(h, page);
1277         } else if (h->surplus_huge_pages_node[nid]) {
1278                 /* remove the page from active list */
1279                 list_del(&page->lru);
1280                 update_and_free_page(h, page);
1281                 h->surplus_huge_pages--;
1282                 h->surplus_huge_pages_node[nid]--;
1283         } else {
1284                 arch_clear_hugepage_flags(page);
1285                 enqueue_huge_page(h, page);
1286         }
1287         spin_unlock(&hugetlb_lock);
1288 }
1289
1290 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1291 {
1292         INIT_LIST_HEAD(&page->lru);
1293         set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1294         spin_lock(&hugetlb_lock);
1295         set_hugetlb_cgroup(page, NULL);
1296         h->nr_huge_pages++;
1297         h->nr_huge_pages_node[nid]++;
1298         spin_unlock(&hugetlb_lock);
1299 }
1300
1301 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1302 {
1303         int i;
1304         int nr_pages = 1 << order;
1305         struct page *p = page + 1;
1306
1307         /* we rely on prep_new_huge_page to set the destructor */
1308         set_compound_order(page, order);
1309         __ClearPageReserved(page);
1310         __SetPageHead(page);
1311         for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1312                 /*
1313                  * For gigantic hugepages allocated through bootmem at
1314                  * boot, it's safer to be consistent with the not-gigantic
1315                  * hugepages and clear the PG_reserved bit from all tail pages
1316                  * too.  Otherwse drivers using get_user_pages() to access tail
1317                  * pages may get the reference counting wrong if they see
1318                  * PG_reserved set on a tail page (despite the head page not
1319                  * having PG_reserved set).  Enforcing this consistency between
1320                  * head and tail pages allows drivers to optimize away a check
1321                  * on the head page when they need know if put_page() is needed
1322                  * after get_user_pages().
1323                  */
1324                 __ClearPageReserved(p);
1325                 set_page_count(p, 0);
1326                 set_compound_head(p, page);
1327         }
1328         atomic_set(compound_mapcount_ptr(page), -1);
1329 }
1330
1331 /*
1332  * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1333  * transparent huge pages.  See the PageTransHuge() documentation for more
1334  * details.
1335  */
1336 int PageHuge(struct page *page)
1337 {
1338         if (!PageCompound(page))
1339                 return 0;
1340
1341         page = compound_head(page);
1342         return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1343 }
1344 EXPORT_SYMBOL_GPL(PageHuge);
1345
1346 /*
1347  * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1348  * normal or transparent huge pages.
1349  */
1350 int PageHeadHuge(struct page *page_head)
1351 {
1352         if (!PageHead(page_head))
1353                 return 0;
1354
1355         return get_compound_page_dtor(page_head) == free_huge_page;
1356 }
1357
1358 pgoff_t __basepage_index(struct page *page)
1359 {
1360         struct page *page_head = compound_head(page);
1361         pgoff_t index = page_index(page_head);
1362         unsigned long compound_idx;
1363
1364         if (!PageHuge(page_head))
1365                 return page_index(page);
1366
1367         if (compound_order(page_head) >= MAX_ORDER)
1368                 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1369         else
1370                 compound_idx = page - page_head;
1371
1372         return (index << compound_order(page_head)) + compound_idx;
1373 }
1374
1375 static struct page *alloc_buddy_huge_page(struct hstate *h,
1376                 gfp_t gfp_mask, int nid, nodemask_t *nmask)
1377 {
1378         int order = huge_page_order(h);
1379         struct page *page;
1380
1381         gfp_mask |= __GFP_COMP|__GFP_RETRY_MAYFAIL|__GFP_NOWARN;
1382         if (nid == NUMA_NO_NODE)
1383                 nid = numa_mem_id();
1384         page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
1385         if (page)
1386                 __count_vm_event(HTLB_BUDDY_PGALLOC);
1387         else
1388                 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1389
1390         return page;
1391 }
1392
1393 /*
1394  * Common helper to allocate a fresh hugetlb page. All specific allocators
1395  * should use this function to get new hugetlb pages
1396  */
1397 static struct page *alloc_fresh_huge_page(struct hstate *h,
1398                 gfp_t gfp_mask, int nid, nodemask_t *nmask)
1399 {
1400         struct page *page;
1401
1402         if (hstate_is_gigantic(h))
1403                 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1404         else
1405                 page = alloc_buddy_huge_page(h, gfp_mask,
1406                                 nid, nmask);
1407         if (!page)
1408                 return NULL;
1409
1410         if (hstate_is_gigantic(h))
1411                 prep_compound_gigantic_page(page, huge_page_order(h));
1412         prep_new_huge_page(h, page, page_to_nid(page));
1413
1414         return page;
1415 }
1416
1417 /*
1418  * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1419  * manner.
1420  */
1421 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1422 {
1423         struct page *page;
1424         int nr_nodes, node;
1425         gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1426
1427         for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1428                 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed);
1429                 if (page)
1430                         break;
1431         }
1432
1433         if (!page)
1434                 return 0;
1435
1436         put_page(page); /* free it into the hugepage allocator */
1437
1438         return 1;
1439 }
1440
1441 /*
1442  * Free huge page from pool from next node to free.
1443  * Attempt to keep persistent huge pages more or less
1444  * balanced over allowed nodes.
1445  * Called with hugetlb_lock locked.
1446  */
1447 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1448                                                          bool acct_surplus)
1449 {
1450         int nr_nodes, node;
1451         int ret = 0;
1452
1453         for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1454                 /*
1455                  * If we're returning unused surplus pages, only examine
1456                  * nodes with surplus pages.
1457                  */
1458                 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1459                     !list_empty(&h->hugepage_freelists[node])) {
1460                         struct page *page =
1461                                 list_entry(h->hugepage_freelists[node].next,
1462                                           struct page, lru);
1463                         list_del(&page->lru);
1464                         h->free_huge_pages--;
1465                         h->free_huge_pages_node[node]--;
1466                         if (acct_surplus) {
1467                                 h->surplus_huge_pages--;
1468                                 h->surplus_huge_pages_node[node]--;
1469                         }
1470                         update_and_free_page(h, page);
1471                         ret = 1;
1472                         break;
1473                 }
1474         }
1475
1476         return ret;
1477 }
1478
1479 /*
1480  * Dissolve a given free hugepage into free buddy pages. This function does
1481  * nothing for in-use (including surplus) hugepages. Returns -EBUSY if the
1482  * number of free hugepages would be reduced below the number of reserved
1483  * hugepages.
1484  */
1485 int dissolve_free_huge_page(struct page *page)
1486 {
1487         int rc = 0;
1488
1489         spin_lock(&hugetlb_lock);
1490         if (PageHuge(page) && !page_count(page)) {
1491                 struct page *head = compound_head(page);
1492                 struct hstate *h = page_hstate(head);
1493                 int nid = page_to_nid(head);
1494                 if (h->free_huge_pages - h->resv_huge_pages == 0) {
1495                         rc = -EBUSY;
1496                         goto out;
1497                 }
1498                 /*
1499                  * Move PageHWPoison flag from head page to the raw error page,
1500                  * which makes any subpages rather than the error page reusable.
1501                  */
1502                 if (PageHWPoison(head) && page != head) {
1503                         SetPageHWPoison(page);
1504                         ClearPageHWPoison(head);
1505                 }
1506                 list_del(&head->lru);
1507                 h->free_huge_pages--;
1508                 h->free_huge_pages_node[nid]--;
1509                 h->max_huge_pages--;
1510                 update_and_free_page(h, head);
1511         }
1512 out:
1513         spin_unlock(&hugetlb_lock);
1514         return rc;
1515 }
1516
1517 /*
1518  * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1519  * make specified memory blocks removable from the system.
1520  * Note that this will dissolve a free gigantic hugepage completely, if any
1521  * part of it lies within the given range.
1522  * Also note that if dissolve_free_huge_page() returns with an error, all
1523  * free hugepages that were dissolved before that error are lost.
1524  */
1525 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1526 {
1527         unsigned long pfn;
1528         struct page *page;
1529         int rc = 0;
1530
1531         if (!hugepages_supported())
1532                 return rc;
1533
1534         for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1535                 page = pfn_to_page(pfn);
1536                 if (PageHuge(page) && !page_count(page)) {
1537                         rc = dissolve_free_huge_page(page);
1538                         if (rc)
1539                                 break;
1540                 }
1541         }
1542
1543         return rc;
1544 }
1545
1546 /*
1547  * Allocates a fresh surplus page from the page allocator.
1548  */
1549 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
1550                 int nid, nodemask_t *nmask)
1551 {
1552         struct page *page = NULL;
1553
1554         if (hstate_is_gigantic(h))
1555                 return NULL;
1556
1557         spin_lock(&hugetlb_lock);
1558         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1559                 goto out_unlock;
1560         spin_unlock(&hugetlb_lock);
1561
1562         page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask);
1563         if (!page)
1564                 return NULL;
1565
1566         spin_lock(&hugetlb_lock);
1567         /*
1568          * We could have raced with the pool size change.
1569          * Double check that and simply deallocate the new page
1570          * if we would end up overcommiting the surpluses. Abuse
1571          * temporary page to workaround the nasty free_huge_page
1572          * codeflow
1573          */
1574         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1575                 SetPageHugeTemporary(page);
1576                 put_page(page);
1577                 page = NULL;
1578         } else {
1579                 h->surplus_huge_pages++;
1580                 h->surplus_huge_pages_node[page_to_nid(page)]++;
1581         }
1582
1583 out_unlock:
1584         spin_unlock(&hugetlb_lock);
1585
1586         return page;
1587 }
1588
1589 static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
1590                 int nid, nodemask_t *nmask)
1591 {
1592         struct page *page;
1593
1594         if (hstate_is_gigantic(h))
1595                 return NULL;
1596
1597         page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask);
1598         if (!page)
1599                 return NULL;
1600
1601         /*
1602          * We do not account these pages as surplus because they are only
1603          * temporary and will be released properly on the last reference
1604          */
1605         SetPageHugeTemporary(page);
1606
1607         return page;
1608 }
1609
1610 /*
1611  * Use the VMA's mpolicy to allocate a huge page from the buddy.
1612  */
1613 static
1614 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
1615                 struct vm_area_struct *vma, unsigned long addr)
1616 {
1617         struct page *page;
1618         struct mempolicy *mpol;
1619         gfp_t gfp_mask = htlb_alloc_mask(h);
1620         int nid;
1621         nodemask_t *nodemask;
1622
1623         nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1624         page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
1625         mpol_cond_put(mpol);
1626
1627         return page;
1628 }
1629
1630 /* page migration callback function */
1631 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1632 {
1633         gfp_t gfp_mask = htlb_alloc_mask(h);
1634         struct page *page = NULL;
1635
1636         if (nid != NUMA_NO_NODE)
1637                 gfp_mask |= __GFP_THISNODE;
1638
1639         spin_lock(&hugetlb_lock);
1640         if (h->free_huge_pages - h->resv_huge_pages > 0)
1641                 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL);
1642         spin_unlock(&hugetlb_lock);
1643
1644         if (!page)
1645                 page = alloc_migrate_huge_page(h, gfp_mask, nid, NULL);
1646
1647         return page;
1648 }
1649
1650 /* page migration callback function */
1651 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
1652                 nodemask_t *nmask)
1653 {
1654         gfp_t gfp_mask = htlb_alloc_mask(h);
1655
1656         spin_lock(&hugetlb_lock);
1657         if (h->free_huge_pages - h->resv_huge_pages > 0) {
1658                 struct page *page;
1659
1660                 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1661                 if (page) {
1662                         spin_unlock(&hugetlb_lock);
1663                         return page;
1664                 }
1665         }
1666         spin_unlock(&hugetlb_lock);
1667
1668         return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
1669 }
1670
1671 /* mempolicy aware migration callback */
1672 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
1673                 unsigned long address)
1674 {
1675         struct mempolicy *mpol;
1676         nodemask_t *nodemask;
1677         struct page *page;
1678         gfp_t gfp_mask;
1679         int node;
1680
1681         gfp_mask = htlb_alloc_mask(h);
1682         node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1683         page = alloc_huge_page_nodemask(h, node, nodemask);
1684         mpol_cond_put(mpol);
1685
1686         return page;
1687 }
1688
1689 /*
1690  * Increase the hugetlb pool such that it can accommodate a reservation
1691  * of size 'delta'.
1692  */
1693 static int gather_surplus_pages(struct hstate *h, int delta)
1694 {
1695         struct list_head surplus_list;
1696         struct page *page, *tmp;
1697         int ret, i;
1698         int needed, allocated;
1699         bool alloc_ok = true;
1700
1701         needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1702         if (needed <= 0) {
1703                 h->resv_huge_pages += delta;
1704                 return 0;
1705         }
1706
1707         allocated = 0;
1708         INIT_LIST_HEAD(&surplus_list);
1709
1710         ret = -ENOMEM;
1711 retry:
1712         spin_unlock(&hugetlb_lock);
1713         for (i = 0; i < needed; i++) {
1714                 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
1715                                 NUMA_NO_NODE, NULL);
1716                 if (!page) {
1717                         alloc_ok = false;
1718                         break;
1719                 }
1720                 list_add(&page->lru, &surplus_list);
1721                 cond_resched();
1722         }
1723         allocated += i;
1724
1725         /*
1726          * After retaking hugetlb_lock, we need to recalculate 'needed'
1727          * because either resv_huge_pages or free_huge_pages may have changed.
1728          */
1729         spin_lock(&hugetlb_lock);
1730         needed = (h->resv_huge_pages + delta) -
1731                         (h->free_huge_pages + allocated);
1732         if (needed > 0) {
1733                 if (alloc_ok)
1734                         goto retry;
1735                 /*
1736                  * We were not able to allocate enough pages to
1737                  * satisfy the entire reservation so we free what
1738                  * we've allocated so far.
1739                  */
1740                 goto free;
1741         }
1742         /*
1743          * The surplus_list now contains _at_least_ the number of extra pages
1744          * needed to accommodate the reservation.  Add the appropriate number
1745          * of pages to the hugetlb pool and free the extras back to the buddy
1746          * allocator.  Commit the entire reservation here to prevent another
1747          * process from stealing the pages as they are added to the pool but
1748          * before they are reserved.
1749          */
1750         needed += allocated;
1751         h->resv_huge_pages += delta;
1752         ret = 0;
1753
1754         /* Free the needed pages to the hugetlb pool */
1755         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1756                 if ((--needed) < 0)
1757                         break;
1758                 /*
1759                  * This page is now managed by the hugetlb allocator and has
1760                  * no users -- drop the buddy allocator's reference.
1761                  */
1762                 put_page_testzero(page);
1763                 VM_BUG_ON_PAGE(page_count(page), page);
1764                 enqueue_huge_page(h, page);
1765         }
1766 free:
1767         spin_unlock(&hugetlb_lock);
1768
1769         /* Free unnecessary surplus pages to the buddy allocator */
1770         list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1771                 put_page(page);
1772         spin_lock(&hugetlb_lock);
1773
1774         return ret;
1775 }
1776
1777 /*
1778  * This routine has two main purposes:
1779  * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1780  *    in unused_resv_pages.  This corresponds to the prior adjustments made
1781  *    to the associated reservation map.
1782  * 2) Free any unused surplus pages that may have been allocated to satisfy
1783  *    the reservation.  As many as unused_resv_pages may be freed.
1784  *
1785  * Called with hugetlb_lock held.  However, the lock could be dropped (and
1786  * reacquired) during calls to cond_resched_lock.  Whenever dropping the lock,
1787  * we must make sure nobody else can claim pages we are in the process of
1788  * freeing.  Do this by ensuring resv_huge_page always is greater than the
1789  * number of huge pages we plan to free when dropping the lock.
1790  */
1791 static void return_unused_surplus_pages(struct hstate *h,
1792                                         unsigned long unused_resv_pages)
1793 {
1794         unsigned long nr_pages;
1795
1796         /* Cannot return gigantic pages currently */
1797         if (hstate_is_gigantic(h))
1798                 goto out;
1799
1800         /*
1801          * Part (or even all) of the reservation could have been backed
1802          * by pre-allocated pages. Only free surplus pages.
1803          */
1804         nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1805
1806         /*
1807          * We want to release as many surplus pages as possible, spread
1808          * evenly across all nodes with memory. Iterate across these nodes
1809          * until we can no longer free unreserved surplus pages. This occurs
1810          * when the nodes with surplus pages have no free pages.
1811          * free_pool_huge_page() will balance the the freed pages across the
1812          * on-line nodes with memory and will handle the hstate accounting.
1813          *
1814          * Note that we decrement resv_huge_pages as we free the pages.  If
1815          * we drop the lock, resv_huge_pages will still be sufficiently large
1816          * to cover subsequent pages we may free.
1817          */
1818         while (nr_pages--) {
1819                 h->resv_huge_pages--;
1820                 unused_resv_pages--;
1821                 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1822                         goto out;
1823                 cond_resched_lock(&hugetlb_lock);
1824         }
1825
1826 out:
1827         /* Fully uncommit the reservation */
1828         h->resv_huge_pages -= unused_resv_pages;
1829 }
1830
1831
1832 /*
1833  * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1834  * are used by the huge page allocation routines to manage reservations.
1835  *
1836  * vma_needs_reservation is called to determine if the huge page at addr
1837  * within the vma has an associated reservation.  If a reservation is
1838  * needed, the value 1 is returned.  The caller is then responsible for
1839  * managing the global reservation and subpool usage counts.  After
1840  * the huge page has been allocated, vma_commit_reservation is called
1841  * to add the page to the reservation map.  If the page allocation fails,
1842  * the reservation must be ended instead of committed.  vma_end_reservation
1843  * is called in such cases.
1844  *
1845  * In the normal case, vma_commit_reservation returns the same value
1846  * as the preceding vma_needs_reservation call.  The only time this
1847  * is not the case is if a reserve map was changed between calls.  It
1848  * is the responsibility of the caller to notice the difference and
1849  * take appropriate action.
1850  *
1851  * vma_add_reservation is used in error paths where a reservation must
1852  * be restored when a newly allocated huge page must be freed.  It is
1853  * to be called after calling vma_needs_reservation to determine if a
1854  * reservation exists.
1855  */
1856 enum vma_resv_mode {
1857         VMA_NEEDS_RESV,
1858         VMA_COMMIT_RESV,
1859         VMA_END_RESV,
1860         VMA_ADD_RESV,
1861 };
1862 static long __vma_reservation_common(struct hstate *h,
1863                                 struct vm_area_struct *vma, unsigned long addr,
1864                                 enum vma_resv_mode mode)
1865 {
1866         struct resv_map *resv;
1867         pgoff_t idx;
1868         long ret;
1869
1870         resv = vma_resv_map(vma);
1871         if (!resv)
1872                 return 1;
1873
1874         idx = vma_hugecache_offset(h, vma, addr);
1875         switch (mode) {
1876         case VMA_NEEDS_RESV:
1877                 ret = region_chg(resv, idx, idx + 1);
1878                 break;
1879         case VMA_COMMIT_RESV:
1880                 ret = region_add(resv, idx, idx + 1);
1881                 break;
1882         case VMA_END_RESV:
1883                 region_abort(resv, idx, idx + 1);
1884                 ret = 0;
1885                 break;
1886         case VMA_ADD_RESV:
1887                 if (vma->vm_flags & VM_MAYSHARE)
1888                         ret = region_add(resv, idx, idx + 1);
1889                 else {
1890                         region_abort(resv, idx, idx + 1);
1891                         ret = region_del(resv, idx, idx + 1);
1892                 }
1893                 break;
1894         default:
1895                 BUG();
1896         }
1897
1898         if (vma->vm_flags & VM_MAYSHARE)
1899                 return ret;
1900         else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
1901                 /*
1902                  * In most cases, reserves always exist for private mappings.
1903                  * However, a file associated with mapping could have been
1904                  * hole punched or truncated after reserves were consumed.
1905                  * As subsequent fault on such a range will not use reserves.
1906                  * Subtle - The reserve map for private mappings has the
1907                  * opposite meaning than that of shared mappings.  If NO
1908                  * entry is in the reserve map, it means a reservation exists.
1909                  * If an entry exists in the reserve map, it means the
1910                  * reservation has already been consumed.  As a result, the
1911                  * return value of this routine is the opposite of the
1912                  * value returned from reserve map manipulation routines above.
1913                  */
1914                 if (ret)
1915                         return 0;
1916                 else
1917                         return 1;
1918         }
1919         else
1920                 return ret < 0 ? ret : 0;
1921 }
1922
1923 static long vma_needs_reservation(struct hstate *h,
1924                         struct vm_area_struct *vma, unsigned long addr)
1925 {
1926         return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
1927 }
1928
1929 static long vma_commit_reservation(struct hstate *h,
1930                         struct vm_area_struct *vma, unsigned long addr)
1931 {
1932         return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1933 }
1934
1935 static void vma_end_reservation(struct hstate *h,
1936                         struct vm_area_struct *vma, unsigned long addr)
1937 {
1938         (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
1939 }
1940
1941 static long vma_add_reservation(struct hstate *h,
1942                         struct vm_area_struct *vma, unsigned long addr)
1943 {
1944         return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
1945 }
1946
1947 /*
1948  * This routine is called to restore a reservation on error paths.  In the
1949  * specific error paths, a huge page was allocated (via alloc_huge_page)
1950  * and is about to be freed.  If a reservation for the page existed,
1951  * alloc_huge_page would have consumed the reservation and set PagePrivate
1952  * in the newly allocated page.  When the page is freed via free_huge_page,
1953  * the global reservation count will be incremented if PagePrivate is set.
1954  * However, free_huge_page can not adjust the reserve map.  Adjust the
1955  * reserve map here to be consistent with global reserve count adjustments
1956  * to be made by free_huge_page.
1957  */
1958 static void restore_reserve_on_error(struct hstate *h,
1959                         struct vm_area_struct *vma, unsigned long address,
1960                         struct page *page)
1961 {
1962         if (unlikely(PagePrivate(page))) {
1963                 long rc = vma_needs_reservation(h, vma, address);
1964
1965                 if (unlikely(rc < 0)) {
1966                         /*
1967                          * Rare out of memory condition in reserve map
1968                          * manipulation.  Clear PagePrivate so that
1969                          * global reserve count will not be incremented
1970                          * by free_huge_page.  This will make it appear
1971                          * as though the reservation for this page was
1972                          * consumed.  This may prevent the task from
1973                          * faulting in the page at a later time.  This
1974                          * is better than inconsistent global huge page
1975                          * accounting of reserve counts.
1976                          */
1977                         ClearPagePrivate(page);
1978                 } else if (rc) {
1979                         rc = vma_add_reservation(h, vma, address);
1980                         if (unlikely(rc < 0))
1981                                 /*
1982                                  * See above comment about rare out of
1983                                  * memory condition.
1984                                  */
1985                                 ClearPagePrivate(page);
1986                 } else
1987                         vma_end_reservation(h, vma, address);
1988         }
1989 }
1990
1991 struct page *alloc_huge_page(struct vm_area_struct *vma,
1992                                     unsigned long addr, int avoid_reserve)
1993 {
1994         struct hugepage_subpool *spool = subpool_vma(vma);
1995         struct hstate *h = hstate_vma(vma);
1996         struct page *page;
1997         long map_chg, map_commit;
1998         long gbl_chg;
1999         int ret, idx;
2000         struct hugetlb_cgroup *h_cg;
2001
2002         idx = hstate_index(h);
2003         /*
2004          * Examine the region/reserve map to determine if the process
2005          * has a reservation for the page to be allocated.  A return
2006          * code of zero indicates a reservation exists (no change).
2007          */
2008         map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2009         if (map_chg < 0)
2010                 return ERR_PTR(-ENOMEM);
2011
2012         /*
2013          * Processes that did not create the mapping will have no
2014          * reserves as indicated by the region/reserve map. Check
2015          * that the allocation will not exceed the subpool limit.
2016          * Allocations for MAP_NORESERVE mappings also need to be
2017          * checked against any subpool limit.
2018          */
2019         if (map_chg || avoid_reserve) {
2020                 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2021                 if (gbl_chg < 0) {
2022                         vma_end_reservation(h, vma, addr);
2023                         return ERR_PTR(-ENOSPC);
2024                 }
2025
2026                 /*
2027                  * Even though there was no reservation in the region/reserve
2028                  * map, there could be reservations associated with the
2029                  * subpool that can be used.  This would be indicated if the
2030                  * return value of hugepage_subpool_get_pages() is zero.
2031                  * However, if avoid_reserve is specified we still avoid even
2032                  * the subpool reservations.
2033                  */
2034                 if (avoid_reserve)
2035                         gbl_chg = 1;
2036         }
2037
2038         ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2039         if (ret)
2040                 goto out_subpool_put;
2041
2042         spin_lock(&hugetlb_lock);
2043         /*
2044          * glb_chg is passed to indicate whether or not a page must be taken
2045          * from the global free pool (global change).  gbl_chg == 0 indicates
2046          * a reservation exists for the allocation.
2047          */
2048         page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2049         if (!page) {
2050                 spin_unlock(&hugetlb_lock);
2051                 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2052                 if (!page)
2053                         goto out_uncharge_cgroup;
2054                 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2055                         SetPagePrivate(page);
2056                         h->resv_huge_pages--;
2057                 }
2058                 spin_lock(&hugetlb_lock);
2059                 list_move(&page->lru, &h->hugepage_activelist);
2060                 /* Fall through */
2061         }
2062         hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2063         spin_unlock(&hugetlb_lock);
2064
2065         set_page_private(page, (unsigned long)spool);
2066
2067         map_commit = vma_commit_reservation(h, vma, addr);
2068         if (unlikely(map_chg > map_commit)) {
2069                 /*
2070                  * The page was added to the reservation map between
2071                  * vma_needs_reservation and vma_commit_reservation.
2072                  * This indicates a race with hugetlb_reserve_pages.
2073                  * Adjust for the subpool count incremented above AND
2074                  * in hugetlb_reserve_pages for the same page.  Also,
2075                  * the reservation count added in hugetlb_reserve_pages
2076                  * no longer applies.
2077                  */
2078                 long rsv_adjust;
2079
2080                 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2081                 hugetlb_acct_memory(h, -rsv_adjust);
2082         }
2083         return page;
2084
2085 out_uncharge_cgroup:
2086         hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2087 out_subpool_put:
2088         if (map_chg || avoid_reserve)
2089                 hugepage_subpool_put_pages(spool, 1);
2090         vma_end_reservation(h, vma, addr);
2091         return ERR_PTR(-ENOSPC);
2092 }
2093
2094 int alloc_bootmem_huge_page(struct hstate *h)
2095         __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2096 int __alloc_bootmem_huge_page(struct hstate *h)
2097 {
2098         struct huge_bootmem_page *m;
2099         int nr_nodes, node;
2100
2101         for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2102                 void *addr;
2103
2104                 addr = memblock_virt_alloc_try_nid_nopanic(
2105                                 huge_page_size(h), huge_page_size(h),
2106                                 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
2107                 if (addr) {
2108                         /*
2109                          * Use the beginning of the huge page to store the
2110                          * huge_bootmem_page struct (until gather_bootmem
2111                          * puts them into the mem_map).
2112                          */
2113                         m = addr;
2114                         goto found;
2115                 }
2116         }
2117         return 0;
2118
2119 found:
2120         BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2121         /* Put them into a private list first because mem_map is not up yet */
2122         list_add(&m->list, &huge_boot_pages);
2123         m->hstate = h;
2124         return 1;
2125 }
2126
2127 static void __init prep_compound_huge_page(struct page *page,
2128                 unsigned int order)
2129 {
2130         if (unlikely(order > (MAX_ORDER - 1)))
2131                 prep_compound_gigantic_page(page, order);
2132         else
2133                 prep_compound_page(page, order);
2134 }
2135
2136 /* Put bootmem huge pages into the standard lists after mem_map is up */
2137 static void __init gather_bootmem_prealloc(void)
2138 {
2139         struct huge_bootmem_page *m;
2140
2141         list_for_each_entry(m, &huge_boot_pages, list) {
2142                 struct hstate *h = m->hstate;
2143                 struct page *page;
2144
2145 #ifdef CONFIG_HIGHMEM
2146                 page = pfn_to_page(m->phys >> PAGE_SHIFT);
2147                 memblock_free_late(__pa(m),
2148                                    sizeof(struct huge_bootmem_page));
2149 #else
2150                 page = virt_to_page(m);
2151 #endif
2152                 WARN_ON(page_count(page) != 1);
2153                 prep_compound_huge_page(page, h->order);
2154                 WARN_ON(PageReserved(page));
2155                 prep_new_huge_page(h, page, page_to_nid(page));
2156                 put_page(page); /* free it into the hugepage allocator */
2157
2158                 /*
2159                  * If we had gigantic hugepages allocated at boot time, we need
2160                  * to restore the 'stolen' pages to totalram_pages in order to
2161                  * fix confusing memory reports from free(1) and another
2162                  * side-effects, like CommitLimit going negative.
2163                  */
2164                 if (hstate_is_gigantic(h))
2165                         adjust_managed_page_count(page, 1 << h->order);
2166         }
2167 }
2168
2169 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2170 {
2171         unsigned long i;
2172
2173         for (i = 0; i < h->max_huge_pages; ++i) {
2174                 if (hstate_is_gigantic(h)) {
2175                         if (!alloc_bootmem_huge_page(h))
2176                                 break;
2177                 } else if (!alloc_pool_huge_page(h,
2178                                          &node_states[N_MEMORY]))
2179                         break;
2180                 cond_resched();
2181         }
2182         if (i < h->max_huge_pages) {
2183                 char buf[32];
2184
2185                 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2186                 pr_warn("HugeTLB: allocating %lu of page size %s failed.  Only allocated %lu hugepages.\n",
2187                         h->max_huge_pages, buf, i);
2188                 h->max_huge_pages = i;
2189         }
2190 }
2191
2192 static void __init hugetlb_init_hstates(void)
2193 {
2194         struct hstate *h;
2195
2196         for_each_hstate(h) {
2197                 if (minimum_order > huge_page_order(h))
2198                         minimum_order = huge_page_order(h);
2199
2200                 /* oversize hugepages were init'ed in early boot */
2201                 if (!hstate_is_gigantic(h))
2202                         hugetlb_hstate_alloc_pages(h);
2203         }
2204         VM_BUG_ON(minimum_order == UINT_MAX);
2205 }
2206
2207 static void __init report_hugepages(void)
2208 {
2209         struct hstate *h;
2210
2211         for_each_hstate(h) {
2212                 char buf[32];
2213
2214                 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2215                 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2216                         buf, h->free_huge_pages);
2217         }
2218 }
2219
2220 #ifdef CONFIG_HIGHMEM
2221 static void try_to_free_low(struct hstate *h, unsigned long count,
2222                                                 nodemask_t *nodes_allowed)
2223 {
2224         int i;
2225
2226         if (hstate_is_gigantic(h))
2227                 return;
2228
2229         for_each_node_mask(i, *nodes_allowed) {
2230                 struct page *page, *next;
2231                 struct list_head *freel = &h->hugepage_freelists[i];
2232                 list_for_each_entry_safe(page, next, freel, lru) {
2233                         if (count >= h->nr_huge_pages)
2234                                 return;
2235                         if (PageHighMem(page))
2236                                 continue;
2237                         list_del(&page->lru);
2238                         update_and_free_page(h, page);
2239                         h->free_huge_pages--;
2240                         h->free_huge_pages_node[page_to_nid(page)]--;
2241                 }
2242         }
2243 }
2244 #else
2245 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2246                                                 nodemask_t *nodes_allowed)
2247 {
2248 }
2249 #endif
2250
2251 /*
2252  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
2253  * balanced by operating on them in a round-robin fashion.
2254  * Returns 1 if an adjustment was made.
2255  */
2256 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2257                                 int delta)
2258 {
2259         int nr_nodes, node;
2260
2261         VM_BUG_ON(delta != -1 && delta != 1);
2262
2263         if (delta < 0) {
2264                 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2265                         if (h->surplus_huge_pages_node[node])
2266                                 goto found;
2267                 }
2268         } else {
2269                 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2270                         if (h->surplus_huge_pages_node[node] <
2271                                         h->nr_huge_pages_node[node])
2272                                 goto found;
2273                 }
2274         }
2275         return 0;
2276
2277 found:
2278         h->surplus_huge_pages += delta;
2279         h->surplus_huge_pages_node[node] += delta;
2280         return 1;
2281 }
2282
2283 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2284 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
2285                                                 nodemask_t *nodes_allowed)
2286 {
2287         unsigned long min_count, ret;
2288
2289         if (hstate_is_gigantic(h) && !gigantic_page_supported())
2290                 return h->max_huge_pages;
2291
2292         /*
2293          * Increase the pool size
2294          * First take pages out of surplus state.  Then make up the
2295          * remaining difference by allocating fresh huge pages.
2296          *
2297          * We might race with alloc_surplus_huge_page() here and be unable
2298          * to convert a surplus huge page to a normal huge page. That is
2299          * not critical, though, it just means the overall size of the
2300          * pool might be one hugepage larger than it needs to be, but
2301          * within all the constraints specified by the sysctls.
2302          */
2303         spin_lock(&hugetlb_lock);
2304         while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2305                 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2306                         break;
2307         }
2308
2309         while (count > persistent_huge_pages(h)) {
2310                 /*
2311                  * If this allocation races such that we no longer need the
2312                  * page, free_huge_page will handle it by freeing the page
2313                  * and reducing the surplus.
2314                  */
2315                 spin_unlock(&hugetlb_lock);
2316
2317                 /* yield cpu to avoid soft lockup */
2318                 cond_resched();
2319
2320                 ret = alloc_pool_huge_page(h, nodes_allowed);
2321                 spin_lock(&hugetlb_lock);
2322                 if (!ret)
2323                         goto out;
2324
2325                 /* Bail for signals. Probably ctrl-c from user */
2326                 if (signal_pending(current))
2327                         goto out;
2328         }
2329
2330         /*
2331          * Decrease the pool size
2332          * First return free pages to the buddy allocator (being careful
2333          * to keep enough around to satisfy reservations).  Then place
2334          * pages into surplus state as needed so the pool will shrink
2335          * to the desired size as pages become free.
2336          *
2337          * By placing pages into the surplus state independent of the
2338          * overcommit value, we are allowing the surplus pool size to
2339          * exceed overcommit. There are few sane options here. Since
2340          * alloc_surplus_huge_page() is checking the global counter,
2341          * though, we'll note that we're not allowed to exceed surplus
2342          * and won't grow the pool anywhere else. Not until one of the
2343          * sysctls are changed, or the surplus pages go out of use.
2344          */
2345         min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2346         min_count = max(count, min_count);
2347         try_to_free_low(h, min_count, nodes_allowed);
2348         while (min_count < persistent_huge_pages(h)) {
2349                 if (!free_pool_huge_page(h, nodes_allowed, 0))
2350                         break;
2351                 cond_resched_lock(&hugetlb_lock);
2352         }
2353         while (count < persistent_huge_pages(h)) {
2354                 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2355                         break;
2356         }
2357 out:
2358         ret = persistent_huge_pages(h);
2359         spin_unlock(&hugetlb_lock);
2360         return ret;
2361 }
2362
2363 #define HSTATE_ATTR_RO(_name) \
2364         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2365
2366 #define HSTATE_ATTR(_name) \
2367         static struct kobj_attribute _name##_attr = \
2368                 __ATTR(_name, 0644, _name##_show, _name##_store)
2369
2370 static struct kobject *hugepages_kobj;
2371 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2372
2373 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2374
2375 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2376 {
2377         int i;
2378
2379         for (i = 0; i < HUGE_MAX_HSTATE; i++)
2380                 if (hstate_kobjs[i] == kobj) {
2381                         if (nidp)
2382                                 *nidp = NUMA_NO_NODE;
2383                         return &hstates[i];
2384                 }
2385
2386         return kobj_to_node_hstate(kobj, nidp);
2387 }
2388
2389 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2390                                         struct kobj_attribute *attr, char *buf)
2391 {
2392         struct hstate *h;
2393         unsigned long nr_huge_pages;
2394         int nid;
2395
2396         h = kobj_to_hstate(kobj, &nid);
2397         if (nid == NUMA_NO_NODE)
2398                 nr_huge_pages = h->nr_huge_pages;
2399         else
2400                 nr_huge_pages = h->nr_huge_pages_node[nid];
2401
2402         return sprintf(buf, "%lu\n", nr_huge_pages);
2403 }
2404
2405 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2406                                            struct hstate *h, int nid,
2407                                            unsigned long count, size_t len)
2408 {
2409         int err;
2410         NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
2411
2412         if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
2413                 err = -EINVAL;
2414                 goto out;
2415         }
2416
2417         if (nid == NUMA_NO_NODE) {
2418                 /*
2419                  * global hstate attribute
2420                  */
2421                 if (!(obey_mempolicy &&
2422                                 init_nodemask_of_mempolicy(nodes_allowed))) {
2423                         NODEMASK_FREE(nodes_allowed);
2424                         nodes_allowed = &node_states[N_MEMORY];
2425                 }
2426         } else if (nodes_allowed) {
2427                 /*
2428                  * per node hstate attribute: adjust count to global,
2429                  * but restrict alloc/free to the specified node.
2430                  */
2431                 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2432                 init_nodemask_of_node(nodes_allowed, nid);
2433         } else
2434                 nodes_allowed = &node_states[N_MEMORY];
2435
2436         h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
2437
2438         if (nodes_allowed != &node_states[N_MEMORY])
2439                 NODEMASK_FREE(nodes_allowed);
2440
2441         return len;
2442 out:
2443         NODEMASK_FREE(nodes_allowed);
2444         return err;
2445 }
2446
2447 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2448                                          struct kobject *kobj, const char *buf,
2449                                          size_t len)
2450 {
2451         struct hstate *h;
2452         unsigned long count;
2453         int nid;
2454         int err;
2455
2456         err = kstrtoul(buf, 10, &count);
2457         if (err)
2458                 return err;
2459
2460         h = kobj_to_hstate(kobj, &nid);
2461         return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2462 }
2463
2464 static ssize_t nr_hugepages_show(struct kobject *kobj,
2465                                        struct kobj_attribute *attr, char *buf)
2466 {
2467         return nr_hugepages_show_common(kobj, attr, buf);
2468 }
2469
2470 static ssize_t nr_hugepages_store(struct kobject *kobj,
2471                struct kobj_attribute *attr, const char *buf, size_t len)
2472 {
2473         return nr_hugepages_store_common(false, kobj, buf, len);
2474 }
2475 HSTATE_ATTR(nr_hugepages);
2476
2477 #ifdef CONFIG_NUMA
2478
2479 /*
2480  * hstate attribute for optionally mempolicy-based constraint on persistent
2481  * huge page alloc/free.
2482  */
2483 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2484                                        struct kobj_attribute *attr, char *buf)
2485 {
2486         return nr_hugepages_show_common(kobj, attr, buf);
2487 }
2488
2489 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2490                struct kobj_attribute *attr, const char *buf, size_t len)
2491 {
2492         return nr_hugepages_store_common(true, kobj, buf, len);
2493 }
2494 HSTATE_ATTR(nr_hugepages_mempolicy);
2495 #endif
2496
2497
2498 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2499                                         struct kobj_attribute *attr, char *buf)
2500 {
2501         struct hstate *h = kobj_to_hstate(kobj, NULL);
2502         return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2503 }
2504
2505 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2506                 struct kobj_attribute *attr, const char *buf, size_t count)
2507 {
2508         int err;
2509         unsigned long input;
2510         struct hstate *h = kobj_to_hstate(kobj, NULL);
2511
2512         if (hstate_is_gigantic(h))
2513                 return -EINVAL;
2514
2515         err = kstrtoul(buf, 10, &input);
2516         if (err)
2517                 return err;
2518
2519         spin_lock(&hugetlb_lock);
2520         h->nr_overcommit_huge_pages = input;
2521         spin_unlock(&hugetlb_lock);
2522
2523         return count;
2524 }
2525 HSTATE_ATTR(nr_overcommit_hugepages);
2526
2527 static ssize_t free_hugepages_show(struct kobject *kobj,
2528                                         struct kobj_attribute *attr, char *buf)
2529 {
2530         struct hstate *h;
2531         unsigned long free_huge_pages;
2532         int nid;
2533
2534         h = kobj_to_hstate(kobj, &nid);
2535         if (nid == NUMA_NO_NODE)
2536                 free_huge_pages = h->free_huge_pages;
2537         else
2538                 free_huge_pages = h->free_huge_pages_node[nid];
2539
2540         return sprintf(buf, "%lu\n", free_huge_pages);
2541 }
2542 HSTATE_ATTR_RO(free_hugepages);
2543
2544 static ssize_t resv_hugepages_show(struct kobject *kobj,
2545                                         struct kobj_attribute *attr, char *buf)
2546 {
2547         struct hstate *h = kobj_to_hstate(kobj, NULL);
2548         return sprintf(buf, "%lu\n", h->resv_huge_pages);
2549 }
2550 HSTATE_ATTR_RO(resv_hugepages);
2551
2552 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2553                                         struct kobj_attribute *attr, char *buf)
2554 {
2555         struct hstate *h;
2556         unsigned long surplus_huge_pages;
2557         int nid;
2558
2559         h = kobj_to_hstate(kobj, &nid);
2560         if (nid == NUMA_NO_NODE)
2561                 surplus_huge_pages = h->surplus_huge_pages;
2562         else
2563                 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2564
2565         return sprintf(buf, "%lu\n", surplus_huge_pages);
2566 }
2567 HSTATE_ATTR_RO(surplus_hugepages);
2568
2569 static struct attribute *hstate_attrs[] = {
2570         &nr_hugepages_attr.attr,
2571         &nr_overcommit_hugepages_attr.attr,
2572         &free_hugepages_attr.attr,
2573         &resv_hugepages_attr.attr,
2574         &surplus_hugepages_attr.attr,
2575 #ifdef CONFIG_NUMA
2576         &nr_hugepages_mempolicy_attr.attr,
2577 #endif
2578         NULL,
2579 };
2580
2581 static const struct attribute_group hstate_attr_group = {
2582         .attrs = hstate_attrs,
2583 };
2584
2585 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2586                                     struct kobject **hstate_kobjs,
2587                                     const struct attribute_group *hstate_attr_group)
2588 {
2589         int retval;
2590         int hi = hstate_index(h);
2591
2592         hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2593         if (!hstate_kobjs[hi])
2594                 return -ENOMEM;
2595
2596         retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2597         if (retval)
2598                 kobject_put(hstate_kobjs[hi]);
2599
2600         return retval;
2601 }
2602
2603 static void __init hugetlb_sysfs_init(void)
2604 {
2605         struct hstate *h;
2606         int err;
2607
2608         hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2609         if (!hugepages_kobj)
2610                 return;
2611
2612         for_each_hstate(h) {
2613                 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2614                                          hstate_kobjs, &hstate_attr_group);
2615                 if (err)
2616                         pr_err("Hugetlb: Unable to add hstate %s", h->name);
2617         }
2618 }
2619
2620 #ifdef CONFIG_NUMA
2621
2622 /*
2623  * node_hstate/s - associate per node hstate attributes, via their kobjects,
2624  * with node devices in node_devices[] using a parallel array.  The array
2625  * index of a node device or _hstate == node id.
2626  * This is here to avoid any static dependency of the node device driver, in
2627  * the base kernel, on the hugetlb module.
2628  */
2629 struct node_hstate {
2630         struct kobject          *hugepages_kobj;
2631         struct kobject          *hstate_kobjs[HUGE_MAX_HSTATE];
2632 };
2633 static struct node_hstate node_hstates[MAX_NUMNODES];
2634
2635 /*
2636  * A subset of global hstate attributes for node devices
2637  */
2638 static struct attribute *per_node_hstate_attrs[] = {
2639         &nr_hugepages_attr.attr,
2640         &free_hugepages_attr.attr,
2641         &surplus_hugepages_attr.attr,
2642         NULL,
2643 };
2644
2645 static const struct attribute_group per_node_hstate_attr_group = {
2646         .attrs = per_node_hstate_attrs,
2647 };
2648
2649 /*
2650  * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2651  * Returns node id via non-NULL nidp.
2652  */
2653 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2654 {
2655         int nid;
2656
2657         for (nid = 0; nid < nr_node_ids; nid++) {
2658                 struct node_hstate *nhs = &node_hstates[nid];
2659                 int i;
2660                 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2661                         if (nhs->hstate_kobjs[i] == kobj) {
2662                                 if (nidp)
2663                                         *nidp = nid;
2664                                 return &hstates[i];
2665                         }
2666         }
2667
2668         BUG();
2669         return NULL;
2670 }
2671
2672 /*
2673  * Unregister hstate attributes from a single node device.
2674  * No-op if no hstate attributes attached.
2675  */
2676 static void hugetlb_unregister_node(struct node *node)
2677 {
2678         struct hstate *h;
2679         struct node_hstate *nhs = &node_hstates[node->dev.id];
2680
2681         if (!nhs->hugepages_kobj)
2682                 return;         /* no hstate attributes */
2683
2684         for_each_hstate(h) {
2685                 int idx = hstate_index(h);
2686                 if (nhs->hstate_kobjs[idx]) {
2687                         kobject_put(nhs->hstate_kobjs[idx]);
2688                         nhs->hstate_kobjs[idx] = NULL;
2689                 }
2690         }
2691
2692         kobject_put(nhs->hugepages_kobj);
2693         nhs->hugepages_kobj = NULL;
2694 }
2695
2696
2697 /*
2698  * Register hstate attributes for a single node device.
2699  * No-op if attributes already registered.
2700  */
2701 static void hugetlb_register_node(struct node *node)
2702 {
2703         struct hstate *h;
2704         struct node_hstate *nhs = &node_hstates[node->dev.id];
2705         int err;
2706
2707         if (nhs->hugepages_kobj)
2708                 return;         /* already allocated */
2709
2710         nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2711                                                         &node->dev.kobj);
2712         if (!nhs->hugepages_kobj)
2713                 return;
2714
2715         for_each_hstate(h) {
2716                 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2717                                                 nhs->hstate_kobjs,
2718                                                 &per_node_hstate_attr_group);
2719                 if (err) {
2720                         pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2721                                 h->name, node->dev.id);
2722                         hugetlb_unregister_node(node);
2723                         break;
2724                 }
2725         }
2726 }
2727
2728 /*
2729  * hugetlb init time:  register hstate attributes for all registered node
2730  * devices of nodes that have memory.  All on-line nodes should have
2731  * registered their associated device by this time.
2732  */
2733 static void __init hugetlb_register_all_nodes(void)
2734 {
2735         int nid;
2736
2737         for_each_node_state(nid, N_MEMORY) {
2738                 struct node *node = node_devices[nid];
2739                 if (node->dev.id == nid)
2740                         hugetlb_register_node(node);
2741         }
2742
2743         /*
2744          * Let the node device driver know we're here so it can
2745          * [un]register hstate attributes on node hotplug.
2746          */
2747         register_hugetlbfs_with_node(hugetlb_register_node,
2748                                      hugetlb_unregister_node);
2749 }
2750 #else   /* !CONFIG_NUMA */
2751
2752 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2753 {
2754         BUG();
2755         if (nidp)
2756                 *nidp = -1;
2757         return NULL;
2758 }
2759
2760 static void hugetlb_register_all_nodes(void) { }
2761
2762 #endif
2763
2764 static int __init hugetlb_init(void)
2765 {
2766         int i;
2767
2768         if (!hugepages_supported())
2769                 return 0;
2770
2771         if (!size_to_hstate(default_hstate_size)) {
2772                 if (default_hstate_size != 0) {
2773                         pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n",
2774                                default_hstate_size, HPAGE_SIZE);
2775                 }
2776
2777                 default_hstate_size = HPAGE_SIZE;
2778                 if (!size_to_hstate(default_hstate_size))
2779                         hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2780         }
2781         default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2782         if (default_hstate_max_huge_pages) {
2783                 if (!default_hstate.max_huge_pages)
2784                         default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2785         }
2786
2787         hugetlb_init_hstates();
2788         gather_bootmem_prealloc();
2789         report_hugepages();
2790
2791         hugetlb_sysfs_init();
2792         hugetlb_register_all_nodes();
2793         hugetlb_cgroup_file_init();
2794
2795 #ifdef CONFIG_SMP
2796         num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2797 #else
2798         num_fault_mutexes = 1;
2799 #endif
2800         hugetlb_fault_mutex_table =
2801                 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
2802                               GFP_KERNEL);
2803         BUG_ON(!hugetlb_fault_mutex_table);
2804
2805         for (i = 0; i < num_fault_mutexes; i++)
2806                 mutex_init(&hugetlb_fault_mutex_table[i]);
2807         return 0;
2808 }
2809 subsys_initcall(hugetlb_init);
2810
2811 /* Should be called on processing a hugepagesz=... option */
2812 void __init hugetlb_bad_size(void)
2813 {
2814         parsed_valid_hugepagesz = false;
2815 }
2816
2817 void __init hugetlb_add_hstate(unsigned int order)
2818 {
2819         struct hstate *h;
2820         unsigned long i;
2821
2822         if (size_to_hstate(PAGE_SIZE << order)) {
2823                 pr_warn("hugepagesz= specified twice, ignoring\n");
2824                 return;
2825         }
2826         BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2827         BUG_ON(order == 0);
2828         h = &hstates[hugetlb_max_hstate++];
2829         h->order = order;
2830         h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2831         h->nr_huge_pages = 0;
2832         h->free_huge_pages = 0;
2833         for (i = 0; i < MAX_NUMNODES; ++i)
2834                 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2835         INIT_LIST_HEAD(&h->hugepage_activelist);
2836         h->next_nid_to_alloc = first_memory_node;
2837         h->next_nid_to_free = first_memory_node;
2838         snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2839                                         huge_page_size(h)/1024);
2840
2841         parsed_hstate = h;
2842 }
2843
2844 static int __init hugetlb_nrpages_setup(char *s)
2845 {
2846         unsigned long *mhp;
2847         static unsigned long *last_mhp;
2848
2849         if (!parsed_valid_hugepagesz) {
2850                 pr_warn("hugepages = %s preceded by "
2851                         "an unsupported hugepagesz, ignoring\n", s);
2852                 parsed_valid_hugepagesz = true;
2853                 return 1;
2854         }
2855         /*
2856          * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2857          * so this hugepages= parameter goes to the "default hstate".
2858          */
2859         else if (!hugetlb_max_hstate)
2860                 mhp = &default_hstate_max_huge_pages;
2861         else
2862                 mhp = &parsed_hstate->max_huge_pages;
2863
2864         if (mhp == last_mhp) {
2865                 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
2866                 return 1;
2867         }
2868
2869         if (sscanf(s, "%lu", mhp) <= 0)
2870                 *mhp = 0;
2871
2872         /*
2873          * Global state is always initialized later in hugetlb_init.
2874          * But we need to allocate >= MAX_ORDER hstates here early to still
2875          * use the bootmem allocator.
2876          */
2877         if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2878                 hugetlb_hstate_alloc_pages(parsed_hstate);
2879
2880         last_mhp = mhp;
2881
2882         return 1;
2883 }
2884 __setup("hugepages=", hugetlb_nrpages_setup);
2885
2886 static int __init hugetlb_default_setup(char *s)
2887 {
2888         default_hstate_size = memparse(s, &s);
2889         return 1;
2890 }
2891 __setup("default_hugepagesz=", hugetlb_default_setup);
2892
2893 static unsigned int cpuset_mems_nr(unsigned int *array)
2894 {
2895         int node;
2896         unsigned int nr = 0;
2897
2898         for_each_node_mask(node, cpuset_current_mems_allowed)
2899                 nr += array[node];
2900
2901         return nr;
2902 }
2903
2904 #ifdef CONFIG_SYSCTL
2905 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2906                          struct ctl_table *table, int write,
2907                          void __user *buffer, size_t *length, loff_t *ppos)
2908 {
2909         struct hstate *h = &default_hstate;
2910         unsigned long tmp = h->max_huge_pages;
2911         int ret;
2912
2913         if (!hugepages_supported())
2914                 return -EOPNOTSUPP;
2915
2916         table->data = &tmp;
2917         table->maxlen = sizeof(unsigned long);
2918         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2919         if (ret)
2920                 goto out;
2921
2922         if (write)
2923                 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2924                                                   NUMA_NO_NODE, tmp, *length);
2925 out:
2926         return ret;
2927 }
2928
2929 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2930                           void __user *buffer, size_t *length, loff_t *ppos)
2931 {
2932
2933         return hugetlb_sysctl_handler_common(false, table, write,
2934                                                         buffer, length, ppos);
2935 }
2936
2937 #ifdef CONFIG_NUMA
2938 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2939                           void __user *buffer, size_t *length, loff_t *ppos)
2940 {
2941         return hugetlb_sysctl_handler_common(true, table, write,
2942                                                         buffer, length, ppos);
2943 }
2944 #endif /* CONFIG_NUMA */
2945
2946 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2947                         void __user *buffer,
2948                         size_t *length, loff_t *ppos)
2949 {
2950         struct hstate *h = &default_hstate;
2951         unsigned long tmp;
2952         int ret;
2953
2954         if (!hugepages_supported())
2955                 return -EOPNOTSUPP;
2956
2957         tmp = h->nr_overcommit_huge_pages;
2958
2959         if (write && hstate_is_gigantic(h))
2960                 return -EINVAL;
2961
2962         table->data = &tmp;
2963         table->maxlen = sizeof(unsigned long);
2964         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2965         if (ret)
2966                 goto out;
2967
2968         if (write) {
2969                 spin_lock(&hugetlb_lock);
2970                 h->nr_overcommit_huge_pages = tmp;
2971                 spin_unlock(&hugetlb_lock);
2972         }
2973 out:
2974         return ret;
2975 }
2976
2977 #endif /* CONFIG_SYSCTL */
2978
2979 void hugetlb_report_meminfo(struct seq_file *m)
2980 {
2981         struct hstate *h;
2982         unsigned long total = 0;
2983
2984         if (!hugepages_supported())
2985                 return;
2986
2987         for_each_hstate(h) {
2988                 unsigned long count = h->nr_huge_pages;
2989
2990                 total += (PAGE_SIZE << huge_page_order(h)) * count;
2991
2992                 if (h == &default_hstate)
2993                         seq_printf(m,
2994                                    "HugePages_Total:   %5lu\n"
2995                                    "HugePages_Free:    %5lu\n"
2996                                    "HugePages_Rsvd:    %5lu\n"
2997                                    "HugePages_Surp:    %5lu\n"
2998                                    "Hugepagesize:   %8lu kB\n",
2999                                    count,
3000                                    h->free_huge_pages,
3001                                    h->resv_huge_pages,
3002                                    h->surplus_huge_pages,
3003                                    (PAGE_SIZE << huge_page_order(h)) / 1024);
3004         }
3005
3006         seq_printf(m, "Hugetlb:        %8lu kB\n", total / 1024);
3007 }
3008
3009 int hugetlb_report_node_meminfo(int nid, char *buf)
3010 {
3011         struct hstate *h = &default_hstate;
3012         if (!hugepages_supported())
3013                 return 0;
3014         return sprintf(buf,
3015                 "Node %d HugePages_Total: %5u\n"
3016                 "Node %d HugePages_Free:  %5u\n"
3017                 "Node %d HugePages_Surp:  %5u\n",
3018                 nid, h->nr_huge_pages_node[nid],
3019                 nid, h->free_huge_pages_node[nid],
3020                 nid, h->surplus_huge_pages_node[nid]);
3021 }
3022
3023 void hugetlb_show_meminfo(void)
3024 {
3025         struct hstate *h;
3026         int nid;
3027
3028         if (!hugepages_supported())
3029                 return;
3030
3031         for_each_node_state(nid, N_MEMORY)
3032                 for_each_hstate(h)
3033                         pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3034                                 nid,
3035                                 h->nr_huge_pages_node[nid],
3036                                 h->free_huge_pages_node[nid],
3037                                 h->surplus_huge_pages_node[nid],
3038                                 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3039 }
3040
3041 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3042 {
3043         seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3044                    atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3045 }
3046
3047 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3048 unsigned long hugetlb_total_pages(void)
3049 {
3050         struct hstate *h;
3051         unsigned long nr_total_pages = 0;
3052
3053         for_each_hstate(h)
3054                 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3055         return nr_total_pages;
3056 }
3057
3058 static int hugetlb_acct_memory(struct hstate *h, long delta)
3059 {
3060         int ret = -ENOMEM;
3061
3062         spin_lock(&hugetlb_lock);
3063         /*
3064          * When cpuset is configured, it breaks the strict hugetlb page
3065          * reservation as the accounting is done on a global variable. Such
3066          * reservation is completely rubbish in the presence of cpuset because
3067          * the reservation is not checked against page availability for the
3068          * current cpuset. Application can still potentially OOM'ed by kernel
3069          * with lack of free htlb page in cpuset that the task is in.
3070          * Attempt to enforce strict accounting with cpuset is almost
3071          * impossible (or too ugly) because cpuset is too fluid that
3072          * task or memory node can be dynamically moved between cpusets.
3073          *
3074          * The change of semantics for shared hugetlb mapping with cpuset is
3075          * undesirable. However, in order to preserve some of the semantics,
3076          * we fall back to check against current free page availability as
3077          * a best attempt and hopefully to minimize the impact of changing
3078          * semantics that cpuset has.
3079          */
3080         if (delta > 0) {
3081                 if (gather_surplus_pages(h, delta) < 0)
3082                         goto out;
3083
3084                 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
3085                         return_unused_surplus_pages(h, delta);
3086                         goto out;
3087                 }
3088         }
3089
3090         ret = 0;
3091         if (delta < 0)
3092                 return_unused_surplus_pages(h, (unsigned long) -delta);
3093
3094 out:
3095         spin_unlock(&hugetlb_lock);
3096         return ret;
3097 }
3098
3099 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3100 {
3101         struct resv_map *resv = vma_resv_map(vma);
3102
3103         /*
3104          * This new VMA should share its siblings reservation map if present.
3105          * The VMA will only ever have a valid reservation map pointer where
3106          * it is being copied for another still existing VMA.  As that VMA
3107          * has a reference to the reservation map it cannot disappear until
3108          * after this open call completes.  It is therefore safe to take a
3109          * new reference here without additional locking.
3110          */
3111         if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3112                 kref_get(&resv->refs);
3113 }
3114
3115 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3116 {
3117         struct hstate *h = hstate_vma(vma);
3118         struct resv_map *resv = vma_resv_map(vma);
3119         struct hugepage_subpool *spool = subpool_vma(vma);
3120         unsigned long reserve, start, end;
3121         long gbl_reserve;
3122
3123         if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3124                 return;
3125
3126         start = vma_hugecache_offset(h, vma, vma->vm_start);
3127         end = vma_hugecache_offset(h, vma, vma->vm_end);
3128
3129         reserve = (end - start) - region_count(resv, start, end);
3130
3131         kref_put(&resv->refs, resv_map_release);
3132
3133         if (reserve) {
3134                 /*
3135                  * Decrement reserve counts.  The global reserve count may be
3136                  * adjusted if the subpool has a minimum size.
3137                  */
3138                 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3139                 hugetlb_acct_memory(h, -gbl_reserve);
3140         }
3141 }
3142
3143 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3144 {
3145         if (addr & ~(huge_page_mask(hstate_vma(vma))))
3146                 return -EINVAL;
3147         return 0;
3148 }
3149
3150 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
3151 {
3152         struct hstate *hstate = hstate_vma(vma);
3153
3154         return 1UL << huge_page_shift(hstate);
3155 }
3156
3157 /*
3158  * We cannot handle pagefaults against hugetlb pages at all.  They cause
3159  * handle_mm_fault() to try to instantiate regular-sized pages in the
3160  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
3161  * this far.
3162  */
3163 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
3164 {
3165         BUG();
3166         return 0;
3167 }
3168
3169 const struct vm_operations_struct hugetlb_vm_ops = {
3170         .fault = hugetlb_vm_op_fault,
3171         .open = hugetlb_vm_op_open,
3172         .close = hugetlb_vm_op_close,
3173         .split = hugetlb_vm_op_split,
3174         .pagesize = hugetlb_vm_op_pagesize,
3175 };
3176
3177 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3178                                 int writable)
3179 {
3180         pte_t entry;
3181
3182         if (writable) {
3183                 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3184                                          vma->vm_page_prot)));
3185         } else {
3186                 entry = huge_pte_wrprotect(mk_huge_pte(page,
3187                                            vma->vm_page_prot));
3188         }
3189         entry = pte_mkyoung(entry);
3190         entry = pte_mkhuge(entry);
3191         entry = arch_make_huge_pte(entry, vma, page, writable);
3192
3193         return entry;
3194 }
3195
3196 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3197                                    unsigned long address, pte_t *ptep)
3198 {
3199         pte_t entry;
3200
3201         entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3202         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3203                 update_mmu_cache(vma, address, ptep);
3204 }
3205
3206 bool is_hugetlb_entry_migration(pte_t pte)
3207 {
3208         swp_entry_t swp;
3209
3210         if (huge_pte_none(pte) || pte_present(pte))
3211                 return false;
3212         swp = pte_to_swp_entry(pte);
3213         if (non_swap_entry(swp) && is_migration_entry(swp))
3214                 return true;
3215         else
3216                 return false;
3217 }
3218
3219 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3220 {
3221         swp_entry_t swp;
3222
3223         if (huge_pte_none(pte) || pte_present(pte))
3224                 return 0;
3225         swp = pte_to_swp_entry(pte);
3226         if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3227                 return 1;
3228         else
3229                 return 0;
3230 }
3231
3232 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3233                             struct vm_area_struct *vma)
3234 {
3235         pte_t *src_pte, *dst_pte, entry;
3236         struct page *ptepage;
3237         unsigned long addr;
3238         int cow;
3239         struct hstate *h = hstate_vma(vma);
3240         unsigned long sz = huge_page_size(h);
3241         unsigned long mmun_start;       /* For mmu_notifiers */
3242         unsigned long mmun_end;         /* For mmu_notifiers */
3243         int ret = 0;
3244
3245         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3246
3247         mmun_start = vma->vm_start;
3248         mmun_end = vma->vm_end;
3249         if (cow)
3250                 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
3251
3252         for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3253                 spinlock_t *src_ptl, *dst_ptl;
3254                 src_pte = huge_pte_offset(src, addr, sz);
3255                 if (!src_pte)
3256                         continue;
3257                 dst_pte = huge_pte_alloc(dst, addr, sz);
3258                 if (!dst_pte) {
3259                         ret = -ENOMEM;
3260                         break;
3261                 }
3262
3263                 /* If the pagetables are shared don't copy or take references */
3264                 if (dst_pte == src_pte)
3265                         continue;
3266
3267                 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3268                 src_ptl = huge_pte_lockptr(h, src, src_pte);
3269                 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3270                 entry = huge_ptep_get(src_pte);
3271                 if (huge_pte_none(entry)) { /* skip none entry */
3272                         ;
3273                 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3274                                     is_hugetlb_entry_hwpoisoned(entry))) {
3275                         swp_entry_t swp_entry = pte_to_swp_entry(entry);
3276
3277                         if (is_write_migration_entry(swp_entry) && cow) {
3278                                 /*
3279                                  * COW mappings require pages in both
3280                                  * parent and child to be set to read.
3281                                  */
3282                                 make_migration_entry_read(&swp_entry);
3283                                 entry = swp_entry_to_pte(swp_entry);
3284                                 set_huge_swap_pte_at(src, addr, src_pte,
3285                                                      entry, sz);
3286                         }
3287                         set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3288                 } else {
3289                         if (cow) {
3290                                 /*
3291                                  * No need to notify as we are downgrading page
3292                                  * table protection not changing it to point
3293                                  * to a new page.
3294                                  *
3295                                  * See Documentation/vm/mmu_notifier.rst
3296                                  */
3297                                 huge_ptep_set_wrprotect(src, addr, src_pte);
3298                         }
3299                         entry = huge_ptep_get(src_pte);
3300                         ptepage = pte_page(entry);
3301                         get_page(ptepage);
3302                         page_dup_rmap(ptepage, true);
3303                         set_huge_pte_at(dst, addr, dst_pte, entry);
3304                         hugetlb_count_add(pages_per_huge_page(h), dst);
3305                 }
3306                 spin_unlock(src_ptl);
3307                 spin_unlock(dst_ptl);
3308         }
3309
3310         if (cow)
3311                 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
3312
3313         return ret;
3314 }
3315
3316 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3317                             unsigned long start, unsigned long end,
3318                             struct page *ref_page)
3319 {
3320         struct mm_struct *mm = vma->vm_mm;
3321         unsigned long address;
3322         pte_t *ptep;
3323         pte_t pte;
3324         spinlock_t *ptl;
3325         struct page *page;
3326         struct hstate *h = hstate_vma(vma);
3327         unsigned long sz = huge_page_size(h);
3328         const unsigned long mmun_start = start; /* For mmu_notifiers */
3329         const unsigned long mmun_end   = end;   /* For mmu_notifiers */
3330
3331         WARN_ON(!is_vm_hugetlb_page(vma));
3332         BUG_ON(start & ~huge_page_mask(h));
3333         BUG_ON(end & ~huge_page_mask(h));
3334
3335         /*
3336          * This is a hugetlb vma, all the pte entries should point
3337          * to huge page.
3338          */
3339         tlb_remove_check_page_size_change(tlb, sz);
3340         tlb_start_vma(tlb, vma);
3341         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3342         address = start;
3343         for (; address < end; address += sz) {
3344                 ptep = huge_pte_offset(mm, address, sz);
3345                 if (!ptep)
3346                         continue;
3347
3348                 ptl = huge_pte_lock(h, mm, ptep);
3349                 if (huge_pmd_unshare(mm, &address, ptep)) {
3350                         spin_unlock(ptl);
3351                         continue;
3352                 }
3353
3354                 pte = huge_ptep_get(ptep);
3355                 if (huge_pte_none(pte)) {
3356                         spin_unlock(ptl);
3357                         continue;
3358                 }
3359
3360                 /*
3361                  * Migrating hugepage or HWPoisoned hugepage is already
3362                  * unmapped and its refcount is dropped, so just clear pte here.
3363                  */
3364                 if (unlikely(!pte_present(pte))) {
3365                         huge_pte_clear(mm, address, ptep, sz);
3366                         spin_unlock(ptl);
3367                         continue;
3368                 }
3369
3370                 page = pte_page(pte);
3371                 /*
3372                  * If a reference page is supplied, it is because a specific
3373                  * page is being unmapped, not a range. Ensure the page we
3374                  * are about to unmap is the actual page of interest.
3375                  */
3376                 if (ref_page) {
3377                         if (page != ref_page) {
3378                                 spin_unlock(ptl);
3379                                 continue;
3380                         }
3381                         /*
3382                          * Mark the VMA as having unmapped its page so that
3383                          * future faults in this VMA will fail rather than
3384                          * looking like data was lost
3385                          */
3386                         set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3387                 }
3388
3389                 pte = huge_ptep_get_and_clear(mm, address, ptep);
3390                 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
3391                 if (huge_pte_dirty(pte))
3392                         set_page_dirty(page);
3393
3394                 hugetlb_count_sub(pages_per_huge_page(h), mm);
3395                 page_remove_rmap(page, true);
3396
3397                 spin_unlock(ptl);
3398                 tlb_remove_page_size(tlb, page, huge_page_size(h));
3399                 /*
3400                  * Bail out after unmapping reference page if supplied
3401                  */
3402                 if (ref_page)
3403                         break;
3404         }
3405         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3406         tlb_end_vma(tlb, vma);
3407 }
3408
3409 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3410                           struct vm_area_struct *vma, unsigned long start,
3411                           unsigned long end, struct page *ref_page)
3412 {
3413         __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3414
3415         /*
3416          * Clear this flag so that x86's huge_pmd_share page_table_shareable
3417          * test will fail on a vma being torn down, and not grab a page table
3418          * on its way out.  We're lucky that the flag has such an appropriate
3419          * name, and can in fact be safely cleared here. We could clear it
3420          * before the __unmap_hugepage_range above, but all that's necessary
3421          * is to clear it before releasing the i_mmap_rwsem. This works
3422          * because in the context this is called, the VMA is about to be
3423          * destroyed and the i_mmap_rwsem is held.
3424          */
3425         vma->vm_flags &= ~VM_MAYSHARE;
3426 }
3427
3428 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3429                           unsigned long end, struct page *ref_page)
3430 {
3431         struct mm_struct *mm;
3432         struct mmu_gather tlb;
3433
3434         mm = vma->vm_mm;
3435
3436         tlb_gather_mmu(&tlb, mm, start, end);
3437         __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3438         tlb_finish_mmu(&tlb, start, end);
3439 }
3440
3441 /*
3442  * This is called when the original mapper is failing to COW a MAP_PRIVATE
3443  * mappping it owns the reserve page for. The intention is to unmap the page
3444  * from other VMAs and let the children be SIGKILLed if they are faulting the
3445  * same region.
3446  */
3447 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3448                               struct page *page, unsigned long address)
3449 {
3450         struct hstate *h = hstate_vma(vma);
3451         struct vm_area_struct *iter_vma;
3452         struct address_space *mapping;
3453         pgoff_t pgoff;
3454
3455         /*
3456          * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3457          * from page cache lookup which is in HPAGE_SIZE units.
3458          */
3459         address = address & huge_page_mask(h);
3460         pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3461                         vma->vm_pgoff;
3462         mapping = vma->vm_file->f_mapping;
3463
3464         /*
3465          * Take the mapping lock for the duration of the table walk. As
3466          * this mapping should be shared between all the VMAs,
3467          * __unmap_hugepage_range() is called as the lock is already held
3468          */
3469         i_mmap_lock_write(mapping);
3470         vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3471                 /* Do not unmap the current VMA */
3472                 if (iter_vma == vma)
3473                         continue;
3474
3475                 /*
3476                  * Shared VMAs have their own reserves and do not affect
3477                  * MAP_PRIVATE accounting but it is possible that a shared
3478                  * VMA is using the same page so check and skip such VMAs.
3479                  */
3480                 if (iter_vma->vm_flags & VM_MAYSHARE)
3481                         continue;
3482
3483                 /*
3484                  * Unmap the page from other VMAs without their own reserves.
3485                  * They get marked to be SIGKILLed if they fault in these
3486                  * areas. This is because a future no-page fault on this VMA
3487                  * could insert a zeroed page instead of the data existing
3488                  * from the time of fork. This would look like data corruption
3489                  */
3490                 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3491                         unmap_hugepage_range(iter_vma, address,
3492                                              address + huge_page_size(h), page);
3493         }
3494         i_mmap_unlock_write(mapping);
3495 }
3496
3497 /*
3498  * Hugetlb_cow() should be called with page lock of the original hugepage held.
3499  * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3500  * cannot race with other handlers or page migration.
3501  * Keep the pte_same checks anyway to make transition from the mutex easier.
3502  */
3503 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3504                        unsigned long address, pte_t *ptep,
3505                        struct page *pagecache_page, spinlock_t *ptl)
3506 {
3507         pte_t pte;
3508         struct hstate *h = hstate_vma(vma);
3509         struct page *old_page, *new_page;
3510         int ret = 0, outside_reserve = 0;
3511         unsigned long mmun_start;       /* For mmu_notifiers */
3512         unsigned long mmun_end;         /* For mmu_notifiers */
3513
3514         pte = huge_ptep_get(ptep);
3515         old_page = pte_page(pte);
3516
3517 retry_avoidcopy:
3518         /* If no-one else is actually using this page, avoid the copy
3519          * and just make the page writable */
3520         if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3521                 page_move_anon_rmap(old_page, vma);
3522                 set_huge_ptep_writable(vma, address, ptep);
3523                 return 0;
3524         }
3525
3526         /*
3527          * If the process that created a MAP_PRIVATE mapping is about to
3528          * perform a COW due to a shared page count, attempt to satisfy
3529          * the allocation without using the existing reserves. The pagecache
3530          * page is used to determine if the reserve at this address was
3531          * consumed or not. If reserves were used, a partial faulted mapping
3532          * at the time of fork() could consume its reserves on COW instead
3533          * of the full address range.
3534          */
3535         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3536                         old_page != pagecache_page)
3537                 outside_reserve = 1;
3538
3539         get_page(old_page);
3540
3541         /*
3542          * Drop page table lock as buddy allocator may be called. It will
3543          * be acquired again before returning to the caller, as expected.
3544          */
3545         spin_unlock(ptl);
3546         new_page = alloc_huge_page(vma, address, outside_reserve);
3547
3548         if (IS_ERR(new_page)) {
3549                 /*
3550                  * If a process owning a MAP_PRIVATE mapping fails to COW,
3551                  * it is due to references held by a child and an insufficient
3552                  * huge page pool. To guarantee the original mappers
3553                  * reliability, unmap the page from child processes. The child
3554                  * may get SIGKILLed if it later faults.
3555                  */
3556                 if (outside_reserve) {
3557                         put_page(old_page);
3558                         BUG_ON(huge_pte_none(pte));
3559                         unmap_ref_private(mm, vma, old_page, address);
3560                         BUG_ON(huge_pte_none(pte));
3561                         spin_lock(ptl);
3562                         ptep = huge_pte_offset(mm, address & huge_page_mask(h),
3563                                                huge_page_size(h));
3564                         if (likely(ptep &&
3565                                    pte_same(huge_ptep_get(ptep), pte)))
3566                                 goto retry_avoidcopy;
3567                         /*
3568                          * race occurs while re-acquiring page table
3569                          * lock, and our job is done.
3570                          */
3571                         return 0;
3572                 }
3573
3574                 ret = (PTR_ERR(new_page) == -ENOMEM) ?
3575                         VM_FAULT_OOM : VM_FAULT_SIGBUS;
3576                 goto out_release_old;
3577         }
3578
3579         /*
3580          * When the original hugepage is shared one, it does not have
3581          * anon_vma prepared.
3582          */
3583         if (unlikely(anon_vma_prepare(vma))) {
3584                 ret = VM_FAULT_OOM;
3585                 goto out_release_all;
3586         }
3587
3588         copy_user_huge_page(new_page, old_page, address, vma,
3589                             pages_per_huge_page(h));
3590         __SetPageUptodate(new_page);
3591         set_page_huge_active(new_page);
3592
3593         mmun_start = address & huge_page_mask(h);
3594         mmun_end = mmun_start + huge_page_size(h);
3595         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3596
3597         /*
3598          * Retake the page table lock to check for racing updates
3599          * before the page tables are altered
3600          */
3601         spin_lock(ptl);
3602         ptep = huge_pte_offset(mm, address & huge_page_mask(h),
3603                                huge_page_size(h));
3604         if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3605                 ClearPagePrivate(new_page);
3606
3607                 /* Break COW */
3608                 huge_ptep_clear_flush(vma, address, ptep);
3609                 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
3610                 set_huge_pte_at(mm, address, ptep,
3611                                 make_huge_pte(vma, new_page, 1));
3612                 page_remove_rmap(old_page, true);
3613                 hugepage_add_new_anon_rmap(new_page, vma, address);
3614                 /* Make the old page be freed below */
3615                 new_page = old_page;
3616         }
3617         spin_unlock(ptl);
3618         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3619 out_release_all:
3620         restore_reserve_on_error(h, vma, address, new_page);
3621         put_page(new_page);
3622 out_release_old:
3623         put_page(old_page);
3624
3625         spin_lock(ptl); /* Caller expects lock to be held */
3626         return ret;
3627 }
3628
3629 /* Return the pagecache page at a given address within a VMA */
3630 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3631                         struct vm_area_struct *vma, unsigned long address)
3632 {
3633         struct address_space *mapping;
3634         pgoff_t idx;
3635
3636         mapping = vma->vm_file->f_mapping;
3637         idx = vma_hugecache_offset(h, vma, address);
3638
3639         return find_lock_page(mapping, idx);
3640 }
3641
3642 /*
3643  * Return whether there is a pagecache page to back given address within VMA.
3644  * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3645  */
3646 static bool hugetlbfs_pagecache_present(struct hstate *h,
3647                         struct vm_area_struct *vma, unsigned long address)
3648 {
3649         struct address_space *mapping;
3650         pgoff_t idx;
3651         struct page *page;
3652
3653         mapping = vma->vm_file->f_mapping;
3654         idx = vma_hugecache_offset(h, vma, address);
3655
3656         page = find_get_page(mapping, idx);
3657         if (page)
3658                 put_page(page);
3659         return page != NULL;
3660 }
3661
3662 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3663                            pgoff_t idx)
3664 {
3665         struct inode *inode = mapping->host;
3666         struct hstate *h = hstate_inode(inode);
3667         int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3668
3669         if (err)
3670                 return err;
3671         ClearPagePrivate(page);
3672
3673         spin_lock(&inode->i_lock);
3674         inode->i_blocks += blocks_per_huge_page(h);
3675         spin_unlock(&inode->i_lock);
3676         return 0;
3677 }
3678
3679 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
3680                            struct address_space *mapping, pgoff_t idx,
3681                            unsigned long address, pte_t *ptep, unsigned int flags)
3682 {
3683         struct hstate *h = hstate_vma(vma);
3684         int ret = VM_FAULT_SIGBUS;
3685         int anon_rmap = 0;
3686         unsigned long size;
3687         struct page *page;
3688         pte_t new_pte;
3689         spinlock_t *ptl;
3690         unsigned long haddr = address & huge_page_mask(h);
3691
3692         /*
3693          * Currently, we are forced to kill the process in the event the
3694          * original mapper has unmapped pages from the child due to a failed
3695          * COW. Warn that such a situation has occurred as it may not be obvious
3696          */
3697         if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3698                 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
3699                            current->pid);
3700                 return ret;
3701         }
3702
3703         /*
3704          * Use page lock to guard against racing truncation
3705          * before we get page_table_lock.
3706          */
3707 retry:
3708         page = find_lock_page(mapping, idx);
3709         if (!page) {
3710                 size = i_size_read(mapping->host) >> huge_page_shift(h);
3711                 if (idx >= size)
3712                         goto out;
3713
3714                 /*
3715                  * Check for page in userfault range
3716                  */
3717                 if (userfaultfd_missing(vma)) {
3718                         u32 hash;
3719                         struct vm_fault vmf = {
3720                                 .vma = vma,
3721                                 .address = haddr,
3722                                 .flags = flags,
3723                                 /*
3724                                  * Hard to debug if it ends up being
3725                                  * used by a callee that assumes
3726                                  * something about the other
3727                                  * uninitialized fields... same as in
3728                                  * memory.c
3729                                  */
3730                         };
3731
3732                         /*
3733                          * hugetlb_fault_mutex must be dropped before
3734                          * handling userfault.  Reacquire after handling
3735                          * fault to make calling code simpler.
3736                          */
3737                         hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping,
3738                                                         idx, haddr);
3739                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3740                         ret = handle_userfault(&vmf, VM_UFFD_MISSING);
3741                         mutex_lock(&hugetlb_fault_mutex_table[hash]);
3742                         goto out;
3743                 }
3744
3745                 page = alloc_huge_page(vma, haddr, 0);
3746                 if (IS_ERR(page)) {
3747                         ret = PTR_ERR(page);
3748                         if (ret == -ENOMEM)
3749                                 ret = VM_FAULT_OOM;
3750                         else
3751                                 ret = VM_FAULT_SIGBUS;
3752                         goto out;
3753                 }
3754                 clear_huge_page(page, address, pages_per_huge_page(h));
3755                 __SetPageUptodate(page);
3756                 set_page_huge_active(page);
3757
3758                 if (vma->vm_flags & VM_MAYSHARE) {
3759                         int err = huge_add_to_page_cache(page, mapping, idx);
3760                         if (err) {
3761                                 put_page(page);
3762                                 if (err == -EEXIST)
3763                                         goto retry;
3764                                 goto out;
3765                         }
3766                 } else {
3767                         lock_page(page);
3768                         if (unlikely(anon_vma_prepare(vma))) {
3769                                 ret = VM_FAULT_OOM;
3770                                 goto backout_unlocked;
3771                         }
3772                         anon_rmap = 1;
3773                 }
3774         } else {
3775                 /*
3776                  * If memory error occurs between mmap() and fault, some process
3777                  * don't have hwpoisoned swap entry for errored virtual address.
3778                  * So we need to block hugepage fault by PG_hwpoison bit check.
3779                  */
3780                 if (unlikely(PageHWPoison(page))) {
3781                         ret = VM_FAULT_HWPOISON |
3782                                 VM_FAULT_SET_HINDEX(hstate_index(h));
3783                         goto backout_unlocked;
3784                 }
3785         }
3786
3787         /*
3788          * If we are going to COW a private mapping later, we examine the
3789          * pending reservations for this page now. This will ensure that
3790          * any allocations necessary to record that reservation occur outside
3791          * the spinlock.
3792          */
3793         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3794                 if (vma_needs_reservation(h, vma, haddr) < 0) {
3795                         ret = VM_FAULT_OOM;
3796                         goto backout_unlocked;
3797                 }
3798                 /* Just decrements count, does not deallocate */
3799                 vma_end_reservation(h, vma, haddr);
3800         }
3801
3802         ptl = huge_pte_lock(h, mm, ptep);
3803         size = i_size_read(mapping->host) >> huge_page_shift(h);
3804         if (idx >= size)
3805                 goto backout;
3806
3807         ret = 0;
3808         if (!huge_pte_none(huge_ptep_get(ptep)))
3809                 goto backout;
3810
3811         if (anon_rmap) {
3812                 ClearPagePrivate(page);
3813                 hugepage_add_new_anon_rmap(page, vma, haddr);
3814         } else
3815                 page_dup_rmap(page, true);
3816         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3817                                 && (vma->vm_flags & VM_SHARED)));
3818         set_huge_pte_at(mm, haddr, ptep, new_pte);
3819
3820         hugetlb_count_add(pages_per_huge_page(h), mm);
3821         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3822                 /* Optimization, do the COW without a second fault */
3823                 ret = hugetlb_cow(mm, vma, haddr, ptep, page, ptl);
3824         }
3825
3826         spin_unlock(ptl);
3827         unlock_page(page);
3828 out:
3829         return ret;
3830
3831 backout:
3832         spin_unlock(ptl);
3833 backout_unlocked:
3834         unlock_page(page);
3835         restore_reserve_on_error(h, vma, haddr, page);
3836         put_page(page);
3837         goto out;
3838 }
3839
3840 #ifdef CONFIG_SMP
3841 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3842                             struct vm_area_struct *vma,
3843                             struct address_space *mapping,
3844                             pgoff_t idx, unsigned long address)
3845 {
3846         unsigned long key[2];
3847         u32 hash;
3848
3849         if (vma->vm_flags & VM_SHARED) {
3850                 key[0] = (unsigned long) mapping;
3851                 key[1] = idx;
3852         } else {
3853                 key[0] = (unsigned long) mm;
3854                 key[1] = address >> huge_page_shift(h);
3855         }
3856
3857         hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3858
3859         return hash & (num_fault_mutexes - 1);
3860 }
3861 #else
3862 /*
3863  * For uniprocesor systems we always use a single mutex, so just
3864  * return 0 and avoid the hashing overhead.
3865  */
3866 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3867                             struct vm_area_struct *vma,
3868                             struct address_space *mapping,
3869                             pgoff_t idx, unsigned long address)
3870 {
3871         return 0;
3872 }
3873 #endif
3874
3875 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3876                         unsigned long address, unsigned int flags)
3877 {
3878         pte_t *ptep, entry;
3879         spinlock_t *ptl;
3880         int ret;
3881         u32 hash;
3882         pgoff_t idx;
3883         struct page *page = NULL;
3884         struct page *pagecache_page = NULL;
3885         struct hstate *h = hstate_vma(vma);
3886         struct address_space *mapping;
3887         int need_wait_lock = 0;
3888         unsigned long haddr = address & huge_page_mask(h);
3889
3890         ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
3891         if (ptep) {
3892                 entry = huge_ptep_get(ptep);
3893                 if (unlikely(is_hugetlb_entry_migration(entry))) {
3894                         migration_entry_wait_huge(vma, mm, ptep);
3895                         return 0;
3896                 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
3897                         return VM_FAULT_HWPOISON_LARGE |
3898                                 VM_FAULT_SET_HINDEX(hstate_index(h));
3899         } else {
3900                 ptep = huge_pte_alloc(mm, haddr, huge_page_size(h));
3901                 if (!ptep)
3902                         return VM_FAULT_OOM;
3903         }
3904
3905         mapping = vma->vm_file->f_mapping;
3906         idx = vma_hugecache_offset(h, vma, haddr);
3907
3908         /*
3909          * Serialize hugepage allocation and instantiation, so that we don't
3910          * get spurious allocation failures if two CPUs race to instantiate
3911          * the same page in the page cache.
3912          */
3913         hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, haddr);
3914         mutex_lock(&hugetlb_fault_mutex_table[hash]);
3915
3916         entry = huge_ptep_get(ptep);
3917         if (huge_pte_none(entry)) {
3918                 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
3919                 goto out_mutex;
3920         }
3921
3922         ret = 0;
3923
3924         /*
3925          * entry could be a migration/hwpoison entry at this point, so this
3926          * check prevents the kernel from going below assuming that we have
3927          * a active hugepage in pagecache. This goto expects the 2nd page fault,
3928          * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3929          * handle it.
3930          */
3931         if (!pte_present(entry))
3932                 goto out_mutex;
3933
3934         /*
3935          * If we are going to COW the mapping later, we examine the pending
3936          * reservations for this page now. This will ensure that any
3937          * allocations necessary to record that reservation occur outside the
3938          * spinlock. For private mappings, we also lookup the pagecache
3939          * page now as it is used to determine if a reservation has been
3940          * consumed.
3941          */
3942         if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
3943                 if (vma_needs_reservation(h, vma, haddr) < 0) {
3944                         ret = VM_FAULT_OOM;
3945                         goto out_mutex;
3946                 }
3947                 /* Just decrements count, does not deallocate */
3948                 vma_end_reservation(h, vma, haddr);
3949
3950                 if (!(vma->vm_flags & VM_MAYSHARE))
3951                         pagecache_page = hugetlbfs_pagecache_page(h,
3952                                                                 vma, haddr);
3953         }
3954
3955         ptl = huge_pte_lock(h, mm, ptep);
3956
3957         /* Check for a racing update before calling hugetlb_cow */
3958         if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
3959                 goto out_ptl;
3960
3961         /*
3962          * hugetlb_cow() requires page locks of pte_page(entry) and
3963          * pagecache_page, so here we need take the former one
3964          * when page != pagecache_page or !pagecache_page.
3965          */
3966         page = pte_page(entry);
3967         if (page != pagecache_page)
3968                 if (!trylock_page(page)) {
3969                         need_wait_lock = 1;
3970                         goto out_ptl;
3971                 }
3972
3973         get_page(page);
3974
3975         if (flags & FAULT_FLAG_WRITE) {
3976                 if (!huge_pte_write(entry)) {
3977                         ret = hugetlb_cow(mm, vma, haddr, ptep,
3978                                           pagecache_page, ptl);
3979                         goto out_put_page;
3980                 }
3981                 entry = huge_pte_mkdirty(entry);
3982         }
3983         entry = pte_mkyoung(entry);
3984         if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
3985                                                 flags & FAULT_FLAG_WRITE))
3986                 update_mmu_cache(vma, haddr, ptep);
3987 out_put_page:
3988         if (page != pagecache_page)
3989                 unlock_page(page);
3990         put_page(page);
3991 out_ptl:
3992         spin_unlock(ptl);
3993
3994         if (pagecache_page) {
3995                 unlock_page(pagecache_page);
3996                 put_page(pagecache_page);
3997         }
3998 out_mutex:
3999         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4000         /*
4001          * Generally it's safe to hold refcount during waiting page lock. But
4002          * here we just wait to defer the next page fault to avoid busy loop and
4003          * the page is not used after unlocked before returning from the current
4004          * page fault. So we are safe from accessing freed page, even if we wait
4005          * here without taking refcount.
4006          */
4007         if (need_wait_lock)
4008                 wait_on_page_locked(page);
4009         return ret;
4010 }
4011
4012 /*
4013  * Used by userfaultfd UFFDIO_COPY.  Based on mcopy_atomic_pte with
4014  * modifications for huge pages.
4015  */
4016 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4017                             pte_t *dst_pte,
4018                             struct vm_area_struct *dst_vma,
4019                             unsigned long dst_addr,
4020                             unsigned long src_addr,
4021                             struct page **pagep)
4022 {
4023         struct address_space *mapping;
4024         pgoff_t idx;
4025         unsigned long size;
4026         int vm_shared = dst_vma->vm_flags & VM_SHARED;
4027         struct hstate *h = hstate_vma(dst_vma);
4028         pte_t _dst_pte;
4029         spinlock_t *ptl;
4030         int ret;
4031         struct page *page;
4032
4033         if (!*pagep) {
4034                 ret = -ENOMEM;
4035                 page = alloc_huge_page(dst_vma, dst_addr, 0);
4036                 if (IS_ERR(page))
4037                         goto out;
4038
4039                 ret = copy_huge_page_from_user(page,
4040                                                 (const void __user *) src_addr,
4041                                                 pages_per_huge_page(h), false);
4042
4043                 /* fallback to copy_from_user outside mmap_sem */
4044                 if (unlikely(ret)) {
4045                         ret = -EFAULT;
4046                         *pagep = page;
4047                         /* don't free the page */
4048                         goto out;
4049                 }
4050         } else {
4051                 page = *pagep;
4052                 *pagep = NULL;
4053         }
4054
4055         /*
4056          * The memory barrier inside __SetPageUptodate makes sure that
4057          * preceding stores to the page contents become visible before
4058          * the set_pte_at() write.
4059          */
4060         __SetPageUptodate(page);
4061         set_page_huge_active(page);
4062
4063         mapping = dst_vma->vm_file->f_mapping;
4064         idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4065
4066         /*
4067          * If shared, add to page cache
4068          */
4069         if (vm_shared) {
4070                 size = i_size_read(mapping->host) >> huge_page_shift(h);
4071                 ret = -EFAULT;
4072                 if (idx >= size)
4073                         goto out_release_nounlock;
4074
4075                 /*
4076                  * Serialization between remove_inode_hugepages() and
4077                  * huge_add_to_page_cache() below happens through the
4078                  * hugetlb_fault_mutex_table that here must be hold by
4079                  * the caller.
4080                  */
4081                 ret = huge_add_to_page_cache(page, mapping, idx);
4082                 if (ret)
4083                         goto out_release_nounlock;
4084         }
4085
4086         ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4087         spin_lock(ptl);
4088
4089         /*
4090          * Recheck the i_size after holding PT lock to make sure not
4091          * to leave any page mapped (as page_mapped()) beyond the end
4092          * of the i_size (remove_inode_hugepages() is strict about
4093          * enforcing that). If we bail out here, we'll also leave a
4094          * page in the radix tree in the vm_shared case beyond the end
4095          * of the i_size, but remove_inode_hugepages() will take care
4096          * of it as soon as we drop the hugetlb_fault_mutex_table.
4097          */
4098         size = i_size_read(mapping->host) >> huge_page_shift(h);
4099         ret = -EFAULT;
4100         if (idx >= size)
4101                 goto out_release_unlock;
4102
4103         ret = -EEXIST;
4104         if (!huge_pte_none(huge_ptep_get(dst_pte)))
4105                 goto out_release_unlock;
4106
4107         if (vm_shared) {
4108                 page_dup_rmap(page, true);
4109         } else {
4110                 ClearPagePrivate(page);
4111                 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4112         }
4113
4114         _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4115         if (dst_vma->vm_flags & VM_WRITE)
4116                 _dst_pte = huge_pte_mkdirty(_dst_pte);
4117         _dst_pte = pte_mkyoung(_dst_pte);
4118
4119         set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4120
4121         (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4122                                         dst_vma->vm_flags & VM_WRITE);
4123         hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4124
4125         /* No need to invalidate - it was non-present before */
4126         update_mmu_cache(dst_vma, dst_addr, dst_pte);
4127
4128         spin_unlock(ptl);
4129         if (vm_shared)
4130                 unlock_page(page);
4131         ret = 0;
4132 out:
4133         return ret;
4134 out_release_unlock:
4135         spin_unlock(ptl);
4136         if (vm_shared)
4137                 unlock_page(page);
4138 out_release_nounlock:
4139         put_page(page);
4140         goto out;
4141 }
4142
4143 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4144                          struct page **pages, struct vm_area_struct **vmas,
4145                          unsigned long *position, unsigned long *nr_pages,
4146                          long i, unsigned int flags, int *nonblocking)
4147 {
4148         unsigned long pfn_offset;
4149         unsigned long vaddr = *position;
4150         unsigned long remainder = *nr_pages;
4151         struct hstate *h = hstate_vma(vma);
4152         int err = -EFAULT;
4153
4154         while (vaddr < vma->vm_end && remainder) {
4155                 pte_t *pte;
4156                 spinlock_t *ptl = NULL;
4157                 int absent;
4158                 struct page *page;
4159
4160                 /*
4161                  * If we have a pending SIGKILL, don't keep faulting pages and
4162                  * potentially allocating memory.
4163                  */
4164                 if (unlikely(fatal_signal_pending(current))) {
4165                         remainder = 0;
4166                         break;
4167                 }
4168
4169                 /*
4170                  * Some archs (sparc64, sh*) have multiple pte_ts to
4171                  * each hugepage.  We have to make sure we get the
4172                  * first, for the page indexing below to work.
4173                  *
4174                  * Note that page table lock is not held when pte is null.
4175                  */
4176                 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4177                                       huge_page_size(h));
4178                 if (pte)
4179                         ptl = huge_pte_lock(h, mm, pte);
4180                 absent = !pte || huge_pte_none(huge_ptep_get(pte));
4181
4182                 /*
4183                  * When coredumping, it suits get_dump_page if we just return
4184                  * an error where there's an empty slot with no huge pagecache
4185                  * to back it.  This way, we avoid allocating a hugepage, and
4186                  * the sparse dumpfile avoids allocating disk blocks, but its
4187                  * huge holes still show up with zeroes where they need to be.
4188                  */
4189                 if (absent && (flags & FOLL_DUMP) &&
4190                     !hugetlbfs_pagecache_present(h, vma, vaddr)) {
4191                         if (pte)
4192                                 spin_unlock(ptl);
4193                         remainder = 0;
4194                         break;
4195                 }
4196
4197                 /*
4198                  * We need call hugetlb_fault for both hugepages under migration
4199                  * (in which case hugetlb_fault waits for the migration,) and
4200                  * hwpoisoned hugepages (in which case we need to prevent the
4201                  * caller from accessing to them.) In order to do this, we use
4202                  * here is_swap_pte instead of is_hugetlb_entry_migration and
4203                  * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4204                  * both cases, and because we can't follow correct pages
4205                  * directly from any kind of swap entries.
4206                  */
4207                 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4208                     ((flags & FOLL_WRITE) &&
4209                       !huge_pte_write(huge_ptep_get(pte)))) {
4210                         int ret;
4211                         unsigned int fault_flags = 0;
4212
4213                         if (pte)
4214                                 spin_unlock(ptl);
4215                         if (flags & FOLL_WRITE)
4216                                 fault_flags |= FAULT_FLAG_WRITE;
4217                         if (nonblocking)
4218                                 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
4219                         if (flags & FOLL_NOWAIT)
4220                                 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4221                                         FAULT_FLAG_RETRY_NOWAIT;
4222                         if (flags & FOLL_TRIED) {
4223                                 VM_WARN_ON_ONCE(fault_flags &
4224                                                 FAULT_FLAG_ALLOW_RETRY);
4225                                 fault_flags |= FAULT_FLAG_TRIED;
4226                         }
4227                         ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4228                         if (ret & VM_FAULT_ERROR) {
4229                                 err = vm_fault_to_errno(ret, flags);
4230                                 remainder = 0;
4231                                 break;
4232                         }
4233                         if (ret & VM_FAULT_RETRY) {
4234                                 if (nonblocking)
4235                                         *nonblocking = 0;
4236                                 *nr_pages = 0;
4237                                 /*
4238                                  * VM_FAULT_RETRY must not return an
4239                                  * error, it will return zero
4240                                  * instead.
4241                                  *
4242                                  * No need to update "position" as the
4243                                  * caller will not check it after
4244                                  * *nr_pages is set to 0.
4245                                  */
4246                                 return i;
4247                         }
4248                         continue;
4249                 }
4250
4251                 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4252                 page = pte_page(huge_ptep_get(pte));
4253 same_page:
4254                 if (pages) {
4255                         pages[i] = mem_map_offset(page, pfn_offset);
4256                         get_page(pages[i]);
4257                 }
4258
4259                 if (vmas)
4260                         vmas[i] = vma;
4261
4262                 vaddr += PAGE_SIZE;
4263                 ++pfn_offset;
4264                 --remainder;
4265                 ++i;
4266                 if (vaddr < vma->vm_end && remainder &&
4267                                 pfn_offset < pages_per_huge_page(h)) {
4268                         /*
4269                          * We use pfn_offset to avoid touching the pageframes
4270                          * of this compound page.
4271                          */
4272                         goto same_page;
4273                 }
4274                 spin_unlock(ptl);
4275         }
4276         *nr_pages = remainder;
4277         /*
4278          * setting position is actually required only if remainder is
4279          * not zero but it's faster not to add a "if (remainder)"
4280          * branch.
4281          */
4282         *position = vaddr;
4283
4284         return i ? i : err;
4285 }
4286
4287 #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4288 /*
4289  * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4290  * implement this.
4291  */
4292 #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4293 #endif
4294
4295 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
4296                 unsigned long address, unsigned long end, pgprot_t newprot)
4297 {
4298         struct mm_struct *mm = vma->vm_mm;
4299         unsigned long start = address;
4300         pte_t *ptep;
4301         pte_t pte;
4302         struct hstate *h = hstate_vma(vma);
4303         unsigned long pages = 0;
4304
4305         BUG_ON(address >= end);
4306         flush_cache_range(vma, address, end);
4307
4308         mmu_notifier_invalidate_range_start(mm, start, end);
4309         i_mmap_lock_write(vma->vm_file->f_mapping);
4310         for (; address < end; address += huge_page_size(h)) {
4311                 spinlock_t *ptl;
4312                 ptep = huge_pte_offset(mm, address, huge_page_size(h));
4313                 if (!ptep)
4314                         continue;
4315                 ptl = huge_pte_lock(h, mm, ptep);
4316                 if (huge_pmd_unshare(mm, &address, ptep)) {
4317                         pages++;
4318                         spin_unlock(ptl);
4319                         continue;
4320                 }
4321                 pte = huge_ptep_get(ptep);
4322                 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
4323                         spin_unlock(ptl);
4324                         continue;
4325                 }
4326                 if (unlikely(is_hugetlb_entry_migration(pte))) {
4327                         swp_entry_t entry = pte_to_swp_entry(pte);
4328
4329                         if (is_write_migration_entry(entry)) {
4330                                 pte_t newpte;
4331
4332                                 make_migration_entry_read(&entry);
4333                                 newpte = swp_entry_to_pte(entry);
4334                                 set_huge_swap_pte_at(mm, address, ptep,
4335                                                      newpte, huge_page_size(h));
4336                                 pages++;
4337                         }
4338                         spin_unlock(ptl);
4339                         continue;
4340                 }
4341                 if (!huge_pte_none(pte)) {
4342                         pte = huge_ptep_get_and_clear(mm, address, ptep);
4343                         pte = pte_mkhuge(huge_pte_modify(pte, newprot));
4344                         pte = arch_make_huge_pte(pte, vma, NULL, 0);
4345                         set_huge_pte_at(mm, address, ptep, pte);
4346                         pages++;
4347                 }
4348                 spin_unlock(ptl);
4349         }
4350         /*
4351          * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
4352          * may have cleared our pud entry and done put_page on the page table:
4353          * once we release i_mmap_rwsem, another task can do the final put_page
4354          * and that page table be reused and filled with junk.
4355          */
4356         flush_hugetlb_tlb_range(vma, start, end);
4357         /*
4358          * No need to call mmu_notifier_invalidate_range() we are downgrading
4359          * page table protection not changing it to point to a new page.
4360          *
4361          * See Documentation/vm/mmu_notifier.rst
4362          */
4363         i_mmap_unlock_write(vma->vm_file->f_mapping);
4364         mmu_notifier_invalidate_range_end(mm, start, end);
4365
4366         return pages << h->order;
4367 }
4368
4369 int hugetlb_reserve_pages(struct inode *inode,
4370                                         long from, long to,
4371                                         struct vm_area_struct *vma,
4372                                         vm_flags_t vm_flags)
4373 {
4374         long ret, chg;
4375         struct hstate *h = hstate_inode(inode);
4376         struct hugepage_subpool *spool = subpool_inode(inode);
4377         struct resv_map *resv_map;
4378         long gbl_reserve;
4379
4380         /* This should never happen */
4381         if (from > to) {
4382                 VM_WARN(1, "%s called with a negative range\n", __func__);
4383                 return -EINVAL;
4384         }
4385
4386         /*
4387          * Only apply hugepage reservation if asked. At fault time, an
4388          * attempt will be made for VM_NORESERVE to allocate a page
4389          * without using reserves
4390          */
4391         if (vm_flags & VM_NORESERVE)
4392                 return 0;
4393
4394         /*
4395          * Shared mappings base their reservation on the number of pages that
4396          * are already allocated on behalf of the file. Private mappings need
4397          * to reserve the full area even if read-only as mprotect() may be
4398          * called to make the mapping read-write. Assume !vma is a shm mapping
4399          */
4400         if (!vma || vma->vm_flags & VM_MAYSHARE) {
4401                 resv_map = inode_resv_map(inode);
4402
4403                 chg = region_chg(resv_map, from, to);
4404
4405         } else {
4406                 resv_map = resv_map_alloc();
4407                 if (!resv_map)
4408                         return -ENOMEM;
4409
4410                 chg = to - from;
4411
4412                 set_vma_resv_map(vma, resv_map);
4413                 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4414         }
4415
4416         if (chg < 0) {
4417                 ret = chg;
4418                 goto out_err;
4419         }
4420
4421         /*
4422          * There must be enough pages in the subpool for the mapping. If
4423          * the subpool has a minimum size, there may be some global
4424          * reservations already in place (gbl_reserve).
4425          */
4426         gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4427         if (gbl_reserve < 0) {
4428                 ret = -ENOSPC;
4429                 goto out_err;
4430         }
4431
4432         /*
4433          * Check enough hugepages are available for the reservation.
4434          * Hand the pages back to the subpool if there are not
4435          */
4436         ret = hugetlb_acct_memory(h, gbl_reserve);
4437         if (ret < 0) {
4438                 /* put back original number of pages, chg */
4439                 (void)hugepage_subpool_put_pages(spool, chg);
4440                 goto out_err;
4441         }
4442
4443         /*
4444          * Account for the reservations made. Shared mappings record regions
4445          * that have reservations as they are shared by multiple VMAs.
4446          * When the last VMA disappears, the region map says how much
4447          * the reservation was and the page cache tells how much of
4448          * the reservation was consumed. Private mappings are per-VMA and
4449          * only the consumed reservations are tracked. When the VMA
4450          * disappears, the original reservation is the VMA size and the
4451          * consumed reservations are stored in the map. Hence, nothing
4452          * else has to be done for private mappings here
4453          */
4454         if (!vma || vma->vm_flags & VM_MAYSHARE) {
4455                 long add = region_add(resv_map, from, to);
4456
4457                 if (unlikely(chg > add)) {
4458                         /*
4459                          * pages in this range were added to the reserve
4460                          * map between region_chg and region_add.  This
4461                          * indicates a race with alloc_huge_page.  Adjust
4462                          * the subpool and reserve counts modified above
4463                          * based on the difference.
4464                          */
4465                         long rsv_adjust;
4466
4467                         rsv_adjust = hugepage_subpool_put_pages(spool,
4468                                                                 chg - add);
4469                         hugetlb_acct_memory(h, -rsv_adjust);
4470                 }
4471         }
4472         return 0;
4473 out_err:
4474         if (!vma || vma->vm_flags & VM_MAYSHARE)
4475                 /* Don't call region_abort if region_chg failed */
4476                 if (chg >= 0)
4477                         region_abort(resv_map, from, to);
4478         if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4479                 kref_put(&resv_map->refs, resv_map_release);
4480         return ret;
4481 }
4482
4483 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4484                                                                 long freed)
4485 {
4486         struct hstate *h = hstate_inode(inode);
4487         struct resv_map *resv_map = inode_resv_map(inode);
4488         long chg = 0;
4489         struct hugepage_subpool *spool = subpool_inode(inode);
4490         long gbl_reserve;
4491
4492         if (resv_map) {
4493                 chg = region_del(resv_map, start, end);
4494                 /*
4495                  * region_del() can fail in the rare case where a region
4496                  * must be split and another region descriptor can not be
4497                  * allocated.  If end == LONG_MAX, it will not fail.
4498                  */
4499                 if (chg < 0)
4500                         return chg;
4501         }
4502
4503         spin_lock(&inode->i_lock);
4504         inode->i_blocks -= (blocks_per_huge_page(h) * freed);
4505         spin_unlock(&inode->i_lock);
4506
4507         /*
4508          * If the subpool has a minimum size, the number of global
4509          * reservations to be released may be adjusted.
4510          */
4511         gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4512         hugetlb_acct_memory(h, -gbl_reserve);
4513
4514         return 0;
4515 }
4516
4517 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4518 static unsigned long page_table_shareable(struct vm_area_struct *svma,
4519                                 struct vm_area_struct *vma,
4520                                 unsigned long addr, pgoff_t idx)
4521 {
4522         unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4523                                 svma->vm_start;
4524         unsigned long sbase = saddr & PUD_MASK;
4525         unsigned long s_end = sbase + PUD_SIZE;
4526
4527         /* Allow segments to share if only one is marked locked */
4528         unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4529         unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
4530
4531         /*
4532          * match the virtual addresses, permission and the alignment of the
4533          * page table page.
4534          */
4535         if (pmd_index(addr) != pmd_index(saddr) ||
4536             vm_flags != svm_flags ||
4537             sbase < svma->vm_start || svma->vm_end < s_end)
4538                 return 0;
4539
4540         return saddr;
4541 }
4542
4543 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
4544 {
4545         unsigned long base = addr & PUD_MASK;
4546         unsigned long end = base + PUD_SIZE;
4547
4548         /*
4549          * check on proper vm_flags and page table alignment
4550          */
4551         if (vma->vm_flags & VM_MAYSHARE &&
4552             vma->vm_start <= base && end <= vma->vm_end)
4553                 return true;
4554         return false;
4555 }
4556
4557 /*
4558  * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4559  * and returns the corresponding pte. While this is not necessary for the
4560  * !shared pmd case because we can allocate the pmd later as well, it makes the
4561  * code much cleaner. pmd allocation is essential for the shared case because
4562  * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4563  * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4564  * bad pmd for sharing.
4565  */
4566 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4567 {
4568         struct vm_area_struct *vma = find_vma(mm, addr);
4569         struct address_space *mapping = vma->vm_file->f_mapping;
4570         pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4571                         vma->vm_pgoff;
4572         struct vm_area_struct *svma;
4573         unsigned long saddr;
4574         pte_t *spte = NULL;
4575         pte_t *pte;
4576         spinlock_t *ptl;
4577
4578         if (!vma_shareable(vma, addr))
4579                 return (pte_t *)pmd_alloc(mm, pud, addr);
4580
4581         i_mmap_lock_write(mapping);
4582         vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4583                 if (svma == vma)
4584                         continue;
4585
4586                 saddr = page_table_shareable(svma, vma, addr, idx);
4587                 if (saddr) {
4588                         spte = huge_pte_offset(svma->vm_mm, saddr,
4589                                                vma_mmu_pagesize(svma));
4590                         if (spte) {
4591                                 get_page(virt_to_page(spte));
4592                                 break;
4593                         }
4594                 }
4595         }
4596
4597         if (!spte)
4598                 goto out;
4599
4600         ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
4601         if (pud_none(*pud)) {
4602                 pud_populate(mm, pud,
4603                                 (pmd_t *)((unsigned long)spte & PAGE_MASK));
4604                 mm_inc_nr_pmds(mm);
4605         } else {
4606                 put_page(virt_to_page(spte));
4607         }
4608         spin_unlock(ptl);
4609 out:
4610         pte = (pte_t *)pmd_alloc(mm, pud, addr);
4611         i_mmap_unlock_write(mapping);
4612         return pte;
4613 }
4614
4615 /*
4616  * unmap huge page backed by shared pte.
4617  *
4618  * Hugetlb pte page is ref counted at the time of mapping.  If pte is shared
4619  * indicated by page_count > 1, unmap is achieved by clearing pud and
4620  * decrementing the ref count. If count == 1, the pte page is not shared.
4621  *
4622  * called with page table lock held.
4623  *
4624  * returns: 1 successfully unmapped a shared pte page
4625  *          0 the underlying pte page is not shared, or it is the last user
4626  */
4627 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4628 {
4629         pgd_t *pgd = pgd_offset(mm, *addr);
4630         p4d_t *p4d = p4d_offset(pgd, *addr);
4631         pud_t *pud = pud_offset(p4d, *addr);
4632
4633         BUG_ON(page_count(virt_to_page(ptep)) == 0);
4634         if (page_count(virt_to_page(ptep)) == 1)
4635                 return 0;
4636
4637         pud_clear(pud);
4638         put_page(virt_to_page(ptep));
4639         mm_dec_nr_pmds(mm);
4640         *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4641         return 1;
4642 }
4643 #define want_pmd_share()        (1)
4644 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4645 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4646 {
4647         return NULL;
4648 }
4649
4650 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4651 {
4652         return 0;
4653 }
4654 #define want_pmd_share()        (0)
4655 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4656
4657 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4658 pte_t *huge_pte_alloc(struct mm_struct *mm,
4659                         unsigned long addr, unsigned long sz)
4660 {
4661         pgd_t *pgd;
4662         p4d_t *p4d;
4663         pud_t *pud;
4664         pte_t *pte = NULL;
4665
4666         pgd = pgd_offset(mm, addr);
4667         p4d = p4d_alloc(mm, pgd, addr);
4668         if (!p4d)
4669                 return NULL;
4670         pud = pud_alloc(mm, p4d, addr);
4671         if (pud) {
4672                 if (sz == PUD_SIZE) {
4673                         pte = (pte_t *)pud;
4674                 } else {
4675                         BUG_ON(sz != PMD_SIZE);
4676                         if (want_pmd_share() && pud_none(*pud))
4677                                 pte = huge_pmd_share(mm, addr, pud);
4678                         else
4679                                 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4680                 }
4681         }
4682         BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
4683
4684         return pte;
4685 }
4686
4687 /*
4688  * huge_pte_offset() - Walk the page table to resolve the hugepage
4689  * entry at address @addr
4690  *
4691  * Return: Pointer to page table or swap entry (PUD or PMD) for
4692  * address @addr, or NULL if a p*d_none() entry is encountered and the
4693  * size @sz doesn't match the hugepage size at this level of the page
4694  * table.
4695  */
4696 pte_t *huge_pte_offset(struct mm_struct *mm,
4697                        unsigned long addr, unsigned long sz)
4698 {
4699         pgd_t *pgd;
4700         p4d_t *p4d;
4701         pud_t *pud;
4702         pmd_t *pmd;
4703
4704         pgd = pgd_offset(mm, addr);
4705         if (!pgd_present(*pgd))
4706                 return NULL;
4707         p4d = p4d_offset(pgd, addr);
4708         if (!p4d_present(*p4d))
4709                 return NULL;
4710
4711         pud = pud_offset(p4d, addr);
4712         if (sz != PUD_SIZE && pud_none(*pud))
4713                 return NULL;
4714         /* hugepage or swap? */
4715         if (pud_huge(*pud) || !pud_present(*pud))
4716                 return (pte_t *)pud;
4717
4718         pmd = pmd_offset(pud, addr);
4719         if (sz != PMD_SIZE && pmd_none(*pmd))
4720                 return NULL;
4721         /* hugepage or swap? */
4722         if (pmd_huge(*pmd) || !pmd_present(*pmd))
4723                 return (pte_t *)pmd;
4724
4725         return NULL;
4726 }
4727
4728 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4729
4730 /*
4731  * These functions are overwritable if your architecture needs its own
4732  * behavior.
4733  */
4734 struct page * __weak
4735 follow_huge_addr(struct mm_struct *mm, unsigned long address,
4736                               int write)
4737 {
4738         return ERR_PTR(-EINVAL);
4739 }
4740
4741 struct page * __weak
4742 follow_huge_pd(struct vm_area_struct *vma,
4743                unsigned long address, hugepd_t hpd, int flags, int pdshift)
4744 {
4745         WARN(1, "hugepd follow called with no support for hugepage directory format\n");
4746         return NULL;
4747 }
4748
4749 struct page * __weak
4750 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
4751                 pmd_t *pmd, int flags)
4752 {
4753         struct page *page = NULL;
4754         spinlock_t *ptl;
4755         pte_t pte;
4756 retry:
4757         ptl = pmd_lockptr(mm, pmd);
4758         spin_lock(ptl);
4759         /*
4760          * make sure that the address range covered by this pmd is not
4761          * unmapped from other threads.
4762          */
4763         if (!pmd_huge(*pmd))
4764                 goto out;
4765         pte = huge_ptep_get((pte_t *)pmd);
4766         if (pte_present(pte)) {
4767                 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
4768                 if (flags & FOLL_GET)
4769                         get_page(page);
4770         } else {
4771                 if (is_hugetlb_entry_migration(pte)) {
4772                         spin_unlock(ptl);
4773                         __migration_entry_wait(mm, (pte_t *)pmd, ptl);
4774                         goto retry;
4775                 }
4776                 /*
4777                  * hwpoisoned entry is treated as no_page_table in
4778                  * follow_page_mask().
4779                  */
4780         }
4781 out:
4782         spin_unlock(ptl);
4783         return page;
4784 }
4785
4786 struct page * __weak
4787 follow_huge_pud(struct mm_struct *mm, unsigned long address,
4788                 pud_t *pud, int flags)
4789 {
4790         if (flags & FOLL_GET)
4791                 return NULL;
4792
4793         return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
4794 }
4795
4796 struct page * __weak
4797 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
4798 {
4799         if (flags & FOLL_GET)
4800                 return NULL;
4801
4802         return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
4803 }
4804
4805 bool isolate_huge_page(struct page *page, struct list_head *list)
4806 {
4807         bool ret = true;
4808
4809         VM_BUG_ON_PAGE(!PageHead(page), page);
4810         spin_lock(&hugetlb_lock);
4811         if (!page_huge_active(page) || !get_page_unless_zero(page)) {
4812                 ret = false;
4813                 goto unlock;
4814         }
4815         clear_page_huge_active(page);
4816         list_move_tail(&page->lru, list);
4817 unlock:
4818         spin_unlock(&hugetlb_lock);
4819         return ret;
4820 }
4821
4822 void putback_active_hugepage(struct page *page)
4823 {
4824         VM_BUG_ON_PAGE(!PageHead(page), page);
4825         spin_lock(&hugetlb_lock);
4826         set_page_huge_active(page);
4827         list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
4828         spin_unlock(&hugetlb_lock);
4829         put_page(page);
4830 }
4831
4832 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
4833 {
4834         struct hstate *h = page_hstate(oldpage);
4835
4836         hugetlb_cgroup_migrate(oldpage, newpage);
4837         set_page_owner_migrate_reason(newpage, reason);
4838
4839         /*
4840          * transfer temporary state of the new huge page. This is
4841          * reverse to other transitions because the newpage is going to
4842          * be final while the old one will be freed so it takes over
4843          * the temporary status.
4844          *
4845          * Also note that we have to transfer the per-node surplus state
4846          * here as well otherwise the global surplus count will not match
4847          * the per-node's.
4848          */
4849         if (PageHugeTemporary(newpage)) {
4850                 int old_nid = page_to_nid(oldpage);
4851                 int new_nid = page_to_nid(newpage);
4852
4853                 SetPageHugeTemporary(oldpage);
4854                 ClearPageHugeTemporary(newpage);
4855
4856                 spin_lock(&hugetlb_lock);
4857                 if (h->surplus_huge_pages_node[old_nid]) {
4858                         h->surplus_huge_pages_node[old_nid]--;
4859                         h->surplus_huge_pages_node[new_nid]++;
4860                 }
4861                 spin_unlock(&hugetlb_lock);
4862         }
4863 }