]> git.kernelconcepts.de Git - karo-tx-linux.git/blob - mm/hugetlb.c
Merge branch 'irq-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel...
[karo-tx-linux.git] / mm / hugetlb.c
1 /*
2  * Generic hugetlb support.
3  * (C) Nadia Yvette Chambers, April 2004
4  */
5 #include <linux/list.h>
6 #include <linux/init.h>
7 #include <linux/module.h>
8 #include <linux/mm.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/compiler.h>
17 #include <linux/cpuset.h>
18 #include <linux/mutex.h>
19 #include <linux/bootmem.h>
20 #include <linux/sysfs.h>
21 #include <linux/slab.h>
22 #include <linux/rmap.h>
23 #include <linux/swap.h>
24 #include <linux/swapops.h>
25 #include <linux/page-isolation.h>
26 #include <linux/jhash.h>
27
28 #include <asm/page.h>
29 #include <asm/pgtable.h>
30 #include <asm/tlb.h>
31
32 #include <linux/io.h>
33 #include <linux/hugetlb.h>
34 #include <linux/hugetlb_cgroup.h>
35 #include <linux/node.h>
36 #include "internal.h"
37
38 int hugepages_treat_as_movable;
39
40 int hugetlb_max_hstate __read_mostly;
41 unsigned int default_hstate_idx;
42 struct hstate hstates[HUGE_MAX_HSTATE];
43 /*
44  * Minimum page order among possible hugepage sizes, set to a proper value
45  * at boot time.
46  */
47 static unsigned int minimum_order __read_mostly = UINT_MAX;
48
49 __initdata LIST_HEAD(huge_boot_pages);
50
51 /* for command line parsing */
52 static struct hstate * __initdata parsed_hstate;
53 static unsigned long __initdata default_hstate_max_huge_pages;
54 static unsigned long __initdata default_hstate_size;
55
56 /*
57  * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
58  * free_huge_pages, and surplus_huge_pages.
59  */
60 DEFINE_SPINLOCK(hugetlb_lock);
61
62 /*
63  * Serializes faults on the same logical page.  This is used to
64  * prevent spurious OOMs when the hugepage pool is fully utilized.
65  */
66 static int num_fault_mutexes;
67 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
68
69 /* Forward declaration */
70 static int hugetlb_acct_memory(struct hstate *h, long delta);
71
72 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
73 {
74         bool free = (spool->count == 0) && (spool->used_hpages == 0);
75
76         spin_unlock(&spool->lock);
77
78         /* If no pages are used, and no other handles to the subpool
79          * remain, give up any reservations mased on minimum size and
80          * free the subpool */
81         if (free) {
82                 if (spool->min_hpages != -1)
83                         hugetlb_acct_memory(spool->hstate,
84                                                 -spool->min_hpages);
85                 kfree(spool);
86         }
87 }
88
89 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
90                                                 long min_hpages)
91 {
92         struct hugepage_subpool *spool;
93
94         spool = kzalloc(sizeof(*spool), GFP_KERNEL);
95         if (!spool)
96                 return NULL;
97
98         spin_lock_init(&spool->lock);
99         spool->count = 1;
100         spool->max_hpages = max_hpages;
101         spool->hstate = h;
102         spool->min_hpages = min_hpages;
103
104         if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
105                 kfree(spool);
106                 return NULL;
107         }
108         spool->rsv_hpages = min_hpages;
109
110         return spool;
111 }
112
113 void hugepage_put_subpool(struct hugepage_subpool *spool)
114 {
115         spin_lock(&spool->lock);
116         BUG_ON(!spool->count);
117         spool->count--;
118         unlock_or_release_subpool(spool);
119 }
120
121 /*
122  * Subpool accounting for allocating and reserving pages.
123  * Return -ENOMEM if there are not enough resources to satisfy the
124  * the request.  Otherwise, return the number of pages by which the
125  * global pools must be adjusted (upward).  The returned value may
126  * only be different than the passed value (delta) in the case where
127  * a subpool minimum size must be manitained.
128  */
129 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
130                                       long delta)
131 {
132         long ret = delta;
133
134         if (!spool)
135                 return ret;
136
137         spin_lock(&spool->lock);
138
139         if (spool->max_hpages != -1) {          /* maximum size accounting */
140                 if ((spool->used_hpages + delta) <= spool->max_hpages)
141                         spool->used_hpages += delta;
142                 else {
143                         ret = -ENOMEM;
144                         goto unlock_ret;
145                 }
146         }
147
148         if (spool->min_hpages != -1) {          /* minimum size accounting */
149                 if (delta > spool->rsv_hpages) {
150                         /*
151                          * Asking for more reserves than those already taken on
152                          * behalf of subpool.  Return difference.
153                          */
154                         ret = delta - spool->rsv_hpages;
155                         spool->rsv_hpages = 0;
156                 } else {
157                         ret = 0;        /* reserves already accounted for */
158                         spool->rsv_hpages -= delta;
159                 }
160         }
161
162 unlock_ret:
163         spin_unlock(&spool->lock);
164         return ret;
165 }
166
167 /*
168  * Subpool accounting for freeing and unreserving pages.
169  * Return the number of global page reservations that must be dropped.
170  * The return value may only be different than the passed value (delta)
171  * in the case where a subpool minimum size must be maintained.
172  */
173 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
174                                        long delta)
175 {
176         long ret = delta;
177
178         if (!spool)
179                 return delta;
180
181         spin_lock(&spool->lock);
182
183         if (spool->max_hpages != -1)            /* maximum size accounting */
184                 spool->used_hpages -= delta;
185
186         if (spool->min_hpages != -1) {          /* minimum size accounting */
187                 if (spool->rsv_hpages + delta <= spool->min_hpages)
188                         ret = 0;
189                 else
190                         ret = spool->rsv_hpages + delta - spool->min_hpages;
191
192                 spool->rsv_hpages += delta;
193                 if (spool->rsv_hpages > spool->min_hpages)
194                         spool->rsv_hpages = spool->min_hpages;
195         }
196
197         /*
198          * If hugetlbfs_put_super couldn't free spool due to an outstanding
199          * quota reference, free it now.
200          */
201         unlock_or_release_subpool(spool);
202
203         return ret;
204 }
205
206 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
207 {
208         return HUGETLBFS_SB(inode->i_sb)->spool;
209 }
210
211 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
212 {
213         return subpool_inode(file_inode(vma->vm_file));
214 }
215
216 /*
217  * Region tracking -- allows tracking of reservations and instantiated pages
218  *                    across the pages in a mapping.
219  *
220  * The region data structures are embedded into a resv_map and protected
221  * by a resv_map's lock.  The set of regions within the resv_map represent
222  * reservations for huge pages, or huge pages that have already been
223  * instantiated within the map.  The from and to elements are huge page
224  * indicies into the associated mapping.  from indicates the starting index
225  * of the region.  to represents the first index past the end of  the region.
226  *
227  * For example, a file region structure with from == 0 and to == 4 represents
228  * four huge pages in a mapping.  It is important to note that the to element
229  * represents the first element past the end of the region. This is used in
230  * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
231  *
232  * Interval notation of the form [from, to) will be used to indicate that
233  * the endpoint from is inclusive and to is exclusive.
234  */
235 struct file_region {
236         struct list_head link;
237         long from;
238         long to;
239 };
240
241 /*
242  * Add the huge page range represented by [f, t) to the reserve
243  * map.  In the normal case, existing regions will be expanded
244  * to accommodate the specified range.  Sufficient regions should
245  * exist for expansion due to the previous call to region_chg
246  * with the same range.  However, it is possible that region_del
247  * could have been called after region_chg and modifed the map
248  * in such a way that no region exists to be expanded.  In this
249  * case, pull a region descriptor from the cache associated with
250  * the map and use that for the new range.
251  *
252  * Return the number of new huge pages added to the map.  This
253  * number is greater than or equal to zero.
254  */
255 static long region_add(struct resv_map *resv, long f, long t)
256 {
257         struct list_head *head = &resv->regions;
258         struct file_region *rg, *nrg, *trg;
259         long add = 0;
260
261         spin_lock(&resv->lock);
262         /* Locate the region we are either in or before. */
263         list_for_each_entry(rg, head, link)
264                 if (f <= rg->to)
265                         break;
266
267         /*
268          * If no region exists which can be expanded to include the
269          * specified range, the list must have been modified by an
270          * interleving call to region_del().  Pull a region descriptor
271          * from the cache and use it for this range.
272          */
273         if (&rg->link == head || t < rg->from) {
274                 VM_BUG_ON(resv->region_cache_count <= 0);
275
276                 resv->region_cache_count--;
277                 nrg = list_first_entry(&resv->region_cache, struct file_region,
278                                         link);
279                 list_del(&nrg->link);
280
281                 nrg->from = f;
282                 nrg->to = t;
283                 list_add(&nrg->link, rg->link.prev);
284
285                 add += t - f;
286                 goto out_locked;
287         }
288
289         /* Round our left edge to the current segment if it encloses us. */
290         if (f > rg->from)
291                 f = rg->from;
292
293         /* Check for and consume any regions we now overlap with. */
294         nrg = rg;
295         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
296                 if (&rg->link == head)
297                         break;
298                 if (rg->from > t)
299                         break;
300
301                 /* If this area reaches higher then extend our area to
302                  * include it completely.  If this is not the first area
303                  * which we intend to reuse, free it. */
304                 if (rg->to > t)
305                         t = rg->to;
306                 if (rg != nrg) {
307                         /* Decrement return value by the deleted range.
308                          * Another range will span this area so that by
309                          * end of routine add will be >= zero
310                          */
311                         add -= (rg->to - rg->from);
312                         list_del(&rg->link);
313                         kfree(rg);
314                 }
315         }
316
317         add += (nrg->from - f);         /* Added to beginning of region */
318         nrg->from = f;
319         add += t - nrg->to;             /* Added to end of region */
320         nrg->to = t;
321
322 out_locked:
323         resv->adds_in_progress--;
324         spin_unlock(&resv->lock);
325         VM_BUG_ON(add < 0);
326         return add;
327 }
328
329 /*
330  * Examine the existing reserve map and determine how many
331  * huge pages in the specified range [f, t) are NOT currently
332  * represented.  This routine is called before a subsequent
333  * call to region_add that will actually modify the reserve
334  * map to add the specified range [f, t).  region_chg does
335  * not change the number of huge pages represented by the
336  * map.  However, if the existing regions in the map can not
337  * be expanded to represent the new range, a new file_region
338  * structure is added to the map as a placeholder.  This is
339  * so that the subsequent region_add call will have all the
340  * regions it needs and will not fail.
341  *
342  * Upon entry, region_chg will also examine the cache of region descriptors
343  * associated with the map.  If there are not enough descriptors cached, one
344  * will be allocated for the in progress add operation.
345  *
346  * Returns the number of huge pages that need to be added to the existing
347  * reservation map for the range [f, t).  This number is greater or equal to
348  * zero.  -ENOMEM is returned if a new file_region structure or cache entry
349  * is needed and can not be allocated.
350  */
351 static long region_chg(struct resv_map *resv, long f, long t)
352 {
353         struct list_head *head = &resv->regions;
354         struct file_region *rg, *nrg = NULL;
355         long chg = 0;
356
357 retry:
358         spin_lock(&resv->lock);
359 retry_locked:
360         resv->adds_in_progress++;
361
362         /*
363          * Check for sufficient descriptors in the cache to accommodate
364          * the number of in progress add operations.
365          */
366         if (resv->adds_in_progress > resv->region_cache_count) {
367                 struct file_region *trg;
368
369                 VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1);
370                 /* Must drop lock to allocate a new descriptor. */
371                 resv->adds_in_progress--;
372                 spin_unlock(&resv->lock);
373
374                 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
375                 if (!trg) {
376                         kfree(nrg);
377                         return -ENOMEM;
378                 }
379
380                 spin_lock(&resv->lock);
381                 list_add(&trg->link, &resv->region_cache);
382                 resv->region_cache_count++;
383                 goto retry_locked;
384         }
385
386         /* Locate the region we are before or in. */
387         list_for_each_entry(rg, head, link)
388                 if (f <= rg->to)
389                         break;
390
391         /* If we are below the current region then a new region is required.
392          * Subtle, allocate a new region at the position but make it zero
393          * size such that we can guarantee to record the reservation. */
394         if (&rg->link == head || t < rg->from) {
395                 if (!nrg) {
396                         resv->adds_in_progress--;
397                         spin_unlock(&resv->lock);
398                         nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
399                         if (!nrg)
400                                 return -ENOMEM;
401
402                         nrg->from = f;
403                         nrg->to   = f;
404                         INIT_LIST_HEAD(&nrg->link);
405                         goto retry;
406                 }
407
408                 list_add(&nrg->link, rg->link.prev);
409                 chg = t - f;
410                 goto out_nrg;
411         }
412
413         /* Round our left edge to the current segment if it encloses us. */
414         if (f > rg->from)
415                 f = rg->from;
416         chg = t - f;
417
418         /* Check for and consume any regions we now overlap with. */
419         list_for_each_entry(rg, rg->link.prev, link) {
420                 if (&rg->link == head)
421                         break;
422                 if (rg->from > t)
423                         goto out;
424
425                 /* We overlap with this area, if it extends further than
426                  * us then we must extend ourselves.  Account for its
427                  * existing reservation. */
428                 if (rg->to > t) {
429                         chg += rg->to - t;
430                         t = rg->to;
431                 }
432                 chg -= rg->to - rg->from;
433         }
434
435 out:
436         spin_unlock(&resv->lock);
437         /*  We already know we raced and no longer need the new region */
438         kfree(nrg);
439         return chg;
440 out_nrg:
441         spin_unlock(&resv->lock);
442         return chg;
443 }
444
445 /*
446  * Abort the in progress add operation.  The adds_in_progress field
447  * of the resv_map keeps track of the operations in progress between
448  * calls to region_chg and region_add.  Operations are sometimes
449  * aborted after the call to region_chg.  In such cases, region_abort
450  * is called to decrement the adds_in_progress counter.
451  *
452  * NOTE: The range arguments [f, t) are not needed or used in this
453  * routine.  They are kept to make reading the calling code easier as
454  * arguments will match the associated region_chg call.
455  */
456 static void region_abort(struct resv_map *resv, long f, long t)
457 {
458         spin_lock(&resv->lock);
459         VM_BUG_ON(!resv->region_cache_count);
460         resv->adds_in_progress--;
461         spin_unlock(&resv->lock);
462 }
463
464 /*
465  * Delete the specified range [f, t) from the reserve map.  If the
466  * t parameter is LONG_MAX, this indicates that ALL regions after f
467  * should be deleted.  Locate the regions which intersect [f, t)
468  * and either trim, delete or split the existing regions.
469  *
470  * Returns the number of huge pages deleted from the reserve map.
471  * In the normal case, the return value is zero or more.  In the
472  * case where a region must be split, a new region descriptor must
473  * be allocated.  If the allocation fails, -ENOMEM will be returned.
474  * NOTE: If the parameter t == LONG_MAX, then we will never split
475  * a region and possibly return -ENOMEM.  Callers specifying
476  * t == LONG_MAX do not need to check for -ENOMEM error.
477  */
478 static long region_del(struct resv_map *resv, long f, long t)
479 {
480         struct list_head *head = &resv->regions;
481         struct file_region *rg, *trg;
482         struct file_region *nrg = NULL;
483         long del = 0;
484
485 retry:
486         spin_lock(&resv->lock);
487         list_for_each_entry_safe(rg, trg, head, link) {
488                 /*
489                  * Skip regions before the range to be deleted.  file_region
490                  * ranges are normally of the form [from, to).  However, there
491                  * may be a "placeholder" entry in the map which is of the form
492                  * (from, to) with from == to.  Check for placeholder entries
493                  * at the beginning of the range to be deleted.
494                  */
495                 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
496                         continue;
497
498                 if (rg->from >= t)
499                         break;
500
501                 if (f > rg->from && t < rg->to) { /* Must split region */
502                         /*
503                          * Check for an entry in the cache before dropping
504                          * lock and attempting allocation.
505                          */
506                         if (!nrg &&
507                             resv->region_cache_count > resv->adds_in_progress) {
508                                 nrg = list_first_entry(&resv->region_cache,
509                                                         struct file_region,
510                                                         link);
511                                 list_del(&nrg->link);
512                                 resv->region_cache_count--;
513                         }
514
515                         if (!nrg) {
516                                 spin_unlock(&resv->lock);
517                                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
518                                 if (!nrg)
519                                         return -ENOMEM;
520                                 goto retry;
521                         }
522
523                         del += t - f;
524
525                         /* New entry for end of split region */
526                         nrg->from = t;
527                         nrg->to = rg->to;
528                         INIT_LIST_HEAD(&nrg->link);
529
530                         /* Original entry is trimmed */
531                         rg->to = f;
532
533                         list_add(&nrg->link, &rg->link);
534                         nrg = NULL;
535                         break;
536                 }
537
538                 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
539                         del += rg->to - rg->from;
540                         list_del(&rg->link);
541                         kfree(rg);
542                         continue;
543                 }
544
545                 if (f <= rg->from) {    /* Trim beginning of region */
546                         del += t - rg->from;
547                         rg->from = t;
548                 } else {                /* Trim end of region */
549                         del += rg->to - f;
550                         rg->to = f;
551                 }
552         }
553
554         spin_unlock(&resv->lock);
555         kfree(nrg);
556         return del;
557 }
558
559 /*
560  * A rare out of memory error was encountered which prevented removal of
561  * the reserve map region for a page.  The huge page itself was free'ed
562  * and removed from the page cache.  This routine will adjust the subpool
563  * usage count, and the global reserve count if needed.  By incrementing
564  * these counts, the reserve map entry which could not be deleted will
565  * appear as a "reserved" entry instead of simply dangling with incorrect
566  * counts.
567  */
568 void hugetlb_fix_reserve_counts(struct inode *inode, bool restore_reserve)
569 {
570         struct hugepage_subpool *spool = subpool_inode(inode);
571         long rsv_adjust;
572
573         rsv_adjust = hugepage_subpool_get_pages(spool, 1);
574         if (restore_reserve && rsv_adjust) {
575                 struct hstate *h = hstate_inode(inode);
576
577                 hugetlb_acct_memory(h, 1);
578         }
579 }
580
581 /*
582  * Count and return the number of huge pages in the reserve map
583  * that intersect with the range [f, t).
584  */
585 static long region_count(struct resv_map *resv, long f, long t)
586 {
587         struct list_head *head = &resv->regions;
588         struct file_region *rg;
589         long chg = 0;
590
591         spin_lock(&resv->lock);
592         /* Locate each segment we overlap with, and count that overlap. */
593         list_for_each_entry(rg, head, link) {
594                 long seg_from;
595                 long seg_to;
596
597                 if (rg->to <= f)
598                         continue;
599                 if (rg->from >= t)
600                         break;
601
602                 seg_from = max(rg->from, f);
603                 seg_to = min(rg->to, t);
604
605                 chg += seg_to - seg_from;
606         }
607         spin_unlock(&resv->lock);
608
609         return chg;
610 }
611
612 /*
613  * Convert the address within this vma to the page offset within
614  * the mapping, in pagecache page units; huge pages here.
615  */
616 static pgoff_t vma_hugecache_offset(struct hstate *h,
617                         struct vm_area_struct *vma, unsigned long address)
618 {
619         return ((address - vma->vm_start) >> huge_page_shift(h)) +
620                         (vma->vm_pgoff >> huge_page_order(h));
621 }
622
623 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
624                                      unsigned long address)
625 {
626         return vma_hugecache_offset(hstate_vma(vma), vma, address);
627 }
628
629 /*
630  * Return the size of the pages allocated when backing a VMA. In the majority
631  * cases this will be same size as used by the page table entries.
632  */
633 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
634 {
635         struct hstate *hstate;
636
637         if (!is_vm_hugetlb_page(vma))
638                 return PAGE_SIZE;
639
640         hstate = hstate_vma(vma);
641
642         return 1UL << huge_page_shift(hstate);
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 version of this
650  * function is required.
651  */
652 #ifndef vma_mmu_pagesize
653 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
654 {
655         return vma_kernel_pagesize(vma);
656 }
657 #endif
658
659 /*
660  * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
661  * bits of the reservation map pointer, which are always clear due to
662  * alignment.
663  */
664 #define HPAGE_RESV_OWNER    (1UL << 0)
665 #define HPAGE_RESV_UNMAPPED (1UL << 1)
666 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
667
668 /*
669  * These helpers are used to track how many pages are reserved for
670  * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
671  * is guaranteed to have their future faults succeed.
672  *
673  * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
674  * the reserve counters are updated with the hugetlb_lock held. It is safe
675  * to reset the VMA at fork() time as it is not in use yet and there is no
676  * chance of the global counters getting corrupted as a result of the values.
677  *
678  * The private mapping reservation is represented in a subtly different
679  * manner to a shared mapping.  A shared mapping has a region map associated
680  * with the underlying file, this region map represents the backing file
681  * pages which have ever had a reservation assigned which this persists even
682  * after the page is instantiated.  A private mapping has a region map
683  * associated with the original mmap which is attached to all VMAs which
684  * reference it, this region map represents those offsets which have consumed
685  * reservation ie. where pages have been instantiated.
686  */
687 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
688 {
689         return (unsigned long)vma->vm_private_data;
690 }
691
692 static void set_vma_private_data(struct vm_area_struct *vma,
693                                                         unsigned long value)
694 {
695         vma->vm_private_data = (void *)value;
696 }
697
698 struct resv_map *resv_map_alloc(void)
699 {
700         struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
701         struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
702
703         if (!resv_map || !rg) {
704                 kfree(resv_map);
705                 kfree(rg);
706                 return NULL;
707         }
708
709         kref_init(&resv_map->refs);
710         spin_lock_init(&resv_map->lock);
711         INIT_LIST_HEAD(&resv_map->regions);
712
713         resv_map->adds_in_progress = 0;
714
715         INIT_LIST_HEAD(&resv_map->region_cache);
716         list_add(&rg->link, &resv_map->region_cache);
717         resv_map->region_cache_count = 1;
718
719         return resv_map;
720 }
721
722 void resv_map_release(struct kref *ref)
723 {
724         struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
725         struct list_head *head = &resv_map->region_cache;
726         struct file_region *rg, *trg;
727
728         /* Clear out any active regions before we release the map. */
729         region_del(resv_map, 0, LONG_MAX);
730
731         /* ... and any entries left in the cache */
732         list_for_each_entry_safe(rg, trg, head, link) {
733                 list_del(&rg->link);
734                 kfree(rg);
735         }
736
737         VM_BUG_ON(resv_map->adds_in_progress);
738
739         kfree(resv_map);
740 }
741
742 static inline struct resv_map *inode_resv_map(struct inode *inode)
743 {
744         return inode->i_mapping->private_data;
745 }
746
747 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
748 {
749         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
750         if (vma->vm_flags & VM_MAYSHARE) {
751                 struct address_space *mapping = vma->vm_file->f_mapping;
752                 struct inode *inode = mapping->host;
753
754                 return inode_resv_map(inode);
755
756         } else {
757                 return (struct resv_map *)(get_vma_private_data(vma) &
758                                                         ~HPAGE_RESV_MASK);
759         }
760 }
761
762 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
763 {
764         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
765         VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
766
767         set_vma_private_data(vma, (get_vma_private_data(vma) &
768                                 HPAGE_RESV_MASK) | (unsigned long)map);
769 }
770
771 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
772 {
773         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
774         VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
775
776         set_vma_private_data(vma, get_vma_private_data(vma) | flags);
777 }
778
779 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
780 {
781         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
782
783         return (get_vma_private_data(vma) & flag) != 0;
784 }
785
786 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
787 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
788 {
789         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
790         if (!(vma->vm_flags & VM_MAYSHARE))
791                 vma->vm_private_data = (void *)0;
792 }
793
794 /* Returns true if the VMA has associated reserve pages */
795 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
796 {
797         if (vma->vm_flags & VM_NORESERVE) {
798                 /*
799                  * This address is already reserved by other process(chg == 0),
800                  * so, we should decrement reserved count. Without decrementing,
801                  * reserve count remains after releasing inode, because this
802                  * allocated page will go into page cache and is regarded as
803                  * coming from reserved pool in releasing step.  Currently, we
804                  * don't have any other solution to deal with this situation
805                  * properly, so add work-around here.
806                  */
807                 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
808                         return true;
809                 else
810                         return false;
811         }
812
813         /* Shared mappings always use reserves */
814         if (vma->vm_flags & VM_MAYSHARE) {
815                 /*
816                  * We know VM_NORESERVE is not set.  Therefore, there SHOULD
817                  * be a region map for all pages.  The only situation where
818                  * there is no region map is if a hole was punched via
819                  * fallocate.  In this case, there really are no reverves to
820                  * use.  This situation is indicated if chg != 0.
821                  */
822                 if (chg)
823                         return false;
824                 else
825                         return true;
826         }
827
828         /*
829          * Only the process that called mmap() has reserves for
830          * private mappings.
831          */
832         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
833                 return true;
834
835         return false;
836 }
837
838 static void enqueue_huge_page(struct hstate *h, struct page *page)
839 {
840         int nid = page_to_nid(page);
841         list_move(&page->lru, &h->hugepage_freelists[nid]);
842         h->free_huge_pages++;
843         h->free_huge_pages_node[nid]++;
844 }
845
846 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
847 {
848         struct page *page;
849
850         list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
851                 if (!is_migrate_isolate_page(page))
852                         break;
853         /*
854          * if 'non-isolated free hugepage' not found on the list,
855          * the allocation fails.
856          */
857         if (&h->hugepage_freelists[nid] == &page->lru)
858                 return NULL;
859         list_move(&page->lru, &h->hugepage_activelist);
860         set_page_refcounted(page);
861         h->free_huge_pages--;
862         h->free_huge_pages_node[nid]--;
863         return page;
864 }
865
866 /* Movability of hugepages depends on migration support. */
867 static inline gfp_t htlb_alloc_mask(struct hstate *h)
868 {
869         if (hugepages_treat_as_movable || hugepage_migration_supported(h))
870                 return GFP_HIGHUSER_MOVABLE;
871         else
872                 return GFP_HIGHUSER;
873 }
874
875 static struct page *dequeue_huge_page_vma(struct hstate *h,
876                                 struct vm_area_struct *vma,
877                                 unsigned long address, int avoid_reserve,
878                                 long chg)
879 {
880         struct page *page = NULL;
881         struct mempolicy *mpol;
882         nodemask_t *nodemask;
883         struct zonelist *zonelist;
884         struct zone *zone;
885         struct zoneref *z;
886         unsigned int cpuset_mems_cookie;
887
888         /*
889          * A child process with MAP_PRIVATE mappings created by their parent
890          * have no page reserves. This check ensures that reservations are
891          * not "stolen". The child may still get SIGKILLed
892          */
893         if (!vma_has_reserves(vma, chg) &&
894                         h->free_huge_pages - h->resv_huge_pages == 0)
895                 goto err;
896
897         /* If reserves cannot be used, ensure enough pages are in the pool */
898         if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
899                 goto err;
900
901 retry_cpuset:
902         cpuset_mems_cookie = read_mems_allowed_begin();
903         zonelist = huge_zonelist(vma, address,
904                                         htlb_alloc_mask(h), &mpol, &nodemask);
905
906         for_each_zone_zonelist_nodemask(zone, z, zonelist,
907                                                 MAX_NR_ZONES - 1, nodemask) {
908                 if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) {
909                         page = dequeue_huge_page_node(h, zone_to_nid(zone));
910                         if (page) {
911                                 if (avoid_reserve)
912                                         break;
913                                 if (!vma_has_reserves(vma, chg))
914                                         break;
915
916                                 SetPagePrivate(page);
917                                 h->resv_huge_pages--;
918                                 break;
919                         }
920                 }
921         }
922
923         mpol_cond_put(mpol);
924         if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
925                 goto retry_cpuset;
926         return page;
927
928 err:
929         return NULL;
930 }
931
932 /*
933  * common helper functions for hstate_next_node_to_{alloc|free}.
934  * We may have allocated or freed a huge page based on a different
935  * nodes_allowed previously, so h->next_node_to_{alloc|free} might
936  * be outside of *nodes_allowed.  Ensure that we use an allowed
937  * node for alloc or free.
938  */
939 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
940 {
941         nid = next_node(nid, *nodes_allowed);
942         if (nid == MAX_NUMNODES)
943                 nid = first_node(*nodes_allowed);
944         VM_BUG_ON(nid >= MAX_NUMNODES);
945
946         return nid;
947 }
948
949 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
950 {
951         if (!node_isset(nid, *nodes_allowed))
952                 nid = next_node_allowed(nid, nodes_allowed);
953         return nid;
954 }
955
956 /*
957  * returns the previously saved node ["this node"] from which to
958  * allocate a persistent huge page for the pool and advance the
959  * next node from which to allocate, handling wrap at end of node
960  * mask.
961  */
962 static int hstate_next_node_to_alloc(struct hstate *h,
963                                         nodemask_t *nodes_allowed)
964 {
965         int nid;
966
967         VM_BUG_ON(!nodes_allowed);
968
969         nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
970         h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
971
972         return nid;
973 }
974
975 /*
976  * helper for free_pool_huge_page() - return the previously saved
977  * node ["this node"] from which to free a huge page.  Advance the
978  * next node id whether or not we find a free huge page to free so
979  * that the next attempt to free addresses the next node.
980  */
981 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
982 {
983         int nid;
984
985         VM_BUG_ON(!nodes_allowed);
986
987         nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
988         h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
989
990         return nid;
991 }
992
993 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask)           \
994         for (nr_nodes = nodes_weight(*mask);                            \
995                 nr_nodes > 0 &&                                         \
996                 ((node = hstate_next_node_to_alloc(hs, mask)) || 1);    \
997                 nr_nodes--)
998
999 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask)            \
1000         for (nr_nodes = nodes_weight(*mask);                            \
1001                 nr_nodes > 0 &&                                         \
1002                 ((node = hstate_next_node_to_free(hs, mask)) || 1);     \
1003                 nr_nodes--)
1004
1005 #if defined(CONFIG_CMA) && defined(CONFIG_X86_64)
1006 static void destroy_compound_gigantic_page(struct page *page,
1007                                         unsigned int order)
1008 {
1009         int i;
1010         int nr_pages = 1 << order;
1011         struct page *p = page + 1;
1012
1013         for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1014                 clear_compound_head(p);
1015                 set_page_refcounted(p);
1016         }
1017
1018         set_compound_order(page, 0);
1019         __ClearPageHead(page);
1020 }
1021
1022 static void free_gigantic_page(struct page *page, unsigned int order)
1023 {
1024         free_contig_range(page_to_pfn(page), 1 << order);
1025 }
1026
1027 static int __alloc_gigantic_page(unsigned long start_pfn,
1028                                 unsigned long nr_pages)
1029 {
1030         unsigned long end_pfn = start_pfn + nr_pages;
1031         return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
1032 }
1033
1034 static bool pfn_range_valid_gigantic(unsigned long start_pfn,
1035                                 unsigned long nr_pages)
1036 {
1037         unsigned long i, end_pfn = start_pfn + nr_pages;
1038         struct page *page;
1039
1040         for (i = start_pfn; i < end_pfn; i++) {
1041                 if (!pfn_valid(i))
1042                         return false;
1043
1044                 page = pfn_to_page(i);
1045
1046                 if (PageReserved(page))
1047                         return false;
1048
1049                 if (page_count(page) > 0)
1050                         return false;
1051
1052                 if (PageHuge(page))
1053                         return false;
1054         }
1055
1056         return true;
1057 }
1058
1059 static bool zone_spans_last_pfn(const struct zone *zone,
1060                         unsigned long start_pfn, unsigned long nr_pages)
1061 {
1062         unsigned long last_pfn = start_pfn + nr_pages - 1;
1063         return zone_spans_pfn(zone, last_pfn);
1064 }
1065
1066 static struct page *alloc_gigantic_page(int nid, unsigned int order)
1067 {
1068         unsigned long nr_pages = 1 << order;
1069         unsigned long ret, pfn, flags;
1070         struct zone *z;
1071
1072         z = NODE_DATA(nid)->node_zones;
1073         for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) {
1074                 spin_lock_irqsave(&z->lock, flags);
1075
1076                 pfn = ALIGN(z->zone_start_pfn, nr_pages);
1077                 while (zone_spans_last_pfn(z, pfn, nr_pages)) {
1078                         if (pfn_range_valid_gigantic(pfn, nr_pages)) {
1079                                 /*
1080                                  * We release the zone lock here because
1081                                  * alloc_contig_range() will also lock the zone
1082                                  * at some point. If there's an allocation
1083                                  * spinning on this lock, it may win the race
1084                                  * and cause alloc_contig_range() to fail...
1085                                  */
1086                                 spin_unlock_irqrestore(&z->lock, flags);
1087                                 ret = __alloc_gigantic_page(pfn, nr_pages);
1088                                 if (!ret)
1089                                         return pfn_to_page(pfn);
1090                                 spin_lock_irqsave(&z->lock, flags);
1091                         }
1092                         pfn += nr_pages;
1093                 }
1094
1095                 spin_unlock_irqrestore(&z->lock, flags);
1096         }
1097
1098         return NULL;
1099 }
1100
1101 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1102 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1103
1104 static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid)
1105 {
1106         struct page *page;
1107
1108         page = alloc_gigantic_page(nid, huge_page_order(h));
1109         if (page) {
1110                 prep_compound_gigantic_page(page, huge_page_order(h));
1111                 prep_new_huge_page(h, page, nid);
1112         }
1113
1114         return page;
1115 }
1116
1117 static int alloc_fresh_gigantic_page(struct hstate *h,
1118                                 nodemask_t *nodes_allowed)
1119 {
1120         struct page *page = NULL;
1121         int nr_nodes, node;
1122
1123         for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1124                 page = alloc_fresh_gigantic_page_node(h, node);
1125                 if (page)
1126                         return 1;
1127         }
1128
1129         return 0;
1130 }
1131
1132 static inline bool gigantic_page_supported(void) { return true; }
1133 #else
1134 static inline bool gigantic_page_supported(void) { return false; }
1135 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1136 static inline void destroy_compound_gigantic_page(struct page *page,
1137                                                 unsigned int order) { }
1138 static inline int alloc_fresh_gigantic_page(struct hstate *h,
1139                                         nodemask_t *nodes_allowed) { return 0; }
1140 #endif
1141
1142 static void update_and_free_page(struct hstate *h, struct page *page)
1143 {
1144         int i;
1145
1146         if (hstate_is_gigantic(h) && !gigantic_page_supported())
1147                 return;
1148
1149         h->nr_huge_pages--;
1150         h->nr_huge_pages_node[page_to_nid(page)]--;
1151         for (i = 0; i < pages_per_huge_page(h); i++) {
1152                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1153                                 1 << PG_referenced | 1 << PG_dirty |
1154                                 1 << PG_active | 1 << PG_private |
1155                                 1 << PG_writeback);
1156         }
1157         VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1158         set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1159         set_page_refcounted(page);
1160         if (hstate_is_gigantic(h)) {
1161                 destroy_compound_gigantic_page(page, huge_page_order(h));
1162                 free_gigantic_page(page, huge_page_order(h));
1163         } else {
1164                 __free_pages(page, huge_page_order(h));
1165         }
1166 }
1167
1168 struct hstate *size_to_hstate(unsigned long size)
1169 {
1170         struct hstate *h;
1171
1172         for_each_hstate(h) {
1173                 if (huge_page_size(h) == size)
1174                         return h;
1175         }
1176         return NULL;
1177 }
1178
1179 /*
1180  * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1181  * to hstate->hugepage_activelist.)
1182  *
1183  * This function can be called for tail pages, but never returns true for them.
1184  */
1185 bool page_huge_active(struct page *page)
1186 {
1187         VM_BUG_ON_PAGE(!PageHuge(page), page);
1188         return PageHead(page) && PagePrivate(&page[1]);
1189 }
1190
1191 /* never called for tail page */
1192 static void set_page_huge_active(struct page *page)
1193 {
1194         VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1195         SetPagePrivate(&page[1]);
1196 }
1197
1198 static void clear_page_huge_active(struct page *page)
1199 {
1200         VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1201         ClearPagePrivate(&page[1]);
1202 }
1203
1204 void free_huge_page(struct page *page)
1205 {
1206         /*
1207          * Can't pass hstate in here because it is called from the
1208          * compound page destructor.
1209          */
1210         struct hstate *h = page_hstate(page);
1211         int nid = page_to_nid(page);
1212         struct hugepage_subpool *spool =
1213                 (struct hugepage_subpool *)page_private(page);
1214         bool restore_reserve;
1215
1216         set_page_private(page, 0);
1217         page->mapping = NULL;
1218         BUG_ON(page_count(page));
1219         BUG_ON(page_mapcount(page));
1220         restore_reserve = PagePrivate(page);
1221         ClearPagePrivate(page);
1222
1223         /*
1224          * A return code of zero implies that the subpool will be under its
1225          * minimum size if the reservation is not restored after page is free.
1226          * Therefore, force restore_reserve operation.
1227          */
1228         if (hugepage_subpool_put_pages(spool, 1) == 0)
1229                 restore_reserve = true;
1230
1231         spin_lock(&hugetlb_lock);
1232         clear_page_huge_active(page);
1233         hugetlb_cgroup_uncharge_page(hstate_index(h),
1234                                      pages_per_huge_page(h), page);
1235         if (restore_reserve)
1236                 h->resv_huge_pages++;
1237
1238         if (h->surplus_huge_pages_node[nid]) {
1239                 /* remove the page from active list */
1240                 list_del(&page->lru);
1241                 update_and_free_page(h, page);
1242                 h->surplus_huge_pages--;
1243                 h->surplus_huge_pages_node[nid]--;
1244         } else {
1245                 arch_clear_hugepage_flags(page);
1246                 enqueue_huge_page(h, page);
1247         }
1248         spin_unlock(&hugetlb_lock);
1249 }
1250
1251 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1252 {
1253         INIT_LIST_HEAD(&page->lru);
1254         set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1255         spin_lock(&hugetlb_lock);
1256         set_hugetlb_cgroup(page, NULL);
1257         h->nr_huge_pages++;
1258         h->nr_huge_pages_node[nid]++;
1259         spin_unlock(&hugetlb_lock);
1260         put_page(page); /* free it into the hugepage allocator */
1261 }
1262
1263 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1264 {
1265         int i;
1266         int nr_pages = 1 << order;
1267         struct page *p = page + 1;
1268
1269         /* we rely on prep_new_huge_page to set the destructor */
1270         set_compound_order(page, order);
1271         __SetPageHead(page);
1272         __ClearPageReserved(page);
1273         for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1274                 /*
1275                  * For gigantic hugepages allocated through bootmem at
1276                  * boot, it's safer to be consistent with the not-gigantic
1277                  * hugepages and clear the PG_reserved bit from all tail pages
1278                  * too.  Otherwse drivers using get_user_pages() to access tail
1279                  * pages may get the reference counting wrong if they see
1280                  * PG_reserved set on a tail page (despite the head page not
1281                  * having PG_reserved set).  Enforcing this consistency between
1282                  * head and tail pages allows drivers to optimize away a check
1283                  * on the head page when they need know if put_page() is needed
1284                  * after get_user_pages().
1285                  */
1286                 __ClearPageReserved(p);
1287                 set_page_count(p, 0);
1288                 set_compound_head(p, page);
1289         }
1290 }
1291
1292 /*
1293  * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1294  * transparent huge pages.  See the PageTransHuge() documentation for more
1295  * details.
1296  */
1297 int PageHuge(struct page *page)
1298 {
1299         if (!PageCompound(page))
1300                 return 0;
1301
1302         page = compound_head(page);
1303         return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1304 }
1305 EXPORT_SYMBOL_GPL(PageHuge);
1306
1307 /*
1308  * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1309  * normal or transparent huge pages.
1310  */
1311 int PageHeadHuge(struct page *page_head)
1312 {
1313         if (!PageHead(page_head))
1314                 return 0;
1315
1316         return get_compound_page_dtor(page_head) == free_huge_page;
1317 }
1318
1319 pgoff_t __basepage_index(struct page *page)
1320 {
1321         struct page *page_head = compound_head(page);
1322         pgoff_t index = page_index(page_head);
1323         unsigned long compound_idx;
1324
1325         if (!PageHuge(page_head))
1326                 return page_index(page);
1327
1328         if (compound_order(page_head) >= MAX_ORDER)
1329                 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1330         else
1331                 compound_idx = page - page_head;
1332
1333         return (index << compound_order(page_head)) + compound_idx;
1334 }
1335
1336 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
1337 {
1338         struct page *page;
1339
1340         page = __alloc_pages_node(nid,
1341                 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
1342                                                 __GFP_REPEAT|__GFP_NOWARN,
1343                 huge_page_order(h));
1344         if (page) {
1345                 prep_new_huge_page(h, page, nid);
1346         }
1347
1348         return page;
1349 }
1350
1351 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1352 {
1353         struct page *page;
1354         int nr_nodes, node;
1355         int ret = 0;
1356
1357         for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1358                 page = alloc_fresh_huge_page_node(h, node);
1359                 if (page) {
1360                         ret = 1;
1361                         break;
1362                 }
1363         }
1364
1365         if (ret)
1366                 count_vm_event(HTLB_BUDDY_PGALLOC);
1367         else
1368                 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1369
1370         return ret;
1371 }
1372
1373 /*
1374  * Free huge page from pool from next node to free.
1375  * Attempt to keep persistent huge pages more or less
1376  * balanced over allowed nodes.
1377  * Called with hugetlb_lock locked.
1378  */
1379 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1380                                                          bool acct_surplus)
1381 {
1382         int nr_nodes, node;
1383         int ret = 0;
1384
1385         for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1386                 /*
1387                  * If we're returning unused surplus pages, only examine
1388                  * nodes with surplus pages.
1389                  */
1390                 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1391                     !list_empty(&h->hugepage_freelists[node])) {
1392                         struct page *page =
1393                                 list_entry(h->hugepage_freelists[node].next,
1394                                           struct page, lru);
1395                         list_del(&page->lru);
1396                         h->free_huge_pages--;
1397                         h->free_huge_pages_node[node]--;
1398                         if (acct_surplus) {
1399                                 h->surplus_huge_pages--;
1400                                 h->surplus_huge_pages_node[node]--;
1401                         }
1402                         update_and_free_page(h, page);
1403                         ret = 1;
1404                         break;
1405                 }
1406         }
1407
1408         return ret;
1409 }
1410
1411 /*
1412  * Dissolve a given free hugepage into free buddy pages. This function does
1413  * nothing for in-use (including surplus) hugepages.
1414  */
1415 static void dissolve_free_huge_page(struct page *page)
1416 {
1417         spin_lock(&hugetlb_lock);
1418         if (PageHuge(page) && !page_count(page)) {
1419                 struct hstate *h = page_hstate(page);
1420                 int nid = page_to_nid(page);
1421                 list_del(&page->lru);
1422                 h->free_huge_pages--;
1423                 h->free_huge_pages_node[nid]--;
1424                 update_and_free_page(h, page);
1425         }
1426         spin_unlock(&hugetlb_lock);
1427 }
1428
1429 /*
1430  * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1431  * make specified memory blocks removable from the system.
1432  * Note that start_pfn should aligned with (minimum) hugepage size.
1433  */
1434 void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1435 {
1436         unsigned long pfn;
1437
1438         if (!hugepages_supported())
1439                 return;
1440
1441         VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << minimum_order));
1442         for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order)
1443                 dissolve_free_huge_page(pfn_to_page(pfn));
1444 }
1445
1446 /*
1447  * There are 3 ways this can get called:
1448  * 1. With vma+addr: we use the VMA's memory policy
1449  * 2. With !vma, but nid=NUMA_NO_NODE:  We try to allocate a huge
1450  *    page from any node, and let the buddy allocator itself figure
1451  *    it out.
1452  * 3. With !vma, but nid!=NUMA_NO_NODE.  We allocate a huge page
1453  *    strictly from 'nid'
1454  */
1455 static struct page *__hugetlb_alloc_buddy_huge_page(struct hstate *h,
1456                 struct vm_area_struct *vma, unsigned long addr, int nid)
1457 {
1458         int order = huge_page_order(h);
1459         gfp_t gfp = htlb_alloc_mask(h)|__GFP_COMP|__GFP_REPEAT|__GFP_NOWARN;
1460         unsigned int cpuset_mems_cookie;
1461
1462         /*
1463          * We need a VMA to get a memory policy.  If we do not
1464          * have one, we use the 'nid' argument.
1465          *
1466          * The mempolicy stuff below has some non-inlined bits
1467          * and calls ->vm_ops.  That makes it hard to optimize at
1468          * compile-time, even when NUMA is off and it does
1469          * nothing.  This helps the compiler optimize it out.
1470          */
1471         if (!IS_ENABLED(CONFIG_NUMA) || !vma) {
1472                 /*
1473                  * If a specific node is requested, make sure to
1474                  * get memory from there, but only when a node
1475                  * is explicitly specified.
1476                  */
1477                 if (nid != NUMA_NO_NODE)
1478                         gfp |= __GFP_THISNODE;
1479                 /*
1480                  * Make sure to call something that can handle
1481                  * nid=NUMA_NO_NODE
1482                  */
1483                 return alloc_pages_node(nid, gfp, order);
1484         }
1485
1486         /*
1487          * OK, so we have a VMA.  Fetch the mempolicy and try to
1488          * allocate a huge page with it.  We will only reach this
1489          * when CONFIG_NUMA=y.
1490          */
1491         do {
1492                 struct page *page;
1493                 struct mempolicy *mpol;
1494                 struct zonelist *zl;
1495                 nodemask_t *nodemask;
1496
1497                 cpuset_mems_cookie = read_mems_allowed_begin();
1498                 zl = huge_zonelist(vma, addr, gfp, &mpol, &nodemask);
1499                 mpol_cond_put(mpol);
1500                 page = __alloc_pages_nodemask(gfp, order, zl, nodemask);
1501                 if (page)
1502                         return page;
1503         } while (read_mems_allowed_retry(cpuset_mems_cookie));
1504
1505         return NULL;
1506 }
1507
1508 /*
1509  * There are two ways to allocate a huge page:
1510  * 1. When you have a VMA and an address (like a fault)
1511  * 2. When you have no VMA (like when setting /proc/.../nr_hugepages)
1512  *
1513  * 'vma' and 'addr' are only for (1).  'nid' is always NUMA_NO_NODE in
1514  * this case which signifies that the allocation should be done with
1515  * respect for the VMA's memory policy.
1516  *
1517  * For (2), we ignore 'vma' and 'addr' and use 'nid' exclusively. This
1518  * implies that memory policies will not be taken in to account.
1519  */
1520 static struct page *__alloc_buddy_huge_page(struct hstate *h,
1521                 struct vm_area_struct *vma, unsigned long addr, int nid)
1522 {
1523         struct page *page;
1524         unsigned int r_nid;
1525
1526         if (hstate_is_gigantic(h))
1527                 return NULL;
1528
1529         /*
1530          * Make sure that anyone specifying 'nid' is not also specifying a VMA.
1531          * This makes sure the caller is picking _one_ of the modes with which
1532          * we can call this function, not both.
1533          */
1534         if (vma || (addr != -1)) {
1535                 VM_WARN_ON_ONCE(addr == -1);
1536                 VM_WARN_ON_ONCE(nid != NUMA_NO_NODE);
1537         }
1538         /*
1539          * Assume we will successfully allocate the surplus page to
1540          * prevent racing processes from causing the surplus to exceed
1541          * overcommit
1542          *
1543          * This however introduces a different race, where a process B
1544          * tries to grow the static hugepage pool while alloc_pages() is
1545          * called by process A. B will only examine the per-node
1546          * counters in determining if surplus huge pages can be
1547          * converted to normal huge pages in adjust_pool_surplus(). A
1548          * won't be able to increment the per-node counter, until the
1549          * lock is dropped by B, but B doesn't drop hugetlb_lock until
1550          * no more huge pages can be converted from surplus to normal
1551          * state (and doesn't try to convert again). Thus, we have a
1552          * case where a surplus huge page exists, the pool is grown, and
1553          * the surplus huge page still exists after, even though it
1554          * should just have been converted to a normal huge page. This
1555          * does not leak memory, though, as the hugepage will be freed
1556          * once it is out of use. It also does not allow the counters to
1557          * go out of whack in adjust_pool_surplus() as we don't modify
1558          * the node values until we've gotten the hugepage and only the
1559          * per-node value is checked there.
1560          */
1561         spin_lock(&hugetlb_lock);
1562         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1563                 spin_unlock(&hugetlb_lock);
1564                 return NULL;
1565         } else {
1566                 h->nr_huge_pages++;
1567                 h->surplus_huge_pages++;
1568         }
1569         spin_unlock(&hugetlb_lock);
1570
1571         page = __hugetlb_alloc_buddy_huge_page(h, vma, addr, nid);
1572
1573         spin_lock(&hugetlb_lock);
1574         if (page) {
1575                 INIT_LIST_HEAD(&page->lru);
1576                 r_nid = page_to_nid(page);
1577                 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1578                 set_hugetlb_cgroup(page, NULL);
1579                 /*
1580                  * We incremented the global counters already
1581                  */
1582                 h->nr_huge_pages_node[r_nid]++;
1583                 h->surplus_huge_pages_node[r_nid]++;
1584                 __count_vm_event(HTLB_BUDDY_PGALLOC);
1585         } else {
1586                 h->nr_huge_pages--;
1587                 h->surplus_huge_pages--;
1588                 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1589         }
1590         spin_unlock(&hugetlb_lock);
1591
1592         return page;
1593 }
1594
1595 /*
1596  * Allocate a huge page from 'nid'.  Note, 'nid' may be
1597  * NUMA_NO_NODE, which means that it may be allocated
1598  * anywhere.
1599  */
1600 static
1601 struct page *__alloc_buddy_huge_page_no_mpol(struct hstate *h, int nid)
1602 {
1603         unsigned long addr = -1;
1604
1605         return __alloc_buddy_huge_page(h, NULL, addr, nid);
1606 }
1607
1608 /*
1609  * Use the VMA's mpolicy to allocate a huge page from the buddy.
1610  */
1611 static
1612 struct page *__alloc_buddy_huge_page_with_mpol(struct hstate *h,
1613                 struct vm_area_struct *vma, unsigned long addr)
1614 {
1615         return __alloc_buddy_huge_page(h, vma, addr, NUMA_NO_NODE);
1616 }
1617
1618 /*
1619  * This allocation function is useful in the context where vma is irrelevant.
1620  * E.g. soft-offlining uses this function because it only cares physical
1621  * address of error page.
1622  */
1623 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1624 {
1625         struct page *page = NULL;
1626
1627         spin_lock(&hugetlb_lock);
1628         if (h->free_huge_pages - h->resv_huge_pages > 0)
1629                 page = dequeue_huge_page_node(h, nid);
1630         spin_unlock(&hugetlb_lock);
1631
1632         if (!page)
1633                 page = __alloc_buddy_huge_page_no_mpol(h, nid);
1634
1635         return page;
1636 }
1637
1638 /*
1639  * Increase the hugetlb pool such that it can accommodate a reservation
1640  * of size 'delta'.
1641  */
1642 static int gather_surplus_pages(struct hstate *h, int delta)
1643 {
1644         struct list_head surplus_list;
1645         struct page *page, *tmp;
1646         int ret, i;
1647         int needed, allocated;
1648         bool alloc_ok = true;
1649
1650         needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1651         if (needed <= 0) {
1652                 h->resv_huge_pages += delta;
1653                 return 0;
1654         }
1655
1656         allocated = 0;
1657         INIT_LIST_HEAD(&surplus_list);
1658
1659         ret = -ENOMEM;
1660 retry:
1661         spin_unlock(&hugetlb_lock);
1662         for (i = 0; i < needed; i++) {
1663                 page = __alloc_buddy_huge_page_no_mpol(h, NUMA_NO_NODE);
1664                 if (!page) {
1665                         alloc_ok = false;
1666                         break;
1667                 }
1668                 list_add(&page->lru, &surplus_list);
1669         }
1670         allocated += i;
1671
1672         /*
1673          * After retaking hugetlb_lock, we need to recalculate 'needed'
1674          * because either resv_huge_pages or free_huge_pages may have changed.
1675          */
1676         spin_lock(&hugetlb_lock);
1677         needed = (h->resv_huge_pages + delta) -
1678                         (h->free_huge_pages + allocated);
1679         if (needed > 0) {
1680                 if (alloc_ok)
1681                         goto retry;
1682                 /*
1683                  * We were not able to allocate enough pages to
1684                  * satisfy the entire reservation so we free what
1685                  * we've allocated so far.
1686                  */
1687                 goto free;
1688         }
1689         /*
1690          * The surplus_list now contains _at_least_ the number of extra pages
1691          * needed to accommodate the reservation.  Add the appropriate number
1692          * of pages to the hugetlb pool and free the extras back to the buddy
1693          * allocator.  Commit the entire reservation here to prevent another
1694          * process from stealing the pages as they are added to the pool but
1695          * before they are reserved.
1696          */
1697         needed += allocated;
1698         h->resv_huge_pages += delta;
1699         ret = 0;
1700
1701         /* Free the needed pages to the hugetlb pool */
1702         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1703                 if ((--needed) < 0)
1704                         break;
1705                 /*
1706                  * This page is now managed by the hugetlb allocator and has
1707                  * no users -- drop the buddy allocator's reference.
1708                  */
1709                 put_page_testzero(page);
1710                 VM_BUG_ON_PAGE(page_count(page), page);
1711                 enqueue_huge_page(h, page);
1712         }
1713 free:
1714         spin_unlock(&hugetlb_lock);
1715
1716         /* Free unnecessary surplus pages to the buddy allocator */
1717         list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1718                 put_page(page);
1719         spin_lock(&hugetlb_lock);
1720
1721         return ret;
1722 }
1723
1724 /*
1725  * When releasing a hugetlb pool reservation, any surplus pages that were
1726  * allocated to satisfy the reservation must be explicitly freed if they were
1727  * never used.
1728  * Called with hugetlb_lock held.
1729  */
1730 static void return_unused_surplus_pages(struct hstate *h,
1731                                         unsigned long unused_resv_pages)
1732 {
1733         unsigned long nr_pages;
1734
1735         /* Uncommit the reservation */
1736         h->resv_huge_pages -= unused_resv_pages;
1737
1738         /* Cannot return gigantic pages currently */
1739         if (hstate_is_gigantic(h))
1740                 return;
1741
1742         nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1743
1744         /*
1745          * We want to release as many surplus pages as possible, spread
1746          * evenly across all nodes with memory. Iterate across these nodes
1747          * until we can no longer free unreserved surplus pages. This occurs
1748          * when the nodes with surplus pages have no free pages.
1749          * free_pool_huge_page() will balance the the freed pages across the
1750          * on-line nodes with memory and will handle the hstate accounting.
1751          */
1752         while (nr_pages--) {
1753                 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1754                         break;
1755                 cond_resched_lock(&hugetlb_lock);
1756         }
1757 }
1758
1759
1760 /*
1761  * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1762  * are used by the huge page allocation routines to manage reservations.
1763  *
1764  * vma_needs_reservation is called to determine if the huge page at addr
1765  * within the vma has an associated reservation.  If a reservation is
1766  * needed, the value 1 is returned.  The caller is then responsible for
1767  * managing the global reservation and subpool usage counts.  After
1768  * the huge page has been allocated, vma_commit_reservation is called
1769  * to add the page to the reservation map.  If the page allocation fails,
1770  * the reservation must be ended instead of committed.  vma_end_reservation
1771  * is called in such cases.
1772  *
1773  * In the normal case, vma_commit_reservation returns the same value
1774  * as the preceding vma_needs_reservation call.  The only time this
1775  * is not the case is if a reserve map was changed between calls.  It
1776  * is the responsibility of the caller to notice the difference and
1777  * take appropriate action.
1778  */
1779 enum vma_resv_mode {
1780         VMA_NEEDS_RESV,
1781         VMA_COMMIT_RESV,
1782         VMA_END_RESV,
1783 };
1784 static long __vma_reservation_common(struct hstate *h,
1785                                 struct vm_area_struct *vma, unsigned long addr,
1786                                 enum vma_resv_mode mode)
1787 {
1788         struct resv_map *resv;
1789         pgoff_t idx;
1790         long ret;
1791
1792         resv = vma_resv_map(vma);
1793         if (!resv)
1794                 return 1;
1795
1796         idx = vma_hugecache_offset(h, vma, addr);
1797         switch (mode) {
1798         case VMA_NEEDS_RESV:
1799                 ret = region_chg(resv, idx, idx + 1);
1800                 break;
1801         case VMA_COMMIT_RESV:
1802                 ret = region_add(resv, idx, idx + 1);
1803                 break;
1804         case VMA_END_RESV:
1805                 region_abort(resv, idx, idx + 1);
1806                 ret = 0;
1807                 break;
1808         default:
1809                 BUG();
1810         }
1811
1812         if (vma->vm_flags & VM_MAYSHARE)
1813                 return ret;
1814         else
1815                 return ret < 0 ? ret : 0;
1816 }
1817
1818 static long vma_needs_reservation(struct hstate *h,
1819                         struct vm_area_struct *vma, unsigned long addr)
1820 {
1821         return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
1822 }
1823
1824 static long vma_commit_reservation(struct hstate *h,
1825                         struct vm_area_struct *vma, unsigned long addr)
1826 {
1827         return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1828 }
1829
1830 static void vma_end_reservation(struct hstate *h,
1831                         struct vm_area_struct *vma, unsigned long addr)
1832 {
1833         (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
1834 }
1835
1836 struct page *alloc_huge_page(struct vm_area_struct *vma,
1837                                     unsigned long addr, int avoid_reserve)
1838 {
1839         struct hugepage_subpool *spool = subpool_vma(vma);
1840         struct hstate *h = hstate_vma(vma);
1841         struct page *page;
1842         long map_chg, map_commit;
1843         long gbl_chg;
1844         int ret, idx;
1845         struct hugetlb_cgroup *h_cg;
1846
1847         idx = hstate_index(h);
1848         /*
1849          * Examine the region/reserve map to determine if the process
1850          * has a reservation for the page to be allocated.  A return
1851          * code of zero indicates a reservation exists (no change).
1852          */
1853         map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
1854         if (map_chg < 0)
1855                 return ERR_PTR(-ENOMEM);
1856
1857         /*
1858          * Processes that did not create the mapping will have no
1859          * reserves as indicated by the region/reserve map. Check
1860          * that the allocation will not exceed the subpool limit.
1861          * Allocations for MAP_NORESERVE mappings also need to be
1862          * checked against any subpool limit.
1863          */
1864         if (map_chg || avoid_reserve) {
1865                 gbl_chg = hugepage_subpool_get_pages(spool, 1);
1866                 if (gbl_chg < 0) {
1867                         vma_end_reservation(h, vma, addr);
1868                         return ERR_PTR(-ENOSPC);
1869                 }
1870
1871                 /*
1872                  * Even though there was no reservation in the region/reserve
1873                  * map, there could be reservations associated with the
1874                  * subpool that can be used.  This would be indicated if the
1875                  * return value of hugepage_subpool_get_pages() is zero.
1876                  * However, if avoid_reserve is specified we still avoid even
1877                  * the subpool reservations.
1878                  */
1879                 if (avoid_reserve)
1880                         gbl_chg = 1;
1881         }
1882
1883         ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1884         if (ret)
1885                 goto out_subpool_put;
1886
1887         spin_lock(&hugetlb_lock);
1888         /*
1889          * glb_chg is passed to indicate whether or not a page must be taken
1890          * from the global free pool (global change).  gbl_chg == 0 indicates
1891          * a reservation exists for the allocation.
1892          */
1893         page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
1894         if (!page) {
1895                 spin_unlock(&hugetlb_lock);
1896                 page = __alloc_buddy_huge_page_with_mpol(h, vma, addr);
1897                 if (!page)
1898                         goto out_uncharge_cgroup;
1899                 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
1900                         SetPagePrivate(page);
1901                         h->resv_huge_pages--;
1902                 }
1903                 spin_lock(&hugetlb_lock);
1904                 list_move(&page->lru, &h->hugepage_activelist);
1905                 /* Fall through */
1906         }
1907         hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1908         spin_unlock(&hugetlb_lock);
1909
1910         set_page_private(page, (unsigned long)spool);
1911
1912         map_commit = vma_commit_reservation(h, vma, addr);
1913         if (unlikely(map_chg > map_commit)) {
1914                 /*
1915                  * The page was added to the reservation map between
1916                  * vma_needs_reservation and vma_commit_reservation.
1917                  * This indicates a race with hugetlb_reserve_pages.
1918                  * Adjust for the subpool count incremented above AND
1919                  * in hugetlb_reserve_pages for the same page.  Also,
1920                  * the reservation count added in hugetlb_reserve_pages
1921                  * no longer applies.
1922                  */
1923                 long rsv_adjust;
1924
1925                 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
1926                 hugetlb_acct_memory(h, -rsv_adjust);
1927         }
1928         return page;
1929
1930 out_uncharge_cgroup:
1931         hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
1932 out_subpool_put:
1933         if (map_chg || avoid_reserve)
1934                 hugepage_subpool_put_pages(spool, 1);
1935         vma_end_reservation(h, vma, addr);
1936         return ERR_PTR(-ENOSPC);
1937 }
1938
1939 /*
1940  * alloc_huge_page()'s wrapper which simply returns the page if allocation
1941  * succeeds, otherwise NULL. This function is called from new_vma_page(),
1942  * where no ERR_VALUE is expected to be returned.
1943  */
1944 struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
1945                                 unsigned long addr, int avoid_reserve)
1946 {
1947         struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
1948         if (IS_ERR(page))
1949                 page = NULL;
1950         return page;
1951 }
1952
1953 int __weak alloc_bootmem_huge_page(struct hstate *h)
1954 {
1955         struct huge_bootmem_page *m;
1956         int nr_nodes, node;
1957
1958         for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
1959                 void *addr;
1960
1961                 addr = memblock_virt_alloc_try_nid_nopanic(
1962                                 huge_page_size(h), huge_page_size(h),
1963                                 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
1964                 if (addr) {
1965                         /*
1966                          * Use the beginning of the huge page to store the
1967                          * huge_bootmem_page struct (until gather_bootmem
1968                          * puts them into the mem_map).
1969                          */
1970                         m = addr;
1971                         goto found;
1972                 }
1973         }
1974         return 0;
1975
1976 found:
1977         BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
1978         /* Put them into a private list first because mem_map is not up yet */
1979         list_add(&m->list, &huge_boot_pages);
1980         m->hstate = h;
1981         return 1;
1982 }
1983
1984 static void __init prep_compound_huge_page(struct page *page,
1985                 unsigned int order)
1986 {
1987         if (unlikely(order > (MAX_ORDER - 1)))
1988                 prep_compound_gigantic_page(page, order);
1989         else
1990                 prep_compound_page(page, order);
1991 }
1992
1993 /* Put bootmem huge pages into the standard lists after mem_map is up */
1994 static void __init gather_bootmem_prealloc(void)
1995 {
1996         struct huge_bootmem_page *m;
1997
1998         list_for_each_entry(m, &huge_boot_pages, list) {
1999                 struct hstate *h = m->hstate;
2000                 struct page *page;
2001
2002 #ifdef CONFIG_HIGHMEM
2003                 page = pfn_to_page(m->phys >> PAGE_SHIFT);
2004                 memblock_free_late(__pa(m),
2005                                    sizeof(struct huge_bootmem_page));
2006 #else
2007                 page = virt_to_page(m);
2008 #endif
2009                 WARN_ON(page_count(page) != 1);
2010                 prep_compound_huge_page(page, h->order);
2011                 WARN_ON(PageReserved(page));
2012                 prep_new_huge_page(h, page, page_to_nid(page));
2013                 /*
2014                  * If we had gigantic hugepages allocated at boot time, we need
2015                  * to restore the 'stolen' pages to totalram_pages in order to
2016                  * fix confusing memory reports from free(1) and another
2017                  * side-effects, like CommitLimit going negative.
2018                  */
2019                 if (hstate_is_gigantic(h))
2020                         adjust_managed_page_count(page, 1 << h->order);
2021         }
2022 }
2023
2024 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2025 {
2026         unsigned long i;
2027
2028         for (i = 0; i < h->max_huge_pages; ++i) {
2029                 if (hstate_is_gigantic(h)) {
2030                         if (!alloc_bootmem_huge_page(h))
2031                                 break;
2032                 } else if (!alloc_fresh_huge_page(h,
2033                                          &node_states[N_MEMORY]))
2034                         break;
2035         }
2036         h->max_huge_pages = i;
2037 }
2038
2039 static void __init hugetlb_init_hstates(void)
2040 {
2041         struct hstate *h;
2042
2043         for_each_hstate(h) {
2044                 if (minimum_order > huge_page_order(h))
2045                         minimum_order = huge_page_order(h);
2046
2047                 /* oversize hugepages were init'ed in early boot */
2048                 if (!hstate_is_gigantic(h))
2049                         hugetlb_hstate_alloc_pages(h);
2050         }
2051         VM_BUG_ON(minimum_order == UINT_MAX);
2052 }
2053
2054 static char * __init memfmt(char *buf, unsigned long n)
2055 {
2056         if (n >= (1UL << 30))
2057                 sprintf(buf, "%lu GB", n >> 30);
2058         else if (n >= (1UL << 20))
2059                 sprintf(buf, "%lu MB", n >> 20);
2060         else
2061                 sprintf(buf, "%lu KB", n >> 10);
2062         return buf;
2063 }
2064
2065 static void __init report_hugepages(void)
2066 {
2067         struct hstate *h;
2068
2069         for_each_hstate(h) {
2070                 char buf[32];
2071                 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2072                         memfmt(buf, huge_page_size(h)),
2073                         h->free_huge_pages);
2074         }
2075 }
2076
2077 #ifdef CONFIG_HIGHMEM
2078 static void try_to_free_low(struct hstate *h, unsigned long count,
2079                                                 nodemask_t *nodes_allowed)
2080 {
2081         int i;
2082
2083         if (hstate_is_gigantic(h))
2084                 return;
2085
2086         for_each_node_mask(i, *nodes_allowed) {
2087                 struct page *page, *next;
2088                 struct list_head *freel = &h->hugepage_freelists[i];
2089                 list_for_each_entry_safe(page, next, freel, lru) {
2090                         if (count >= h->nr_huge_pages)
2091                                 return;
2092                         if (PageHighMem(page))
2093                                 continue;
2094                         list_del(&page->lru);
2095                         update_and_free_page(h, page);
2096                         h->free_huge_pages--;
2097                         h->free_huge_pages_node[page_to_nid(page)]--;
2098                 }
2099         }
2100 }
2101 #else
2102 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2103                                                 nodemask_t *nodes_allowed)
2104 {
2105 }
2106 #endif
2107
2108 /*
2109  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
2110  * balanced by operating on them in a round-robin fashion.
2111  * Returns 1 if an adjustment was made.
2112  */
2113 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2114                                 int delta)
2115 {
2116         int nr_nodes, node;
2117
2118         VM_BUG_ON(delta != -1 && delta != 1);
2119
2120         if (delta < 0) {
2121                 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2122                         if (h->surplus_huge_pages_node[node])
2123                                 goto found;
2124                 }
2125         } else {
2126                 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2127                         if (h->surplus_huge_pages_node[node] <
2128                                         h->nr_huge_pages_node[node])
2129                                 goto found;
2130                 }
2131         }
2132         return 0;
2133
2134 found:
2135         h->surplus_huge_pages += delta;
2136         h->surplus_huge_pages_node[node] += delta;
2137         return 1;
2138 }
2139
2140 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2141 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
2142                                                 nodemask_t *nodes_allowed)
2143 {
2144         unsigned long min_count, ret;
2145
2146         if (hstate_is_gigantic(h) && !gigantic_page_supported())
2147                 return h->max_huge_pages;
2148
2149         /*
2150          * Increase the pool size
2151          * First take pages out of surplus state.  Then make up the
2152          * remaining difference by allocating fresh huge pages.
2153          *
2154          * We might race with __alloc_buddy_huge_page() here and be unable
2155          * to convert a surplus huge page to a normal huge page. That is
2156          * not critical, though, it just means the overall size of the
2157          * pool might be one hugepage larger than it needs to be, but
2158          * within all the constraints specified by the sysctls.
2159          */
2160         spin_lock(&hugetlb_lock);
2161         while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2162                 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2163                         break;
2164         }
2165
2166         while (count > persistent_huge_pages(h)) {
2167                 /*
2168                  * If this allocation races such that we no longer need the
2169                  * page, free_huge_page will handle it by freeing the page
2170                  * and reducing the surplus.
2171                  */
2172                 spin_unlock(&hugetlb_lock);
2173                 if (hstate_is_gigantic(h))
2174                         ret = alloc_fresh_gigantic_page(h, nodes_allowed);
2175                 else
2176                         ret = alloc_fresh_huge_page(h, nodes_allowed);
2177                 spin_lock(&hugetlb_lock);
2178                 if (!ret)
2179                         goto out;
2180
2181                 /* Bail for signals. Probably ctrl-c from user */
2182                 if (signal_pending(current))
2183                         goto out;
2184         }
2185
2186         /*
2187          * Decrease the pool size
2188          * First return free pages to the buddy allocator (being careful
2189          * to keep enough around to satisfy reservations).  Then place
2190          * pages into surplus state as needed so the pool will shrink
2191          * to the desired size as pages become free.
2192          *
2193          * By placing pages into the surplus state independent of the
2194          * overcommit value, we are allowing the surplus pool size to
2195          * exceed overcommit. There are few sane options here. Since
2196          * __alloc_buddy_huge_page() is checking the global counter,
2197          * though, we'll note that we're not allowed to exceed surplus
2198          * and won't grow the pool anywhere else. Not until one of the
2199          * sysctls are changed, or the surplus pages go out of use.
2200          */
2201         min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2202         min_count = max(count, min_count);
2203         try_to_free_low(h, min_count, nodes_allowed);
2204         while (min_count < persistent_huge_pages(h)) {
2205                 if (!free_pool_huge_page(h, nodes_allowed, 0))
2206                         break;
2207                 cond_resched_lock(&hugetlb_lock);
2208         }
2209         while (count < persistent_huge_pages(h)) {
2210                 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2211                         break;
2212         }
2213 out:
2214         ret = persistent_huge_pages(h);
2215         spin_unlock(&hugetlb_lock);
2216         return ret;
2217 }
2218
2219 #define HSTATE_ATTR_RO(_name) \
2220         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2221
2222 #define HSTATE_ATTR(_name) \
2223         static struct kobj_attribute _name##_attr = \
2224                 __ATTR(_name, 0644, _name##_show, _name##_store)
2225
2226 static struct kobject *hugepages_kobj;
2227 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2228
2229 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2230
2231 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2232 {
2233         int i;
2234
2235         for (i = 0; i < HUGE_MAX_HSTATE; i++)
2236                 if (hstate_kobjs[i] == kobj) {
2237                         if (nidp)
2238                                 *nidp = NUMA_NO_NODE;
2239                         return &hstates[i];
2240                 }
2241
2242         return kobj_to_node_hstate(kobj, nidp);
2243 }
2244
2245 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2246                                         struct kobj_attribute *attr, char *buf)
2247 {
2248         struct hstate *h;
2249         unsigned long nr_huge_pages;
2250         int nid;
2251
2252         h = kobj_to_hstate(kobj, &nid);
2253         if (nid == NUMA_NO_NODE)
2254                 nr_huge_pages = h->nr_huge_pages;
2255         else
2256                 nr_huge_pages = h->nr_huge_pages_node[nid];
2257
2258         return sprintf(buf, "%lu\n", nr_huge_pages);
2259 }
2260
2261 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2262                                            struct hstate *h, int nid,
2263                                            unsigned long count, size_t len)
2264 {
2265         int err;
2266         NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
2267
2268         if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
2269                 err = -EINVAL;
2270                 goto out;
2271         }
2272
2273         if (nid == NUMA_NO_NODE) {
2274                 /*
2275                  * global hstate attribute
2276                  */
2277                 if (!(obey_mempolicy &&
2278                                 init_nodemask_of_mempolicy(nodes_allowed))) {
2279                         NODEMASK_FREE(nodes_allowed);
2280                         nodes_allowed = &node_states[N_MEMORY];
2281                 }
2282         } else if (nodes_allowed) {
2283                 /*
2284                  * per node hstate attribute: adjust count to global,
2285                  * but restrict alloc/free to the specified node.
2286                  */
2287                 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2288                 init_nodemask_of_node(nodes_allowed, nid);
2289         } else
2290                 nodes_allowed = &node_states[N_MEMORY];
2291
2292         h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
2293
2294         if (nodes_allowed != &node_states[N_MEMORY])
2295                 NODEMASK_FREE(nodes_allowed);
2296
2297         return len;
2298 out:
2299         NODEMASK_FREE(nodes_allowed);
2300         return err;
2301 }
2302
2303 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2304                                          struct kobject *kobj, const char *buf,
2305                                          size_t len)
2306 {
2307         struct hstate *h;
2308         unsigned long count;
2309         int nid;
2310         int err;
2311
2312         err = kstrtoul(buf, 10, &count);
2313         if (err)
2314                 return err;
2315
2316         h = kobj_to_hstate(kobj, &nid);
2317         return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2318 }
2319
2320 static ssize_t nr_hugepages_show(struct kobject *kobj,
2321                                        struct kobj_attribute *attr, char *buf)
2322 {
2323         return nr_hugepages_show_common(kobj, attr, buf);
2324 }
2325
2326 static ssize_t nr_hugepages_store(struct kobject *kobj,
2327                struct kobj_attribute *attr, const char *buf, size_t len)
2328 {
2329         return nr_hugepages_store_common(false, kobj, buf, len);
2330 }
2331 HSTATE_ATTR(nr_hugepages);
2332
2333 #ifdef CONFIG_NUMA
2334
2335 /*
2336  * hstate attribute for optionally mempolicy-based constraint on persistent
2337  * huge page alloc/free.
2338  */
2339 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2340                                        struct kobj_attribute *attr, char *buf)
2341 {
2342         return nr_hugepages_show_common(kobj, attr, buf);
2343 }
2344
2345 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2346                struct kobj_attribute *attr, const char *buf, size_t len)
2347 {
2348         return nr_hugepages_store_common(true, kobj, buf, len);
2349 }
2350 HSTATE_ATTR(nr_hugepages_mempolicy);
2351 #endif
2352
2353
2354 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2355                                         struct kobj_attribute *attr, char *buf)
2356 {
2357         struct hstate *h = kobj_to_hstate(kobj, NULL);
2358         return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2359 }
2360
2361 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2362                 struct kobj_attribute *attr, const char *buf, size_t count)
2363 {
2364         int err;
2365         unsigned long input;
2366         struct hstate *h = kobj_to_hstate(kobj, NULL);
2367
2368         if (hstate_is_gigantic(h))
2369                 return -EINVAL;
2370
2371         err = kstrtoul(buf, 10, &input);
2372         if (err)
2373                 return err;
2374
2375         spin_lock(&hugetlb_lock);
2376         h->nr_overcommit_huge_pages = input;
2377         spin_unlock(&hugetlb_lock);
2378
2379         return count;
2380 }
2381 HSTATE_ATTR(nr_overcommit_hugepages);
2382
2383 static ssize_t free_hugepages_show(struct kobject *kobj,
2384                                         struct kobj_attribute *attr, char *buf)
2385 {
2386         struct hstate *h;
2387         unsigned long free_huge_pages;
2388         int nid;
2389
2390         h = kobj_to_hstate(kobj, &nid);
2391         if (nid == NUMA_NO_NODE)
2392                 free_huge_pages = h->free_huge_pages;
2393         else
2394                 free_huge_pages = h->free_huge_pages_node[nid];
2395
2396         return sprintf(buf, "%lu\n", free_huge_pages);
2397 }
2398 HSTATE_ATTR_RO(free_hugepages);
2399
2400 static ssize_t resv_hugepages_show(struct kobject *kobj,
2401                                         struct kobj_attribute *attr, char *buf)
2402 {
2403         struct hstate *h = kobj_to_hstate(kobj, NULL);
2404         return sprintf(buf, "%lu\n", h->resv_huge_pages);
2405 }
2406 HSTATE_ATTR_RO(resv_hugepages);
2407
2408 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2409                                         struct kobj_attribute *attr, char *buf)
2410 {
2411         struct hstate *h;
2412         unsigned long surplus_huge_pages;
2413         int nid;
2414
2415         h = kobj_to_hstate(kobj, &nid);
2416         if (nid == NUMA_NO_NODE)
2417                 surplus_huge_pages = h->surplus_huge_pages;
2418         else
2419                 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2420
2421         return sprintf(buf, "%lu\n", surplus_huge_pages);
2422 }
2423 HSTATE_ATTR_RO(surplus_hugepages);
2424
2425 static struct attribute *hstate_attrs[] = {
2426         &nr_hugepages_attr.attr,
2427         &nr_overcommit_hugepages_attr.attr,
2428         &free_hugepages_attr.attr,
2429         &resv_hugepages_attr.attr,
2430         &surplus_hugepages_attr.attr,
2431 #ifdef CONFIG_NUMA
2432         &nr_hugepages_mempolicy_attr.attr,
2433 #endif
2434         NULL,
2435 };
2436
2437 static struct attribute_group hstate_attr_group = {
2438         .attrs = hstate_attrs,
2439 };
2440
2441 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2442                                     struct kobject **hstate_kobjs,
2443                                     struct attribute_group *hstate_attr_group)
2444 {
2445         int retval;
2446         int hi = hstate_index(h);
2447
2448         hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2449         if (!hstate_kobjs[hi])
2450                 return -ENOMEM;
2451
2452         retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2453         if (retval)
2454                 kobject_put(hstate_kobjs[hi]);
2455
2456         return retval;
2457 }
2458
2459 static void __init hugetlb_sysfs_init(void)
2460 {
2461         struct hstate *h;
2462         int err;
2463
2464         hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2465         if (!hugepages_kobj)
2466                 return;
2467
2468         for_each_hstate(h) {
2469                 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2470                                          hstate_kobjs, &hstate_attr_group);
2471                 if (err)
2472                         pr_err("Hugetlb: Unable to add hstate %s", h->name);
2473         }
2474 }
2475
2476 #ifdef CONFIG_NUMA
2477
2478 /*
2479  * node_hstate/s - associate per node hstate attributes, via their kobjects,
2480  * with node devices in node_devices[] using a parallel array.  The array
2481  * index of a node device or _hstate == node id.
2482  * This is here to avoid any static dependency of the node device driver, in
2483  * the base kernel, on the hugetlb module.
2484  */
2485 struct node_hstate {
2486         struct kobject          *hugepages_kobj;
2487         struct kobject          *hstate_kobjs[HUGE_MAX_HSTATE];
2488 };
2489 static struct node_hstate node_hstates[MAX_NUMNODES];
2490
2491 /*
2492  * A subset of global hstate attributes for node devices
2493  */
2494 static struct attribute *per_node_hstate_attrs[] = {
2495         &nr_hugepages_attr.attr,
2496         &free_hugepages_attr.attr,
2497         &surplus_hugepages_attr.attr,
2498         NULL,
2499 };
2500
2501 static struct attribute_group per_node_hstate_attr_group = {
2502         .attrs = per_node_hstate_attrs,
2503 };
2504
2505 /*
2506  * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2507  * Returns node id via non-NULL nidp.
2508  */
2509 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2510 {
2511         int nid;
2512
2513         for (nid = 0; nid < nr_node_ids; nid++) {
2514                 struct node_hstate *nhs = &node_hstates[nid];
2515                 int i;
2516                 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2517                         if (nhs->hstate_kobjs[i] == kobj) {
2518                                 if (nidp)
2519                                         *nidp = nid;
2520                                 return &hstates[i];
2521                         }
2522         }
2523
2524         BUG();
2525         return NULL;
2526 }
2527
2528 /*
2529  * Unregister hstate attributes from a single node device.
2530  * No-op if no hstate attributes attached.
2531  */
2532 static void hugetlb_unregister_node(struct node *node)
2533 {
2534         struct hstate *h;
2535         struct node_hstate *nhs = &node_hstates[node->dev.id];
2536
2537         if (!nhs->hugepages_kobj)
2538                 return;         /* no hstate attributes */
2539
2540         for_each_hstate(h) {
2541                 int idx = hstate_index(h);
2542                 if (nhs->hstate_kobjs[idx]) {
2543                         kobject_put(nhs->hstate_kobjs[idx]);
2544                         nhs->hstate_kobjs[idx] = NULL;
2545                 }
2546         }
2547
2548         kobject_put(nhs->hugepages_kobj);
2549         nhs->hugepages_kobj = NULL;
2550 }
2551
2552 /*
2553  * hugetlb module exit:  unregister hstate attributes from node devices
2554  * that have them.
2555  */
2556 static void hugetlb_unregister_all_nodes(void)
2557 {
2558         int nid;
2559
2560         /*
2561          * disable node device registrations.
2562          */
2563         register_hugetlbfs_with_node(NULL, NULL);
2564
2565         /*
2566          * remove hstate attributes from any nodes that have them.
2567          */
2568         for (nid = 0; nid < nr_node_ids; nid++)
2569                 hugetlb_unregister_node(node_devices[nid]);
2570 }
2571
2572 /*
2573  * Register hstate attributes for a single node device.
2574  * No-op if attributes already registered.
2575  */
2576 static void hugetlb_register_node(struct node *node)
2577 {
2578         struct hstate *h;
2579         struct node_hstate *nhs = &node_hstates[node->dev.id];
2580         int err;
2581
2582         if (nhs->hugepages_kobj)
2583                 return;         /* already allocated */
2584
2585         nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2586                                                         &node->dev.kobj);
2587         if (!nhs->hugepages_kobj)
2588                 return;
2589
2590         for_each_hstate(h) {
2591                 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2592                                                 nhs->hstate_kobjs,
2593                                                 &per_node_hstate_attr_group);
2594                 if (err) {
2595                         pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2596                                 h->name, node->dev.id);
2597                         hugetlb_unregister_node(node);
2598                         break;
2599                 }
2600         }
2601 }
2602
2603 /*
2604  * hugetlb init time:  register hstate attributes for all registered node
2605  * devices of nodes that have memory.  All on-line nodes should have
2606  * registered their associated device by this time.
2607  */
2608 static void __init hugetlb_register_all_nodes(void)
2609 {
2610         int nid;
2611
2612         for_each_node_state(nid, N_MEMORY) {
2613                 struct node *node = node_devices[nid];
2614                 if (node->dev.id == nid)
2615                         hugetlb_register_node(node);
2616         }
2617
2618         /*
2619          * Let the node device driver know we're here so it can
2620          * [un]register hstate attributes on node hotplug.
2621          */
2622         register_hugetlbfs_with_node(hugetlb_register_node,
2623                                      hugetlb_unregister_node);
2624 }
2625 #else   /* !CONFIG_NUMA */
2626
2627 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2628 {
2629         BUG();
2630         if (nidp)
2631                 *nidp = -1;
2632         return NULL;
2633 }
2634
2635 static void hugetlb_unregister_all_nodes(void) { }
2636
2637 static void hugetlb_register_all_nodes(void) { }
2638
2639 #endif
2640
2641 static void __exit hugetlb_exit(void)
2642 {
2643         struct hstate *h;
2644
2645         hugetlb_unregister_all_nodes();
2646
2647         for_each_hstate(h) {
2648                 kobject_put(hstate_kobjs[hstate_index(h)]);
2649         }
2650
2651         kobject_put(hugepages_kobj);
2652         kfree(hugetlb_fault_mutex_table);
2653 }
2654 module_exit(hugetlb_exit);
2655
2656 static int __init hugetlb_init(void)
2657 {
2658         int i;
2659
2660         if (!hugepages_supported())
2661                 return 0;
2662
2663         if (!size_to_hstate(default_hstate_size)) {
2664                 default_hstate_size = HPAGE_SIZE;
2665                 if (!size_to_hstate(default_hstate_size))
2666                         hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2667         }
2668         default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2669         if (default_hstate_max_huge_pages)
2670                 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2671
2672         hugetlb_init_hstates();
2673         gather_bootmem_prealloc();
2674         report_hugepages();
2675
2676         hugetlb_sysfs_init();
2677         hugetlb_register_all_nodes();
2678         hugetlb_cgroup_file_init();
2679
2680 #ifdef CONFIG_SMP
2681         num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2682 #else
2683         num_fault_mutexes = 1;
2684 #endif
2685         hugetlb_fault_mutex_table =
2686                 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
2687         BUG_ON(!hugetlb_fault_mutex_table);
2688
2689         for (i = 0; i < num_fault_mutexes; i++)
2690                 mutex_init(&hugetlb_fault_mutex_table[i]);
2691         return 0;
2692 }
2693 module_init(hugetlb_init);
2694
2695 /* Should be called on processing a hugepagesz=... option */
2696 void __init hugetlb_add_hstate(unsigned int order)
2697 {
2698         struct hstate *h;
2699         unsigned long i;
2700
2701         if (size_to_hstate(PAGE_SIZE << order)) {
2702                 pr_warning("hugepagesz= specified twice, ignoring\n");
2703                 return;
2704         }
2705         BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2706         BUG_ON(order == 0);
2707         h = &hstates[hugetlb_max_hstate++];
2708         h->order = order;
2709         h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2710         h->nr_huge_pages = 0;
2711         h->free_huge_pages = 0;
2712         for (i = 0; i < MAX_NUMNODES; ++i)
2713                 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2714         INIT_LIST_HEAD(&h->hugepage_activelist);
2715         h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
2716         h->next_nid_to_free = first_node(node_states[N_MEMORY]);
2717         snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2718                                         huge_page_size(h)/1024);
2719
2720         parsed_hstate = h;
2721 }
2722
2723 static int __init hugetlb_nrpages_setup(char *s)
2724 {
2725         unsigned long *mhp;
2726         static unsigned long *last_mhp;
2727
2728         /*
2729          * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2730          * so this hugepages= parameter goes to the "default hstate".
2731          */
2732         if (!hugetlb_max_hstate)
2733                 mhp = &default_hstate_max_huge_pages;
2734         else
2735                 mhp = &parsed_hstate->max_huge_pages;
2736
2737         if (mhp == last_mhp) {
2738                 pr_warning("hugepages= specified twice without "
2739                            "interleaving hugepagesz=, ignoring\n");
2740                 return 1;
2741         }
2742
2743         if (sscanf(s, "%lu", mhp) <= 0)
2744                 *mhp = 0;
2745
2746         /*
2747          * Global state is always initialized later in hugetlb_init.
2748          * But we need to allocate >= MAX_ORDER hstates here early to still
2749          * use the bootmem allocator.
2750          */
2751         if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2752                 hugetlb_hstate_alloc_pages(parsed_hstate);
2753
2754         last_mhp = mhp;
2755
2756         return 1;
2757 }
2758 __setup("hugepages=", hugetlb_nrpages_setup);
2759
2760 static int __init hugetlb_default_setup(char *s)
2761 {
2762         default_hstate_size = memparse(s, &s);
2763         return 1;
2764 }
2765 __setup("default_hugepagesz=", hugetlb_default_setup);
2766
2767 static unsigned int cpuset_mems_nr(unsigned int *array)
2768 {
2769         int node;
2770         unsigned int nr = 0;
2771
2772         for_each_node_mask(node, cpuset_current_mems_allowed)
2773                 nr += array[node];
2774
2775         return nr;
2776 }
2777
2778 #ifdef CONFIG_SYSCTL
2779 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2780                          struct ctl_table *table, int write,
2781                          void __user *buffer, size_t *length, loff_t *ppos)
2782 {
2783         struct hstate *h = &default_hstate;
2784         unsigned long tmp = h->max_huge_pages;
2785         int ret;
2786
2787         if (!hugepages_supported())
2788                 return -ENOTSUPP;
2789
2790         table->data = &tmp;
2791         table->maxlen = sizeof(unsigned long);
2792         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2793         if (ret)
2794                 goto out;
2795
2796         if (write)
2797                 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2798                                                   NUMA_NO_NODE, tmp, *length);
2799 out:
2800         return ret;
2801 }
2802
2803 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2804                           void __user *buffer, size_t *length, loff_t *ppos)
2805 {
2806
2807         return hugetlb_sysctl_handler_common(false, table, write,
2808                                                         buffer, length, ppos);
2809 }
2810
2811 #ifdef CONFIG_NUMA
2812 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2813                           void __user *buffer, size_t *length, loff_t *ppos)
2814 {
2815         return hugetlb_sysctl_handler_common(true, table, write,
2816                                                         buffer, length, ppos);
2817 }
2818 #endif /* CONFIG_NUMA */
2819
2820 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2821                         void __user *buffer,
2822                         size_t *length, loff_t *ppos)
2823 {
2824         struct hstate *h = &default_hstate;
2825         unsigned long tmp;
2826         int ret;
2827
2828         if (!hugepages_supported())
2829                 return -ENOTSUPP;
2830
2831         tmp = h->nr_overcommit_huge_pages;
2832
2833         if (write && hstate_is_gigantic(h))
2834                 return -EINVAL;
2835
2836         table->data = &tmp;
2837         table->maxlen = sizeof(unsigned long);
2838         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2839         if (ret)
2840                 goto out;
2841
2842         if (write) {
2843                 spin_lock(&hugetlb_lock);
2844                 h->nr_overcommit_huge_pages = tmp;
2845                 spin_unlock(&hugetlb_lock);
2846         }
2847 out:
2848         return ret;
2849 }
2850
2851 #endif /* CONFIG_SYSCTL */
2852
2853 void hugetlb_report_meminfo(struct seq_file *m)
2854 {
2855         struct hstate *h = &default_hstate;
2856         if (!hugepages_supported())
2857                 return;
2858         seq_printf(m,
2859                         "HugePages_Total:   %5lu\n"
2860                         "HugePages_Free:    %5lu\n"
2861                         "HugePages_Rsvd:    %5lu\n"
2862                         "HugePages_Surp:    %5lu\n"
2863                         "Hugepagesize:   %8lu kB\n",
2864                         h->nr_huge_pages,
2865                         h->free_huge_pages,
2866                         h->resv_huge_pages,
2867                         h->surplus_huge_pages,
2868                         1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2869 }
2870
2871 int hugetlb_report_node_meminfo(int nid, char *buf)
2872 {
2873         struct hstate *h = &default_hstate;
2874         if (!hugepages_supported())
2875                 return 0;
2876         return sprintf(buf,
2877                 "Node %d HugePages_Total: %5u\n"
2878                 "Node %d HugePages_Free:  %5u\n"
2879                 "Node %d HugePages_Surp:  %5u\n",
2880                 nid, h->nr_huge_pages_node[nid],
2881                 nid, h->free_huge_pages_node[nid],
2882                 nid, h->surplus_huge_pages_node[nid]);
2883 }
2884
2885 void hugetlb_show_meminfo(void)
2886 {
2887         struct hstate *h;
2888         int nid;
2889
2890         if (!hugepages_supported())
2891                 return;
2892
2893         for_each_node_state(nid, N_MEMORY)
2894                 for_each_hstate(h)
2895                         pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2896                                 nid,
2897                                 h->nr_huge_pages_node[nid],
2898                                 h->free_huge_pages_node[nid],
2899                                 h->surplus_huge_pages_node[nid],
2900                                 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2901 }
2902
2903 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
2904 {
2905         seq_printf(m, "HugetlbPages:\t%8lu kB\n",
2906                    atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
2907 }
2908
2909 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2910 unsigned long hugetlb_total_pages(void)
2911 {
2912         struct hstate *h;
2913         unsigned long nr_total_pages = 0;
2914
2915         for_each_hstate(h)
2916                 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2917         return nr_total_pages;
2918 }
2919
2920 static int hugetlb_acct_memory(struct hstate *h, long delta)
2921 {
2922         int ret = -ENOMEM;
2923
2924         spin_lock(&hugetlb_lock);
2925         /*
2926          * When cpuset is configured, it breaks the strict hugetlb page
2927          * reservation as the accounting is done on a global variable. Such
2928          * reservation is completely rubbish in the presence of cpuset because
2929          * the reservation is not checked against page availability for the
2930          * current cpuset. Application can still potentially OOM'ed by kernel
2931          * with lack of free htlb page in cpuset that the task is in.
2932          * Attempt to enforce strict accounting with cpuset is almost
2933          * impossible (or too ugly) because cpuset is too fluid that
2934          * task or memory node can be dynamically moved between cpusets.
2935          *
2936          * The change of semantics for shared hugetlb mapping with cpuset is
2937          * undesirable. However, in order to preserve some of the semantics,
2938          * we fall back to check against current free page availability as
2939          * a best attempt and hopefully to minimize the impact of changing
2940          * semantics that cpuset has.
2941          */
2942         if (delta > 0) {
2943                 if (gather_surplus_pages(h, delta) < 0)
2944                         goto out;
2945
2946                 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2947                         return_unused_surplus_pages(h, delta);
2948                         goto out;
2949                 }
2950         }
2951
2952         ret = 0;
2953         if (delta < 0)
2954                 return_unused_surplus_pages(h, (unsigned long) -delta);
2955
2956 out:
2957         spin_unlock(&hugetlb_lock);
2958         return ret;
2959 }
2960
2961 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2962 {
2963         struct resv_map *resv = vma_resv_map(vma);
2964
2965         /*
2966          * This new VMA should share its siblings reservation map if present.
2967          * The VMA will only ever have a valid reservation map pointer where
2968          * it is being copied for another still existing VMA.  As that VMA
2969          * has a reference to the reservation map it cannot disappear until
2970          * after this open call completes.  It is therefore safe to take a
2971          * new reference here without additional locking.
2972          */
2973         if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2974                 kref_get(&resv->refs);
2975 }
2976
2977 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2978 {
2979         struct hstate *h = hstate_vma(vma);
2980         struct resv_map *resv = vma_resv_map(vma);
2981         struct hugepage_subpool *spool = subpool_vma(vma);
2982         unsigned long reserve, start, end;
2983         long gbl_reserve;
2984
2985         if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2986                 return;
2987
2988         start = vma_hugecache_offset(h, vma, vma->vm_start);
2989         end = vma_hugecache_offset(h, vma, vma->vm_end);
2990
2991         reserve = (end - start) - region_count(resv, start, end);
2992
2993         kref_put(&resv->refs, resv_map_release);
2994
2995         if (reserve) {
2996                 /*
2997                  * Decrement reserve counts.  The global reserve count may be
2998                  * adjusted if the subpool has a minimum size.
2999                  */
3000                 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3001                 hugetlb_acct_memory(h, -gbl_reserve);
3002         }
3003 }
3004
3005 /*
3006  * We cannot handle pagefaults against hugetlb pages at all.  They cause
3007  * handle_mm_fault() to try to instantiate regular-sized pages in the
3008  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
3009  * this far.
3010  */
3011 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3012 {
3013         BUG();
3014         return 0;
3015 }
3016
3017 const struct vm_operations_struct hugetlb_vm_ops = {
3018         .fault = hugetlb_vm_op_fault,
3019         .open = hugetlb_vm_op_open,
3020         .close = hugetlb_vm_op_close,
3021 };
3022
3023 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3024                                 int writable)
3025 {
3026         pte_t entry;
3027
3028         if (writable) {
3029                 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3030                                          vma->vm_page_prot)));
3031         } else {
3032                 entry = huge_pte_wrprotect(mk_huge_pte(page,
3033                                            vma->vm_page_prot));
3034         }
3035         entry = pte_mkyoung(entry);
3036         entry = pte_mkhuge(entry);
3037         entry = arch_make_huge_pte(entry, vma, page, writable);
3038
3039         return entry;
3040 }
3041
3042 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3043                                    unsigned long address, pte_t *ptep)
3044 {
3045         pte_t entry;
3046
3047         entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3048         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3049                 update_mmu_cache(vma, address, ptep);
3050 }
3051
3052 static int is_hugetlb_entry_migration(pte_t pte)
3053 {
3054         swp_entry_t swp;
3055
3056         if (huge_pte_none(pte) || pte_present(pte))
3057                 return 0;
3058         swp = pte_to_swp_entry(pte);
3059         if (non_swap_entry(swp) && is_migration_entry(swp))
3060                 return 1;
3061         else
3062                 return 0;
3063 }
3064
3065 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3066 {
3067         swp_entry_t swp;
3068
3069         if (huge_pte_none(pte) || pte_present(pte))
3070                 return 0;
3071         swp = pte_to_swp_entry(pte);
3072         if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3073                 return 1;
3074         else
3075                 return 0;
3076 }
3077
3078 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3079                             struct vm_area_struct *vma)
3080 {
3081         pte_t *src_pte, *dst_pte, entry;
3082         struct page *ptepage;
3083         unsigned long addr;
3084         int cow;
3085         struct hstate *h = hstate_vma(vma);
3086         unsigned long sz = huge_page_size(h);
3087         unsigned long mmun_start;       /* For mmu_notifiers */
3088         unsigned long mmun_end;         /* For mmu_notifiers */
3089         int ret = 0;
3090
3091         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3092
3093         mmun_start = vma->vm_start;
3094         mmun_end = vma->vm_end;
3095         if (cow)
3096                 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
3097
3098         for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3099                 spinlock_t *src_ptl, *dst_ptl;
3100                 src_pte = huge_pte_offset(src, addr);
3101                 if (!src_pte)
3102                         continue;
3103                 dst_pte = huge_pte_alloc(dst, addr, sz);
3104                 if (!dst_pte) {
3105                         ret = -ENOMEM;
3106                         break;
3107                 }
3108
3109                 /* If the pagetables are shared don't copy or take references */
3110                 if (dst_pte == src_pte)
3111                         continue;
3112
3113                 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3114                 src_ptl = huge_pte_lockptr(h, src, src_pte);
3115                 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3116                 entry = huge_ptep_get(src_pte);
3117                 if (huge_pte_none(entry)) { /* skip none entry */
3118                         ;
3119                 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3120                                     is_hugetlb_entry_hwpoisoned(entry))) {
3121                         swp_entry_t swp_entry = pte_to_swp_entry(entry);
3122
3123                         if (is_write_migration_entry(swp_entry) && cow) {
3124                                 /*
3125                                  * COW mappings require pages in both
3126                                  * parent and child to be set to read.
3127                                  */
3128                                 make_migration_entry_read(&swp_entry);
3129                                 entry = swp_entry_to_pte(swp_entry);
3130                                 set_huge_pte_at(src, addr, src_pte, entry);
3131                         }
3132                         set_huge_pte_at(dst, addr, dst_pte, entry);
3133                 } else {
3134                         if (cow) {
3135                                 huge_ptep_set_wrprotect(src, addr, src_pte);
3136                                 mmu_notifier_invalidate_range(src, mmun_start,
3137                                                                    mmun_end);
3138                         }
3139                         entry = huge_ptep_get(src_pte);
3140                         ptepage = pte_page(entry);
3141                         get_page(ptepage);
3142                         page_dup_rmap(ptepage);
3143                         set_huge_pte_at(dst, addr, dst_pte, entry);
3144                         hugetlb_count_add(pages_per_huge_page(h), dst);
3145                 }
3146                 spin_unlock(src_ptl);
3147                 spin_unlock(dst_ptl);
3148         }
3149
3150         if (cow)
3151                 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
3152
3153         return ret;
3154 }
3155
3156 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3157                             unsigned long start, unsigned long end,
3158                             struct page *ref_page)
3159 {
3160         int force_flush = 0;
3161         struct mm_struct *mm = vma->vm_mm;
3162         unsigned long address;
3163         pte_t *ptep;
3164         pte_t pte;
3165         spinlock_t *ptl;
3166         struct page *page;
3167         struct hstate *h = hstate_vma(vma);
3168         unsigned long sz = huge_page_size(h);
3169         const unsigned long mmun_start = start; /* For mmu_notifiers */
3170         const unsigned long mmun_end   = end;   /* For mmu_notifiers */
3171
3172         WARN_ON(!is_vm_hugetlb_page(vma));
3173         BUG_ON(start & ~huge_page_mask(h));
3174         BUG_ON(end & ~huge_page_mask(h));
3175
3176         tlb_start_vma(tlb, vma);
3177         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3178         address = start;
3179 again:
3180         for (; address < end; address += sz) {
3181                 ptep = huge_pte_offset(mm, address);
3182                 if (!ptep)
3183                         continue;
3184
3185                 ptl = huge_pte_lock(h, mm, ptep);
3186                 if (huge_pmd_unshare(mm, &address, ptep))
3187                         goto unlock;
3188
3189                 pte = huge_ptep_get(ptep);
3190                 if (huge_pte_none(pte))
3191                         goto unlock;
3192
3193                 /*
3194                  * Migrating hugepage or HWPoisoned hugepage is already
3195                  * unmapped and its refcount is dropped, so just clear pte here.
3196                  */
3197                 if (unlikely(!pte_present(pte))) {
3198                         huge_pte_clear(mm, address, ptep);
3199                         goto unlock;
3200                 }
3201
3202                 page = pte_page(pte);
3203                 /*
3204                  * If a reference page is supplied, it is because a specific
3205                  * page is being unmapped, not a range. Ensure the page we
3206                  * are about to unmap is the actual page of interest.
3207                  */
3208                 if (ref_page) {
3209                         if (page != ref_page)
3210                                 goto unlock;
3211
3212                         /*
3213                          * Mark the VMA as having unmapped its page so that
3214                          * future faults in this VMA will fail rather than
3215                          * looking like data was lost
3216                          */
3217                         set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3218                 }
3219
3220                 pte = huge_ptep_get_and_clear(mm, address, ptep);
3221                 tlb_remove_tlb_entry(tlb, ptep, address);
3222                 if (huge_pte_dirty(pte))
3223                         set_page_dirty(page);
3224
3225                 hugetlb_count_sub(pages_per_huge_page(h), mm);
3226                 page_remove_rmap(page);
3227                 force_flush = !__tlb_remove_page(tlb, page);
3228                 if (force_flush) {
3229                         address += sz;
3230                         spin_unlock(ptl);
3231                         break;
3232                 }
3233                 /* Bail out after unmapping reference page if supplied */
3234                 if (ref_page) {
3235                         spin_unlock(ptl);
3236                         break;
3237                 }
3238 unlock:
3239                 spin_unlock(ptl);
3240         }
3241         /*
3242          * mmu_gather ran out of room to batch pages, we break out of
3243          * the PTE lock to avoid doing the potential expensive TLB invalidate
3244          * and page-free while holding it.
3245          */
3246         if (force_flush) {
3247                 force_flush = 0;
3248                 tlb_flush_mmu(tlb);
3249                 if (address < end && !ref_page)
3250                         goto again;
3251         }
3252         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3253         tlb_end_vma(tlb, vma);
3254 }
3255
3256 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3257                           struct vm_area_struct *vma, unsigned long start,
3258                           unsigned long end, struct page *ref_page)
3259 {
3260         __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3261
3262         /*
3263          * Clear this flag so that x86's huge_pmd_share page_table_shareable
3264          * test will fail on a vma being torn down, and not grab a page table
3265          * on its way out.  We're lucky that the flag has such an appropriate
3266          * name, and can in fact be safely cleared here. We could clear it
3267          * before the __unmap_hugepage_range above, but all that's necessary
3268          * is to clear it before releasing the i_mmap_rwsem. This works
3269          * because in the context this is called, the VMA is about to be
3270          * destroyed and the i_mmap_rwsem is held.
3271          */
3272         vma->vm_flags &= ~VM_MAYSHARE;
3273 }
3274
3275 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3276                           unsigned long end, struct page *ref_page)
3277 {
3278         struct mm_struct *mm;
3279         struct mmu_gather tlb;
3280
3281         mm = vma->vm_mm;
3282
3283         tlb_gather_mmu(&tlb, mm, start, end);
3284         __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3285         tlb_finish_mmu(&tlb, start, end);
3286 }
3287
3288 /*
3289  * This is called when the original mapper is failing to COW a MAP_PRIVATE
3290  * mappping it owns the reserve page for. The intention is to unmap the page
3291  * from other VMAs and let the children be SIGKILLed if they are faulting the
3292  * same region.
3293  */
3294 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3295                               struct page *page, unsigned long address)
3296 {
3297         struct hstate *h = hstate_vma(vma);
3298         struct vm_area_struct *iter_vma;
3299         struct address_space *mapping;
3300         pgoff_t pgoff;
3301
3302         /*
3303          * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3304          * from page cache lookup which is in HPAGE_SIZE units.
3305          */
3306         address = address & huge_page_mask(h);
3307         pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3308                         vma->vm_pgoff;
3309         mapping = file_inode(vma->vm_file)->i_mapping;
3310
3311         /*
3312          * Take the mapping lock for the duration of the table walk. As
3313          * this mapping should be shared between all the VMAs,
3314          * __unmap_hugepage_range() is called as the lock is already held
3315          */
3316         i_mmap_lock_write(mapping);
3317         vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3318                 /* Do not unmap the current VMA */
3319                 if (iter_vma == vma)
3320                         continue;
3321
3322                 /*
3323                  * Shared VMAs have their own reserves and do not affect
3324                  * MAP_PRIVATE accounting but it is possible that a shared
3325                  * VMA is using the same page so check and skip such VMAs.
3326                  */
3327                 if (iter_vma->vm_flags & VM_MAYSHARE)
3328                         continue;
3329
3330                 /*
3331                  * Unmap the page from other VMAs without their own reserves.
3332                  * They get marked to be SIGKILLed if they fault in these
3333                  * areas. This is because a future no-page fault on this VMA
3334                  * could insert a zeroed page instead of the data existing
3335                  * from the time of fork. This would look like data corruption
3336                  */
3337                 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3338                         unmap_hugepage_range(iter_vma, address,
3339                                              address + huge_page_size(h), page);
3340         }
3341         i_mmap_unlock_write(mapping);
3342 }
3343
3344 /*
3345  * Hugetlb_cow() should be called with page lock of the original hugepage held.
3346  * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3347  * cannot race with other handlers or page migration.
3348  * Keep the pte_same checks anyway to make transition from the mutex easier.
3349  */
3350 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3351                         unsigned long address, pte_t *ptep, pte_t pte,
3352                         struct page *pagecache_page, spinlock_t *ptl)
3353 {
3354         struct hstate *h = hstate_vma(vma);
3355         struct page *old_page, *new_page;
3356         int ret = 0, outside_reserve = 0;
3357         unsigned long mmun_start;       /* For mmu_notifiers */
3358         unsigned long mmun_end;         /* For mmu_notifiers */
3359
3360         old_page = pte_page(pte);
3361
3362 retry_avoidcopy:
3363         /* If no-one else is actually using this page, avoid the copy
3364          * and just make the page writable */
3365         if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3366                 page_move_anon_rmap(old_page, vma, address);
3367                 set_huge_ptep_writable(vma, address, ptep);
3368                 return 0;
3369         }
3370
3371         /*
3372          * If the process that created a MAP_PRIVATE mapping is about to
3373          * perform a COW due to a shared page count, attempt to satisfy
3374          * the allocation without using the existing reserves. The pagecache
3375          * page is used to determine if the reserve at this address was
3376          * consumed or not. If reserves were used, a partial faulted mapping
3377          * at the time of fork() could consume its reserves on COW instead
3378          * of the full address range.
3379          */
3380         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3381                         old_page != pagecache_page)
3382                 outside_reserve = 1;
3383
3384         page_cache_get(old_page);
3385
3386         /*
3387          * Drop page table lock as buddy allocator may be called. It will
3388          * be acquired again before returning to the caller, as expected.
3389          */
3390         spin_unlock(ptl);
3391         new_page = alloc_huge_page(vma, address, outside_reserve);
3392
3393         if (IS_ERR(new_page)) {
3394                 /*
3395                  * If a process owning a MAP_PRIVATE mapping fails to COW,
3396                  * it is due to references held by a child and an insufficient
3397                  * huge page pool. To guarantee the original mappers
3398                  * reliability, unmap the page from child processes. The child
3399                  * may get SIGKILLed if it later faults.
3400                  */
3401                 if (outside_reserve) {
3402                         page_cache_release(old_page);
3403                         BUG_ON(huge_pte_none(pte));
3404                         unmap_ref_private(mm, vma, old_page, address);
3405                         BUG_ON(huge_pte_none(pte));
3406                         spin_lock(ptl);
3407                         ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3408                         if (likely(ptep &&
3409                                    pte_same(huge_ptep_get(ptep), pte)))
3410                                 goto retry_avoidcopy;
3411                         /*
3412                          * race occurs while re-acquiring page table
3413                          * lock, and our job is done.
3414                          */
3415                         return 0;
3416                 }
3417
3418                 ret = (PTR_ERR(new_page) == -ENOMEM) ?
3419                         VM_FAULT_OOM : VM_FAULT_SIGBUS;
3420                 goto out_release_old;
3421         }
3422
3423         /*
3424          * When the original hugepage is shared one, it does not have
3425          * anon_vma prepared.
3426          */
3427         if (unlikely(anon_vma_prepare(vma))) {
3428                 ret = VM_FAULT_OOM;
3429                 goto out_release_all;
3430         }
3431
3432         copy_user_huge_page(new_page, old_page, address, vma,
3433                             pages_per_huge_page(h));
3434         __SetPageUptodate(new_page);
3435         set_page_huge_active(new_page);
3436
3437         mmun_start = address & huge_page_mask(h);
3438         mmun_end = mmun_start + huge_page_size(h);
3439         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3440
3441         /*
3442          * Retake the page table lock to check for racing updates
3443          * before the page tables are altered
3444          */
3445         spin_lock(ptl);
3446         ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3447         if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3448                 ClearPagePrivate(new_page);
3449
3450                 /* Break COW */
3451                 huge_ptep_clear_flush(vma, address, ptep);
3452                 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
3453                 set_huge_pte_at(mm, address, ptep,
3454                                 make_huge_pte(vma, new_page, 1));
3455                 page_remove_rmap(old_page);
3456                 hugepage_add_new_anon_rmap(new_page, vma, address);
3457                 /* Make the old page be freed below */
3458                 new_page = old_page;
3459         }
3460         spin_unlock(ptl);
3461         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3462 out_release_all:
3463         page_cache_release(new_page);
3464 out_release_old:
3465         page_cache_release(old_page);
3466
3467         spin_lock(ptl); /* Caller expects lock to be held */
3468         return ret;
3469 }
3470
3471 /* Return the pagecache page at a given address within a VMA */
3472 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3473                         struct vm_area_struct *vma, unsigned long address)
3474 {
3475         struct address_space *mapping;
3476         pgoff_t idx;
3477
3478         mapping = vma->vm_file->f_mapping;
3479         idx = vma_hugecache_offset(h, vma, address);
3480
3481         return find_lock_page(mapping, idx);
3482 }
3483
3484 /*
3485  * Return whether there is a pagecache page to back given address within VMA.
3486  * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3487  */
3488 static bool hugetlbfs_pagecache_present(struct hstate *h,
3489                         struct vm_area_struct *vma, unsigned long address)
3490 {
3491         struct address_space *mapping;
3492         pgoff_t idx;
3493         struct page *page;
3494
3495         mapping = vma->vm_file->f_mapping;
3496         idx = vma_hugecache_offset(h, vma, address);
3497
3498         page = find_get_page(mapping, idx);
3499         if (page)
3500                 put_page(page);
3501         return page != NULL;
3502 }
3503
3504 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3505                            pgoff_t idx)
3506 {
3507         struct inode *inode = mapping->host;
3508         struct hstate *h = hstate_inode(inode);
3509         int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3510
3511         if (err)
3512                 return err;
3513         ClearPagePrivate(page);
3514
3515         spin_lock(&inode->i_lock);
3516         inode->i_blocks += blocks_per_huge_page(h);
3517         spin_unlock(&inode->i_lock);
3518         return 0;
3519 }
3520
3521 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
3522                            struct address_space *mapping, pgoff_t idx,
3523                            unsigned long address, pte_t *ptep, unsigned int flags)
3524 {
3525         struct hstate *h = hstate_vma(vma);
3526         int ret = VM_FAULT_SIGBUS;
3527         int anon_rmap = 0;
3528         unsigned long size;
3529         struct page *page;
3530         pte_t new_pte;
3531         spinlock_t *ptl;
3532
3533         /*
3534          * Currently, we are forced to kill the process in the event the
3535          * original mapper has unmapped pages from the child due to a failed
3536          * COW. Warn that such a situation has occurred as it may not be obvious
3537          */
3538         if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3539                 pr_warning("PID %d killed due to inadequate hugepage pool\n",
3540                            current->pid);
3541                 return ret;
3542         }
3543
3544         /*
3545          * Use page lock to guard against racing truncation
3546          * before we get page_table_lock.
3547          */
3548 retry:
3549         page = find_lock_page(mapping, idx);
3550         if (!page) {
3551                 size = i_size_read(mapping->host) >> huge_page_shift(h);
3552                 if (idx >= size)
3553                         goto out;
3554                 page = alloc_huge_page(vma, address, 0);
3555                 if (IS_ERR(page)) {
3556                         ret = PTR_ERR(page);
3557                         if (ret == -ENOMEM)
3558                                 ret = VM_FAULT_OOM;
3559                         else
3560                                 ret = VM_FAULT_SIGBUS;
3561                         goto out;
3562                 }
3563                 clear_huge_page(page, address, pages_per_huge_page(h));
3564                 __SetPageUptodate(page);
3565                 set_page_huge_active(page);
3566
3567                 if (vma->vm_flags & VM_MAYSHARE) {
3568                         int err = huge_add_to_page_cache(page, mapping, idx);
3569                         if (err) {
3570                                 put_page(page);
3571                                 if (err == -EEXIST)
3572                                         goto retry;
3573                                 goto out;
3574                         }
3575                 } else {
3576                         lock_page(page);
3577                         if (unlikely(anon_vma_prepare(vma))) {
3578                                 ret = VM_FAULT_OOM;
3579                                 goto backout_unlocked;
3580                         }
3581                         anon_rmap = 1;
3582                 }
3583         } else {
3584                 /*
3585                  * If memory error occurs between mmap() and fault, some process
3586                  * don't have hwpoisoned swap entry for errored virtual address.
3587                  * So we need to block hugepage fault by PG_hwpoison bit check.
3588                  */
3589                 if (unlikely(PageHWPoison(page))) {
3590                         ret = VM_FAULT_HWPOISON |
3591                                 VM_FAULT_SET_HINDEX(hstate_index(h));
3592                         goto backout_unlocked;
3593                 }
3594         }
3595
3596         /*
3597          * If we are going to COW a private mapping later, we examine the
3598          * pending reservations for this page now. This will ensure that
3599          * any allocations necessary to record that reservation occur outside
3600          * the spinlock.
3601          */
3602         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3603                 if (vma_needs_reservation(h, vma, address) < 0) {
3604                         ret = VM_FAULT_OOM;
3605                         goto backout_unlocked;
3606                 }
3607                 /* Just decrements count, does not deallocate */
3608                 vma_end_reservation(h, vma, address);
3609         }
3610
3611         ptl = huge_pte_lockptr(h, mm, ptep);
3612         spin_lock(ptl);
3613         size = i_size_read(mapping->host) >> huge_page_shift(h);
3614         if (idx >= size)
3615                 goto backout;
3616
3617         ret = 0;
3618         if (!huge_pte_none(huge_ptep_get(ptep)))
3619                 goto backout;
3620
3621         if (anon_rmap) {
3622                 ClearPagePrivate(page);
3623                 hugepage_add_new_anon_rmap(page, vma, address);
3624         } else
3625                 page_dup_rmap(page);
3626         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3627                                 && (vma->vm_flags & VM_SHARED)));
3628         set_huge_pte_at(mm, address, ptep, new_pte);
3629
3630         hugetlb_count_add(pages_per_huge_page(h), mm);
3631         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3632                 /* Optimization, do the COW without a second fault */
3633                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl);
3634         }
3635
3636         spin_unlock(ptl);
3637         unlock_page(page);
3638 out:
3639         return ret;
3640
3641 backout:
3642         spin_unlock(ptl);
3643 backout_unlocked:
3644         unlock_page(page);
3645         put_page(page);
3646         goto out;
3647 }
3648
3649 #ifdef CONFIG_SMP
3650 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3651                             struct vm_area_struct *vma,
3652                             struct address_space *mapping,
3653                             pgoff_t idx, unsigned long address)
3654 {
3655         unsigned long key[2];
3656         u32 hash;
3657
3658         if (vma->vm_flags & VM_SHARED) {
3659                 key[0] = (unsigned long) mapping;
3660                 key[1] = idx;
3661         } else {
3662                 key[0] = (unsigned long) mm;
3663                 key[1] = address >> huge_page_shift(h);
3664         }
3665
3666         hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3667
3668         return hash & (num_fault_mutexes - 1);
3669 }
3670 #else
3671 /*
3672  * For uniprocesor systems we always use a single mutex, so just
3673  * return 0 and avoid the hashing overhead.
3674  */
3675 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3676                             struct vm_area_struct *vma,
3677                             struct address_space *mapping,
3678                             pgoff_t idx, unsigned long address)
3679 {
3680         return 0;
3681 }
3682 #endif
3683
3684 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3685                         unsigned long address, unsigned int flags)
3686 {
3687         pte_t *ptep, entry;
3688         spinlock_t *ptl;
3689         int ret;
3690         u32 hash;
3691         pgoff_t idx;
3692         struct page *page = NULL;
3693         struct page *pagecache_page = NULL;
3694         struct hstate *h = hstate_vma(vma);
3695         struct address_space *mapping;
3696         int need_wait_lock = 0;
3697
3698         address &= huge_page_mask(h);
3699
3700         ptep = huge_pte_offset(mm, address);
3701         if (ptep) {
3702                 entry = huge_ptep_get(ptep);
3703                 if (unlikely(is_hugetlb_entry_migration(entry))) {
3704                         migration_entry_wait_huge(vma, mm, ptep);
3705                         return 0;
3706                 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
3707                         return VM_FAULT_HWPOISON_LARGE |
3708                                 VM_FAULT_SET_HINDEX(hstate_index(h));
3709         } else {
3710                 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
3711                 if (!ptep)
3712                         return VM_FAULT_OOM;
3713         }
3714
3715         mapping = vma->vm_file->f_mapping;
3716         idx = vma_hugecache_offset(h, vma, address);
3717
3718         /*
3719          * Serialize hugepage allocation and instantiation, so that we don't
3720          * get spurious allocation failures if two CPUs race to instantiate
3721          * the same page in the page cache.
3722          */
3723         hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address);
3724         mutex_lock(&hugetlb_fault_mutex_table[hash]);
3725
3726         entry = huge_ptep_get(ptep);
3727         if (huge_pte_none(entry)) {
3728                 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
3729                 goto out_mutex;
3730         }
3731
3732         ret = 0;
3733
3734         /*
3735          * entry could be a migration/hwpoison entry at this point, so this
3736          * check prevents the kernel from going below assuming that we have
3737          * a active hugepage in pagecache. This goto expects the 2nd page fault,
3738          * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3739          * handle it.
3740          */
3741         if (!pte_present(entry))
3742                 goto out_mutex;
3743
3744         /*
3745          * If we are going to COW the mapping later, we examine the pending
3746          * reservations for this page now. This will ensure that any
3747          * allocations necessary to record that reservation occur outside the
3748          * spinlock. For private mappings, we also lookup the pagecache
3749          * page now as it is used to determine if a reservation has been
3750          * consumed.
3751          */
3752         if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
3753                 if (vma_needs_reservation(h, vma, address) < 0) {
3754                         ret = VM_FAULT_OOM;
3755                         goto out_mutex;
3756                 }
3757                 /* Just decrements count, does not deallocate */
3758                 vma_end_reservation(h, vma, address);
3759
3760                 if (!(vma->vm_flags & VM_MAYSHARE))
3761                         pagecache_page = hugetlbfs_pagecache_page(h,
3762                                                                 vma, address);
3763         }
3764
3765         ptl = huge_pte_lock(h, mm, ptep);
3766
3767         /* Check for a racing update before calling hugetlb_cow */
3768         if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
3769                 goto out_ptl;
3770
3771         /*
3772          * hugetlb_cow() requires page locks of pte_page(entry) and
3773          * pagecache_page, so here we need take the former one
3774          * when page != pagecache_page or !pagecache_page.
3775          */
3776         page = pte_page(entry);
3777         if (page != pagecache_page)
3778                 if (!trylock_page(page)) {
3779                         need_wait_lock = 1;
3780                         goto out_ptl;
3781                 }
3782
3783         get_page(page);
3784
3785         if (flags & FAULT_FLAG_WRITE) {
3786                 if (!huge_pte_write(entry)) {
3787                         ret = hugetlb_cow(mm, vma, address, ptep, entry,
3788                                         pagecache_page, ptl);
3789                         goto out_put_page;
3790                 }
3791                 entry = huge_pte_mkdirty(entry);
3792         }
3793         entry = pte_mkyoung(entry);
3794         if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3795                                                 flags & FAULT_FLAG_WRITE))
3796                 update_mmu_cache(vma, address, ptep);
3797 out_put_page:
3798         if (page != pagecache_page)
3799                 unlock_page(page);
3800         put_page(page);
3801 out_ptl:
3802         spin_unlock(ptl);
3803
3804         if (pagecache_page) {
3805                 unlock_page(pagecache_page);
3806                 put_page(pagecache_page);
3807         }
3808 out_mutex:
3809         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3810         /*
3811          * Generally it's safe to hold refcount during waiting page lock. But
3812          * here we just wait to defer the next page fault to avoid busy loop and
3813          * the page is not used after unlocked before returning from the current
3814          * page fault. So we are safe from accessing freed page, even if we wait
3815          * here without taking refcount.
3816          */
3817         if (need_wait_lock)
3818                 wait_on_page_locked(page);
3819         return ret;
3820 }
3821
3822 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3823                          struct page **pages, struct vm_area_struct **vmas,
3824                          unsigned long *position, unsigned long *nr_pages,
3825                          long i, unsigned int flags)
3826 {
3827         unsigned long pfn_offset;
3828         unsigned long vaddr = *position;
3829         unsigned long remainder = *nr_pages;
3830         struct hstate *h = hstate_vma(vma);
3831
3832         while (vaddr < vma->vm_end && remainder) {
3833                 pte_t *pte;
3834                 spinlock_t *ptl = NULL;
3835                 int absent;
3836                 struct page *page;
3837
3838                 /*
3839                  * If we have a pending SIGKILL, don't keep faulting pages and
3840                  * potentially allocating memory.
3841                  */
3842                 if (unlikely(fatal_signal_pending(current))) {
3843                         remainder = 0;
3844                         break;
3845                 }
3846
3847                 /*
3848                  * Some archs (sparc64, sh*) have multiple pte_ts to
3849                  * each hugepage.  We have to make sure we get the
3850                  * first, for the page indexing below to work.
3851                  *
3852                  * Note that page table lock is not held when pte is null.
3853                  */
3854                 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
3855                 if (pte)
3856                         ptl = huge_pte_lock(h, mm, pte);
3857                 absent = !pte || huge_pte_none(huge_ptep_get(pte));
3858
3859                 /*
3860                  * When coredumping, it suits get_dump_page if we just return
3861                  * an error where there's an empty slot with no huge pagecache
3862                  * to back it.  This way, we avoid allocating a hugepage, and
3863                  * the sparse dumpfile avoids allocating disk blocks, but its
3864                  * huge holes still show up with zeroes where they need to be.
3865                  */
3866                 if (absent && (flags & FOLL_DUMP) &&
3867                     !hugetlbfs_pagecache_present(h, vma, vaddr)) {
3868                         if (pte)
3869                                 spin_unlock(ptl);
3870                         remainder = 0;
3871                         break;
3872                 }
3873
3874                 /*
3875                  * We need call hugetlb_fault for both hugepages under migration
3876                  * (in which case hugetlb_fault waits for the migration,) and
3877                  * hwpoisoned hugepages (in which case we need to prevent the
3878                  * caller from accessing to them.) In order to do this, we use
3879                  * here is_swap_pte instead of is_hugetlb_entry_migration and
3880                  * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3881                  * both cases, and because we can't follow correct pages
3882                  * directly from any kind of swap entries.
3883                  */
3884                 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
3885                     ((flags & FOLL_WRITE) &&
3886                       !huge_pte_write(huge_ptep_get(pte)))) {
3887                         int ret;
3888
3889                         if (pte)
3890                                 spin_unlock(ptl);
3891                         ret = hugetlb_fault(mm, vma, vaddr,
3892                                 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
3893                         if (!(ret & VM_FAULT_ERROR))
3894                                 continue;
3895
3896                         remainder = 0;
3897                         break;
3898                 }
3899
3900                 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
3901                 page = pte_page(huge_ptep_get(pte));
3902 same_page:
3903                 if (pages) {
3904                         pages[i] = mem_map_offset(page, pfn_offset);
3905                         get_page_foll(pages[i]);
3906                 }
3907
3908                 if (vmas)
3909                         vmas[i] = vma;
3910
3911                 vaddr += PAGE_SIZE;
3912                 ++pfn_offset;
3913                 --remainder;
3914                 ++i;
3915                 if (vaddr < vma->vm_end && remainder &&
3916                                 pfn_offset < pages_per_huge_page(h)) {
3917                         /*
3918                          * We use pfn_offset to avoid touching the pageframes
3919                          * of this compound page.
3920                          */
3921                         goto same_page;
3922                 }
3923                 spin_unlock(ptl);
3924         }
3925         *nr_pages = remainder;
3926         *position = vaddr;
3927
3928         return i ? i : -EFAULT;
3929 }
3930
3931 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
3932                 unsigned long address, unsigned long end, pgprot_t newprot)
3933 {
3934         struct mm_struct *mm = vma->vm_mm;
3935         unsigned long start = address;
3936         pte_t *ptep;
3937         pte_t pte;
3938         struct hstate *h = hstate_vma(vma);
3939         unsigned long pages = 0;
3940
3941         BUG_ON(address >= end);
3942         flush_cache_range(vma, address, end);
3943
3944         mmu_notifier_invalidate_range_start(mm, start, end);
3945         i_mmap_lock_write(vma->vm_file->f_mapping);
3946         for (; address < end; address += huge_page_size(h)) {
3947                 spinlock_t *ptl;
3948                 ptep = huge_pte_offset(mm, address);
3949                 if (!ptep)
3950                         continue;
3951                 ptl = huge_pte_lock(h, mm, ptep);
3952                 if (huge_pmd_unshare(mm, &address, ptep)) {
3953                         pages++;
3954                         spin_unlock(ptl);
3955                         continue;
3956                 }
3957                 pte = huge_ptep_get(ptep);
3958                 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
3959                         spin_unlock(ptl);
3960                         continue;
3961                 }
3962                 if (unlikely(is_hugetlb_entry_migration(pte))) {
3963                         swp_entry_t entry = pte_to_swp_entry(pte);
3964
3965                         if (is_write_migration_entry(entry)) {
3966                                 pte_t newpte;
3967
3968                                 make_migration_entry_read(&entry);
3969                                 newpte = swp_entry_to_pte(entry);
3970                                 set_huge_pte_at(mm, address, ptep, newpte);
3971                                 pages++;
3972                         }
3973                         spin_unlock(ptl);
3974                         continue;
3975                 }
3976                 if (!huge_pte_none(pte)) {
3977                         pte = huge_ptep_get_and_clear(mm, address, ptep);
3978                         pte = pte_mkhuge(huge_pte_modify(pte, newprot));
3979                         pte = arch_make_huge_pte(pte, vma, NULL, 0);
3980                         set_huge_pte_at(mm, address, ptep, pte);
3981                         pages++;
3982                 }
3983                 spin_unlock(ptl);
3984         }
3985         /*
3986          * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
3987          * may have cleared our pud entry and done put_page on the page table:
3988          * once we release i_mmap_rwsem, another task can do the final put_page
3989          * and that page table be reused and filled with junk.
3990          */
3991         flush_tlb_range(vma, start, end);
3992         mmu_notifier_invalidate_range(mm, start, end);
3993         i_mmap_unlock_write(vma->vm_file->f_mapping);
3994         mmu_notifier_invalidate_range_end(mm, start, end);
3995
3996         return pages << h->order;
3997 }
3998
3999 int hugetlb_reserve_pages(struct inode *inode,
4000                                         long from, long to,
4001                                         struct vm_area_struct *vma,
4002                                         vm_flags_t vm_flags)
4003 {
4004         long ret, chg;
4005         struct hstate *h = hstate_inode(inode);
4006         struct hugepage_subpool *spool = subpool_inode(inode);
4007         struct resv_map *resv_map;
4008         long gbl_reserve;
4009
4010         /*
4011          * Only apply hugepage reservation if asked. At fault time, an
4012          * attempt will be made for VM_NORESERVE to allocate a page
4013          * without using reserves
4014          */
4015         if (vm_flags & VM_NORESERVE)
4016                 return 0;
4017
4018         /*
4019          * Shared mappings base their reservation on the number of pages that
4020          * are already allocated on behalf of the file. Private mappings need
4021          * to reserve the full area even if read-only as mprotect() may be
4022          * called to make the mapping read-write. Assume !vma is a shm mapping
4023          */
4024         if (!vma || vma->vm_flags & VM_MAYSHARE) {
4025                 resv_map = inode_resv_map(inode);
4026
4027                 chg = region_chg(resv_map, from, to);
4028
4029         } else {
4030                 resv_map = resv_map_alloc();
4031                 if (!resv_map)
4032                         return -ENOMEM;
4033
4034                 chg = to - from;
4035
4036                 set_vma_resv_map(vma, resv_map);
4037                 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4038         }
4039
4040         if (chg < 0) {
4041                 ret = chg;
4042                 goto out_err;
4043         }
4044
4045         /*
4046          * There must be enough pages in the subpool for the mapping. If
4047          * the subpool has a minimum size, there may be some global
4048          * reservations already in place (gbl_reserve).
4049          */
4050         gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4051         if (gbl_reserve < 0) {
4052                 ret = -ENOSPC;
4053                 goto out_err;
4054         }
4055
4056         /*
4057          * Check enough hugepages are available for the reservation.
4058          * Hand the pages back to the subpool if there are not
4059          */
4060         ret = hugetlb_acct_memory(h, gbl_reserve);
4061         if (ret < 0) {
4062                 /* put back original number of pages, chg */
4063                 (void)hugepage_subpool_put_pages(spool, chg);
4064                 goto out_err;
4065         }
4066
4067         /*
4068          * Account for the reservations made. Shared mappings record regions
4069          * that have reservations as they are shared by multiple VMAs.
4070          * When the last VMA disappears, the region map says how much
4071          * the reservation was and the page cache tells how much of
4072          * the reservation was consumed. Private mappings are per-VMA and
4073          * only the consumed reservations are tracked. When the VMA
4074          * disappears, the original reservation is the VMA size and the
4075          * consumed reservations are stored in the map. Hence, nothing
4076          * else has to be done for private mappings here
4077          */
4078         if (!vma || vma->vm_flags & VM_MAYSHARE) {
4079                 long add = region_add(resv_map, from, to);
4080
4081                 if (unlikely(chg > add)) {
4082                         /*
4083                          * pages in this range were added to the reserve
4084                          * map between region_chg and region_add.  This
4085                          * indicates a race with alloc_huge_page.  Adjust
4086                          * the subpool and reserve counts modified above
4087                          * based on the difference.
4088                          */
4089                         long rsv_adjust;
4090
4091                         rsv_adjust = hugepage_subpool_put_pages(spool,
4092                                                                 chg - add);
4093                         hugetlb_acct_memory(h, -rsv_adjust);
4094                 }
4095         }
4096         return 0;
4097 out_err:
4098         if (!vma || vma->vm_flags & VM_MAYSHARE)
4099                 region_abort(resv_map, from, to);
4100         if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4101                 kref_put(&resv_map->refs, resv_map_release);
4102         return ret;
4103 }
4104
4105 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4106                                                                 long freed)
4107 {
4108         struct hstate *h = hstate_inode(inode);
4109         struct resv_map *resv_map = inode_resv_map(inode);
4110         long chg = 0;
4111         struct hugepage_subpool *spool = subpool_inode(inode);
4112         long gbl_reserve;
4113
4114         if (resv_map) {
4115                 chg = region_del(resv_map, start, end);
4116                 /*
4117                  * region_del() can fail in the rare case where a region
4118                  * must be split and another region descriptor can not be
4119                  * allocated.  If end == LONG_MAX, it will not fail.
4120                  */
4121                 if (chg < 0)
4122                         return chg;
4123         }
4124
4125         spin_lock(&inode->i_lock);
4126         inode->i_blocks -= (blocks_per_huge_page(h) * freed);
4127         spin_unlock(&inode->i_lock);
4128
4129         /*
4130          * If the subpool has a minimum size, the number of global
4131          * reservations to be released may be adjusted.
4132          */
4133         gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4134         hugetlb_acct_memory(h, -gbl_reserve);
4135
4136         return 0;
4137 }
4138
4139 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4140 static unsigned long page_table_shareable(struct vm_area_struct *svma,
4141                                 struct vm_area_struct *vma,
4142                                 unsigned long addr, pgoff_t idx)
4143 {
4144         unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4145                                 svma->vm_start;
4146         unsigned long sbase = saddr & PUD_MASK;
4147         unsigned long s_end = sbase + PUD_SIZE;
4148
4149         /* Allow segments to share if only one is marked locked */
4150         unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4151         unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
4152
4153         /*
4154          * match the virtual addresses, permission and the alignment of the
4155          * page table page.
4156          */
4157         if (pmd_index(addr) != pmd_index(saddr) ||
4158             vm_flags != svm_flags ||
4159             sbase < svma->vm_start || svma->vm_end < s_end)
4160                 return 0;
4161
4162         return saddr;
4163 }
4164
4165 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
4166 {
4167         unsigned long base = addr & PUD_MASK;
4168         unsigned long end = base + PUD_SIZE;
4169
4170         /*
4171          * check on proper vm_flags and page table alignment
4172          */
4173         if (vma->vm_flags & VM_MAYSHARE &&
4174             vma->vm_start <= base && end <= vma->vm_end)
4175                 return true;
4176         return false;
4177 }
4178
4179 /*
4180  * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4181  * and returns the corresponding pte. While this is not necessary for the
4182  * !shared pmd case because we can allocate the pmd later as well, it makes the
4183  * code much cleaner. pmd allocation is essential for the shared case because
4184  * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4185  * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4186  * bad pmd for sharing.
4187  */
4188 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4189 {
4190         struct vm_area_struct *vma = find_vma(mm, addr);
4191         struct address_space *mapping = vma->vm_file->f_mapping;
4192         pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4193                         vma->vm_pgoff;
4194         struct vm_area_struct *svma;
4195         unsigned long saddr;
4196         pte_t *spte = NULL;
4197         pte_t *pte;
4198         spinlock_t *ptl;
4199
4200         if (!vma_shareable(vma, addr))
4201                 return (pte_t *)pmd_alloc(mm, pud, addr);
4202
4203         i_mmap_lock_write(mapping);
4204         vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4205                 if (svma == vma)
4206                         continue;
4207
4208                 saddr = page_table_shareable(svma, vma, addr, idx);
4209                 if (saddr) {
4210                         spte = huge_pte_offset(svma->vm_mm, saddr);
4211                         if (spte) {
4212                                 mm_inc_nr_pmds(mm);
4213                                 get_page(virt_to_page(spte));
4214                                 break;
4215                         }
4216                 }
4217         }
4218
4219         if (!spte)
4220                 goto out;
4221
4222         ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte);
4223         spin_lock(ptl);
4224         if (pud_none(*pud)) {
4225                 pud_populate(mm, pud,
4226                                 (pmd_t *)((unsigned long)spte & PAGE_MASK));
4227         } else {
4228                 put_page(virt_to_page(spte));
4229                 mm_inc_nr_pmds(mm);
4230         }
4231         spin_unlock(ptl);
4232 out:
4233         pte = (pte_t *)pmd_alloc(mm, pud, addr);
4234         i_mmap_unlock_write(mapping);
4235         return pte;
4236 }
4237
4238 /*
4239  * unmap huge page backed by shared pte.
4240  *
4241  * Hugetlb pte page is ref counted at the time of mapping.  If pte is shared
4242  * indicated by page_count > 1, unmap is achieved by clearing pud and
4243  * decrementing the ref count. If count == 1, the pte page is not shared.
4244  *
4245  * called with page table lock held.
4246  *
4247  * returns: 1 successfully unmapped a shared pte page
4248  *          0 the underlying pte page is not shared, or it is the last user
4249  */
4250 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4251 {
4252         pgd_t *pgd = pgd_offset(mm, *addr);
4253         pud_t *pud = pud_offset(pgd, *addr);
4254
4255         BUG_ON(page_count(virt_to_page(ptep)) == 0);
4256         if (page_count(virt_to_page(ptep)) == 1)
4257                 return 0;
4258
4259         pud_clear(pud);
4260         put_page(virt_to_page(ptep));
4261         mm_dec_nr_pmds(mm);
4262         *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4263         return 1;
4264 }
4265 #define want_pmd_share()        (1)
4266 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4267 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4268 {
4269         return NULL;
4270 }
4271
4272 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4273 {
4274         return 0;
4275 }
4276 #define want_pmd_share()        (0)
4277 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4278
4279 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4280 pte_t *huge_pte_alloc(struct mm_struct *mm,
4281                         unsigned long addr, unsigned long sz)
4282 {
4283         pgd_t *pgd;
4284         pud_t *pud;
4285         pte_t *pte = NULL;
4286
4287         pgd = pgd_offset(mm, addr);
4288         pud = pud_alloc(mm, pgd, addr);
4289         if (pud) {
4290                 if (sz == PUD_SIZE) {
4291                         pte = (pte_t *)pud;
4292                 } else {
4293                         BUG_ON(sz != PMD_SIZE);
4294                         if (want_pmd_share() && pud_none(*pud))
4295                                 pte = huge_pmd_share(mm, addr, pud);
4296                         else
4297                                 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4298                 }
4299         }
4300         BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
4301
4302         return pte;
4303 }
4304
4305 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
4306 {
4307         pgd_t *pgd;
4308         pud_t *pud;
4309         pmd_t *pmd = NULL;
4310
4311         pgd = pgd_offset(mm, addr);
4312         if (pgd_present(*pgd)) {
4313                 pud = pud_offset(pgd, addr);
4314                 if (pud_present(*pud)) {
4315                         if (pud_huge(*pud))
4316                                 return (pte_t *)pud;
4317                         pmd = pmd_offset(pud, addr);
4318                 }
4319         }
4320         return (pte_t *) pmd;
4321 }
4322
4323 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4324
4325 /*
4326  * These functions are overwritable if your architecture needs its own
4327  * behavior.
4328  */
4329 struct page * __weak
4330 follow_huge_addr(struct mm_struct *mm, unsigned long address,
4331                               int write)
4332 {
4333         return ERR_PTR(-EINVAL);
4334 }
4335
4336 struct page * __weak
4337 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
4338                 pmd_t *pmd, int flags)
4339 {
4340         struct page *page = NULL;
4341         spinlock_t *ptl;
4342 retry:
4343         ptl = pmd_lockptr(mm, pmd);
4344         spin_lock(ptl);
4345         /*
4346          * make sure that the address range covered by this pmd is not
4347          * unmapped from other threads.
4348          */
4349         if (!pmd_huge(*pmd))
4350                 goto out;
4351         if (pmd_present(*pmd)) {
4352                 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
4353                 if (flags & FOLL_GET)
4354                         get_page(page);
4355         } else {
4356                 if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) {
4357                         spin_unlock(ptl);
4358                         __migration_entry_wait(mm, (pte_t *)pmd, ptl);
4359                         goto retry;
4360                 }
4361                 /*
4362                  * hwpoisoned entry is treated as no_page_table in
4363                  * follow_page_mask().
4364                  */
4365         }
4366 out:
4367         spin_unlock(ptl);
4368         return page;
4369 }
4370
4371 struct page * __weak
4372 follow_huge_pud(struct mm_struct *mm, unsigned long address,
4373                 pud_t *pud, int flags)
4374 {
4375         if (flags & FOLL_GET)
4376                 return NULL;
4377
4378         return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
4379 }
4380
4381 #ifdef CONFIG_MEMORY_FAILURE
4382
4383 /*
4384  * This function is called from memory failure code.
4385  * Assume the caller holds page lock of the head page.
4386  */
4387 int dequeue_hwpoisoned_huge_page(struct page *hpage)
4388 {
4389         struct hstate *h = page_hstate(hpage);
4390         int nid = page_to_nid(hpage);
4391         int ret = -EBUSY;
4392
4393         spin_lock(&hugetlb_lock);
4394         /*
4395          * Just checking !page_huge_active is not enough, because that could be
4396          * an isolated/hwpoisoned hugepage (which have >0 refcount).
4397          */
4398         if (!page_huge_active(hpage) && !page_count(hpage)) {
4399                 /*
4400                  * Hwpoisoned hugepage isn't linked to activelist or freelist,
4401                  * but dangling hpage->lru can trigger list-debug warnings
4402                  * (this happens when we call unpoison_memory() on it),
4403                  * so let it point to itself with list_del_init().
4404                  */
4405                 list_del_init(&hpage->lru);
4406                 set_page_refcounted(hpage);
4407                 h->free_huge_pages--;
4408                 h->free_huge_pages_node[nid]--;
4409                 ret = 0;
4410         }
4411         spin_unlock(&hugetlb_lock);
4412         return ret;
4413 }
4414 #endif
4415
4416 bool isolate_huge_page(struct page *page, struct list_head *list)
4417 {
4418         bool ret = true;
4419
4420         VM_BUG_ON_PAGE(!PageHead(page), page);
4421         spin_lock(&hugetlb_lock);
4422         if (!page_huge_active(page) || !get_page_unless_zero(page)) {
4423                 ret = false;
4424                 goto unlock;
4425         }
4426         clear_page_huge_active(page);
4427         list_move_tail(&page->lru, list);
4428 unlock:
4429         spin_unlock(&hugetlb_lock);
4430         return ret;
4431 }
4432
4433 void putback_active_hugepage(struct page *page)
4434 {
4435         VM_BUG_ON_PAGE(!PageHead(page), page);
4436         spin_lock(&hugetlb_lock);
4437         set_page_huge_active(page);
4438         list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
4439         spin_unlock(&hugetlb_lock);
4440         put_page(page);
4441 }