Execution of Livemigration through Yardstick
[kvmfornfv.git] / kernel / 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 page *head = compound_head(page);
1420                 struct hstate *h = page_hstate(head);
1421                 int nid = page_to_nid(head);
1422                 list_del(&head->lru);
1423                 h->free_huge_pages--;
1424                 h->free_huge_pages_node[nid]--;
1425                 update_and_free_page(h, head);
1426         }
1427         spin_unlock(&hugetlb_lock);
1428 }
1429
1430 /*
1431  * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1432  * make specified memory blocks removable from the system.
1433  * Note that this will dissolve a free gigantic hugepage completely, if any
1434  * part of it lies within the given range.
1435  */
1436 void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1437 {
1438         unsigned long pfn;
1439
1440         if (!hugepages_supported())
1441                 return;
1442
1443         for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order)
1444                 dissolve_free_huge_page(pfn_to_page(pfn));
1445 }
1446
1447 /*
1448  * There are 3 ways this can get called:
1449  * 1. With vma+addr: we use the VMA's memory policy
1450  * 2. With !vma, but nid=NUMA_NO_NODE:  We try to allocate a huge
1451  *    page from any node, and let the buddy allocator itself figure
1452  *    it out.
1453  * 3. With !vma, but nid!=NUMA_NO_NODE.  We allocate a huge page
1454  *    strictly from 'nid'
1455  */
1456 static struct page *__hugetlb_alloc_buddy_huge_page(struct hstate *h,
1457                 struct vm_area_struct *vma, unsigned long addr, int nid)
1458 {
1459         int order = huge_page_order(h);
1460         gfp_t gfp = htlb_alloc_mask(h)|__GFP_COMP|__GFP_REPEAT|__GFP_NOWARN;
1461         unsigned int cpuset_mems_cookie;
1462
1463         /*
1464          * We need a VMA to get a memory policy.  If we do not
1465          * have one, we use the 'nid' argument.
1466          *
1467          * The mempolicy stuff below has some non-inlined bits
1468          * and calls ->vm_ops.  That makes it hard to optimize at
1469          * compile-time, even when NUMA is off and it does
1470          * nothing.  This helps the compiler optimize it out.
1471          */
1472         if (!IS_ENABLED(CONFIG_NUMA) || !vma) {
1473                 /*
1474                  * If a specific node is requested, make sure to
1475                  * get memory from there, but only when a node
1476                  * is explicitly specified.
1477                  */
1478                 if (nid != NUMA_NO_NODE)
1479                         gfp |= __GFP_THISNODE;
1480                 /*
1481                  * Make sure to call something that can handle
1482                  * nid=NUMA_NO_NODE
1483                  */
1484                 return alloc_pages_node(nid, gfp, order);
1485         }
1486
1487         /*
1488          * OK, so we have a VMA.  Fetch the mempolicy and try to
1489          * allocate a huge page with it.  We will only reach this
1490          * when CONFIG_NUMA=y.
1491          */
1492         do {
1493                 struct page *page;
1494                 struct mempolicy *mpol;
1495                 struct zonelist *zl;
1496                 nodemask_t *nodemask;
1497
1498                 cpuset_mems_cookie = read_mems_allowed_begin();
1499                 zl = huge_zonelist(vma, addr, gfp, &mpol, &nodemask);
1500                 mpol_cond_put(mpol);
1501                 page = __alloc_pages_nodemask(gfp, order, zl, nodemask);
1502                 if (page)
1503                         return page;
1504         } while (read_mems_allowed_retry(cpuset_mems_cookie));
1505
1506         return NULL;
1507 }
1508
1509 /*
1510  * There are two ways to allocate a huge page:
1511  * 1. When you have a VMA and an address (like a fault)
1512  * 2. When you have no VMA (like when setting /proc/.../nr_hugepages)
1513  *
1514  * 'vma' and 'addr' are only for (1).  'nid' is always NUMA_NO_NODE in
1515  * this case which signifies that the allocation should be done with
1516  * respect for the VMA's memory policy.
1517  *
1518  * For (2), we ignore 'vma' and 'addr' and use 'nid' exclusively. This
1519  * implies that memory policies will not be taken in to account.
1520  */
1521 static struct page *__alloc_buddy_huge_page(struct hstate *h,
1522                 struct vm_area_struct *vma, unsigned long addr, int nid)
1523 {
1524         struct page *page;
1525         unsigned int r_nid;
1526
1527         if (hstate_is_gigantic(h))
1528                 return NULL;
1529
1530         /*
1531          * Make sure that anyone specifying 'nid' is not also specifying a VMA.
1532          * This makes sure the caller is picking _one_ of the modes with which
1533          * we can call this function, not both.
1534          */
1535         if (vma || (addr != -1)) {
1536                 VM_WARN_ON_ONCE(addr == -1);
1537                 VM_WARN_ON_ONCE(nid != NUMA_NO_NODE);
1538         }
1539         /*
1540          * Assume we will successfully allocate the surplus page to
1541          * prevent racing processes from causing the surplus to exceed
1542          * overcommit
1543          *
1544          * This however introduces a different race, where a process B
1545          * tries to grow the static hugepage pool while alloc_pages() is
1546          * called by process A. B will only examine the per-node
1547          * counters in determining if surplus huge pages can be
1548          * converted to normal huge pages in adjust_pool_surplus(). A
1549          * won't be able to increment the per-node counter, until the
1550          * lock is dropped by B, but B doesn't drop hugetlb_lock until
1551          * no more huge pages can be converted from surplus to normal
1552          * state (and doesn't try to convert again). Thus, we have a
1553          * case where a surplus huge page exists, the pool is grown, and
1554          * the surplus huge page still exists after, even though it
1555          * should just have been converted to a normal huge page. This
1556          * does not leak memory, though, as the hugepage will be freed
1557          * once it is out of use. It also does not allow the counters to
1558          * go out of whack in adjust_pool_surplus() as we don't modify
1559          * the node values until we've gotten the hugepage and only the
1560          * per-node value is checked there.
1561          */
1562         spin_lock(&hugetlb_lock);
1563         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1564                 spin_unlock(&hugetlb_lock);
1565                 return NULL;
1566         } else {
1567                 h->nr_huge_pages++;
1568                 h->surplus_huge_pages++;
1569         }
1570         spin_unlock(&hugetlb_lock);
1571
1572         page = __hugetlb_alloc_buddy_huge_page(h, vma, addr, nid);
1573
1574         spin_lock(&hugetlb_lock);
1575         if (page) {
1576                 INIT_LIST_HEAD(&page->lru);
1577                 r_nid = page_to_nid(page);
1578                 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1579                 set_hugetlb_cgroup(page, NULL);
1580                 /*
1581                  * We incremented the global counters already
1582                  */
1583                 h->nr_huge_pages_node[r_nid]++;
1584                 h->surplus_huge_pages_node[r_nid]++;
1585                 __count_vm_event(HTLB_BUDDY_PGALLOC);
1586         } else {
1587                 h->nr_huge_pages--;
1588                 h->surplus_huge_pages--;
1589                 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1590         }
1591         spin_unlock(&hugetlb_lock);
1592
1593         return page;
1594 }
1595
1596 /*
1597  * Allocate a huge page from 'nid'.  Note, 'nid' may be
1598  * NUMA_NO_NODE, which means that it may be allocated
1599  * anywhere.
1600  */
1601 static
1602 struct page *__alloc_buddy_huge_page_no_mpol(struct hstate *h, int nid)
1603 {
1604         unsigned long addr = -1;
1605
1606         return __alloc_buddy_huge_page(h, NULL, addr, nid);
1607 }
1608
1609 /*
1610  * Use the VMA's mpolicy to allocate a huge page from the buddy.
1611  */
1612 static
1613 struct page *__alloc_buddy_huge_page_with_mpol(struct hstate *h,
1614                 struct vm_area_struct *vma, unsigned long addr)
1615 {
1616         return __alloc_buddy_huge_page(h, vma, addr, NUMA_NO_NODE);
1617 }
1618
1619 /*
1620  * This allocation function is useful in the context where vma is irrelevant.
1621  * E.g. soft-offlining uses this function because it only cares physical
1622  * address of error page.
1623  */
1624 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1625 {
1626         struct page *page = NULL;
1627
1628         spin_lock(&hugetlb_lock);
1629         if (h->free_huge_pages - h->resv_huge_pages > 0)
1630                 page = dequeue_huge_page_node(h, nid);
1631         spin_unlock(&hugetlb_lock);
1632
1633         if (!page)
1634                 page = __alloc_buddy_huge_page_no_mpol(h, nid);
1635
1636         return page;
1637 }
1638
1639 /*
1640  * Increase the hugetlb pool such that it can accommodate a reservation
1641  * of size 'delta'.
1642  */
1643 static int gather_surplus_pages(struct hstate *h, int delta)
1644 {
1645         struct list_head surplus_list;
1646         struct page *page, *tmp;
1647         int ret, i;
1648         int needed, allocated;
1649         bool alloc_ok = true;
1650
1651         needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1652         if (needed <= 0) {
1653                 h->resv_huge_pages += delta;
1654                 return 0;
1655         }
1656
1657         allocated = 0;
1658         INIT_LIST_HEAD(&surplus_list);
1659
1660         ret = -ENOMEM;
1661 retry:
1662         spin_unlock(&hugetlb_lock);
1663         for (i = 0; i < needed; i++) {
1664                 page = __alloc_buddy_huge_page_no_mpol(h, NUMA_NO_NODE);
1665                 if (!page) {
1666                         alloc_ok = false;
1667                         break;
1668                 }
1669                 list_add(&page->lru, &surplus_list);
1670         }
1671         allocated += i;
1672
1673         /*
1674          * After retaking hugetlb_lock, we need to recalculate 'needed'
1675          * because either resv_huge_pages or free_huge_pages may have changed.
1676          */
1677         spin_lock(&hugetlb_lock);
1678         needed = (h->resv_huge_pages + delta) -
1679                         (h->free_huge_pages + allocated);
1680         if (needed > 0) {
1681                 if (alloc_ok)
1682                         goto retry;
1683                 /*
1684                  * We were not able to allocate enough pages to
1685                  * satisfy the entire reservation so we free what
1686                  * we've allocated so far.
1687                  */
1688                 goto free;
1689         }
1690         /*
1691          * The surplus_list now contains _at_least_ the number of extra pages
1692          * needed to accommodate the reservation.  Add the appropriate number
1693          * of pages to the hugetlb pool and free the extras back to the buddy
1694          * allocator.  Commit the entire reservation here to prevent another
1695          * process from stealing the pages as they are added to the pool but
1696          * before they are reserved.
1697          */
1698         needed += allocated;
1699         h->resv_huge_pages += delta;
1700         ret = 0;
1701
1702         /* Free the needed pages to the hugetlb pool */
1703         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1704                 if ((--needed) < 0)
1705                         break;
1706                 /*
1707                  * This page is now managed by the hugetlb allocator and has
1708                  * no users -- drop the buddy allocator's reference.
1709                  */
1710                 put_page_testzero(page);
1711                 VM_BUG_ON_PAGE(page_count(page), page);
1712                 enqueue_huge_page(h, page);
1713         }
1714 free:
1715         spin_unlock(&hugetlb_lock);
1716
1717         /* Free unnecessary surplus pages to the buddy allocator */
1718         list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1719                 put_page(page);
1720         spin_lock(&hugetlb_lock);
1721
1722         return ret;
1723 }
1724
1725 /*
1726  * This routine has two main purposes:
1727  * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1728  *    in unused_resv_pages.  This corresponds to the prior adjustments made
1729  *    to the associated reservation map.
1730  * 2) Free any unused surplus pages that may have been allocated to satisfy
1731  *    the reservation.  As many as unused_resv_pages may be freed.
1732  *
1733  * Called with hugetlb_lock held.  However, the lock could be dropped (and
1734  * reacquired) during calls to cond_resched_lock.  Whenever dropping the lock,
1735  * we must make sure nobody else can claim pages we are in the process of
1736  * freeing.  Do this by ensuring resv_huge_page always is greater than the
1737  * number of huge pages we plan to free when dropping the lock.
1738  */
1739 static void return_unused_surplus_pages(struct hstate *h,
1740                                         unsigned long unused_resv_pages)
1741 {
1742         unsigned long nr_pages;
1743
1744         /* Cannot return gigantic pages currently */
1745         if (hstate_is_gigantic(h))
1746                 goto out;
1747
1748         /*
1749          * Part (or even all) of the reservation could have been backed
1750          * by pre-allocated pages. Only free surplus pages.
1751          */
1752         nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1753
1754         /*
1755          * We want to release as many surplus pages as possible, spread
1756          * evenly across all nodes with memory. Iterate across these nodes
1757          * until we can no longer free unreserved surplus pages. This occurs
1758          * when the nodes with surplus pages have no free pages.
1759          * free_pool_huge_page() will balance the the freed pages across the
1760          * on-line nodes with memory and will handle the hstate accounting.
1761          *
1762          * Note that we decrement resv_huge_pages as we free the pages.  If
1763          * we drop the lock, resv_huge_pages will still be sufficiently large
1764          * to cover subsequent pages we may free.
1765          */
1766         while (nr_pages--) {
1767                 h->resv_huge_pages--;
1768                 unused_resv_pages--;
1769                 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1770                         goto out;
1771                 cond_resched_lock(&hugetlb_lock);
1772         }
1773
1774 out:
1775         /* Fully uncommit the reservation */
1776         h->resv_huge_pages -= unused_resv_pages;
1777 }
1778
1779
1780 /*
1781  * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1782  * are used by the huge page allocation routines to manage reservations.
1783  *
1784  * vma_needs_reservation is called to determine if the huge page at addr
1785  * within the vma has an associated reservation.  If a reservation is
1786  * needed, the value 1 is returned.  The caller is then responsible for
1787  * managing the global reservation and subpool usage counts.  After
1788  * the huge page has been allocated, vma_commit_reservation is called
1789  * to add the page to the reservation map.  If the page allocation fails,
1790  * the reservation must be ended instead of committed.  vma_end_reservation
1791  * is called in such cases.
1792  *
1793  * In the normal case, vma_commit_reservation returns the same value
1794  * as the preceding vma_needs_reservation call.  The only time this
1795  * is not the case is if a reserve map was changed between calls.  It
1796  * is the responsibility of the caller to notice the difference and
1797  * take appropriate action.
1798  */
1799 enum vma_resv_mode {
1800         VMA_NEEDS_RESV,
1801         VMA_COMMIT_RESV,
1802         VMA_END_RESV,
1803 };
1804 static long __vma_reservation_common(struct hstate *h,
1805                                 struct vm_area_struct *vma, unsigned long addr,
1806                                 enum vma_resv_mode mode)
1807 {
1808         struct resv_map *resv;
1809         pgoff_t idx;
1810         long ret;
1811
1812         resv = vma_resv_map(vma);
1813         if (!resv)
1814                 return 1;
1815
1816         idx = vma_hugecache_offset(h, vma, addr);
1817         switch (mode) {
1818         case VMA_NEEDS_RESV:
1819                 ret = region_chg(resv, idx, idx + 1);
1820                 break;
1821         case VMA_COMMIT_RESV:
1822                 ret = region_add(resv, idx, idx + 1);
1823                 break;
1824         case VMA_END_RESV:
1825                 region_abort(resv, idx, idx + 1);
1826                 ret = 0;
1827                 break;
1828         default:
1829                 BUG();
1830         }
1831
1832         if (vma->vm_flags & VM_MAYSHARE)
1833                 return ret;
1834         else
1835                 return ret < 0 ? ret : 0;
1836 }
1837
1838 static long vma_needs_reservation(struct hstate *h,
1839                         struct vm_area_struct *vma, unsigned long addr)
1840 {
1841         return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
1842 }
1843
1844 static long vma_commit_reservation(struct hstate *h,
1845                         struct vm_area_struct *vma, unsigned long addr)
1846 {
1847         return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1848 }
1849
1850 static void vma_end_reservation(struct hstate *h,
1851                         struct vm_area_struct *vma, unsigned long addr)
1852 {
1853         (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
1854 }
1855
1856 struct page *alloc_huge_page(struct vm_area_struct *vma,
1857                                     unsigned long addr, int avoid_reserve)
1858 {
1859         struct hugepage_subpool *spool = subpool_vma(vma);
1860         struct hstate *h = hstate_vma(vma);
1861         struct page *page;
1862         long map_chg, map_commit;
1863         long gbl_chg;
1864         int ret, idx;
1865         struct hugetlb_cgroup *h_cg;
1866
1867         idx = hstate_index(h);
1868         /*
1869          * Examine the region/reserve map to determine if the process
1870          * has a reservation for the page to be allocated.  A return
1871          * code of zero indicates a reservation exists (no change).
1872          */
1873         map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
1874         if (map_chg < 0)
1875                 return ERR_PTR(-ENOMEM);
1876
1877         /*
1878          * Processes that did not create the mapping will have no
1879          * reserves as indicated by the region/reserve map. Check
1880          * that the allocation will not exceed the subpool limit.
1881          * Allocations for MAP_NORESERVE mappings also need to be
1882          * checked against any subpool limit.
1883          */
1884         if (map_chg || avoid_reserve) {
1885                 gbl_chg = hugepage_subpool_get_pages(spool, 1);
1886                 if (gbl_chg < 0) {
1887                         vma_end_reservation(h, vma, addr);
1888                         return ERR_PTR(-ENOSPC);
1889                 }
1890
1891                 /*
1892                  * Even though there was no reservation in the region/reserve
1893                  * map, there could be reservations associated with the
1894                  * subpool that can be used.  This would be indicated if the
1895                  * return value of hugepage_subpool_get_pages() is zero.
1896                  * However, if avoid_reserve is specified we still avoid even
1897                  * the subpool reservations.
1898                  */
1899                 if (avoid_reserve)
1900                         gbl_chg = 1;
1901         }
1902
1903         ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1904         if (ret)
1905                 goto out_subpool_put;
1906
1907         spin_lock(&hugetlb_lock);
1908         /*
1909          * glb_chg is passed to indicate whether or not a page must be taken
1910          * from the global free pool (global change).  gbl_chg == 0 indicates
1911          * a reservation exists for the allocation.
1912          */
1913         page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
1914         if (!page) {
1915                 spin_unlock(&hugetlb_lock);
1916                 page = __alloc_buddy_huge_page_with_mpol(h, vma, addr);
1917                 if (!page)
1918                         goto out_uncharge_cgroup;
1919                 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
1920                         SetPagePrivate(page);
1921                         h->resv_huge_pages--;
1922                 }
1923                 spin_lock(&hugetlb_lock);
1924                 list_move(&page->lru, &h->hugepage_activelist);
1925                 /* Fall through */
1926         }
1927         hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1928         spin_unlock(&hugetlb_lock);
1929
1930         set_page_private(page, (unsigned long)spool);
1931
1932         map_commit = vma_commit_reservation(h, vma, addr);
1933         if (unlikely(map_chg > map_commit)) {
1934                 /*
1935                  * The page was added to the reservation map between
1936                  * vma_needs_reservation and vma_commit_reservation.
1937                  * This indicates a race with hugetlb_reserve_pages.
1938                  * Adjust for the subpool count incremented above AND
1939                  * in hugetlb_reserve_pages for the same page.  Also,
1940                  * the reservation count added in hugetlb_reserve_pages
1941                  * no longer applies.
1942                  */
1943                 long rsv_adjust;
1944
1945                 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
1946                 hugetlb_acct_memory(h, -rsv_adjust);
1947         }
1948         return page;
1949
1950 out_uncharge_cgroup:
1951         hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
1952 out_subpool_put:
1953         if (map_chg || avoid_reserve)
1954                 hugepage_subpool_put_pages(spool, 1);
1955         vma_end_reservation(h, vma, addr);
1956         return ERR_PTR(-ENOSPC);
1957 }
1958
1959 /*
1960  * alloc_huge_page()'s wrapper which simply returns the page if allocation
1961  * succeeds, otherwise NULL. This function is called from new_vma_page(),
1962  * where no ERR_VALUE is expected to be returned.
1963  */
1964 struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
1965                                 unsigned long addr, int avoid_reserve)
1966 {
1967         struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
1968         if (IS_ERR(page))
1969                 page = NULL;
1970         return page;
1971 }
1972
1973 int __weak alloc_bootmem_huge_page(struct hstate *h)
1974 {
1975         struct huge_bootmem_page *m;
1976         int nr_nodes, node;
1977
1978         for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
1979                 void *addr;
1980
1981                 addr = memblock_virt_alloc_try_nid_nopanic(
1982                                 huge_page_size(h), huge_page_size(h),
1983                                 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
1984                 if (addr) {
1985                         /*
1986                          * Use the beginning of the huge page to store the
1987                          * huge_bootmem_page struct (until gather_bootmem
1988                          * puts them into the mem_map).
1989                          */
1990                         m = addr;
1991                         goto found;
1992                 }
1993         }
1994         return 0;
1995
1996 found:
1997         BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
1998         /* Put them into a private list first because mem_map is not up yet */
1999         list_add(&m->list, &huge_boot_pages);
2000         m->hstate = h;
2001         return 1;
2002 }
2003
2004 static void __init prep_compound_huge_page(struct page *page,
2005                 unsigned int order)
2006 {
2007         if (unlikely(order > (MAX_ORDER - 1)))
2008                 prep_compound_gigantic_page(page, order);
2009         else
2010                 prep_compound_page(page, order);
2011 }
2012
2013 /* Put bootmem huge pages into the standard lists after mem_map is up */
2014 static void __init gather_bootmem_prealloc(void)
2015 {
2016         struct huge_bootmem_page *m;
2017
2018         list_for_each_entry(m, &huge_boot_pages, list) {
2019                 struct hstate *h = m->hstate;
2020                 struct page *page;
2021
2022 #ifdef CONFIG_HIGHMEM
2023                 page = pfn_to_page(m->phys >> PAGE_SHIFT);
2024                 memblock_free_late(__pa(m),
2025                                    sizeof(struct huge_bootmem_page));
2026 #else
2027                 page = virt_to_page(m);
2028 #endif
2029                 WARN_ON(page_count(page) != 1);
2030                 prep_compound_huge_page(page, h->order);
2031                 WARN_ON(PageReserved(page));
2032                 prep_new_huge_page(h, page, page_to_nid(page));
2033                 /*
2034                  * If we had gigantic hugepages allocated at boot time, we need
2035                  * to restore the 'stolen' pages to totalram_pages in order to
2036                  * fix confusing memory reports from free(1) and another
2037                  * side-effects, like CommitLimit going negative.
2038                  */
2039                 if (hstate_is_gigantic(h))
2040                         adjust_managed_page_count(page, 1 << h->order);
2041         }
2042 }
2043
2044 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2045 {
2046         unsigned long i;
2047
2048         for (i = 0; i < h->max_huge_pages; ++i) {
2049                 if (hstate_is_gigantic(h)) {
2050                         if (!alloc_bootmem_huge_page(h))
2051                                 break;
2052                 } else if (!alloc_fresh_huge_page(h,
2053                                          &node_states[N_MEMORY]))
2054                         break;
2055         }
2056         h->max_huge_pages = i;
2057 }
2058
2059 static void __init hugetlb_init_hstates(void)
2060 {
2061         struct hstate *h;
2062
2063         for_each_hstate(h) {
2064                 if (minimum_order > huge_page_order(h))
2065                         minimum_order = huge_page_order(h);
2066
2067                 /* oversize hugepages were init'ed in early boot */
2068                 if (!hstate_is_gigantic(h))
2069                         hugetlb_hstate_alloc_pages(h);
2070         }
2071         VM_BUG_ON(minimum_order == UINT_MAX);
2072 }
2073
2074 static char * __init memfmt(char *buf, unsigned long n)
2075 {
2076         if (n >= (1UL << 30))
2077                 sprintf(buf, "%lu GB", n >> 30);
2078         else if (n >= (1UL << 20))
2079                 sprintf(buf, "%lu MB", n >> 20);
2080         else
2081                 sprintf(buf, "%lu KB", n >> 10);
2082         return buf;
2083 }
2084
2085 static void __init report_hugepages(void)
2086 {
2087         struct hstate *h;
2088
2089         for_each_hstate(h) {
2090                 char buf[32];
2091                 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2092                         memfmt(buf, huge_page_size(h)),
2093                         h->free_huge_pages);
2094         }
2095 }
2096
2097 #ifdef CONFIG_HIGHMEM
2098 static void try_to_free_low(struct hstate *h, unsigned long count,
2099                                                 nodemask_t *nodes_allowed)
2100 {
2101         int i;
2102
2103         if (hstate_is_gigantic(h))
2104                 return;
2105
2106         for_each_node_mask(i, *nodes_allowed) {
2107                 struct page *page, *next;
2108                 struct list_head *freel = &h->hugepage_freelists[i];
2109                 list_for_each_entry_safe(page, next, freel, lru) {
2110                         if (count >= h->nr_huge_pages)
2111                                 return;
2112                         if (PageHighMem(page))
2113                                 continue;
2114                         list_del(&page->lru);
2115                         update_and_free_page(h, page);
2116                         h->free_huge_pages--;
2117                         h->free_huge_pages_node[page_to_nid(page)]--;
2118                 }
2119         }
2120 }
2121 #else
2122 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2123                                                 nodemask_t *nodes_allowed)
2124 {
2125 }
2126 #endif
2127
2128 /*
2129  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
2130  * balanced by operating on them in a round-robin fashion.
2131  * Returns 1 if an adjustment was made.
2132  */
2133 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2134                                 int delta)
2135 {
2136         int nr_nodes, node;
2137
2138         VM_BUG_ON(delta != -1 && delta != 1);
2139
2140         if (delta < 0) {
2141                 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2142                         if (h->surplus_huge_pages_node[node])
2143                                 goto found;
2144                 }
2145         } else {
2146                 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2147                         if (h->surplus_huge_pages_node[node] <
2148                                         h->nr_huge_pages_node[node])
2149                                 goto found;
2150                 }
2151         }
2152         return 0;
2153
2154 found:
2155         h->surplus_huge_pages += delta;
2156         h->surplus_huge_pages_node[node] += delta;
2157         return 1;
2158 }
2159
2160 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2161 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
2162                                                 nodemask_t *nodes_allowed)
2163 {
2164         unsigned long min_count, ret;
2165
2166         if (hstate_is_gigantic(h) && !gigantic_page_supported())
2167                 return h->max_huge_pages;
2168
2169         /*
2170          * Increase the pool size
2171          * First take pages out of surplus state.  Then make up the
2172          * remaining difference by allocating fresh huge pages.
2173          *
2174          * We might race with __alloc_buddy_huge_page() here and be unable
2175          * to convert a surplus huge page to a normal huge page. That is
2176          * not critical, though, it just means the overall size of the
2177          * pool might be one hugepage larger than it needs to be, but
2178          * within all the constraints specified by the sysctls.
2179          */
2180         spin_lock(&hugetlb_lock);
2181         while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2182                 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2183                         break;
2184         }
2185
2186         while (count > persistent_huge_pages(h)) {
2187                 /*
2188                  * If this allocation races such that we no longer need the
2189                  * page, free_huge_page will handle it by freeing the page
2190                  * and reducing the surplus.
2191                  */
2192                 spin_unlock(&hugetlb_lock);
2193
2194                 /* yield cpu to avoid soft lockup */
2195                 cond_resched();
2196
2197                 if (hstate_is_gigantic(h))
2198                         ret = alloc_fresh_gigantic_page(h, nodes_allowed);
2199                 else
2200                         ret = alloc_fresh_huge_page(h, nodes_allowed);
2201                 spin_lock(&hugetlb_lock);
2202                 if (!ret)
2203                         goto out;
2204
2205                 /* Bail for signals. Probably ctrl-c from user */
2206                 if (signal_pending(current))
2207                         goto out;
2208         }
2209
2210         /*
2211          * Decrease the pool size
2212          * First return free pages to the buddy allocator (being careful
2213          * to keep enough around to satisfy reservations).  Then place
2214          * pages into surplus state as needed so the pool will shrink
2215          * to the desired size as pages become free.
2216          *
2217          * By placing pages into the surplus state independent of the
2218          * overcommit value, we are allowing the surplus pool size to
2219          * exceed overcommit. There are few sane options here. Since
2220          * __alloc_buddy_huge_page() is checking the global counter,
2221          * though, we'll note that we're not allowed to exceed surplus
2222          * and won't grow the pool anywhere else. Not until one of the
2223          * sysctls are changed, or the surplus pages go out of use.
2224          */
2225         min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2226         min_count = max(count, min_count);
2227         try_to_free_low(h, min_count, nodes_allowed);
2228         while (min_count < persistent_huge_pages(h)) {
2229                 if (!free_pool_huge_page(h, nodes_allowed, 0))
2230                         break;
2231                 cond_resched_lock(&hugetlb_lock);
2232         }
2233         while (count < persistent_huge_pages(h)) {
2234                 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2235                         break;
2236         }
2237 out:
2238         ret = persistent_huge_pages(h);
2239         spin_unlock(&hugetlb_lock);
2240         return ret;
2241 }
2242
2243 #define HSTATE_ATTR_RO(_name) \
2244         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2245
2246 #define HSTATE_ATTR(_name) \
2247         static struct kobj_attribute _name##_attr = \
2248                 __ATTR(_name, 0644, _name##_show, _name##_store)
2249
2250 static struct kobject *hugepages_kobj;
2251 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2252
2253 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2254
2255 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2256 {
2257         int i;
2258
2259         for (i = 0; i < HUGE_MAX_HSTATE; i++)
2260                 if (hstate_kobjs[i] == kobj) {
2261                         if (nidp)
2262                                 *nidp = NUMA_NO_NODE;
2263                         return &hstates[i];
2264                 }
2265
2266         return kobj_to_node_hstate(kobj, nidp);
2267 }
2268
2269 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2270                                         struct kobj_attribute *attr, char *buf)
2271 {
2272         struct hstate *h;
2273         unsigned long nr_huge_pages;
2274         int nid;
2275
2276         h = kobj_to_hstate(kobj, &nid);
2277         if (nid == NUMA_NO_NODE)
2278                 nr_huge_pages = h->nr_huge_pages;
2279         else
2280                 nr_huge_pages = h->nr_huge_pages_node[nid];
2281
2282         return sprintf(buf, "%lu\n", nr_huge_pages);
2283 }
2284
2285 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2286                                            struct hstate *h, int nid,
2287                                            unsigned long count, size_t len)
2288 {
2289         int err;
2290         NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
2291
2292         if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
2293                 err = -EINVAL;
2294                 goto out;
2295         }
2296
2297         if (nid == NUMA_NO_NODE) {
2298                 /*
2299                  * global hstate attribute
2300                  */
2301                 if (!(obey_mempolicy &&
2302                                 init_nodemask_of_mempolicy(nodes_allowed))) {
2303                         NODEMASK_FREE(nodes_allowed);
2304                         nodes_allowed = &node_states[N_MEMORY];
2305                 }
2306         } else if (nodes_allowed) {
2307                 /*
2308                  * per node hstate attribute: adjust count to global,
2309                  * but restrict alloc/free to the specified node.
2310                  */
2311                 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2312                 init_nodemask_of_node(nodes_allowed, nid);
2313         } else
2314                 nodes_allowed = &node_states[N_MEMORY];
2315
2316         h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
2317
2318         if (nodes_allowed != &node_states[N_MEMORY])
2319                 NODEMASK_FREE(nodes_allowed);
2320
2321         return len;
2322 out:
2323         NODEMASK_FREE(nodes_allowed);
2324         return err;
2325 }
2326
2327 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2328                                          struct kobject *kobj, const char *buf,
2329                                          size_t len)
2330 {
2331         struct hstate *h;
2332         unsigned long count;
2333         int nid;
2334         int err;
2335
2336         err = kstrtoul(buf, 10, &count);
2337         if (err)
2338                 return err;
2339
2340         h = kobj_to_hstate(kobj, &nid);
2341         return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2342 }
2343
2344 static ssize_t nr_hugepages_show(struct kobject *kobj,
2345                                        struct kobj_attribute *attr, char *buf)
2346 {
2347         return nr_hugepages_show_common(kobj, attr, buf);
2348 }
2349
2350 static ssize_t nr_hugepages_store(struct kobject *kobj,
2351                struct kobj_attribute *attr, const char *buf, size_t len)
2352 {
2353         return nr_hugepages_store_common(false, kobj, buf, len);
2354 }
2355 HSTATE_ATTR(nr_hugepages);
2356
2357 #ifdef CONFIG_NUMA
2358
2359 /*
2360  * hstate attribute for optionally mempolicy-based constraint on persistent
2361  * huge page alloc/free.
2362  */
2363 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2364                                        struct kobj_attribute *attr, char *buf)
2365 {
2366         return nr_hugepages_show_common(kobj, attr, buf);
2367 }
2368
2369 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2370                struct kobj_attribute *attr, const char *buf, size_t len)
2371 {
2372         return nr_hugepages_store_common(true, kobj, buf, len);
2373 }
2374 HSTATE_ATTR(nr_hugepages_mempolicy);
2375 #endif
2376
2377
2378 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2379                                         struct kobj_attribute *attr, char *buf)
2380 {
2381         struct hstate *h = kobj_to_hstate(kobj, NULL);
2382         return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2383 }
2384
2385 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2386                 struct kobj_attribute *attr, const char *buf, size_t count)
2387 {
2388         int err;
2389         unsigned long input;
2390         struct hstate *h = kobj_to_hstate(kobj, NULL);
2391
2392         if (hstate_is_gigantic(h))
2393                 return -EINVAL;
2394
2395         err = kstrtoul(buf, 10, &input);
2396         if (err)
2397                 return err;
2398
2399         spin_lock(&hugetlb_lock);
2400         h->nr_overcommit_huge_pages = input;
2401         spin_unlock(&hugetlb_lock);
2402
2403         return count;
2404 }
2405 HSTATE_ATTR(nr_overcommit_hugepages);
2406
2407 static ssize_t free_hugepages_show(struct kobject *kobj,
2408                                         struct kobj_attribute *attr, char *buf)
2409 {
2410         struct hstate *h;
2411         unsigned long free_huge_pages;
2412         int nid;
2413
2414         h = kobj_to_hstate(kobj, &nid);
2415         if (nid == NUMA_NO_NODE)
2416                 free_huge_pages = h->free_huge_pages;
2417         else
2418                 free_huge_pages = h->free_huge_pages_node[nid];
2419
2420         return sprintf(buf, "%lu\n", free_huge_pages);
2421 }
2422 HSTATE_ATTR_RO(free_hugepages);
2423
2424 static ssize_t resv_hugepages_show(struct kobject *kobj,
2425                                         struct kobj_attribute *attr, char *buf)
2426 {
2427         struct hstate *h = kobj_to_hstate(kobj, NULL);
2428         return sprintf(buf, "%lu\n", h->resv_huge_pages);
2429 }
2430 HSTATE_ATTR_RO(resv_hugepages);
2431
2432 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2433                                         struct kobj_attribute *attr, char *buf)
2434 {
2435         struct hstate *h;
2436         unsigned long surplus_huge_pages;
2437         int nid;
2438
2439         h = kobj_to_hstate(kobj, &nid);
2440         if (nid == NUMA_NO_NODE)
2441                 surplus_huge_pages = h->surplus_huge_pages;
2442         else
2443                 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2444
2445         return sprintf(buf, "%lu\n", surplus_huge_pages);
2446 }
2447 HSTATE_ATTR_RO(surplus_hugepages);
2448
2449 static struct attribute *hstate_attrs[] = {
2450         &nr_hugepages_attr.attr,
2451         &nr_overcommit_hugepages_attr.attr,
2452         &free_hugepages_attr.attr,
2453         &resv_hugepages_attr.attr,
2454         &surplus_hugepages_attr.attr,
2455 #ifdef CONFIG_NUMA
2456         &nr_hugepages_mempolicy_attr.attr,
2457 #endif
2458         NULL,
2459 };
2460
2461 static struct attribute_group hstate_attr_group = {
2462         .attrs = hstate_attrs,
2463 };
2464
2465 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2466                                     struct kobject **hstate_kobjs,
2467                                     struct attribute_group *hstate_attr_group)
2468 {
2469         int retval;
2470         int hi = hstate_index(h);
2471
2472         hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2473         if (!hstate_kobjs[hi])
2474                 return -ENOMEM;
2475
2476         retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2477         if (retval)
2478                 kobject_put(hstate_kobjs[hi]);
2479
2480         return retval;
2481 }
2482
2483 static void __init hugetlb_sysfs_init(void)
2484 {
2485         struct hstate *h;
2486         int err;
2487
2488         hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2489         if (!hugepages_kobj)
2490                 return;
2491
2492         for_each_hstate(h) {
2493                 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2494                                          hstate_kobjs, &hstate_attr_group);
2495                 if (err)
2496                         pr_err("Hugetlb: Unable to add hstate %s", h->name);
2497         }
2498 }
2499
2500 #ifdef CONFIG_NUMA
2501
2502 /*
2503  * node_hstate/s - associate per node hstate attributes, via their kobjects,
2504  * with node devices in node_devices[] using a parallel array.  The array
2505  * index of a node device or _hstate == node id.
2506  * This is here to avoid any static dependency of the node device driver, in
2507  * the base kernel, on the hugetlb module.
2508  */
2509 struct node_hstate {
2510         struct kobject          *hugepages_kobj;
2511         struct kobject          *hstate_kobjs[HUGE_MAX_HSTATE];
2512 };
2513 static struct node_hstate node_hstates[MAX_NUMNODES];
2514
2515 /*
2516  * A subset of global hstate attributes for node devices
2517  */
2518 static struct attribute *per_node_hstate_attrs[] = {
2519         &nr_hugepages_attr.attr,
2520         &free_hugepages_attr.attr,
2521         &surplus_hugepages_attr.attr,
2522         NULL,
2523 };
2524
2525 static struct attribute_group per_node_hstate_attr_group = {
2526         .attrs = per_node_hstate_attrs,
2527 };
2528
2529 /*
2530  * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2531  * Returns node id via non-NULL nidp.
2532  */
2533 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2534 {
2535         int nid;
2536
2537         for (nid = 0; nid < nr_node_ids; nid++) {
2538                 struct node_hstate *nhs = &node_hstates[nid];
2539                 int i;
2540                 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2541                         if (nhs->hstate_kobjs[i] == kobj) {
2542                                 if (nidp)
2543                                         *nidp = nid;
2544                                 return &hstates[i];
2545                         }
2546         }
2547
2548         BUG();
2549         return NULL;
2550 }
2551
2552 /*
2553  * Unregister hstate attributes from a single node device.
2554  * No-op if no hstate attributes attached.
2555  */
2556 static void hugetlb_unregister_node(struct node *node)
2557 {
2558         struct hstate *h;
2559         struct node_hstate *nhs = &node_hstates[node->dev.id];
2560
2561         if (!nhs->hugepages_kobj)
2562                 return;         /* no hstate attributes */
2563
2564         for_each_hstate(h) {
2565                 int idx = hstate_index(h);
2566                 if (nhs->hstate_kobjs[idx]) {
2567                         kobject_put(nhs->hstate_kobjs[idx]);
2568                         nhs->hstate_kobjs[idx] = NULL;
2569                 }
2570         }
2571
2572         kobject_put(nhs->hugepages_kobj);
2573         nhs->hugepages_kobj = NULL;
2574 }
2575
2576 /*
2577  * hugetlb module exit:  unregister hstate attributes from node devices
2578  * that have them.
2579  */
2580 static void hugetlb_unregister_all_nodes(void)
2581 {
2582         int nid;
2583
2584         /*
2585          * disable node device registrations.
2586          */
2587         register_hugetlbfs_with_node(NULL, NULL);
2588
2589         /*
2590          * remove hstate attributes from any nodes that have them.
2591          */
2592         for (nid = 0; nid < nr_node_ids; nid++)
2593                 hugetlb_unregister_node(node_devices[nid]);
2594 }
2595
2596 /*
2597  * Register hstate attributes for a single node device.
2598  * No-op if attributes already registered.
2599  */
2600 static void hugetlb_register_node(struct node *node)
2601 {
2602         struct hstate *h;
2603         struct node_hstate *nhs = &node_hstates[node->dev.id];
2604         int err;
2605
2606         if (nhs->hugepages_kobj)
2607                 return;         /* already allocated */
2608
2609         nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2610                                                         &node->dev.kobj);
2611         if (!nhs->hugepages_kobj)
2612                 return;
2613
2614         for_each_hstate(h) {
2615                 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2616                                                 nhs->hstate_kobjs,
2617                                                 &per_node_hstate_attr_group);
2618                 if (err) {
2619                         pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2620                                 h->name, node->dev.id);
2621                         hugetlb_unregister_node(node);
2622                         break;
2623                 }
2624         }
2625 }
2626
2627 /*
2628  * hugetlb init time:  register hstate attributes for all registered node
2629  * devices of nodes that have memory.  All on-line nodes should have
2630  * registered their associated device by this time.
2631  */
2632 static void __init hugetlb_register_all_nodes(void)
2633 {
2634         int nid;
2635
2636         for_each_node_state(nid, N_MEMORY) {
2637                 struct node *node = node_devices[nid];
2638                 if (node->dev.id == nid)
2639                         hugetlb_register_node(node);
2640         }
2641
2642         /*
2643          * Let the node device driver know we're here so it can
2644          * [un]register hstate attributes on node hotplug.
2645          */
2646         register_hugetlbfs_with_node(hugetlb_register_node,
2647                                      hugetlb_unregister_node);
2648 }
2649 #else   /* !CONFIG_NUMA */
2650
2651 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2652 {
2653         BUG();
2654         if (nidp)
2655                 *nidp = -1;
2656         return NULL;
2657 }
2658
2659 static void hugetlb_unregister_all_nodes(void) { }
2660
2661 static void hugetlb_register_all_nodes(void) { }
2662
2663 #endif
2664
2665 static void __exit hugetlb_exit(void)
2666 {
2667         struct hstate *h;
2668
2669         hugetlb_unregister_all_nodes();
2670
2671         for_each_hstate(h) {
2672                 kobject_put(hstate_kobjs[hstate_index(h)]);
2673         }
2674
2675         kobject_put(hugepages_kobj);
2676         kfree(hugetlb_fault_mutex_table);
2677 }
2678 module_exit(hugetlb_exit);
2679
2680 static int __init hugetlb_init(void)
2681 {
2682         int i;
2683
2684         if (!hugepages_supported())
2685                 return 0;
2686
2687         if (!size_to_hstate(default_hstate_size)) {
2688                 default_hstate_size = HPAGE_SIZE;
2689                 if (!size_to_hstate(default_hstate_size))
2690                         hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2691         }
2692         default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2693         if (default_hstate_max_huge_pages)
2694                 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2695
2696         hugetlb_init_hstates();
2697         gather_bootmem_prealloc();
2698         report_hugepages();
2699
2700         hugetlb_sysfs_init();
2701         hugetlb_register_all_nodes();
2702         hugetlb_cgroup_file_init();
2703
2704 #ifdef CONFIG_SMP
2705         num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2706 #else
2707         num_fault_mutexes = 1;
2708 #endif
2709         hugetlb_fault_mutex_table =
2710                 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
2711         BUG_ON(!hugetlb_fault_mutex_table);
2712
2713         for (i = 0; i < num_fault_mutexes; i++)
2714                 mutex_init(&hugetlb_fault_mutex_table[i]);
2715         return 0;
2716 }
2717 module_init(hugetlb_init);
2718
2719 /* Should be called on processing a hugepagesz=... option */
2720 void __init hugetlb_add_hstate(unsigned int order)
2721 {
2722         struct hstate *h;
2723         unsigned long i;
2724
2725         if (size_to_hstate(PAGE_SIZE << order)) {
2726                 pr_warning("hugepagesz= specified twice, ignoring\n");
2727                 return;
2728         }
2729         BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2730         BUG_ON(order == 0);
2731         h = &hstates[hugetlb_max_hstate++];
2732         h->order = order;
2733         h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2734         h->nr_huge_pages = 0;
2735         h->free_huge_pages = 0;
2736         for (i = 0; i < MAX_NUMNODES; ++i)
2737                 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2738         INIT_LIST_HEAD(&h->hugepage_activelist);
2739         h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
2740         h->next_nid_to_free = first_node(node_states[N_MEMORY]);
2741         snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2742                                         huge_page_size(h)/1024);
2743
2744         parsed_hstate = h;
2745 }
2746
2747 static int __init hugetlb_nrpages_setup(char *s)
2748 {
2749         unsigned long *mhp;
2750         static unsigned long *last_mhp;
2751
2752         /*
2753          * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2754          * so this hugepages= parameter goes to the "default hstate".
2755          */
2756         if (!hugetlb_max_hstate)
2757                 mhp = &default_hstate_max_huge_pages;
2758         else
2759                 mhp = &parsed_hstate->max_huge_pages;
2760
2761         if (mhp == last_mhp) {
2762                 pr_warning("hugepages= specified twice without "
2763                            "interleaving hugepagesz=, ignoring\n");
2764                 return 1;
2765         }
2766
2767         if (sscanf(s, "%lu", mhp) <= 0)
2768                 *mhp = 0;
2769
2770         /*
2771          * Global state is always initialized later in hugetlb_init.
2772          * But we need to allocate >= MAX_ORDER hstates here early to still
2773          * use the bootmem allocator.
2774          */
2775         if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2776                 hugetlb_hstate_alloc_pages(parsed_hstate);
2777
2778         last_mhp = mhp;
2779
2780         return 1;
2781 }
2782 __setup("hugepages=", hugetlb_nrpages_setup);
2783
2784 static int __init hugetlb_default_setup(char *s)
2785 {
2786         default_hstate_size = memparse(s, &s);
2787         return 1;
2788 }
2789 __setup("default_hugepagesz=", hugetlb_default_setup);
2790
2791 static unsigned int cpuset_mems_nr(unsigned int *array)
2792 {
2793         int node;
2794         unsigned int nr = 0;
2795
2796         for_each_node_mask(node, cpuset_current_mems_allowed)
2797                 nr += array[node];
2798
2799         return nr;
2800 }
2801
2802 #ifdef CONFIG_SYSCTL
2803 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2804                          struct ctl_table *table, int write,
2805                          void __user *buffer, size_t *length, loff_t *ppos)
2806 {
2807         struct hstate *h = &default_hstate;
2808         unsigned long tmp = h->max_huge_pages;
2809         int ret;
2810
2811         if (!hugepages_supported())
2812                 return -ENOTSUPP;
2813
2814         table->data = &tmp;
2815         table->maxlen = sizeof(unsigned long);
2816         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2817         if (ret)
2818                 goto out;
2819
2820         if (write)
2821                 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2822                                                   NUMA_NO_NODE, tmp, *length);
2823 out:
2824         return ret;
2825 }
2826
2827 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2828                           void __user *buffer, size_t *length, loff_t *ppos)
2829 {
2830
2831         return hugetlb_sysctl_handler_common(false, table, write,
2832                                                         buffer, length, ppos);
2833 }
2834
2835 #ifdef CONFIG_NUMA
2836 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2837                           void __user *buffer, size_t *length, loff_t *ppos)
2838 {
2839         return hugetlb_sysctl_handler_common(true, table, write,
2840                                                         buffer, length, ppos);
2841 }
2842 #endif /* CONFIG_NUMA */
2843
2844 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2845                         void __user *buffer,
2846                         size_t *length, loff_t *ppos)
2847 {
2848         struct hstate *h = &default_hstate;
2849         unsigned long tmp;
2850         int ret;
2851
2852         if (!hugepages_supported())
2853                 return -ENOTSUPP;
2854
2855         tmp = h->nr_overcommit_huge_pages;
2856
2857         if (write && hstate_is_gigantic(h))
2858                 return -EINVAL;
2859
2860         table->data = &tmp;
2861         table->maxlen = sizeof(unsigned long);
2862         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2863         if (ret)
2864                 goto out;
2865
2866         if (write) {
2867                 spin_lock(&hugetlb_lock);
2868                 h->nr_overcommit_huge_pages = tmp;
2869                 spin_unlock(&hugetlb_lock);
2870         }
2871 out:
2872         return ret;
2873 }
2874
2875 #endif /* CONFIG_SYSCTL */
2876
2877 void hugetlb_report_meminfo(struct seq_file *m)
2878 {
2879         struct hstate *h = &default_hstate;
2880         if (!hugepages_supported())
2881                 return;
2882         seq_printf(m,
2883                         "HugePages_Total:   %5lu\n"
2884                         "HugePages_Free:    %5lu\n"
2885                         "HugePages_Rsvd:    %5lu\n"
2886                         "HugePages_Surp:    %5lu\n"
2887                         "Hugepagesize:   %8lu kB\n",
2888                         h->nr_huge_pages,
2889                         h->free_huge_pages,
2890                         h->resv_huge_pages,
2891                         h->surplus_huge_pages,
2892                         1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2893 }
2894
2895 int hugetlb_report_node_meminfo(int nid, char *buf)
2896 {
2897         struct hstate *h = &default_hstate;
2898         if (!hugepages_supported())
2899                 return 0;
2900         return sprintf(buf,
2901                 "Node %d HugePages_Total: %5u\n"
2902                 "Node %d HugePages_Free:  %5u\n"
2903                 "Node %d HugePages_Surp:  %5u\n",
2904                 nid, h->nr_huge_pages_node[nid],
2905                 nid, h->free_huge_pages_node[nid],
2906                 nid, h->surplus_huge_pages_node[nid]);
2907 }
2908
2909 void hugetlb_show_meminfo(void)
2910 {
2911         struct hstate *h;
2912         int nid;
2913
2914         if (!hugepages_supported())
2915                 return;
2916
2917         for_each_node_state(nid, N_MEMORY)
2918                 for_each_hstate(h)
2919                         pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2920                                 nid,
2921                                 h->nr_huge_pages_node[nid],
2922                                 h->free_huge_pages_node[nid],
2923                                 h->surplus_huge_pages_node[nid],
2924                                 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2925 }
2926
2927 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
2928 {
2929         seq_printf(m, "HugetlbPages:\t%8lu kB\n",
2930                    atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
2931 }
2932
2933 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2934 unsigned long hugetlb_total_pages(void)
2935 {
2936         struct hstate *h;
2937         unsigned long nr_total_pages = 0;
2938
2939         for_each_hstate(h)
2940                 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2941         return nr_total_pages;
2942 }
2943
2944 static int hugetlb_acct_memory(struct hstate *h, long delta)
2945 {
2946         int ret = -ENOMEM;
2947
2948         spin_lock(&hugetlb_lock);
2949         /*
2950          * When cpuset is configured, it breaks the strict hugetlb page
2951          * reservation as the accounting is done on a global variable. Such
2952          * reservation is completely rubbish in the presence of cpuset because
2953          * the reservation is not checked against page availability for the
2954          * current cpuset. Application can still potentially OOM'ed by kernel
2955          * with lack of free htlb page in cpuset that the task is in.
2956          * Attempt to enforce strict accounting with cpuset is almost
2957          * impossible (or too ugly) because cpuset is too fluid that
2958          * task or memory node can be dynamically moved between cpusets.
2959          *
2960          * The change of semantics for shared hugetlb mapping with cpuset is
2961          * undesirable. However, in order to preserve some of the semantics,
2962          * we fall back to check against current free page availability as
2963          * a best attempt and hopefully to minimize the impact of changing
2964          * semantics that cpuset has.
2965          */
2966         if (delta > 0) {
2967                 if (gather_surplus_pages(h, delta) < 0)
2968                         goto out;
2969
2970                 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2971                         return_unused_surplus_pages(h, delta);
2972                         goto out;
2973                 }
2974         }
2975
2976         ret = 0;
2977         if (delta < 0)
2978                 return_unused_surplus_pages(h, (unsigned long) -delta);
2979
2980 out:
2981         spin_unlock(&hugetlb_lock);
2982         return ret;
2983 }
2984
2985 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2986 {
2987         struct resv_map *resv = vma_resv_map(vma);
2988
2989         /*
2990          * This new VMA should share its siblings reservation map if present.
2991          * The VMA will only ever have a valid reservation map pointer where
2992          * it is being copied for another still existing VMA.  As that VMA
2993          * has a reference to the reservation map it cannot disappear until
2994          * after this open call completes.  It is therefore safe to take a
2995          * new reference here without additional locking.
2996          */
2997         if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2998                 kref_get(&resv->refs);
2999 }
3000
3001 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3002 {
3003         struct hstate *h = hstate_vma(vma);
3004         struct resv_map *resv = vma_resv_map(vma);
3005         struct hugepage_subpool *spool = subpool_vma(vma);
3006         unsigned long reserve, start, end;
3007         long gbl_reserve;
3008
3009         if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3010                 return;
3011
3012         start = vma_hugecache_offset(h, vma, vma->vm_start);
3013         end = vma_hugecache_offset(h, vma, vma->vm_end);
3014
3015         reserve = (end - start) - region_count(resv, start, end);
3016
3017         kref_put(&resv->refs, resv_map_release);
3018
3019         if (reserve) {
3020                 /*
3021                  * Decrement reserve counts.  The global reserve count may be
3022                  * adjusted if the subpool has a minimum size.
3023                  */
3024                 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3025                 hugetlb_acct_memory(h, -gbl_reserve);
3026         }
3027 }
3028
3029 /*
3030  * We cannot handle pagefaults against hugetlb pages at all.  They cause
3031  * handle_mm_fault() to try to instantiate regular-sized pages in the
3032  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
3033  * this far.
3034  */
3035 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3036 {
3037         BUG();
3038         return 0;
3039 }
3040
3041 const struct vm_operations_struct hugetlb_vm_ops = {
3042         .fault = hugetlb_vm_op_fault,
3043         .open = hugetlb_vm_op_open,
3044         .close = hugetlb_vm_op_close,
3045 };
3046
3047 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3048                                 int writable)
3049 {
3050         pte_t entry;
3051
3052         if (writable) {
3053                 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3054                                          vma->vm_page_prot)));
3055         } else {
3056                 entry = huge_pte_wrprotect(mk_huge_pte(page,
3057                                            vma->vm_page_prot));
3058         }
3059         entry = pte_mkyoung(entry);
3060         entry = pte_mkhuge(entry);
3061         entry = arch_make_huge_pte(entry, vma, page, writable);
3062
3063         return entry;
3064 }
3065
3066 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3067                                    unsigned long address, pte_t *ptep)
3068 {
3069         pte_t entry;
3070
3071         entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3072         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3073                 update_mmu_cache(vma, address, ptep);
3074 }
3075
3076 static int is_hugetlb_entry_migration(pte_t pte)
3077 {
3078         swp_entry_t swp;
3079
3080         if (huge_pte_none(pte) || pte_present(pte))
3081                 return 0;
3082         swp = pte_to_swp_entry(pte);
3083         if (non_swap_entry(swp) && is_migration_entry(swp))
3084                 return 1;
3085         else
3086                 return 0;
3087 }
3088
3089 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3090 {
3091         swp_entry_t swp;
3092
3093         if (huge_pte_none(pte) || pte_present(pte))
3094                 return 0;
3095         swp = pte_to_swp_entry(pte);
3096         if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3097                 return 1;
3098         else
3099                 return 0;
3100 }
3101
3102 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3103                             struct vm_area_struct *vma)
3104 {
3105         pte_t *src_pte, *dst_pte, entry;
3106         struct page *ptepage;
3107         unsigned long addr;
3108         int cow;
3109         struct hstate *h = hstate_vma(vma);
3110         unsigned long sz = huge_page_size(h);
3111         unsigned long mmun_start;       /* For mmu_notifiers */
3112         unsigned long mmun_end;         /* For mmu_notifiers */
3113         int ret = 0;
3114
3115         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3116
3117         mmun_start = vma->vm_start;
3118         mmun_end = vma->vm_end;
3119         if (cow)
3120                 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
3121
3122         for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3123                 spinlock_t *src_ptl, *dst_ptl;
3124                 src_pte = huge_pte_offset(src, addr);
3125                 if (!src_pte)
3126                         continue;
3127                 dst_pte = huge_pte_alloc(dst, addr, sz);
3128                 if (!dst_pte) {
3129                         ret = -ENOMEM;
3130                         break;
3131                 }
3132
3133                 /* If the pagetables are shared don't copy or take references */
3134                 if (dst_pte == src_pte)
3135                         continue;
3136
3137                 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3138                 src_ptl = huge_pte_lockptr(h, src, src_pte);
3139                 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3140                 entry = huge_ptep_get(src_pte);
3141                 if (huge_pte_none(entry)) { /* skip none entry */
3142                         ;
3143                 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3144                                     is_hugetlb_entry_hwpoisoned(entry))) {
3145                         swp_entry_t swp_entry = pte_to_swp_entry(entry);
3146
3147                         if (is_write_migration_entry(swp_entry) && cow) {
3148                                 /*
3149                                  * COW mappings require pages in both
3150                                  * parent and child to be set to read.
3151                                  */
3152                                 make_migration_entry_read(&swp_entry);
3153                                 entry = swp_entry_to_pte(swp_entry);
3154                                 set_huge_pte_at(src, addr, src_pte, entry);
3155                         }
3156                         set_huge_pte_at(dst, addr, dst_pte, entry);
3157                 } else {
3158                         if (cow) {
3159                                 huge_ptep_set_wrprotect(src, addr, src_pte);
3160                                 mmu_notifier_invalidate_range(src, mmun_start,
3161                                                                    mmun_end);
3162                         }
3163                         entry = huge_ptep_get(src_pte);
3164                         ptepage = pte_page(entry);
3165                         get_page(ptepage);
3166                         page_dup_rmap(ptepage);
3167                         set_huge_pte_at(dst, addr, dst_pte, entry);
3168                         hugetlb_count_add(pages_per_huge_page(h), dst);
3169                 }
3170                 spin_unlock(src_ptl);
3171                 spin_unlock(dst_ptl);
3172         }
3173
3174         if (cow)
3175                 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
3176
3177         return ret;
3178 }
3179
3180 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3181                             unsigned long start, unsigned long end,
3182                             struct page *ref_page)
3183 {
3184         int force_flush = 0;
3185         struct mm_struct *mm = vma->vm_mm;
3186         unsigned long address;
3187         pte_t *ptep;
3188         pte_t pte;
3189         spinlock_t *ptl;
3190         struct page *page;
3191         struct hstate *h = hstate_vma(vma);
3192         unsigned long sz = huge_page_size(h);
3193         const unsigned long mmun_start = start; /* For mmu_notifiers */
3194         const unsigned long mmun_end   = end;   /* For mmu_notifiers */
3195
3196         WARN_ON(!is_vm_hugetlb_page(vma));
3197         BUG_ON(start & ~huge_page_mask(h));
3198         BUG_ON(end & ~huge_page_mask(h));
3199
3200         tlb_start_vma(tlb, vma);
3201         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3202         address = start;
3203 again:
3204         for (; address < end; address += sz) {
3205                 ptep = huge_pte_offset(mm, address);
3206                 if (!ptep)
3207                         continue;
3208
3209                 ptl = huge_pte_lock(h, mm, ptep);
3210                 if (huge_pmd_unshare(mm, &address, ptep))
3211                         goto unlock;
3212
3213                 pte = huge_ptep_get(ptep);
3214                 if (huge_pte_none(pte))
3215                         goto unlock;
3216
3217                 /*
3218                  * Migrating hugepage or HWPoisoned hugepage is already
3219                  * unmapped and its refcount is dropped, so just clear pte here.
3220                  */
3221                 if (unlikely(!pte_present(pte))) {
3222                         huge_pte_clear(mm, address, ptep);
3223                         goto unlock;
3224                 }
3225
3226                 page = pte_page(pte);
3227                 /*
3228                  * If a reference page is supplied, it is because a specific
3229                  * page is being unmapped, not a range. Ensure the page we
3230                  * are about to unmap is the actual page of interest.
3231                  */
3232                 if (ref_page) {
3233                         if (page != ref_page)
3234                                 goto unlock;
3235
3236                         /*
3237                          * Mark the VMA as having unmapped its page so that
3238                          * future faults in this VMA will fail rather than
3239                          * looking like data was lost
3240                          */
3241                         set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3242                 }
3243
3244                 pte = huge_ptep_get_and_clear(mm, address, ptep);
3245                 tlb_remove_tlb_entry(tlb, ptep, address);
3246                 if (huge_pte_dirty(pte))
3247                         set_page_dirty(page);
3248
3249                 hugetlb_count_sub(pages_per_huge_page(h), mm);
3250                 page_remove_rmap(page);
3251                 force_flush = !__tlb_remove_page(tlb, page);
3252                 if (force_flush) {
3253                         address += sz;
3254                         spin_unlock(ptl);
3255                         break;
3256                 }
3257                 /* Bail out after unmapping reference page if supplied */
3258                 if (ref_page) {
3259                         spin_unlock(ptl);
3260                         break;
3261                 }
3262 unlock:
3263                 spin_unlock(ptl);
3264         }
3265         /*
3266          * mmu_gather ran out of room to batch pages, we break out of
3267          * the PTE lock to avoid doing the potential expensive TLB invalidate
3268          * and page-free while holding it.
3269          */
3270         if (force_flush) {
3271                 force_flush = 0;
3272                 tlb_flush_mmu(tlb);
3273                 if (address < end && !ref_page)
3274                         goto again;
3275         }
3276         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3277         tlb_end_vma(tlb, vma);
3278 }
3279
3280 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3281                           struct vm_area_struct *vma, unsigned long start,
3282                           unsigned long end, struct page *ref_page)
3283 {
3284         __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3285
3286         /*
3287          * Clear this flag so that x86's huge_pmd_share page_table_shareable
3288          * test will fail on a vma being torn down, and not grab a page table
3289          * on its way out.  We're lucky that the flag has such an appropriate
3290          * name, and can in fact be safely cleared here. We could clear it
3291          * before the __unmap_hugepage_range above, but all that's necessary
3292          * is to clear it before releasing the i_mmap_rwsem. This works
3293          * because in the context this is called, the VMA is about to be
3294          * destroyed and the i_mmap_rwsem is held.
3295          */
3296         vma->vm_flags &= ~VM_MAYSHARE;
3297 }
3298
3299 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3300                           unsigned long end, struct page *ref_page)
3301 {
3302         struct mm_struct *mm;
3303         struct mmu_gather tlb;
3304
3305         mm = vma->vm_mm;
3306
3307         tlb_gather_mmu(&tlb, mm, start, end);
3308         __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3309         tlb_finish_mmu(&tlb, start, end);
3310 }
3311
3312 /*
3313  * This is called when the original mapper is failing to COW a MAP_PRIVATE
3314  * mappping it owns the reserve page for. The intention is to unmap the page
3315  * from other VMAs and let the children be SIGKILLed if they are faulting the
3316  * same region.
3317  */
3318 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3319                               struct page *page, unsigned long address)
3320 {
3321         struct hstate *h = hstate_vma(vma);
3322         struct vm_area_struct *iter_vma;
3323         struct address_space *mapping;
3324         pgoff_t pgoff;
3325
3326         /*
3327          * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3328          * from page cache lookup which is in HPAGE_SIZE units.
3329          */
3330         address = address & huge_page_mask(h);
3331         pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3332                         vma->vm_pgoff;
3333         mapping = file_inode(vma->vm_file)->i_mapping;
3334
3335         /*
3336          * Take the mapping lock for the duration of the table walk. As
3337          * this mapping should be shared between all the VMAs,
3338          * __unmap_hugepage_range() is called as the lock is already held
3339          */
3340         i_mmap_lock_write(mapping);
3341         vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3342                 /* Do not unmap the current VMA */
3343                 if (iter_vma == vma)
3344                         continue;
3345
3346                 /*
3347                  * Shared VMAs have their own reserves and do not affect
3348                  * MAP_PRIVATE accounting but it is possible that a shared
3349                  * VMA is using the same page so check and skip such VMAs.
3350                  */
3351                 if (iter_vma->vm_flags & VM_MAYSHARE)
3352                         continue;
3353
3354                 /*
3355                  * Unmap the page from other VMAs without their own reserves.
3356                  * They get marked to be SIGKILLed if they fault in these
3357                  * areas. This is because a future no-page fault on this VMA
3358                  * could insert a zeroed page instead of the data existing
3359                  * from the time of fork. This would look like data corruption
3360                  */
3361                 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3362                         unmap_hugepage_range(iter_vma, address,
3363                                              address + huge_page_size(h), page);
3364         }
3365         i_mmap_unlock_write(mapping);
3366 }
3367
3368 /*
3369  * Hugetlb_cow() should be called with page lock of the original hugepage held.
3370  * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3371  * cannot race with other handlers or page migration.
3372  * Keep the pte_same checks anyway to make transition from the mutex easier.
3373  */
3374 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3375                         unsigned long address, pte_t *ptep, pte_t pte,
3376                         struct page *pagecache_page, spinlock_t *ptl)
3377 {
3378         struct hstate *h = hstate_vma(vma);
3379         struct page *old_page, *new_page;
3380         int ret = 0, outside_reserve = 0;
3381         unsigned long mmun_start;       /* For mmu_notifiers */
3382         unsigned long mmun_end;         /* For mmu_notifiers */
3383
3384         old_page = pte_page(pte);
3385
3386 retry_avoidcopy:
3387         /* If no-one else is actually using this page, avoid the copy
3388          * and just make the page writable */
3389         if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3390                 page_move_anon_rmap(old_page, vma, address);
3391                 set_huge_ptep_writable(vma, address, ptep);
3392                 return 0;
3393         }
3394
3395         /*
3396          * If the process that created a MAP_PRIVATE mapping is about to
3397          * perform a COW due to a shared page count, attempt to satisfy
3398          * the allocation without using the existing reserves. The pagecache
3399          * page is used to determine if the reserve at this address was
3400          * consumed or not. If reserves were used, a partial faulted mapping
3401          * at the time of fork() could consume its reserves on COW instead
3402          * of the full address range.
3403          */
3404         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3405                         old_page != pagecache_page)
3406                 outside_reserve = 1;
3407
3408         page_cache_get(old_page);
3409
3410         /*
3411          * Drop page table lock as buddy allocator may be called. It will
3412          * be acquired again before returning to the caller, as expected.
3413          */
3414         spin_unlock(ptl);
3415         new_page = alloc_huge_page(vma, address, outside_reserve);
3416
3417         if (IS_ERR(new_page)) {
3418                 /*
3419                  * If a process owning a MAP_PRIVATE mapping fails to COW,
3420                  * it is due to references held by a child and an insufficient
3421                  * huge page pool. To guarantee the original mappers
3422                  * reliability, unmap the page from child processes. The child
3423                  * may get SIGKILLed if it later faults.
3424                  */
3425                 if (outside_reserve) {
3426                         page_cache_release(old_page);
3427                         BUG_ON(huge_pte_none(pte));
3428                         unmap_ref_private(mm, vma, old_page, address);
3429                         BUG_ON(huge_pte_none(pte));
3430                         spin_lock(ptl);
3431                         ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3432                         if (likely(ptep &&
3433                                    pte_same(huge_ptep_get(ptep), pte)))
3434                                 goto retry_avoidcopy;
3435                         /*
3436                          * race occurs while re-acquiring page table
3437                          * lock, and our job is done.
3438                          */
3439                         return 0;
3440                 }
3441
3442                 ret = (PTR_ERR(new_page) == -ENOMEM) ?
3443                         VM_FAULT_OOM : VM_FAULT_SIGBUS;
3444                 goto out_release_old;
3445         }
3446
3447         /*
3448          * When the original hugepage is shared one, it does not have
3449          * anon_vma prepared.
3450          */
3451         if (unlikely(anon_vma_prepare(vma))) {
3452                 ret = VM_FAULT_OOM;
3453                 goto out_release_all;
3454         }
3455
3456         copy_user_huge_page(new_page, old_page, address, vma,
3457                             pages_per_huge_page(h));
3458         __SetPageUptodate(new_page);
3459         set_page_huge_active(new_page);
3460
3461         mmun_start = address & huge_page_mask(h);
3462         mmun_end = mmun_start + huge_page_size(h);
3463         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3464
3465         /*
3466          * Retake the page table lock to check for racing updates
3467          * before the page tables are altered
3468          */
3469         spin_lock(ptl);
3470         ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3471         if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3472                 ClearPagePrivate(new_page);
3473
3474                 /* Break COW */
3475                 huge_ptep_clear_flush(vma, address, ptep);
3476                 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
3477                 set_huge_pte_at(mm, address, ptep,
3478                                 make_huge_pte(vma, new_page, 1));
3479                 page_remove_rmap(old_page);
3480                 hugepage_add_new_anon_rmap(new_page, vma, address);
3481                 /* Make the old page be freed below */
3482                 new_page = old_page;
3483         }
3484         spin_unlock(ptl);
3485         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3486 out_release_all:
3487         page_cache_release(new_page);
3488 out_release_old:
3489         page_cache_release(old_page);
3490
3491         spin_lock(ptl); /* Caller expects lock to be held */
3492         return ret;
3493 }
3494
3495 /* Return the pagecache page at a given address within a VMA */
3496 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3497                         struct vm_area_struct *vma, unsigned long address)
3498 {
3499         struct address_space *mapping;
3500         pgoff_t idx;
3501
3502         mapping = vma->vm_file->f_mapping;
3503         idx = vma_hugecache_offset(h, vma, address);
3504
3505         return find_lock_page(mapping, idx);
3506 }
3507
3508 /*
3509  * Return whether there is a pagecache page to back given address within VMA.
3510  * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3511  */
3512 static bool hugetlbfs_pagecache_present(struct hstate *h,
3513                         struct vm_area_struct *vma, unsigned long address)
3514 {
3515         struct address_space *mapping;
3516         pgoff_t idx;
3517         struct page *page;
3518
3519         mapping = vma->vm_file->f_mapping;
3520         idx = vma_hugecache_offset(h, vma, address);
3521
3522         page = find_get_page(mapping, idx);
3523         if (page)
3524                 put_page(page);
3525         return page != NULL;
3526 }
3527
3528 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3529                            pgoff_t idx)
3530 {
3531         struct inode *inode = mapping->host;
3532         struct hstate *h = hstate_inode(inode);
3533         int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3534
3535         if (err)
3536                 return err;
3537         ClearPagePrivate(page);
3538
3539         spin_lock(&inode->i_lock);
3540         inode->i_blocks += blocks_per_huge_page(h);
3541         spin_unlock(&inode->i_lock);
3542         return 0;
3543 }
3544
3545 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
3546                            struct address_space *mapping, pgoff_t idx,
3547                            unsigned long address, pte_t *ptep, unsigned int flags)
3548 {
3549         struct hstate *h = hstate_vma(vma);
3550         int ret = VM_FAULT_SIGBUS;
3551         int anon_rmap = 0;
3552         unsigned long size;
3553         struct page *page;
3554         pte_t new_pte;
3555         spinlock_t *ptl;
3556
3557         /*
3558          * Currently, we are forced to kill the process in the event the
3559          * original mapper has unmapped pages from the child due to a failed
3560          * COW. Warn that such a situation has occurred as it may not be obvious
3561          */
3562         if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3563                 pr_warning("PID %d killed due to inadequate hugepage pool\n",
3564                            current->pid);
3565                 return ret;
3566         }
3567
3568         /*
3569          * Use page lock to guard against racing truncation
3570          * before we get page_table_lock.
3571          */
3572 retry:
3573         page = find_lock_page(mapping, idx);
3574         if (!page) {
3575                 size = i_size_read(mapping->host) >> huge_page_shift(h);
3576                 if (idx >= size)
3577                         goto out;
3578                 page = alloc_huge_page(vma, address, 0);
3579                 if (IS_ERR(page)) {
3580                         ret = PTR_ERR(page);
3581                         if (ret == -ENOMEM)
3582                                 ret = VM_FAULT_OOM;
3583                         else
3584                                 ret = VM_FAULT_SIGBUS;
3585                         goto out;
3586                 }
3587                 clear_huge_page(page, address, pages_per_huge_page(h));
3588                 __SetPageUptodate(page);
3589                 set_page_huge_active(page);
3590
3591                 if (vma->vm_flags & VM_MAYSHARE) {
3592                         int err = huge_add_to_page_cache(page, mapping, idx);
3593                         if (err) {
3594                                 put_page(page);
3595                                 if (err == -EEXIST)
3596                                         goto retry;
3597                                 goto out;
3598                         }
3599                 } else {
3600                         lock_page(page);
3601                         if (unlikely(anon_vma_prepare(vma))) {
3602                                 ret = VM_FAULT_OOM;
3603                                 goto backout_unlocked;
3604                         }
3605                         anon_rmap = 1;
3606                 }
3607         } else {
3608                 /*
3609                  * If memory error occurs between mmap() and fault, some process
3610                  * don't have hwpoisoned swap entry for errored virtual address.
3611                  * So we need to block hugepage fault by PG_hwpoison bit check.
3612                  */
3613                 if (unlikely(PageHWPoison(page))) {
3614                         ret = VM_FAULT_HWPOISON |
3615                                 VM_FAULT_SET_HINDEX(hstate_index(h));
3616                         goto backout_unlocked;
3617                 }
3618         }
3619
3620         /*
3621          * If we are going to COW a private mapping later, we examine the
3622          * pending reservations for this page now. This will ensure that
3623          * any allocations necessary to record that reservation occur outside
3624          * the spinlock.
3625          */
3626         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3627                 if (vma_needs_reservation(h, vma, address) < 0) {
3628                         ret = VM_FAULT_OOM;
3629                         goto backout_unlocked;
3630                 }
3631                 /* Just decrements count, does not deallocate */
3632                 vma_end_reservation(h, vma, address);
3633         }
3634
3635         ptl = huge_pte_lockptr(h, mm, ptep);
3636         spin_lock(ptl);
3637         size = i_size_read(mapping->host) >> huge_page_shift(h);
3638         if (idx >= size)
3639                 goto backout;
3640
3641         ret = 0;
3642         if (!huge_pte_none(huge_ptep_get(ptep)))
3643                 goto backout;
3644
3645         if (anon_rmap) {
3646                 ClearPagePrivate(page);
3647                 hugepage_add_new_anon_rmap(page, vma, address);
3648         } else
3649                 page_dup_rmap(page);
3650         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3651                                 && (vma->vm_flags & VM_SHARED)));
3652         set_huge_pte_at(mm, address, ptep, new_pte);
3653
3654         hugetlb_count_add(pages_per_huge_page(h), mm);
3655         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3656                 /* Optimization, do the COW without a second fault */
3657                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl);
3658         }
3659
3660         spin_unlock(ptl);
3661         unlock_page(page);
3662 out:
3663         return ret;
3664
3665 backout:
3666         spin_unlock(ptl);
3667 backout_unlocked:
3668         unlock_page(page);
3669         put_page(page);
3670         goto out;
3671 }
3672
3673 #ifdef CONFIG_SMP
3674 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3675                             struct vm_area_struct *vma,
3676                             struct address_space *mapping,
3677                             pgoff_t idx, unsigned long address)
3678 {
3679         unsigned long key[2];
3680         u32 hash;
3681
3682         if (vma->vm_flags & VM_SHARED) {
3683                 key[0] = (unsigned long) mapping;
3684                 key[1] = idx;
3685         } else {
3686                 key[0] = (unsigned long) mm;
3687                 key[1] = address >> huge_page_shift(h);
3688         }
3689
3690         hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3691
3692         return hash & (num_fault_mutexes - 1);
3693 }
3694 #else
3695 /*
3696  * For uniprocesor systems we always use a single mutex, so just
3697  * return 0 and avoid the hashing overhead.
3698  */
3699 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3700                             struct vm_area_struct *vma,
3701                             struct address_space *mapping,
3702                             pgoff_t idx, unsigned long address)
3703 {
3704         return 0;
3705 }
3706 #endif
3707
3708 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3709                         unsigned long address, unsigned int flags)
3710 {
3711         pte_t *ptep, entry;
3712         spinlock_t *ptl;
3713         int ret;
3714         u32 hash;
3715         pgoff_t idx;
3716         struct page *page = NULL;
3717         struct page *pagecache_page = NULL;
3718         struct hstate *h = hstate_vma(vma);
3719         struct address_space *mapping;
3720         int need_wait_lock = 0;
3721
3722         address &= huge_page_mask(h);
3723
3724         ptep = huge_pte_offset(mm, address);
3725         if (ptep) {
3726                 entry = huge_ptep_get(ptep);
3727                 if (unlikely(is_hugetlb_entry_migration(entry))) {
3728                         migration_entry_wait_huge(vma, mm, ptep);
3729                         return 0;
3730                 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
3731                         return VM_FAULT_HWPOISON_LARGE |
3732                                 VM_FAULT_SET_HINDEX(hstate_index(h));
3733         } else {
3734                 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
3735                 if (!ptep)
3736                         return VM_FAULT_OOM;
3737         }
3738
3739         mapping = vma->vm_file->f_mapping;
3740         idx = vma_hugecache_offset(h, vma, address);
3741
3742         /*
3743          * Serialize hugepage allocation and instantiation, so that we don't
3744          * get spurious allocation failures if two CPUs race to instantiate
3745          * the same page in the page cache.
3746          */
3747         hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address);
3748         mutex_lock(&hugetlb_fault_mutex_table[hash]);
3749
3750         entry = huge_ptep_get(ptep);
3751         if (huge_pte_none(entry)) {
3752                 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
3753                 goto out_mutex;
3754         }
3755
3756         ret = 0;
3757
3758         /*
3759          * entry could be a migration/hwpoison entry at this point, so this
3760          * check prevents the kernel from going below assuming that we have
3761          * a active hugepage in pagecache. This goto expects the 2nd page fault,
3762          * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3763          * handle it.
3764          */
3765         if (!pte_present(entry))
3766                 goto out_mutex;
3767
3768         /*
3769          * If we are going to COW the mapping later, we examine the pending
3770          * reservations for this page now. This will ensure that any
3771          * allocations necessary to record that reservation occur outside the
3772          * spinlock. For private mappings, we also lookup the pagecache
3773          * page now as it is used to determine if a reservation has been
3774          * consumed.
3775          */
3776         if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
3777                 if (vma_needs_reservation(h, vma, address) < 0) {
3778                         ret = VM_FAULT_OOM;
3779                         goto out_mutex;
3780                 }
3781                 /* Just decrements count, does not deallocate */
3782                 vma_end_reservation(h, vma, address);
3783
3784                 if (!(vma->vm_flags & VM_MAYSHARE))
3785                         pagecache_page = hugetlbfs_pagecache_page(h,
3786                                                                 vma, address);
3787         }
3788
3789         ptl = huge_pte_lock(h, mm, ptep);
3790
3791         /* Check for a racing update before calling hugetlb_cow */
3792         if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
3793                 goto out_ptl;
3794
3795         /*
3796          * hugetlb_cow() requires page locks of pte_page(entry) and
3797          * pagecache_page, so here we need take the former one
3798          * when page != pagecache_page or !pagecache_page.
3799          */
3800         page = pte_page(entry);
3801         if (page != pagecache_page)
3802                 if (!trylock_page(page)) {
3803                         need_wait_lock = 1;
3804                         goto out_ptl;
3805                 }
3806
3807         get_page(page);
3808
3809         if (flags & FAULT_FLAG_WRITE) {
3810                 if (!huge_pte_write(entry)) {
3811                         ret = hugetlb_cow(mm, vma, address, ptep, entry,
3812                                         pagecache_page, ptl);
3813                         goto out_put_page;
3814                 }
3815                 entry = huge_pte_mkdirty(entry);
3816         }
3817         entry = pte_mkyoung(entry);
3818         if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3819                                                 flags & FAULT_FLAG_WRITE))
3820                 update_mmu_cache(vma, address, ptep);
3821 out_put_page:
3822         if (page != pagecache_page)
3823                 unlock_page(page);
3824         put_page(page);
3825 out_ptl:
3826         spin_unlock(ptl);
3827
3828         if (pagecache_page) {
3829                 unlock_page(pagecache_page);
3830                 put_page(pagecache_page);
3831         }
3832 out_mutex:
3833         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3834         /*
3835          * Generally it's safe to hold refcount during waiting page lock. But
3836          * here we just wait to defer the next page fault to avoid busy loop and
3837          * the page is not used after unlocked before returning from the current
3838          * page fault. So we are safe from accessing freed page, even if we wait
3839          * here without taking refcount.
3840          */
3841         if (need_wait_lock)
3842                 wait_on_page_locked(page);
3843         return ret;
3844 }
3845
3846 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3847                          struct page **pages, struct vm_area_struct **vmas,
3848                          unsigned long *position, unsigned long *nr_pages,
3849                          long i, unsigned int flags)
3850 {
3851         unsigned long pfn_offset;
3852         unsigned long vaddr = *position;
3853         unsigned long remainder = *nr_pages;
3854         struct hstate *h = hstate_vma(vma);
3855
3856         while (vaddr < vma->vm_end && remainder) {
3857                 pte_t *pte;
3858                 spinlock_t *ptl = NULL;
3859                 int absent;
3860                 struct page *page;
3861
3862                 /*
3863                  * If we have a pending SIGKILL, don't keep faulting pages and
3864                  * potentially allocating memory.
3865                  */
3866                 if (unlikely(fatal_signal_pending(current))) {
3867                         remainder = 0;
3868                         break;
3869                 }
3870
3871                 /*
3872                  * Some archs (sparc64, sh*) have multiple pte_ts to
3873                  * each hugepage.  We have to make sure we get the
3874                  * first, for the page indexing below to work.
3875                  *
3876                  * Note that page table lock is not held when pte is null.
3877                  */
3878                 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
3879                 if (pte)
3880                         ptl = huge_pte_lock(h, mm, pte);
3881                 absent = !pte || huge_pte_none(huge_ptep_get(pte));
3882
3883                 /*
3884                  * When coredumping, it suits get_dump_page if we just return
3885                  * an error where there's an empty slot with no huge pagecache
3886                  * to back it.  This way, we avoid allocating a hugepage, and
3887                  * the sparse dumpfile avoids allocating disk blocks, but its
3888                  * huge holes still show up with zeroes where they need to be.
3889                  */
3890                 if (absent && (flags & FOLL_DUMP) &&
3891                     !hugetlbfs_pagecache_present(h, vma, vaddr)) {
3892                         if (pte)
3893                                 spin_unlock(ptl);
3894                         remainder = 0;
3895                         break;
3896                 }
3897
3898                 /*
3899                  * We need call hugetlb_fault for both hugepages under migration
3900                  * (in which case hugetlb_fault waits for the migration,) and
3901                  * hwpoisoned hugepages (in which case we need to prevent the
3902                  * caller from accessing to them.) In order to do this, we use
3903                  * here is_swap_pte instead of is_hugetlb_entry_migration and
3904                  * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3905                  * both cases, and because we can't follow correct pages
3906                  * directly from any kind of swap entries.
3907                  */
3908                 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
3909                     ((flags & FOLL_WRITE) &&
3910                       !huge_pte_write(huge_ptep_get(pte)))) {
3911                         int ret;
3912
3913                         if (pte)
3914                                 spin_unlock(ptl);
3915                         ret = hugetlb_fault(mm, vma, vaddr,
3916                                 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
3917                         if (!(ret & VM_FAULT_ERROR))
3918                                 continue;
3919
3920                         remainder = 0;
3921                         break;
3922                 }
3923
3924                 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
3925                 page = pte_page(huge_ptep_get(pte));
3926 same_page:
3927                 if (pages) {
3928                         pages[i] = mem_map_offset(page, pfn_offset);
3929                         get_page_foll(pages[i]);
3930                 }
3931
3932                 if (vmas)
3933                         vmas[i] = vma;
3934
3935                 vaddr += PAGE_SIZE;
3936                 ++pfn_offset;
3937                 --remainder;
3938                 ++i;
3939                 if (vaddr < vma->vm_end && remainder &&
3940                                 pfn_offset < pages_per_huge_page(h)) {
3941                         /*
3942                          * We use pfn_offset to avoid touching the pageframes
3943                          * of this compound page.
3944                          */
3945                         goto same_page;
3946                 }
3947                 spin_unlock(ptl);
3948         }
3949         *nr_pages = remainder;
3950         *position = vaddr;
3951
3952         return i ? i : -EFAULT;
3953 }
3954
3955 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
3956                 unsigned long address, unsigned long end, pgprot_t newprot)
3957 {
3958         struct mm_struct *mm = vma->vm_mm;
3959         unsigned long start = address;
3960         pte_t *ptep;
3961         pte_t pte;
3962         struct hstate *h = hstate_vma(vma);
3963         unsigned long pages = 0;
3964
3965         BUG_ON(address >= end);
3966         flush_cache_range(vma, address, end);
3967
3968         mmu_notifier_invalidate_range_start(mm, start, end);
3969         i_mmap_lock_write(vma->vm_file->f_mapping);
3970         for (; address < end; address += huge_page_size(h)) {
3971                 spinlock_t *ptl;
3972                 ptep = huge_pte_offset(mm, address);
3973                 if (!ptep)
3974                         continue;
3975                 ptl = huge_pte_lock(h, mm, ptep);
3976                 if (huge_pmd_unshare(mm, &address, ptep)) {
3977                         pages++;
3978                         spin_unlock(ptl);
3979                         continue;
3980                 }
3981                 pte = huge_ptep_get(ptep);
3982                 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
3983                         spin_unlock(ptl);
3984                         continue;
3985                 }
3986                 if (unlikely(is_hugetlb_entry_migration(pte))) {
3987                         swp_entry_t entry = pte_to_swp_entry(pte);
3988
3989                         if (is_write_migration_entry(entry)) {
3990                                 pte_t newpte;
3991
3992                                 make_migration_entry_read(&entry);
3993                                 newpte = swp_entry_to_pte(entry);
3994                                 set_huge_pte_at(mm, address, ptep, newpte);
3995                                 pages++;
3996                         }
3997                         spin_unlock(ptl);
3998                         continue;
3999                 }
4000                 if (!huge_pte_none(pte)) {
4001                         pte = huge_ptep_get_and_clear(mm, address, ptep);
4002                         pte = pte_mkhuge(huge_pte_modify(pte, newprot));
4003                         pte = arch_make_huge_pte(pte, vma, NULL, 0);
4004                         set_huge_pte_at(mm, address, ptep, pte);
4005                         pages++;
4006                 }
4007                 spin_unlock(ptl);
4008         }
4009         /*
4010          * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
4011          * may have cleared our pud entry and done put_page on the page table:
4012          * once we release i_mmap_rwsem, another task can do the final put_page
4013          * and that page table be reused and filled with junk.
4014          */
4015         flush_tlb_range(vma, start, end);
4016         mmu_notifier_invalidate_range(mm, start, end);
4017         i_mmap_unlock_write(vma->vm_file->f_mapping);
4018         mmu_notifier_invalidate_range_end(mm, start, end);
4019
4020         return pages << h->order;
4021 }
4022
4023 int hugetlb_reserve_pages(struct inode *inode,
4024                                         long from, long to,
4025                                         struct vm_area_struct *vma,
4026                                         vm_flags_t vm_flags)
4027 {
4028         long ret, chg;
4029         struct hstate *h = hstate_inode(inode);
4030         struct hugepage_subpool *spool = subpool_inode(inode);
4031         struct resv_map *resv_map;
4032         long gbl_reserve;
4033
4034         /*
4035          * Only apply hugepage reservation if asked. At fault time, an
4036          * attempt will be made for VM_NORESERVE to allocate a page
4037          * without using reserves
4038          */
4039         if (vm_flags & VM_NORESERVE)
4040                 return 0;
4041
4042         /*
4043          * Shared mappings base their reservation on the number of pages that
4044          * are already allocated on behalf of the file. Private mappings need
4045          * to reserve the full area even if read-only as mprotect() may be
4046          * called to make the mapping read-write. Assume !vma is a shm mapping
4047          */
4048         if (!vma || vma->vm_flags & VM_MAYSHARE) {
4049                 resv_map = inode_resv_map(inode);
4050
4051                 chg = region_chg(resv_map, from, to);
4052
4053         } else {
4054                 resv_map = resv_map_alloc();
4055                 if (!resv_map)
4056                         return -ENOMEM;
4057
4058                 chg = to - from;
4059
4060                 set_vma_resv_map(vma, resv_map);
4061                 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4062         }
4063
4064         if (chg < 0) {
4065                 ret = chg;
4066                 goto out_err;
4067         }
4068
4069         /*
4070          * There must be enough pages in the subpool for the mapping. If
4071          * the subpool has a minimum size, there may be some global
4072          * reservations already in place (gbl_reserve).
4073          */
4074         gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4075         if (gbl_reserve < 0) {
4076                 ret = -ENOSPC;
4077                 goto out_err;
4078         }
4079
4080         /*
4081          * Check enough hugepages are available for the reservation.
4082          * Hand the pages back to the subpool if there are not
4083          */
4084         ret = hugetlb_acct_memory(h, gbl_reserve);
4085         if (ret < 0) {
4086                 /* put back original number of pages, chg */
4087                 (void)hugepage_subpool_put_pages(spool, chg);
4088                 goto out_err;
4089         }
4090
4091         /*
4092          * Account for the reservations made. Shared mappings record regions
4093          * that have reservations as they are shared by multiple VMAs.
4094          * When the last VMA disappears, the region map says how much
4095          * the reservation was and the page cache tells how much of
4096          * the reservation was consumed. Private mappings are per-VMA and
4097          * only the consumed reservations are tracked. When the VMA
4098          * disappears, the original reservation is the VMA size and the
4099          * consumed reservations are stored in the map. Hence, nothing
4100          * else has to be done for private mappings here
4101          */
4102         if (!vma || vma->vm_flags & VM_MAYSHARE) {
4103                 long add = region_add(resv_map, from, to);
4104
4105                 if (unlikely(chg > add)) {
4106                         /*
4107                          * pages in this range were added to the reserve
4108                          * map between region_chg and region_add.  This
4109                          * indicates a race with alloc_huge_page.  Adjust
4110                          * the subpool and reserve counts modified above
4111                          * based on the difference.
4112                          */
4113                         long rsv_adjust;
4114
4115                         rsv_adjust = hugepage_subpool_put_pages(spool,
4116                                                                 chg - add);
4117                         hugetlb_acct_memory(h, -rsv_adjust);
4118                 }
4119         }
4120         return 0;
4121 out_err:
4122         if (!vma || vma->vm_flags & VM_MAYSHARE)
4123                 region_abort(resv_map, from, to);
4124         if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4125                 kref_put(&resv_map->refs, resv_map_release);
4126         return ret;
4127 }
4128
4129 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4130                                                                 long freed)
4131 {
4132         struct hstate *h = hstate_inode(inode);
4133         struct resv_map *resv_map = inode_resv_map(inode);
4134         long chg = 0;
4135         struct hugepage_subpool *spool = subpool_inode(inode);
4136         long gbl_reserve;
4137
4138         if (resv_map) {
4139                 chg = region_del(resv_map, start, end);
4140                 /*
4141                  * region_del() can fail in the rare case where a region
4142                  * must be split and another region descriptor can not be
4143                  * allocated.  If end == LONG_MAX, it will not fail.
4144                  */
4145                 if (chg < 0)
4146                         return chg;
4147         }
4148
4149         spin_lock(&inode->i_lock);
4150         inode->i_blocks -= (blocks_per_huge_page(h) * freed);
4151         spin_unlock(&inode->i_lock);
4152
4153         /*
4154          * If the subpool has a minimum size, the number of global
4155          * reservations to be released may be adjusted.
4156          */
4157         gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4158         hugetlb_acct_memory(h, -gbl_reserve);
4159
4160         return 0;
4161 }
4162
4163 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4164 static unsigned long page_table_shareable(struct vm_area_struct *svma,
4165                                 struct vm_area_struct *vma,
4166                                 unsigned long addr, pgoff_t idx)
4167 {
4168         unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4169                                 svma->vm_start;
4170         unsigned long sbase = saddr & PUD_MASK;
4171         unsigned long s_end = sbase + PUD_SIZE;
4172
4173         /* Allow segments to share if only one is marked locked */
4174         unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4175         unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
4176
4177         /*
4178          * match the virtual addresses, permission and the alignment of the
4179          * page table page.
4180          */
4181         if (pmd_index(addr) != pmd_index(saddr) ||
4182             vm_flags != svm_flags ||
4183             sbase < svma->vm_start || svma->vm_end < s_end)
4184                 return 0;
4185
4186         return saddr;
4187 }
4188
4189 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
4190 {
4191         unsigned long base = addr & PUD_MASK;
4192         unsigned long end = base + PUD_SIZE;
4193
4194         /*
4195          * check on proper vm_flags and page table alignment
4196          */
4197         if (vma->vm_flags & VM_MAYSHARE &&
4198             vma->vm_start <= base && end <= vma->vm_end)
4199                 return true;
4200         return false;
4201 }
4202
4203 /*
4204  * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4205  * and returns the corresponding pte. While this is not necessary for the
4206  * !shared pmd case because we can allocate the pmd later as well, it makes the
4207  * code much cleaner. pmd allocation is essential for the shared case because
4208  * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4209  * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4210  * bad pmd for sharing.
4211  */
4212 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4213 {
4214         struct vm_area_struct *vma = find_vma(mm, addr);
4215         struct address_space *mapping = vma->vm_file->f_mapping;
4216         pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4217                         vma->vm_pgoff;
4218         struct vm_area_struct *svma;
4219         unsigned long saddr;
4220         pte_t *spte = NULL;
4221         pte_t *pte;
4222         spinlock_t *ptl;
4223
4224         if (!vma_shareable(vma, addr))
4225                 return (pte_t *)pmd_alloc(mm, pud, addr);
4226
4227         i_mmap_lock_write(mapping);
4228         vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4229                 if (svma == vma)
4230                         continue;
4231
4232                 saddr = page_table_shareable(svma, vma, addr, idx);
4233                 if (saddr) {
4234                         spte = huge_pte_offset(svma->vm_mm, saddr);
4235                         if (spte) {
4236                                 get_page(virt_to_page(spte));
4237                                 break;
4238                         }
4239                 }
4240         }
4241
4242         if (!spte)
4243                 goto out;
4244
4245         ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte);
4246         spin_lock(ptl);
4247         if (pud_none(*pud)) {
4248                 pud_populate(mm, pud,
4249                                 (pmd_t *)((unsigned long)spte & PAGE_MASK));
4250                 mm_inc_nr_pmds(mm);
4251         } else {
4252                 put_page(virt_to_page(spte));
4253         }
4254         spin_unlock(ptl);
4255 out:
4256         pte = (pte_t *)pmd_alloc(mm, pud, addr);
4257         i_mmap_unlock_write(mapping);
4258         return pte;
4259 }
4260
4261 /*
4262  * unmap huge page backed by shared pte.
4263  *
4264  * Hugetlb pte page is ref counted at the time of mapping.  If pte is shared
4265  * indicated by page_count > 1, unmap is achieved by clearing pud and
4266  * decrementing the ref count. If count == 1, the pte page is not shared.
4267  *
4268  * called with page table lock held.
4269  *
4270  * returns: 1 successfully unmapped a shared pte page
4271  *          0 the underlying pte page is not shared, or it is the last user
4272  */
4273 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4274 {
4275         pgd_t *pgd = pgd_offset(mm, *addr);
4276         pud_t *pud = pud_offset(pgd, *addr);
4277
4278         BUG_ON(page_count(virt_to_page(ptep)) == 0);
4279         if (page_count(virt_to_page(ptep)) == 1)
4280                 return 0;
4281
4282         pud_clear(pud);
4283         put_page(virt_to_page(ptep));
4284         mm_dec_nr_pmds(mm);
4285         *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4286         return 1;
4287 }
4288 #define want_pmd_share()        (1)
4289 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4290 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4291 {
4292         return NULL;
4293 }
4294
4295 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4296 {
4297         return 0;
4298 }
4299 #define want_pmd_share()        (0)
4300 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4301
4302 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4303 pte_t *huge_pte_alloc(struct mm_struct *mm,
4304                         unsigned long addr, unsigned long sz)
4305 {
4306         pgd_t *pgd;
4307         pud_t *pud;
4308         pte_t *pte = NULL;
4309
4310         pgd = pgd_offset(mm, addr);
4311         pud = pud_alloc(mm, pgd, addr);
4312         if (pud) {
4313                 if (sz == PUD_SIZE) {
4314                         pte = (pte_t *)pud;
4315                 } else {
4316                         BUG_ON(sz != PMD_SIZE);
4317                         if (want_pmd_share() && pud_none(*pud))
4318                                 pte = huge_pmd_share(mm, addr, pud);
4319                         else
4320                                 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4321                 }
4322         }
4323         BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
4324
4325         return pte;
4326 }
4327
4328 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
4329 {
4330         pgd_t *pgd;
4331         pud_t *pud;
4332         pmd_t *pmd = NULL;
4333
4334         pgd = pgd_offset(mm, addr);
4335         if (pgd_present(*pgd)) {
4336                 pud = pud_offset(pgd, addr);
4337                 if (pud_present(*pud)) {
4338                         if (pud_huge(*pud))
4339                                 return (pte_t *)pud;
4340                         pmd = pmd_offset(pud, addr);
4341                 }
4342         }
4343         return (pte_t *) pmd;
4344 }
4345
4346 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4347
4348 /*
4349  * These functions are overwritable if your architecture needs its own
4350  * behavior.
4351  */
4352 struct page * __weak
4353 follow_huge_addr(struct mm_struct *mm, unsigned long address,
4354                               int write)
4355 {
4356         return ERR_PTR(-EINVAL);
4357 }
4358
4359 struct page * __weak
4360 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
4361                 pmd_t *pmd, int flags)
4362 {
4363         struct page *page = NULL;
4364         spinlock_t *ptl;
4365 retry:
4366         ptl = pmd_lockptr(mm, pmd);
4367         spin_lock(ptl);
4368         /*
4369          * make sure that the address range covered by this pmd is not
4370          * unmapped from other threads.
4371          */
4372         if (!pmd_huge(*pmd))
4373                 goto out;
4374         if (pmd_present(*pmd)) {
4375                 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
4376                 if (flags & FOLL_GET)
4377                         get_page(page);
4378         } else {
4379                 if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) {
4380                         spin_unlock(ptl);
4381                         __migration_entry_wait(mm, (pte_t *)pmd, ptl);
4382                         goto retry;
4383                 }
4384                 /*
4385                  * hwpoisoned entry is treated as no_page_table in
4386                  * follow_page_mask().
4387                  */
4388         }
4389 out:
4390         spin_unlock(ptl);
4391         return page;
4392 }
4393
4394 struct page * __weak
4395 follow_huge_pud(struct mm_struct *mm, unsigned long address,
4396                 pud_t *pud, int flags)
4397 {
4398         if (flags & FOLL_GET)
4399                 return NULL;
4400
4401         return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
4402 }
4403
4404 #ifdef CONFIG_MEMORY_FAILURE
4405
4406 /*
4407  * This function is called from memory failure code.
4408  * Assume the caller holds page lock of the head page.
4409  */
4410 int dequeue_hwpoisoned_huge_page(struct page *hpage)
4411 {
4412         struct hstate *h = page_hstate(hpage);
4413         int nid = page_to_nid(hpage);
4414         int ret = -EBUSY;
4415
4416         spin_lock(&hugetlb_lock);
4417         /*
4418          * Just checking !page_huge_active is not enough, because that could be
4419          * an isolated/hwpoisoned hugepage (which have >0 refcount).
4420          */
4421         if (!page_huge_active(hpage) && !page_count(hpage)) {
4422                 /*
4423                  * Hwpoisoned hugepage isn't linked to activelist or freelist,
4424                  * but dangling hpage->lru can trigger list-debug warnings
4425                  * (this happens when we call unpoison_memory() on it),
4426                  * so let it point to itself with list_del_init().
4427                  */
4428                 list_del_init(&hpage->lru);
4429                 set_page_refcounted(hpage);
4430                 h->free_huge_pages--;
4431                 h->free_huge_pages_node[nid]--;
4432                 ret = 0;
4433         }
4434         spin_unlock(&hugetlb_lock);
4435         return ret;
4436 }
4437 #endif
4438
4439 bool isolate_huge_page(struct page *page, struct list_head *list)
4440 {
4441         bool ret = true;
4442
4443         VM_BUG_ON_PAGE(!PageHead(page), page);
4444         spin_lock(&hugetlb_lock);
4445         if (!page_huge_active(page) || !get_page_unless_zero(page)) {
4446                 ret = false;
4447                 goto unlock;
4448         }
4449         clear_page_huge_active(page);
4450         list_move_tail(&page->lru, list);
4451 unlock:
4452         spin_unlock(&hugetlb_lock);
4453         return ret;
4454 }
4455
4456 void putback_active_hugepage(struct page *page)
4457 {
4458         VM_BUG_ON_PAGE(!PageHead(page), page);
4459         spin_lock(&hugetlb_lock);
4460         set_page_huge_active(page);
4461         list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
4462         spin_unlock(&hugetlb_lock);
4463         put_page(page);
4464 }