176e0318960f93e724280ee37ea1e1b3a66b3e9d
[sfrench/cifs-2.6.git] / mm / hugetlb.c
1 /*
2  * Generic hugetlb support.
3  * (C) Nadia Yvette Chambers, April 2004
4  */
5 #include <linux/list.h>
6 #include <linux/init.h>
7 #include <linux/mm.h>
8 #include <linux/seq_file.h>
9 #include <linux/sysctl.h>
10 #include <linux/highmem.h>
11 #include <linux/mmu_notifier.h>
12 #include <linux/nodemask.h>
13 #include <linux/pagemap.h>
14 #include <linux/mempolicy.h>
15 #include <linux/compiler.h>
16 #include <linux/cpuset.h>
17 #include <linux/mutex.h>
18 #include <linux/bootmem.h>
19 #include <linux/sysfs.h>
20 #include <linux/slab.h>
21 #include <linux/sched/signal.h>
22 #include <linux/rmap.h>
23 #include <linux/swap.h>
24 #include <linux/swapops.h>
25 #include <linux/jhash.h>
26
27 #include <asm/page.h>
28 #include <asm/pgtable.h>
29 #include <asm/tlb.h>
30
31 #include <linux/io.h>
32 #include <linux/hugetlb.h>
33 #include <linux/hugetlb_cgroup.h>
34 #include <linux/node.h>
35 #include <linux/userfaultfd_k.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 static bool __initdata parsed_valid_hugepagesz = true;
56
57 /*
58  * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
59  * free_huge_pages, and surplus_huge_pages.
60  */
61 DEFINE_SPINLOCK(hugetlb_lock);
62
63 /*
64  * Serializes faults on the same logical page.  This is used to
65  * prevent spurious OOMs when the hugepage pool is fully utilized.
66  */
67 static int num_fault_mutexes;
68 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
69
70 /* Forward declaration */
71 static int hugetlb_acct_memory(struct hstate *h, long delta);
72
73 static char * __init memfmt(char *buf, unsigned long n)
74 {
75         if (n >= (1UL << 30))
76                 sprintf(buf, "%lu GB", n >> 30);
77         else if (n >= (1UL << 20))
78                 sprintf(buf, "%lu MB", n >> 20);
79         else
80                 sprintf(buf, "%lu KB", n >> 10);
81         return buf;
82 }
83
84 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
85 {
86         bool free = (spool->count == 0) && (spool->used_hpages == 0);
87
88         spin_unlock(&spool->lock);
89
90         /* If no pages are used, and no other handles to the subpool
91          * remain, give up any reservations mased on minimum size and
92          * free the subpool */
93         if (free) {
94                 if (spool->min_hpages != -1)
95                         hugetlb_acct_memory(spool->hstate,
96                                                 -spool->min_hpages);
97                 kfree(spool);
98         }
99 }
100
101 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
102                                                 long min_hpages)
103 {
104         struct hugepage_subpool *spool;
105
106         spool = kzalloc(sizeof(*spool), GFP_KERNEL);
107         if (!spool)
108                 return NULL;
109
110         spin_lock_init(&spool->lock);
111         spool->count = 1;
112         spool->max_hpages = max_hpages;
113         spool->hstate = h;
114         spool->min_hpages = min_hpages;
115
116         if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
117                 kfree(spool);
118                 return NULL;
119         }
120         spool->rsv_hpages = min_hpages;
121
122         return spool;
123 }
124
125 void hugepage_put_subpool(struct hugepage_subpool *spool)
126 {
127         spin_lock(&spool->lock);
128         BUG_ON(!spool->count);
129         spool->count--;
130         unlock_or_release_subpool(spool);
131 }
132
133 /*
134  * Subpool accounting for allocating and reserving pages.
135  * Return -ENOMEM if there are not enough resources to satisfy the
136  * the request.  Otherwise, return the number of pages by which the
137  * global pools must be adjusted (upward).  The returned value may
138  * only be different than the passed value (delta) in the case where
139  * a subpool minimum size must be manitained.
140  */
141 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
142                                       long delta)
143 {
144         long ret = delta;
145
146         if (!spool)
147                 return ret;
148
149         spin_lock(&spool->lock);
150
151         if (spool->max_hpages != -1) {          /* maximum size accounting */
152                 if ((spool->used_hpages + delta) <= spool->max_hpages)
153                         spool->used_hpages += delta;
154                 else {
155                         ret = -ENOMEM;
156                         goto unlock_ret;
157                 }
158         }
159
160         /* minimum size accounting */
161         if (spool->min_hpages != -1 && spool->rsv_hpages) {
162                 if (delta > spool->rsv_hpages) {
163                         /*
164                          * Asking for more reserves than those already taken on
165                          * behalf of subpool.  Return difference.
166                          */
167                         ret = delta - spool->rsv_hpages;
168                         spool->rsv_hpages = 0;
169                 } else {
170                         ret = 0;        /* reserves already accounted for */
171                         spool->rsv_hpages -= delta;
172                 }
173         }
174
175 unlock_ret:
176         spin_unlock(&spool->lock);
177         return ret;
178 }
179
180 /*
181  * Subpool accounting for freeing and unreserving pages.
182  * Return the number of global page reservations that must be dropped.
183  * The return value may only be different than the passed value (delta)
184  * in the case where a subpool minimum size must be maintained.
185  */
186 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
187                                        long delta)
188 {
189         long ret = delta;
190
191         if (!spool)
192                 return delta;
193
194         spin_lock(&spool->lock);
195
196         if (spool->max_hpages != -1)            /* maximum size accounting */
197                 spool->used_hpages -= delta;
198
199          /* minimum size accounting */
200         if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
201                 if (spool->rsv_hpages + delta <= spool->min_hpages)
202                         ret = 0;
203                 else
204                         ret = spool->rsv_hpages + delta - spool->min_hpages;
205
206                 spool->rsv_hpages += delta;
207                 if (spool->rsv_hpages > spool->min_hpages)
208                         spool->rsv_hpages = spool->min_hpages;
209         }
210
211         /*
212          * If hugetlbfs_put_super couldn't free spool due to an outstanding
213          * quota reference, free it now.
214          */
215         unlock_or_release_subpool(spool);
216
217         return ret;
218 }
219
220 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
221 {
222         return HUGETLBFS_SB(inode->i_sb)->spool;
223 }
224
225 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
226 {
227         return subpool_inode(file_inode(vma->vm_file));
228 }
229
230 /*
231  * Region tracking -- allows tracking of reservations and instantiated pages
232  *                    across the pages in a mapping.
233  *
234  * The region data structures are embedded into a resv_map and protected
235  * by a resv_map's lock.  The set of regions within the resv_map represent
236  * reservations for huge pages, or huge pages that have already been
237  * instantiated within the map.  The from and to elements are huge page
238  * indicies into the associated mapping.  from indicates the starting index
239  * of the region.  to represents the first index past the end of  the region.
240  *
241  * For example, a file region structure with from == 0 and to == 4 represents
242  * four huge pages in a mapping.  It is important to note that the to element
243  * represents the first element past the end of the region. This is used in
244  * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
245  *
246  * Interval notation of the form [from, to) will be used to indicate that
247  * the endpoint from is inclusive and to is exclusive.
248  */
249 struct file_region {
250         struct list_head link;
251         long from;
252         long to;
253 };
254
255 /*
256  * Add the huge page range represented by [f, t) to the reserve
257  * map.  In the normal case, existing regions will be expanded
258  * to accommodate the specified range.  Sufficient regions should
259  * exist for expansion due to the previous call to region_chg
260  * with the same range.  However, it is possible that region_del
261  * could have been called after region_chg and modifed the map
262  * in such a way that no region exists to be expanded.  In this
263  * case, pull a region descriptor from the cache associated with
264  * the map and use that for the new range.
265  *
266  * Return the number of new huge pages added to the map.  This
267  * number is greater than or equal to zero.
268  */
269 static long region_add(struct resv_map *resv, long f, long t)
270 {
271         struct list_head *head = &resv->regions;
272         struct file_region *rg, *nrg, *trg;
273         long add = 0;
274
275         spin_lock(&resv->lock);
276         /* Locate the region we are either in or before. */
277         list_for_each_entry(rg, head, link)
278                 if (f <= rg->to)
279                         break;
280
281         /*
282          * If no region exists which can be expanded to include the
283          * specified range, the list must have been modified by an
284          * interleving call to region_del().  Pull a region descriptor
285          * from the cache and use it for this range.
286          */
287         if (&rg->link == head || t < rg->from) {
288                 VM_BUG_ON(resv->region_cache_count <= 0);
289
290                 resv->region_cache_count--;
291                 nrg = list_first_entry(&resv->region_cache, struct file_region,
292                                         link);
293                 list_del(&nrg->link);
294
295                 nrg->from = f;
296                 nrg->to = t;
297                 list_add(&nrg->link, rg->link.prev);
298
299                 add += t - f;
300                 goto out_locked;
301         }
302
303         /* Round our left edge to the current segment if it encloses us. */
304         if (f > rg->from)
305                 f = rg->from;
306
307         /* Check for and consume any regions we now overlap with. */
308         nrg = rg;
309         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
310                 if (&rg->link == head)
311                         break;
312                 if (rg->from > t)
313                         break;
314
315                 /* If this area reaches higher then extend our area to
316                  * include it completely.  If this is not the first area
317                  * which we intend to reuse, free it. */
318                 if (rg->to > t)
319                         t = rg->to;
320                 if (rg != nrg) {
321                         /* Decrement return value by the deleted range.
322                          * Another range will span this area so that by
323                          * end of routine add will be >= zero
324                          */
325                         add -= (rg->to - rg->from);
326                         list_del(&rg->link);
327                         kfree(rg);
328                 }
329         }
330
331         add += (nrg->from - f);         /* Added to beginning of region */
332         nrg->from = f;
333         add += t - nrg->to;             /* Added to end of region */
334         nrg->to = t;
335
336 out_locked:
337         resv->adds_in_progress--;
338         spin_unlock(&resv->lock);
339         VM_BUG_ON(add < 0);
340         return add;
341 }
342
343 /*
344  * Examine the existing reserve map and determine how many
345  * huge pages in the specified range [f, t) are NOT currently
346  * represented.  This routine is called before a subsequent
347  * call to region_add that will actually modify the reserve
348  * map to add the specified range [f, t).  region_chg does
349  * not change the number of huge pages represented by the
350  * map.  However, if the existing regions in the map can not
351  * be expanded to represent the new range, a new file_region
352  * structure is added to the map as a placeholder.  This is
353  * so that the subsequent region_add call will have all the
354  * regions it needs and will not fail.
355  *
356  * Upon entry, region_chg will also examine the cache of region descriptors
357  * associated with the map.  If there are not enough descriptors cached, one
358  * will be allocated for the in progress add operation.
359  *
360  * Returns the number of huge pages that need to be added to the existing
361  * reservation map for the range [f, t).  This number is greater or equal to
362  * zero.  -ENOMEM is returned if a new file_region structure or cache entry
363  * is needed and can not be allocated.
364  */
365 static long region_chg(struct resv_map *resv, long f, long t)
366 {
367         struct list_head *head = &resv->regions;
368         struct file_region *rg, *nrg = NULL;
369         long chg = 0;
370
371 retry:
372         spin_lock(&resv->lock);
373 retry_locked:
374         resv->adds_in_progress++;
375
376         /*
377          * Check for sufficient descriptors in the cache to accommodate
378          * the number of in progress add operations.
379          */
380         if (resv->adds_in_progress > resv->region_cache_count) {
381                 struct file_region *trg;
382
383                 VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1);
384                 /* Must drop lock to allocate a new descriptor. */
385                 resv->adds_in_progress--;
386                 spin_unlock(&resv->lock);
387
388                 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
389                 if (!trg) {
390                         kfree(nrg);
391                         return -ENOMEM;
392                 }
393
394                 spin_lock(&resv->lock);
395                 list_add(&trg->link, &resv->region_cache);
396                 resv->region_cache_count++;
397                 goto retry_locked;
398         }
399
400         /* Locate the region we are before or in. */
401         list_for_each_entry(rg, head, link)
402                 if (f <= rg->to)
403                         break;
404
405         /* If we are below the current region then a new region is required.
406          * Subtle, allocate a new region at the position but make it zero
407          * size such that we can guarantee to record the reservation. */
408         if (&rg->link == head || t < rg->from) {
409                 if (!nrg) {
410                         resv->adds_in_progress--;
411                         spin_unlock(&resv->lock);
412                         nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
413                         if (!nrg)
414                                 return -ENOMEM;
415
416                         nrg->from = f;
417                         nrg->to   = f;
418                         INIT_LIST_HEAD(&nrg->link);
419                         goto retry;
420                 }
421
422                 list_add(&nrg->link, rg->link.prev);
423                 chg = t - f;
424                 goto out_nrg;
425         }
426
427         /* Round our left edge to the current segment if it encloses us. */
428         if (f > rg->from)
429                 f = rg->from;
430         chg = t - f;
431
432         /* Check for and consume any regions we now overlap with. */
433         list_for_each_entry(rg, rg->link.prev, link) {
434                 if (&rg->link == head)
435                         break;
436                 if (rg->from > t)
437                         goto out;
438
439                 /* We overlap with this area, if it extends further than
440                  * us then we must extend ourselves.  Account for its
441                  * existing reservation. */
442                 if (rg->to > t) {
443                         chg += rg->to - t;
444                         t = rg->to;
445                 }
446                 chg -= rg->to - rg->from;
447         }
448
449 out:
450         spin_unlock(&resv->lock);
451         /*  We already know we raced and no longer need the new region */
452         kfree(nrg);
453         return chg;
454 out_nrg:
455         spin_unlock(&resv->lock);
456         return chg;
457 }
458
459 /*
460  * Abort the in progress add operation.  The adds_in_progress field
461  * of the resv_map keeps track of the operations in progress between
462  * calls to region_chg and region_add.  Operations are sometimes
463  * aborted after the call to region_chg.  In such cases, region_abort
464  * is called to decrement the adds_in_progress counter.
465  *
466  * NOTE: The range arguments [f, t) are not needed or used in this
467  * routine.  They are kept to make reading the calling code easier as
468  * arguments will match the associated region_chg call.
469  */
470 static void region_abort(struct resv_map *resv, long f, long t)
471 {
472         spin_lock(&resv->lock);
473         VM_BUG_ON(!resv->region_cache_count);
474         resv->adds_in_progress--;
475         spin_unlock(&resv->lock);
476 }
477
478 /*
479  * Delete the specified range [f, t) from the reserve map.  If the
480  * t parameter is LONG_MAX, this indicates that ALL regions after f
481  * should be deleted.  Locate the regions which intersect [f, t)
482  * and either trim, delete or split the existing regions.
483  *
484  * Returns the number of huge pages deleted from the reserve map.
485  * In the normal case, the return value is zero or more.  In the
486  * case where a region must be split, a new region descriptor must
487  * be allocated.  If the allocation fails, -ENOMEM will be returned.
488  * NOTE: If the parameter t == LONG_MAX, then we will never split
489  * a region and possibly return -ENOMEM.  Callers specifying
490  * t == LONG_MAX do not need to check for -ENOMEM error.
491  */
492 static long region_del(struct resv_map *resv, long f, long t)
493 {
494         struct list_head *head = &resv->regions;
495         struct file_region *rg, *trg;
496         struct file_region *nrg = NULL;
497         long del = 0;
498
499 retry:
500         spin_lock(&resv->lock);
501         list_for_each_entry_safe(rg, trg, head, link) {
502                 /*
503                  * Skip regions before the range to be deleted.  file_region
504                  * ranges are normally of the form [from, to).  However, there
505                  * may be a "placeholder" entry in the map which is of the form
506                  * (from, to) with from == to.  Check for placeholder entries
507                  * at the beginning of the range to be deleted.
508                  */
509                 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
510                         continue;
511
512                 if (rg->from >= t)
513                         break;
514
515                 if (f > rg->from && t < rg->to) { /* Must split region */
516                         /*
517                          * Check for an entry in the cache before dropping
518                          * lock and attempting allocation.
519                          */
520                         if (!nrg &&
521                             resv->region_cache_count > resv->adds_in_progress) {
522                                 nrg = list_first_entry(&resv->region_cache,
523                                                         struct file_region,
524                                                         link);
525                                 list_del(&nrg->link);
526                                 resv->region_cache_count--;
527                         }
528
529                         if (!nrg) {
530                                 spin_unlock(&resv->lock);
531                                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
532                                 if (!nrg)
533                                         return -ENOMEM;
534                                 goto retry;
535                         }
536
537                         del += t - f;
538
539                         /* New entry for end of split region */
540                         nrg->from = t;
541                         nrg->to = rg->to;
542                         INIT_LIST_HEAD(&nrg->link);
543
544                         /* Original entry is trimmed */
545                         rg->to = f;
546
547                         list_add(&nrg->link, &rg->link);
548                         nrg = NULL;
549                         break;
550                 }
551
552                 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
553                         del += rg->to - rg->from;
554                         list_del(&rg->link);
555                         kfree(rg);
556                         continue;
557                 }
558
559                 if (f <= rg->from) {    /* Trim beginning of region */
560                         del += t - rg->from;
561                         rg->from = t;
562                 } else {                /* Trim end of region */
563                         del += rg->to - f;
564                         rg->to = f;
565                 }
566         }
567
568         spin_unlock(&resv->lock);
569         kfree(nrg);
570         return del;
571 }
572
573 /*
574  * A rare out of memory error was encountered which prevented removal of
575  * the reserve map region for a page.  The huge page itself was free'ed
576  * and removed from the page cache.  This routine will adjust the subpool
577  * usage count, and the global reserve count if needed.  By incrementing
578  * these counts, the reserve map entry which could not be deleted will
579  * appear as a "reserved" entry instead of simply dangling with incorrect
580  * counts.
581  */
582 void hugetlb_fix_reserve_counts(struct inode *inode)
583 {
584         struct hugepage_subpool *spool = subpool_inode(inode);
585         long rsv_adjust;
586
587         rsv_adjust = hugepage_subpool_get_pages(spool, 1);
588         if (rsv_adjust) {
589                 struct hstate *h = hstate_inode(inode);
590
591                 hugetlb_acct_memory(h, 1);
592         }
593 }
594
595 /*
596  * Count and return the number of huge pages in the reserve map
597  * that intersect with the range [f, t).
598  */
599 static long region_count(struct resv_map *resv, long f, long t)
600 {
601         struct list_head *head = &resv->regions;
602         struct file_region *rg;
603         long chg = 0;
604
605         spin_lock(&resv->lock);
606         /* Locate each segment we overlap with, and count that overlap. */
607         list_for_each_entry(rg, head, link) {
608                 long seg_from;
609                 long seg_to;
610
611                 if (rg->to <= f)
612                         continue;
613                 if (rg->from >= t)
614                         break;
615
616                 seg_from = max(rg->from, f);
617                 seg_to = min(rg->to, t);
618
619                 chg += seg_to - seg_from;
620         }
621         spin_unlock(&resv->lock);
622
623         return chg;
624 }
625
626 /*
627  * Convert the address within this vma to the page offset within
628  * the mapping, in pagecache page units; huge pages here.
629  */
630 static pgoff_t vma_hugecache_offset(struct hstate *h,
631                         struct vm_area_struct *vma, unsigned long address)
632 {
633         return ((address - vma->vm_start) >> huge_page_shift(h)) +
634                         (vma->vm_pgoff >> huge_page_order(h));
635 }
636
637 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
638                                      unsigned long address)
639 {
640         return vma_hugecache_offset(hstate_vma(vma), vma, address);
641 }
642 EXPORT_SYMBOL_GPL(linear_hugepage_index);
643
644 /*
645  * Return the size of the pages allocated when backing a VMA. In the majority
646  * cases this will be same size as used by the page table entries.
647  */
648 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
649 {
650         struct hstate *hstate;
651
652         if (!is_vm_hugetlb_page(vma))
653                 return PAGE_SIZE;
654
655         hstate = hstate_vma(vma);
656
657         return 1UL << huge_page_shift(hstate);
658 }
659 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
660
661 /*
662  * Return the page size being used by the MMU to back a VMA. In the majority
663  * of cases, the page size used by the kernel matches the MMU size. On
664  * architectures where it differs, an architecture-specific version of this
665  * function is required.
666  */
667 #ifndef vma_mmu_pagesize
668 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
669 {
670         return vma_kernel_pagesize(vma);
671 }
672 #endif
673
674 /*
675  * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
676  * bits of the reservation map pointer, which are always clear due to
677  * alignment.
678  */
679 #define HPAGE_RESV_OWNER    (1UL << 0)
680 #define HPAGE_RESV_UNMAPPED (1UL << 1)
681 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
682
683 /*
684  * These helpers are used to track how many pages are reserved for
685  * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
686  * is guaranteed to have their future faults succeed.
687  *
688  * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
689  * the reserve counters are updated with the hugetlb_lock held. It is safe
690  * to reset the VMA at fork() time as it is not in use yet and there is no
691  * chance of the global counters getting corrupted as a result of the values.
692  *
693  * The private mapping reservation is represented in a subtly different
694  * manner to a shared mapping.  A shared mapping has a region map associated
695  * with the underlying file, this region map represents the backing file
696  * pages which have ever had a reservation assigned which this persists even
697  * after the page is instantiated.  A private mapping has a region map
698  * associated with the original mmap which is attached to all VMAs which
699  * reference it, this region map represents those offsets which have consumed
700  * reservation ie. where pages have been instantiated.
701  */
702 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
703 {
704         return (unsigned long)vma->vm_private_data;
705 }
706
707 static void set_vma_private_data(struct vm_area_struct *vma,
708                                                         unsigned long value)
709 {
710         vma->vm_private_data = (void *)value;
711 }
712
713 struct resv_map *resv_map_alloc(void)
714 {
715         struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
716         struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
717
718         if (!resv_map || !rg) {
719                 kfree(resv_map);
720                 kfree(rg);
721                 return NULL;
722         }
723
724         kref_init(&resv_map->refs);
725         spin_lock_init(&resv_map->lock);
726         INIT_LIST_HEAD(&resv_map->regions);
727
728         resv_map->adds_in_progress = 0;
729
730         INIT_LIST_HEAD(&resv_map->region_cache);
731         list_add(&rg->link, &resv_map->region_cache);
732         resv_map->region_cache_count = 1;
733
734         return resv_map;
735 }
736
737 void resv_map_release(struct kref *ref)
738 {
739         struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
740         struct list_head *head = &resv_map->region_cache;
741         struct file_region *rg, *trg;
742
743         /* Clear out any active regions before we release the map. */
744         region_del(resv_map, 0, LONG_MAX);
745
746         /* ... and any entries left in the cache */
747         list_for_each_entry_safe(rg, trg, head, link) {
748                 list_del(&rg->link);
749                 kfree(rg);
750         }
751
752         VM_BUG_ON(resv_map->adds_in_progress);
753
754         kfree(resv_map);
755 }
756
757 static inline struct resv_map *inode_resv_map(struct inode *inode)
758 {
759         return inode->i_mapping->private_data;
760 }
761
762 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
763 {
764         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
765         if (vma->vm_flags & VM_MAYSHARE) {
766                 struct address_space *mapping = vma->vm_file->f_mapping;
767                 struct inode *inode = mapping->host;
768
769                 return inode_resv_map(inode);
770
771         } else {
772                 return (struct resv_map *)(get_vma_private_data(vma) &
773                                                         ~HPAGE_RESV_MASK);
774         }
775 }
776
777 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
778 {
779         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
780         VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
781
782         set_vma_private_data(vma, (get_vma_private_data(vma) &
783                                 HPAGE_RESV_MASK) | (unsigned long)map);
784 }
785
786 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
787 {
788         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
789         VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
790
791         set_vma_private_data(vma, get_vma_private_data(vma) | flags);
792 }
793
794 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
795 {
796         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
797
798         return (get_vma_private_data(vma) & flag) != 0;
799 }
800
801 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
802 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
803 {
804         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
805         if (!(vma->vm_flags & VM_MAYSHARE))
806                 vma->vm_private_data = (void *)0;
807 }
808
809 /* Returns true if the VMA has associated reserve pages */
810 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
811 {
812         if (vma->vm_flags & VM_NORESERVE) {
813                 /*
814                  * This address is already reserved by other process(chg == 0),
815                  * so, we should decrement reserved count. Without decrementing,
816                  * reserve count remains after releasing inode, because this
817                  * allocated page will go into page cache and is regarded as
818                  * coming from reserved pool in releasing step.  Currently, we
819                  * don't have any other solution to deal with this situation
820                  * properly, so add work-around here.
821                  */
822                 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
823                         return true;
824                 else
825                         return false;
826         }
827
828         /* Shared mappings always use reserves */
829         if (vma->vm_flags & VM_MAYSHARE) {
830                 /*
831                  * We know VM_NORESERVE is not set.  Therefore, there SHOULD
832                  * be a region map for all pages.  The only situation where
833                  * there is no region map is if a hole was punched via
834                  * fallocate.  In this case, there really are no reverves to
835                  * use.  This situation is indicated if chg != 0.
836                  */
837                 if (chg)
838                         return false;
839                 else
840                         return true;
841         }
842
843         /*
844          * Only the process that called mmap() has reserves for
845          * private mappings.
846          */
847         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
848                 /*
849                  * Like the shared case above, a hole punch or truncate
850                  * could have been performed on the private mapping.
851                  * Examine the value of chg to determine if reserves
852                  * actually exist or were previously consumed.
853                  * Very Subtle - The value of chg comes from a previous
854                  * call to vma_needs_reserves().  The reserve map for
855                  * private mappings has different (opposite) semantics
856                  * than that of shared mappings.  vma_needs_reserves()
857                  * has already taken this difference in semantics into
858                  * account.  Therefore, the meaning of chg is the same
859                  * as in the shared case above.  Code could easily be
860                  * combined, but keeping it separate draws attention to
861                  * subtle differences.
862                  */
863                 if (chg)
864                         return false;
865                 else
866                         return true;
867         }
868
869         return false;
870 }
871
872 static void enqueue_huge_page(struct hstate *h, struct page *page)
873 {
874         int nid = page_to_nid(page);
875         list_move(&page->lru, &h->hugepage_freelists[nid]);
876         h->free_huge_pages++;
877         h->free_huge_pages_node[nid]++;
878 }
879
880 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
881 {
882         struct page *page;
883
884         list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
885                 if (!PageHWPoison(page))
886                         break;
887         /*
888          * if 'non-isolated free hugepage' not found on the list,
889          * the allocation fails.
890          */
891         if (&h->hugepage_freelists[nid] == &page->lru)
892                 return NULL;
893         list_move(&page->lru, &h->hugepage_activelist);
894         set_page_refcounted(page);
895         h->free_huge_pages--;
896         h->free_huge_pages_node[nid]--;
897         return page;
898 }
899
900 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
901 {
902         struct page *page;
903         int node;
904
905         if (nid != NUMA_NO_NODE)
906                 return dequeue_huge_page_node_exact(h, nid);
907
908         for_each_online_node(node) {
909                 page = dequeue_huge_page_node_exact(h, node);
910                 if (page)
911                         return page;
912         }
913         return NULL;
914 }
915
916 /* Movability of hugepages depends on migration support. */
917 static inline gfp_t htlb_alloc_mask(struct hstate *h)
918 {
919         if (hugepages_treat_as_movable || hugepage_migration_supported(h))
920                 return GFP_HIGHUSER_MOVABLE;
921         else
922                 return GFP_HIGHUSER;
923 }
924
925 static struct page *dequeue_huge_page_vma(struct hstate *h,
926                                 struct vm_area_struct *vma,
927                                 unsigned long address, int avoid_reserve,
928                                 long chg)
929 {
930         struct page *page = NULL;
931         struct mempolicy *mpol;
932         nodemask_t *nodemask;
933         gfp_t gfp_mask;
934         int nid;
935         struct zonelist *zonelist;
936         struct zone *zone;
937         struct zoneref *z;
938         unsigned int cpuset_mems_cookie;
939
940         /*
941          * A child process with MAP_PRIVATE mappings created by their parent
942          * have no page reserves. This check ensures that reservations are
943          * not "stolen". The child may still get SIGKILLed
944          */
945         if (!vma_has_reserves(vma, chg) &&
946                         h->free_huge_pages - h->resv_huge_pages == 0)
947                 goto err;
948
949         /* If reserves cannot be used, ensure enough pages are in the pool */
950         if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
951                 goto err;
952
953 retry_cpuset:
954         cpuset_mems_cookie = read_mems_allowed_begin();
955         gfp_mask = htlb_alloc_mask(h);
956         nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
957         zonelist = node_zonelist(nid, gfp_mask);
958
959         for_each_zone_zonelist_nodemask(zone, z, zonelist,
960                                                 MAX_NR_ZONES - 1, nodemask) {
961                 if (cpuset_zone_allowed(zone, gfp_mask)) {
962                         page = dequeue_huge_page_node(h, zone_to_nid(zone));
963                         if (page) {
964                                 if (avoid_reserve)
965                                         break;
966                                 if (!vma_has_reserves(vma, chg))
967                                         break;
968
969                                 SetPagePrivate(page);
970                                 h->resv_huge_pages--;
971                                 break;
972                         }
973                 }
974         }
975
976         mpol_cond_put(mpol);
977         if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
978                 goto retry_cpuset;
979         return page;
980
981 err:
982         return NULL;
983 }
984
985 /*
986  * common helper functions for hstate_next_node_to_{alloc|free}.
987  * We may have allocated or freed a huge page based on a different
988  * nodes_allowed previously, so h->next_node_to_{alloc|free} might
989  * be outside of *nodes_allowed.  Ensure that we use an allowed
990  * node for alloc or free.
991  */
992 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
993 {
994         nid = next_node_in(nid, *nodes_allowed);
995         VM_BUG_ON(nid >= MAX_NUMNODES);
996
997         return nid;
998 }
999
1000 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1001 {
1002         if (!node_isset(nid, *nodes_allowed))
1003                 nid = next_node_allowed(nid, nodes_allowed);
1004         return nid;
1005 }
1006
1007 /*
1008  * returns the previously saved node ["this node"] from which to
1009  * allocate a persistent huge page for the pool and advance the
1010  * next node from which to allocate, handling wrap at end of node
1011  * mask.
1012  */
1013 static int hstate_next_node_to_alloc(struct hstate *h,
1014                                         nodemask_t *nodes_allowed)
1015 {
1016         int nid;
1017
1018         VM_BUG_ON(!nodes_allowed);
1019
1020         nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1021         h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1022
1023         return nid;
1024 }
1025
1026 /*
1027  * helper for free_pool_huge_page() - return the previously saved
1028  * node ["this node"] from which to free a huge page.  Advance the
1029  * next node id whether or not we find a free huge page to free so
1030  * that the next attempt to free addresses the next node.
1031  */
1032 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1033 {
1034         int nid;
1035
1036         VM_BUG_ON(!nodes_allowed);
1037
1038         nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1039         h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1040
1041         return nid;
1042 }
1043
1044 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask)           \
1045         for (nr_nodes = nodes_weight(*mask);                            \
1046                 nr_nodes > 0 &&                                         \
1047                 ((node = hstate_next_node_to_alloc(hs, mask)) || 1);    \
1048                 nr_nodes--)
1049
1050 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask)            \
1051         for (nr_nodes = nodes_weight(*mask);                            \
1052                 nr_nodes > 0 &&                                         \
1053                 ((node = hstate_next_node_to_free(hs, mask)) || 1);     \
1054                 nr_nodes--)
1055
1056 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1057 static void destroy_compound_gigantic_page(struct page *page,
1058                                         unsigned int order)
1059 {
1060         int i;
1061         int nr_pages = 1 << order;
1062         struct page *p = page + 1;
1063
1064         atomic_set(compound_mapcount_ptr(page), 0);
1065         for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1066                 clear_compound_head(p);
1067                 set_page_refcounted(p);
1068         }
1069
1070         set_compound_order(page, 0);
1071         __ClearPageHead(page);
1072 }
1073
1074 static void free_gigantic_page(struct page *page, unsigned int order)
1075 {
1076         free_contig_range(page_to_pfn(page), 1 << order);
1077 }
1078
1079 static int __alloc_gigantic_page(unsigned long start_pfn,
1080                                 unsigned long nr_pages)
1081 {
1082         unsigned long end_pfn = start_pfn + nr_pages;
1083         return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
1084                                   GFP_KERNEL);
1085 }
1086
1087 static bool pfn_range_valid_gigantic(struct zone *z,
1088                         unsigned long start_pfn, unsigned long nr_pages)
1089 {
1090         unsigned long i, end_pfn = start_pfn + nr_pages;
1091         struct page *page;
1092
1093         for (i = start_pfn; i < end_pfn; i++) {
1094                 if (!pfn_valid(i))
1095                         return false;
1096
1097                 page = pfn_to_page(i);
1098
1099                 if (page_zone(page) != z)
1100                         return false;
1101
1102                 if (PageReserved(page))
1103                         return false;
1104
1105                 if (page_count(page) > 0)
1106                         return false;
1107
1108                 if (PageHuge(page))
1109                         return false;
1110         }
1111
1112         return true;
1113 }
1114
1115 static bool zone_spans_last_pfn(const struct zone *zone,
1116                         unsigned long start_pfn, unsigned long nr_pages)
1117 {
1118         unsigned long last_pfn = start_pfn + nr_pages - 1;
1119         return zone_spans_pfn(zone, last_pfn);
1120 }
1121
1122 static struct page *alloc_gigantic_page(int nid, unsigned int order)
1123 {
1124         unsigned long nr_pages = 1 << order;
1125         unsigned long ret, pfn, flags;
1126         struct zone *z;
1127
1128         z = NODE_DATA(nid)->node_zones;
1129         for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) {
1130                 spin_lock_irqsave(&z->lock, flags);
1131
1132                 pfn = ALIGN(z->zone_start_pfn, nr_pages);
1133                 while (zone_spans_last_pfn(z, pfn, nr_pages)) {
1134                         if (pfn_range_valid_gigantic(z, pfn, nr_pages)) {
1135                                 /*
1136                                  * We release the zone lock here because
1137                                  * alloc_contig_range() will also lock the zone
1138                                  * at some point. If there's an allocation
1139                                  * spinning on this lock, it may win the race
1140                                  * and cause alloc_contig_range() to fail...
1141                                  */
1142                                 spin_unlock_irqrestore(&z->lock, flags);
1143                                 ret = __alloc_gigantic_page(pfn, nr_pages);
1144                                 if (!ret)
1145                                         return pfn_to_page(pfn);
1146                                 spin_lock_irqsave(&z->lock, flags);
1147                         }
1148                         pfn += nr_pages;
1149                 }
1150
1151                 spin_unlock_irqrestore(&z->lock, flags);
1152         }
1153
1154         return NULL;
1155 }
1156
1157 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1158 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1159
1160 static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid)
1161 {
1162         struct page *page;
1163
1164         page = alloc_gigantic_page(nid, huge_page_order(h));
1165         if (page) {
1166                 prep_compound_gigantic_page(page, huge_page_order(h));
1167                 prep_new_huge_page(h, page, nid);
1168         }
1169
1170         return page;
1171 }
1172
1173 static int alloc_fresh_gigantic_page(struct hstate *h,
1174                                 nodemask_t *nodes_allowed)
1175 {
1176         struct page *page = NULL;
1177         int nr_nodes, node;
1178
1179         for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1180                 page = alloc_fresh_gigantic_page_node(h, node);
1181                 if (page)
1182                         return 1;
1183         }
1184
1185         return 0;
1186 }
1187
1188 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1189 static inline bool gigantic_page_supported(void) { return false; }
1190 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1191 static inline void destroy_compound_gigantic_page(struct page *page,
1192                                                 unsigned int order) { }
1193 static inline int alloc_fresh_gigantic_page(struct hstate *h,
1194                                         nodemask_t *nodes_allowed) { return 0; }
1195 #endif
1196
1197 static void update_and_free_page(struct hstate *h, struct page *page)
1198 {
1199         int i;
1200
1201         if (hstate_is_gigantic(h) && !gigantic_page_supported())
1202                 return;
1203
1204         h->nr_huge_pages--;
1205         h->nr_huge_pages_node[page_to_nid(page)]--;
1206         for (i = 0; i < pages_per_huge_page(h); i++) {
1207                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1208                                 1 << PG_referenced | 1 << PG_dirty |
1209                                 1 << PG_active | 1 << PG_private |
1210                                 1 << PG_writeback);
1211         }
1212         VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1213         set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1214         set_page_refcounted(page);
1215         if (hstate_is_gigantic(h)) {
1216                 destroy_compound_gigantic_page(page, huge_page_order(h));
1217                 free_gigantic_page(page, huge_page_order(h));
1218         } else {
1219                 __free_pages(page, huge_page_order(h));
1220         }
1221 }
1222
1223 struct hstate *size_to_hstate(unsigned long size)
1224 {
1225         struct hstate *h;
1226
1227         for_each_hstate(h) {
1228                 if (huge_page_size(h) == size)
1229                         return h;
1230         }
1231         return NULL;
1232 }
1233
1234 /*
1235  * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1236  * to hstate->hugepage_activelist.)
1237  *
1238  * This function can be called for tail pages, but never returns true for them.
1239  */
1240 bool page_huge_active(struct page *page)
1241 {
1242         VM_BUG_ON_PAGE(!PageHuge(page), page);
1243         return PageHead(page) && PagePrivate(&page[1]);
1244 }
1245
1246 /* never called for tail page */
1247 static void set_page_huge_active(struct page *page)
1248 {
1249         VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1250         SetPagePrivate(&page[1]);
1251 }
1252
1253 static void clear_page_huge_active(struct page *page)
1254 {
1255         VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1256         ClearPagePrivate(&page[1]);
1257 }
1258
1259 void free_huge_page(struct page *page)
1260 {
1261         /*
1262          * Can't pass hstate in here because it is called from the
1263          * compound page destructor.
1264          */
1265         struct hstate *h = page_hstate(page);
1266         int nid = page_to_nid(page);
1267         struct hugepage_subpool *spool =
1268                 (struct hugepage_subpool *)page_private(page);
1269         bool restore_reserve;
1270
1271         set_page_private(page, 0);
1272         page->mapping = NULL;
1273         VM_BUG_ON_PAGE(page_count(page), page);
1274         VM_BUG_ON_PAGE(page_mapcount(page), page);
1275         restore_reserve = PagePrivate(page);
1276         ClearPagePrivate(page);
1277
1278         /*
1279          * A return code of zero implies that the subpool will be under its
1280          * minimum size if the reservation is not restored after page is free.
1281          * Therefore, force restore_reserve operation.
1282          */
1283         if (hugepage_subpool_put_pages(spool, 1) == 0)
1284                 restore_reserve = true;
1285
1286         spin_lock(&hugetlb_lock);
1287         clear_page_huge_active(page);
1288         hugetlb_cgroup_uncharge_page(hstate_index(h),
1289                                      pages_per_huge_page(h), page);
1290         if (restore_reserve)
1291                 h->resv_huge_pages++;
1292
1293         if (h->surplus_huge_pages_node[nid]) {
1294                 /* remove the page from active list */
1295                 list_del(&page->lru);
1296                 update_and_free_page(h, page);
1297                 h->surplus_huge_pages--;
1298                 h->surplus_huge_pages_node[nid]--;
1299         } else {
1300                 arch_clear_hugepage_flags(page);
1301                 enqueue_huge_page(h, page);
1302         }
1303         spin_unlock(&hugetlb_lock);
1304 }
1305
1306 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1307 {
1308         INIT_LIST_HEAD(&page->lru);
1309         set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1310         spin_lock(&hugetlb_lock);
1311         set_hugetlb_cgroup(page, NULL);
1312         h->nr_huge_pages++;
1313         h->nr_huge_pages_node[nid]++;
1314         spin_unlock(&hugetlb_lock);
1315         put_page(page); /* free it into the hugepage allocator */
1316 }
1317
1318 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1319 {
1320         int i;
1321         int nr_pages = 1 << order;
1322         struct page *p = page + 1;
1323
1324         /* we rely on prep_new_huge_page to set the destructor */
1325         set_compound_order(page, order);
1326         __ClearPageReserved(page);
1327         __SetPageHead(page);
1328         for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1329                 /*
1330                  * For gigantic hugepages allocated through bootmem at
1331                  * boot, it's safer to be consistent with the not-gigantic
1332                  * hugepages and clear the PG_reserved bit from all tail pages
1333                  * too.  Otherwse drivers using get_user_pages() to access tail
1334                  * pages may get the reference counting wrong if they see
1335                  * PG_reserved set on a tail page (despite the head page not
1336                  * having PG_reserved set).  Enforcing this consistency between
1337                  * head and tail pages allows drivers to optimize away a check
1338                  * on the head page when they need know if put_page() is needed
1339                  * after get_user_pages().
1340                  */
1341                 __ClearPageReserved(p);
1342                 set_page_count(p, 0);
1343                 set_compound_head(p, page);
1344         }
1345         atomic_set(compound_mapcount_ptr(page), -1);
1346 }
1347
1348 /*
1349  * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1350  * transparent huge pages.  See the PageTransHuge() documentation for more
1351  * details.
1352  */
1353 int PageHuge(struct page *page)
1354 {
1355         if (!PageCompound(page))
1356                 return 0;
1357
1358         page = compound_head(page);
1359         return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1360 }
1361 EXPORT_SYMBOL_GPL(PageHuge);
1362
1363 /*
1364  * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1365  * normal or transparent huge pages.
1366  */
1367 int PageHeadHuge(struct page *page_head)
1368 {
1369         if (!PageHead(page_head))
1370                 return 0;
1371
1372         return get_compound_page_dtor(page_head) == free_huge_page;
1373 }
1374
1375 pgoff_t __basepage_index(struct page *page)
1376 {
1377         struct page *page_head = compound_head(page);
1378         pgoff_t index = page_index(page_head);
1379         unsigned long compound_idx;
1380
1381         if (!PageHuge(page_head))
1382                 return page_index(page);
1383
1384         if (compound_order(page_head) >= MAX_ORDER)
1385                 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1386         else
1387                 compound_idx = page - page_head;
1388
1389         return (index << compound_order(page_head)) + compound_idx;
1390 }
1391
1392 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
1393 {
1394         struct page *page;
1395
1396         page = __alloc_pages_node(nid,
1397                 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
1398                                                 __GFP_REPEAT|__GFP_NOWARN,
1399                 huge_page_order(h));
1400         if (page) {
1401                 prep_new_huge_page(h, page, nid);
1402         }
1403
1404         return page;
1405 }
1406
1407 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1408 {
1409         struct page *page;
1410         int nr_nodes, node;
1411         int ret = 0;
1412
1413         for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1414                 page = alloc_fresh_huge_page_node(h, node);
1415                 if (page) {
1416                         ret = 1;
1417                         break;
1418                 }
1419         }
1420
1421         if (ret)
1422                 count_vm_event(HTLB_BUDDY_PGALLOC);
1423         else
1424                 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1425
1426         return ret;
1427 }
1428
1429 /*
1430  * Free huge page from pool from next node to free.
1431  * Attempt to keep persistent huge pages more or less
1432  * balanced over allowed nodes.
1433  * Called with hugetlb_lock locked.
1434  */
1435 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1436                                                          bool acct_surplus)
1437 {
1438         int nr_nodes, node;
1439         int ret = 0;
1440
1441         for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1442                 /*
1443                  * If we're returning unused surplus pages, only examine
1444                  * nodes with surplus pages.
1445                  */
1446                 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1447                     !list_empty(&h->hugepage_freelists[node])) {
1448                         struct page *page =
1449                                 list_entry(h->hugepage_freelists[node].next,
1450                                           struct page, lru);
1451                         list_del(&page->lru);
1452                         h->free_huge_pages--;
1453                         h->free_huge_pages_node[node]--;
1454                         if (acct_surplus) {
1455                                 h->surplus_huge_pages--;
1456                                 h->surplus_huge_pages_node[node]--;
1457                         }
1458                         update_and_free_page(h, page);
1459                         ret = 1;
1460                         break;
1461                 }
1462         }
1463
1464         return ret;
1465 }
1466
1467 /*
1468  * Dissolve a given free hugepage into free buddy pages. This function does
1469  * nothing for in-use (including surplus) hugepages. Returns -EBUSY if the
1470  * number of free hugepages would be reduced below the number of reserved
1471  * hugepages.
1472  */
1473 int dissolve_free_huge_page(struct page *page)
1474 {
1475         int rc = 0;
1476
1477         spin_lock(&hugetlb_lock);
1478         if (PageHuge(page) && !page_count(page)) {
1479                 struct page *head = compound_head(page);
1480                 struct hstate *h = page_hstate(head);
1481                 int nid = page_to_nid(head);
1482                 if (h->free_huge_pages - h->resv_huge_pages == 0) {
1483                         rc = -EBUSY;
1484                         goto out;
1485                 }
1486                 /*
1487                  * Move PageHWPoison flag from head page to the raw error page,
1488                  * which makes any subpages rather than the error page reusable.
1489                  */
1490                 if (PageHWPoison(head) && page != head) {
1491                         SetPageHWPoison(page);
1492                         ClearPageHWPoison(head);
1493                 }
1494                 list_del(&head->lru);
1495                 h->free_huge_pages--;
1496                 h->free_huge_pages_node[nid]--;
1497                 h->max_huge_pages--;
1498                 update_and_free_page(h, head);
1499         }
1500 out:
1501         spin_unlock(&hugetlb_lock);
1502         return rc;
1503 }
1504
1505 /*
1506  * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1507  * make specified memory blocks removable from the system.
1508  * Note that this will dissolve a free gigantic hugepage completely, if any
1509  * part of it lies within the given range.
1510  * Also note that if dissolve_free_huge_page() returns with an error, all
1511  * free hugepages that were dissolved before that error are lost.
1512  */
1513 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1514 {
1515         unsigned long pfn;
1516         struct page *page;
1517         int rc = 0;
1518
1519         if (!hugepages_supported())
1520                 return rc;
1521
1522         for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1523                 page = pfn_to_page(pfn);
1524                 if (PageHuge(page) && !page_count(page)) {
1525                         rc = dissolve_free_huge_page(page);
1526                         if (rc)
1527                                 break;
1528                 }
1529         }
1530
1531         return rc;
1532 }
1533
1534 /*
1535  * There are 3 ways this can get called:
1536  * 1. With vma+addr: we use the VMA's memory policy
1537  * 2. With !vma, but nid=NUMA_NO_NODE:  We try to allocate a huge
1538  *    page from any node, and let the buddy allocator itself figure
1539  *    it out.
1540  * 3. With !vma, but nid!=NUMA_NO_NODE.  We allocate a huge page
1541  *    strictly from 'nid'
1542  */
1543 static struct page *__hugetlb_alloc_buddy_huge_page(struct hstate *h,
1544                 struct vm_area_struct *vma, unsigned long addr, int nid)
1545 {
1546         int order = huge_page_order(h);
1547         gfp_t gfp = htlb_alloc_mask(h)|__GFP_COMP|__GFP_REPEAT|__GFP_NOWARN;
1548         unsigned int cpuset_mems_cookie;
1549
1550         /*
1551          * We need a VMA to get a memory policy.  If we do not
1552          * have one, we use the 'nid' argument.
1553          *
1554          * The mempolicy stuff below has some non-inlined bits
1555          * and calls ->vm_ops.  That makes it hard to optimize at
1556          * compile-time, even when NUMA is off and it does
1557          * nothing.  This helps the compiler optimize it out.
1558          */
1559         if (!IS_ENABLED(CONFIG_NUMA) || !vma) {
1560                 /*
1561                  * If a specific node is requested, make sure to
1562                  * get memory from there, but only when a node
1563                  * is explicitly specified.
1564                  */
1565                 if (nid != NUMA_NO_NODE)
1566                         gfp |= __GFP_THISNODE;
1567                 /*
1568                  * Make sure to call something that can handle
1569                  * nid=NUMA_NO_NODE
1570                  */
1571                 return alloc_pages_node(nid, gfp, order);
1572         }
1573
1574         /*
1575          * OK, so we have a VMA.  Fetch the mempolicy and try to
1576          * allocate a huge page with it.  We will only reach this
1577          * when CONFIG_NUMA=y.
1578          */
1579         do {
1580                 struct page *page;
1581                 struct mempolicy *mpol;
1582                 int nid;
1583                 nodemask_t *nodemask;
1584
1585                 cpuset_mems_cookie = read_mems_allowed_begin();
1586                 nid = huge_node(vma, addr, gfp, &mpol, &nodemask);
1587                 mpol_cond_put(mpol);
1588                 page = __alloc_pages_nodemask(gfp, order, nid, nodemask);
1589                 if (page)
1590                         return page;
1591         } while (read_mems_allowed_retry(cpuset_mems_cookie));
1592
1593         return NULL;
1594 }
1595
1596 /*
1597  * There are two ways to allocate a huge page:
1598  * 1. When you have a VMA and an address (like a fault)
1599  * 2. When you have no VMA (like when setting /proc/.../nr_hugepages)
1600  *
1601  * 'vma' and 'addr' are only for (1).  'nid' is always NUMA_NO_NODE in
1602  * this case which signifies that the allocation should be done with
1603  * respect for the VMA's memory policy.
1604  *
1605  * For (2), we ignore 'vma' and 'addr' and use 'nid' exclusively. This
1606  * implies that memory policies will not be taken in to account.
1607  */
1608 static struct page *__alloc_buddy_huge_page(struct hstate *h,
1609                 struct vm_area_struct *vma, unsigned long addr, int nid)
1610 {
1611         struct page *page;
1612         unsigned int r_nid;
1613
1614         if (hstate_is_gigantic(h))
1615                 return NULL;
1616
1617         /*
1618          * Make sure that anyone specifying 'nid' is not also specifying a VMA.
1619          * This makes sure the caller is picking _one_ of the modes with which
1620          * we can call this function, not both.
1621          */
1622         if (vma || (addr != -1)) {
1623                 VM_WARN_ON_ONCE(addr == -1);
1624                 VM_WARN_ON_ONCE(nid != NUMA_NO_NODE);
1625         }
1626         /*
1627          * Assume we will successfully allocate the surplus page to
1628          * prevent racing processes from causing the surplus to exceed
1629          * overcommit
1630          *
1631          * This however introduces a different race, where a process B
1632          * tries to grow the static hugepage pool while alloc_pages() is
1633          * called by process A. B will only examine the per-node
1634          * counters in determining if surplus huge pages can be
1635          * converted to normal huge pages in adjust_pool_surplus(). A
1636          * won't be able to increment the per-node counter, until the
1637          * lock is dropped by B, but B doesn't drop hugetlb_lock until
1638          * no more huge pages can be converted from surplus to normal
1639          * state (and doesn't try to convert again). Thus, we have a
1640          * case where a surplus huge page exists, the pool is grown, and
1641          * the surplus huge page still exists after, even though it
1642          * should just have been converted to a normal huge page. This
1643          * does not leak memory, though, as the hugepage will be freed
1644          * once it is out of use. It also does not allow the counters to
1645          * go out of whack in adjust_pool_surplus() as we don't modify
1646          * the node values until we've gotten the hugepage and only the
1647          * per-node value is checked there.
1648          */
1649         spin_lock(&hugetlb_lock);
1650         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1651                 spin_unlock(&hugetlb_lock);
1652                 return NULL;
1653         } else {
1654                 h->nr_huge_pages++;
1655                 h->surplus_huge_pages++;
1656         }
1657         spin_unlock(&hugetlb_lock);
1658
1659         page = __hugetlb_alloc_buddy_huge_page(h, vma, addr, nid);
1660
1661         spin_lock(&hugetlb_lock);
1662         if (page) {
1663                 INIT_LIST_HEAD(&page->lru);
1664                 r_nid = page_to_nid(page);
1665                 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1666                 set_hugetlb_cgroup(page, NULL);
1667                 /*
1668                  * We incremented the global counters already
1669                  */
1670                 h->nr_huge_pages_node[r_nid]++;
1671                 h->surplus_huge_pages_node[r_nid]++;
1672                 __count_vm_event(HTLB_BUDDY_PGALLOC);
1673         } else {
1674                 h->nr_huge_pages--;
1675                 h->surplus_huge_pages--;
1676                 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1677         }
1678         spin_unlock(&hugetlb_lock);
1679
1680         return page;
1681 }
1682
1683 /*
1684  * Allocate a huge page from 'nid'.  Note, 'nid' may be
1685  * NUMA_NO_NODE, which means that it may be allocated
1686  * anywhere.
1687  */
1688 static
1689 struct page *__alloc_buddy_huge_page_no_mpol(struct hstate *h, int nid)
1690 {
1691         unsigned long addr = -1;
1692
1693         return __alloc_buddy_huge_page(h, NULL, addr, nid);
1694 }
1695
1696 /*
1697  * Use the VMA's mpolicy to allocate a huge page from the buddy.
1698  */
1699 static
1700 struct page *__alloc_buddy_huge_page_with_mpol(struct hstate *h,
1701                 struct vm_area_struct *vma, unsigned long addr)
1702 {
1703         return __alloc_buddy_huge_page(h, vma, addr, NUMA_NO_NODE);
1704 }
1705
1706 /*
1707  * This allocation function is useful in the context where vma is irrelevant.
1708  * E.g. soft-offlining uses this function because it only cares physical
1709  * address of error page.
1710  */
1711 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1712 {
1713         struct page *page = NULL;
1714
1715         spin_lock(&hugetlb_lock);
1716         if (h->free_huge_pages - h->resv_huge_pages > 0)
1717                 page = dequeue_huge_page_node(h, nid);
1718         spin_unlock(&hugetlb_lock);
1719
1720         if (!page)
1721                 page = __alloc_buddy_huge_page_no_mpol(h, nid);
1722
1723         return page;
1724 }
1725
1726 struct page *alloc_huge_page_nodemask(struct hstate *h, const nodemask_t *nmask)
1727 {
1728         struct page *page = NULL;
1729         int node;
1730
1731         spin_lock(&hugetlb_lock);
1732         if (h->free_huge_pages - h->resv_huge_pages > 0) {
1733                 for_each_node_mask(node, *nmask) {
1734                         page = dequeue_huge_page_node_exact(h, node);
1735                         if (page)
1736                                 break;
1737                 }
1738         }
1739         spin_unlock(&hugetlb_lock);
1740         if (page)
1741                 return page;
1742
1743         /* No reservations, try to overcommit */
1744         for_each_node_mask(node, *nmask) {
1745                 page = __alloc_buddy_huge_page_no_mpol(h, node);
1746                 if (page)
1747                         return page;
1748         }
1749
1750         return NULL;
1751 }
1752
1753 /*
1754  * Increase the hugetlb pool such that it can accommodate a reservation
1755  * of size 'delta'.
1756  */
1757 static int gather_surplus_pages(struct hstate *h, int delta)
1758 {
1759         struct list_head surplus_list;
1760         struct page *page, *tmp;
1761         int ret, i;
1762         int needed, allocated;
1763         bool alloc_ok = true;
1764
1765         needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1766         if (needed <= 0) {
1767                 h->resv_huge_pages += delta;
1768                 return 0;
1769         }
1770
1771         allocated = 0;
1772         INIT_LIST_HEAD(&surplus_list);
1773
1774         ret = -ENOMEM;
1775 retry:
1776         spin_unlock(&hugetlb_lock);
1777         for (i = 0; i < needed; i++) {
1778                 page = __alloc_buddy_huge_page_no_mpol(h, NUMA_NO_NODE);
1779                 if (!page) {
1780                         alloc_ok = false;
1781                         break;
1782                 }
1783                 list_add(&page->lru, &surplus_list);
1784                 cond_resched();
1785         }
1786         allocated += i;
1787
1788         /*
1789          * After retaking hugetlb_lock, we need to recalculate 'needed'
1790          * because either resv_huge_pages or free_huge_pages may have changed.
1791          */
1792         spin_lock(&hugetlb_lock);
1793         needed = (h->resv_huge_pages + delta) -
1794                         (h->free_huge_pages + allocated);
1795         if (needed > 0) {
1796                 if (alloc_ok)
1797                         goto retry;
1798                 /*
1799                  * We were not able to allocate enough pages to
1800                  * satisfy the entire reservation so we free what
1801                  * we've allocated so far.
1802                  */
1803                 goto free;
1804         }
1805         /*
1806          * The surplus_list now contains _at_least_ the number of extra pages
1807          * needed to accommodate the reservation.  Add the appropriate number
1808          * of pages to the hugetlb pool and free the extras back to the buddy
1809          * allocator.  Commit the entire reservation here to prevent another
1810          * process from stealing the pages as they are added to the pool but
1811          * before they are reserved.
1812          */
1813         needed += allocated;
1814         h->resv_huge_pages += delta;
1815         ret = 0;
1816
1817         /* Free the needed pages to the hugetlb pool */
1818         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1819                 if ((--needed) < 0)
1820                         break;
1821                 /*
1822                  * This page is now managed by the hugetlb allocator and has
1823                  * no users -- drop the buddy allocator's reference.
1824                  */
1825                 put_page_testzero(page);
1826                 VM_BUG_ON_PAGE(page_count(page), page);
1827                 enqueue_huge_page(h, page);
1828         }
1829 free:
1830         spin_unlock(&hugetlb_lock);
1831
1832         /* Free unnecessary surplus pages to the buddy allocator */
1833         list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1834                 put_page(page);
1835         spin_lock(&hugetlb_lock);
1836
1837         return ret;
1838 }
1839
1840 /*
1841  * This routine has two main purposes:
1842  * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1843  *    in unused_resv_pages.  This corresponds to the prior adjustments made
1844  *    to the associated reservation map.
1845  * 2) Free any unused surplus pages that may have been allocated to satisfy
1846  *    the reservation.  As many as unused_resv_pages may be freed.
1847  *
1848  * Called with hugetlb_lock held.  However, the lock could be dropped (and
1849  * reacquired) during calls to cond_resched_lock.  Whenever dropping the lock,
1850  * we must make sure nobody else can claim pages we are in the process of
1851  * freeing.  Do this by ensuring resv_huge_page always is greater than the
1852  * number of huge pages we plan to free when dropping the lock.
1853  */
1854 static void return_unused_surplus_pages(struct hstate *h,
1855                                         unsigned long unused_resv_pages)
1856 {
1857         unsigned long nr_pages;
1858
1859         /* Cannot return gigantic pages currently */
1860         if (hstate_is_gigantic(h))
1861                 goto out;
1862
1863         /*
1864          * Part (or even all) of the reservation could have been backed
1865          * by pre-allocated pages. Only free surplus pages.
1866          */
1867         nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1868
1869         /*
1870          * We want to release as many surplus pages as possible, spread
1871          * evenly across all nodes with memory. Iterate across these nodes
1872          * until we can no longer free unreserved surplus pages. This occurs
1873          * when the nodes with surplus pages have no free pages.
1874          * free_pool_huge_page() will balance the the freed pages across the
1875          * on-line nodes with memory and will handle the hstate accounting.
1876          *
1877          * Note that we decrement resv_huge_pages as we free the pages.  If
1878          * we drop the lock, resv_huge_pages will still be sufficiently large
1879          * to cover subsequent pages we may free.
1880          */
1881         while (nr_pages--) {
1882                 h->resv_huge_pages--;
1883                 unused_resv_pages--;
1884                 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1885                         goto out;
1886                 cond_resched_lock(&hugetlb_lock);
1887         }
1888
1889 out:
1890         /* Fully uncommit the reservation */
1891         h->resv_huge_pages -= unused_resv_pages;
1892 }
1893
1894
1895 /*
1896  * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1897  * are used by the huge page allocation routines to manage reservations.
1898  *
1899  * vma_needs_reservation is called to determine if the huge page at addr
1900  * within the vma has an associated reservation.  If a reservation is
1901  * needed, the value 1 is returned.  The caller is then responsible for
1902  * managing the global reservation and subpool usage counts.  After
1903  * the huge page has been allocated, vma_commit_reservation is called
1904  * to add the page to the reservation map.  If the page allocation fails,
1905  * the reservation must be ended instead of committed.  vma_end_reservation
1906  * is called in such cases.
1907  *
1908  * In the normal case, vma_commit_reservation returns the same value
1909  * as the preceding vma_needs_reservation call.  The only time this
1910  * is not the case is if a reserve map was changed between calls.  It
1911  * is the responsibility of the caller to notice the difference and
1912  * take appropriate action.
1913  *
1914  * vma_add_reservation is used in error paths where a reservation must
1915  * be restored when a newly allocated huge page must be freed.  It is
1916  * to be called after calling vma_needs_reservation to determine if a
1917  * reservation exists.
1918  */
1919 enum vma_resv_mode {
1920         VMA_NEEDS_RESV,
1921         VMA_COMMIT_RESV,
1922         VMA_END_RESV,
1923         VMA_ADD_RESV,
1924 };
1925 static long __vma_reservation_common(struct hstate *h,
1926                                 struct vm_area_struct *vma, unsigned long addr,
1927                                 enum vma_resv_mode mode)
1928 {
1929         struct resv_map *resv;
1930         pgoff_t idx;
1931         long ret;
1932
1933         resv = vma_resv_map(vma);
1934         if (!resv)
1935                 return 1;
1936
1937         idx = vma_hugecache_offset(h, vma, addr);
1938         switch (mode) {
1939         case VMA_NEEDS_RESV:
1940                 ret = region_chg(resv, idx, idx + 1);
1941                 break;
1942         case VMA_COMMIT_RESV:
1943                 ret = region_add(resv, idx, idx + 1);
1944                 break;
1945         case VMA_END_RESV:
1946                 region_abort(resv, idx, idx + 1);
1947                 ret = 0;
1948                 break;
1949         case VMA_ADD_RESV:
1950                 if (vma->vm_flags & VM_MAYSHARE)
1951                         ret = region_add(resv, idx, idx + 1);
1952                 else {
1953                         region_abort(resv, idx, idx + 1);
1954                         ret = region_del(resv, idx, idx + 1);
1955                 }
1956                 break;
1957         default:
1958                 BUG();
1959         }
1960
1961         if (vma->vm_flags & VM_MAYSHARE)
1962                 return ret;
1963         else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
1964                 /*
1965                  * In most cases, reserves always exist for private mappings.
1966                  * However, a file associated with mapping could have been
1967                  * hole punched or truncated after reserves were consumed.
1968                  * As subsequent fault on such a range will not use reserves.
1969                  * Subtle - The reserve map for private mappings has the
1970                  * opposite meaning than that of shared mappings.  If NO
1971                  * entry is in the reserve map, it means a reservation exists.
1972                  * If an entry exists in the reserve map, it means the
1973                  * reservation has already been consumed.  As a result, the
1974                  * return value of this routine is the opposite of the
1975                  * value returned from reserve map manipulation routines above.
1976                  */
1977                 if (ret)
1978                         return 0;
1979                 else
1980                         return 1;
1981         }
1982         else
1983                 return ret < 0 ? ret : 0;
1984 }
1985
1986 static long vma_needs_reservation(struct hstate *h,
1987                         struct vm_area_struct *vma, unsigned long addr)
1988 {
1989         return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
1990 }
1991
1992 static long vma_commit_reservation(struct hstate *h,
1993                         struct vm_area_struct *vma, unsigned long addr)
1994 {
1995         return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1996 }
1997
1998 static void vma_end_reservation(struct hstate *h,
1999                         struct vm_area_struct *vma, unsigned long addr)
2000 {
2001         (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2002 }
2003
2004 static long vma_add_reservation(struct hstate *h,
2005                         struct vm_area_struct *vma, unsigned long addr)
2006 {
2007         return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2008 }
2009
2010 /*
2011  * This routine is called to restore a reservation on error paths.  In the
2012  * specific error paths, a huge page was allocated (via alloc_huge_page)
2013  * and is about to be freed.  If a reservation for the page existed,
2014  * alloc_huge_page would have consumed the reservation and set PagePrivate
2015  * in the newly allocated page.  When the page is freed via free_huge_page,
2016  * the global reservation count will be incremented if PagePrivate is set.
2017  * However, free_huge_page can not adjust the reserve map.  Adjust the
2018  * reserve map here to be consistent with global reserve count adjustments
2019  * to be made by free_huge_page.
2020  */
2021 static void restore_reserve_on_error(struct hstate *h,
2022                         struct vm_area_struct *vma, unsigned long address,
2023                         struct page *page)
2024 {
2025         if (unlikely(PagePrivate(page))) {
2026                 long rc = vma_needs_reservation(h, vma, address);
2027
2028                 if (unlikely(rc < 0)) {
2029                         /*
2030                          * Rare out of memory condition in reserve map
2031                          * manipulation.  Clear PagePrivate so that
2032                          * global reserve count will not be incremented
2033                          * by free_huge_page.  This will make it appear
2034                          * as though the reservation for this page was
2035                          * consumed.  This may prevent the task from
2036                          * faulting in the page at a later time.  This
2037                          * is better than inconsistent global huge page
2038                          * accounting of reserve counts.
2039                          */
2040                         ClearPagePrivate(page);
2041                 } else if (rc) {
2042                         rc = vma_add_reservation(h, vma, address);
2043                         if (unlikely(rc < 0))
2044                                 /*
2045                                  * See above comment about rare out of
2046                                  * memory condition.
2047                                  */
2048                                 ClearPagePrivate(page);
2049                 } else
2050                         vma_end_reservation(h, vma, address);
2051         }
2052 }
2053
2054 struct page *alloc_huge_page(struct vm_area_struct *vma,
2055                                     unsigned long addr, int avoid_reserve)
2056 {
2057         struct hugepage_subpool *spool = subpool_vma(vma);
2058         struct hstate *h = hstate_vma(vma);
2059         struct page *page;
2060         long map_chg, map_commit;
2061         long gbl_chg;
2062         int ret, idx;
2063         struct hugetlb_cgroup *h_cg;
2064
2065         idx = hstate_index(h);
2066         /*
2067          * Examine the region/reserve map to determine if the process
2068          * has a reservation for the page to be allocated.  A return
2069          * code of zero indicates a reservation exists (no change).
2070          */
2071         map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2072         if (map_chg < 0)
2073                 return ERR_PTR(-ENOMEM);
2074
2075         /*
2076          * Processes that did not create the mapping will have no
2077          * reserves as indicated by the region/reserve map. Check
2078          * that the allocation will not exceed the subpool limit.
2079          * Allocations for MAP_NORESERVE mappings also need to be
2080          * checked against any subpool limit.
2081          */
2082         if (map_chg || avoid_reserve) {
2083                 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2084                 if (gbl_chg < 0) {
2085                         vma_end_reservation(h, vma, addr);
2086                         return ERR_PTR(-ENOSPC);
2087                 }
2088
2089                 /*
2090                  * Even though there was no reservation in the region/reserve
2091                  * map, there could be reservations associated with the
2092                  * subpool that can be used.  This would be indicated if the
2093                  * return value of hugepage_subpool_get_pages() is zero.
2094                  * However, if avoid_reserve is specified we still avoid even
2095                  * the subpool reservations.
2096                  */
2097                 if (avoid_reserve)
2098                         gbl_chg = 1;
2099         }
2100
2101         ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2102         if (ret)
2103                 goto out_subpool_put;
2104
2105         spin_lock(&hugetlb_lock);
2106         /*
2107          * glb_chg is passed to indicate whether or not a page must be taken
2108          * from the global free pool (global change).  gbl_chg == 0 indicates
2109          * a reservation exists for the allocation.
2110          */
2111         page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2112         if (!page) {
2113                 spin_unlock(&hugetlb_lock);
2114                 page = __alloc_buddy_huge_page_with_mpol(h, vma, addr);
2115                 if (!page)
2116                         goto out_uncharge_cgroup;
2117                 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2118                         SetPagePrivate(page);
2119                         h->resv_huge_pages--;
2120                 }
2121                 spin_lock(&hugetlb_lock);
2122                 list_move(&page->lru, &h->hugepage_activelist);
2123                 /* Fall through */
2124         }
2125         hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2126         spin_unlock(&hugetlb_lock);
2127
2128         set_page_private(page, (unsigned long)spool);
2129
2130         map_commit = vma_commit_reservation(h, vma, addr);
2131         if (unlikely(map_chg > map_commit)) {
2132                 /*
2133                  * The page was added to the reservation map between
2134                  * vma_needs_reservation and vma_commit_reservation.
2135                  * This indicates a race with hugetlb_reserve_pages.
2136                  * Adjust for the subpool count incremented above AND
2137                  * in hugetlb_reserve_pages for the same page.  Also,
2138                  * the reservation count added in hugetlb_reserve_pages
2139                  * no longer applies.
2140                  */
2141                 long rsv_adjust;
2142
2143                 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2144                 hugetlb_acct_memory(h, -rsv_adjust);
2145         }
2146         return page;
2147
2148 out_uncharge_cgroup:
2149         hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2150 out_subpool_put:
2151         if (map_chg || avoid_reserve)
2152                 hugepage_subpool_put_pages(spool, 1);
2153         vma_end_reservation(h, vma, addr);
2154         return ERR_PTR(-ENOSPC);
2155 }
2156
2157 /*
2158  * alloc_huge_page()'s wrapper which simply returns the page if allocation
2159  * succeeds, otherwise NULL. This function is called from new_vma_page(),
2160  * where no ERR_VALUE is expected to be returned.
2161  */
2162 struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
2163                                 unsigned long addr, int avoid_reserve)
2164 {
2165         struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
2166         if (IS_ERR(page))
2167                 page = NULL;
2168         return page;
2169 }
2170
2171 int __weak alloc_bootmem_huge_page(struct hstate *h)
2172 {
2173         struct huge_bootmem_page *m;
2174         int nr_nodes, node;
2175
2176         for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2177                 void *addr;
2178
2179                 addr = memblock_virt_alloc_try_nid_nopanic(
2180                                 huge_page_size(h), huge_page_size(h),
2181                                 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
2182                 if (addr) {
2183                         /*
2184                          * Use the beginning of the huge page to store the
2185                          * huge_bootmem_page struct (until gather_bootmem
2186                          * puts them into the mem_map).
2187                          */
2188                         m = addr;
2189                         goto found;
2190                 }
2191         }
2192         return 0;
2193
2194 found:
2195         BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2196         /* Put them into a private list first because mem_map is not up yet */
2197         list_add(&m->list, &huge_boot_pages);
2198         m->hstate = h;
2199         return 1;
2200 }
2201
2202 static void __init prep_compound_huge_page(struct page *page,
2203                 unsigned int order)
2204 {
2205         if (unlikely(order > (MAX_ORDER - 1)))
2206                 prep_compound_gigantic_page(page, order);
2207         else
2208                 prep_compound_page(page, order);
2209 }
2210
2211 /* Put bootmem huge pages into the standard lists after mem_map is up */
2212 static void __init gather_bootmem_prealloc(void)
2213 {
2214         struct huge_bootmem_page *m;
2215
2216         list_for_each_entry(m, &huge_boot_pages, list) {
2217                 struct hstate *h = m->hstate;
2218                 struct page *page;
2219
2220 #ifdef CONFIG_HIGHMEM
2221                 page = pfn_to_page(m->phys >> PAGE_SHIFT);
2222                 memblock_free_late(__pa(m),
2223                                    sizeof(struct huge_bootmem_page));
2224 #else
2225                 page = virt_to_page(m);
2226 #endif
2227                 WARN_ON(page_count(page) != 1);
2228                 prep_compound_huge_page(page, h->order);
2229                 WARN_ON(PageReserved(page));
2230                 prep_new_huge_page(h, page, page_to_nid(page));
2231                 /*
2232                  * If we had gigantic hugepages allocated at boot time, we need
2233                  * to restore the 'stolen' pages to totalram_pages in order to
2234                  * fix confusing memory reports from free(1) and another
2235                  * side-effects, like CommitLimit going negative.
2236                  */
2237                 if (hstate_is_gigantic(h))
2238                         adjust_managed_page_count(page, 1 << h->order);
2239         }
2240 }
2241
2242 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2243 {
2244         unsigned long i;
2245
2246         for (i = 0; i < h->max_huge_pages; ++i) {
2247                 if (hstate_is_gigantic(h)) {
2248                         if (!alloc_bootmem_huge_page(h))
2249                                 break;
2250                 } else if (!alloc_fresh_huge_page(h,
2251                                          &node_states[N_MEMORY]))
2252                         break;
2253                 cond_resched();
2254         }
2255         if (i < h->max_huge_pages) {
2256                 char buf[32];
2257
2258                 memfmt(buf, huge_page_size(h)),
2259                 pr_warn("HugeTLB: allocating %lu of page size %s failed.  Only allocated %lu hugepages.\n",
2260                         h->max_huge_pages, buf, i);
2261                 h->max_huge_pages = i;
2262         }
2263 }
2264
2265 static void __init hugetlb_init_hstates(void)
2266 {
2267         struct hstate *h;
2268
2269         for_each_hstate(h) {
2270                 if (minimum_order > huge_page_order(h))
2271                         minimum_order = huge_page_order(h);
2272
2273                 /* oversize hugepages were init'ed in early boot */
2274                 if (!hstate_is_gigantic(h))
2275                         hugetlb_hstate_alloc_pages(h);
2276         }
2277         VM_BUG_ON(minimum_order == UINT_MAX);
2278 }
2279
2280 static void __init report_hugepages(void)
2281 {
2282         struct hstate *h;
2283
2284         for_each_hstate(h) {
2285                 char buf[32];
2286                 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2287                         memfmt(buf, huge_page_size(h)),
2288                         h->free_huge_pages);
2289         }
2290 }
2291
2292 #ifdef CONFIG_HIGHMEM
2293 static void try_to_free_low(struct hstate *h, unsigned long count,
2294                                                 nodemask_t *nodes_allowed)
2295 {
2296         int i;
2297
2298         if (hstate_is_gigantic(h))
2299                 return;
2300
2301         for_each_node_mask(i, *nodes_allowed) {
2302                 struct page *page, *next;
2303                 struct list_head *freel = &h->hugepage_freelists[i];
2304                 list_for_each_entry_safe(page, next, freel, lru) {
2305                         if (count >= h->nr_huge_pages)
2306                                 return;
2307                         if (PageHighMem(page))
2308                                 continue;
2309                         list_del(&page->lru);
2310                         update_and_free_page(h, page);
2311                         h->free_huge_pages--;
2312                         h->free_huge_pages_node[page_to_nid(page)]--;
2313                 }
2314         }
2315 }
2316 #else
2317 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2318                                                 nodemask_t *nodes_allowed)
2319 {
2320 }
2321 #endif
2322
2323 /*
2324  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
2325  * balanced by operating on them in a round-robin fashion.
2326  * Returns 1 if an adjustment was made.
2327  */
2328 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2329                                 int delta)
2330 {
2331         int nr_nodes, node;
2332
2333         VM_BUG_ON(delta != -1 && delta != 1);
2334
2335         if (delta < 0) {
2336                 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2337                         if (h->surplus_huge_pages_node[node])
2338                                 goto found;
2339                 }
2340         } else {
2341                 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2342                         if (h->surplus_huge_pages_node[node] <
2343                                         h->nr_huge_pages_node[node])
2344                                 goto found;
2345                 }
2346         }
2347         return 0;
2348
2349 found:
2350         h->surplus_huge_pages += delta;
2351         h->surplus_huge_pages_node[node] += delta;
2352         return 1;
2353 }
2354
2355 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2356 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
2357                                                 nodemask_t *nodes_allowed)
2358 {
2359         unsigned long min_count, ret;
2360
2361         if (hstate_is_gigantic(h) && !gigantic_page_supported())
2362                 return h->max_huge_pages;
2363
2364         /*
2365          * Increase the pool size
2366          * First take pages out of surplus state.  Then make up the
2367          * remaining difference by allocating fresh huge pages.
2368          *
2369          * We might race with __alloc_buddy_huge_page() here and be unable
2370          * to convert a surplus huge page to a normal huge page. That is
2371          * not critical, though, it just means the overall size of the
2372          * pool might be one hugepage larger than it needs to be, but
2373          * within all the constraints specified by the sysctls.
2374          */
2375         spin_lock(&hugetlb_lock);
2376         while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2377                 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2378                         break;
2379         }
2380
2381         while (count > persistent_huge_pages(h)) {
2382                 /*
2383                  * If this allocation races such that we no longer need the
2384                  * page, free_huge_page will handle it by freeing the page
2385                  * and reducing the surplus.
2386                  */
2387                 spin_unlock(&hugetlb_lock);
2388
2389                 /* yield cpu to avoid soft lockup */
2390                 cond_resched();
2391
2392                 if (hstate_is_gigantic(h))
2393                         ret = alloc_fresh_gigantic_page(h, nodes_allowed);
2394                 else
2395                         ret = alloc_fresh_huge_page(h, nodes_allowed);
2396                 spin_lock(&hugetlb_lock);
2397                 if (!ret)
2398                         goto out;
2399
2400                 /* Bail for signals. Probably ctrl-c from user */
2401                 if (signal_pending(current))
2402                         goto out;
2403         }
2404
2405         /*
2406          * Decrease the pool size
2407          * First return free pages to the buddy allocator (being careful
2408          * to keep enough around to satisfy reservations).  Then place
2409          * pages into surplus state as needed so the pool will shrink
2410          * to the desired size as pages become free.
2411          *
2412          * By placing pages into the surplus state independent of the
2413          * overcommit value, we are allowing the surplus pool size to
2414          * exceed overcommit. There are few sane options here. Since
2415          * __alloc_buddy_huge_page() is checking the global counter,
2416          * though, we'll note that we're not allowed to exceed surplus
2417          * and won't grow the pool anywhere else. Not until one of the
2418          * sysctls are changed, or the surplus pages go out of use.
2419          */
2420         min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2421         min_count = max(count, min_count);
2422         try_to_free_low(h, min_count, nodes_allowed);
2423         while (min_count < persistent_huge_pages(h)) {
2424                 if (!free_pool_huge_page(h, nodes_allowed, 0))
2425                         break;
2426                 cond_resched_lock(&hugetlb_lock);
2427         }
2428         while (count < persistent_huge_pages(h)) {
2429                 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2430                         break;
2431         }
2432 out:
2433         ret = persistent_huge_pages(h);
2434         spin_unlock(&hugetlb_lock);
2435         return ret;
2436 }
2437
2438 #define HSTATE_ATTR_RO(_name) \
2439         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2440
2441 #define HSTATE_ATTR(_name) \
2442         static struct kobj_attribute _name##_attr = \
2443                 __ATTR(_name, 0644, _name##_show, _name##_store)
2444
2445 static struct kobject *hugepages_kobj;
2446 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2447
2448 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2449
2450 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2451 {
2452         int i;
2453
2454         for (i = 0; i < HUGE_MAX_HSTATE; i++)
2455                 if (hstate_kobjs[i] == kobj) {
2456                         if (nidp)
2457                                 *nidp = NUMA_NO_NODE;
2458                         return &hstates[i];
2459                 }
2460
2461         return kobj_to_node_hstate(kobj, nidp);
2462 }
2463
2464 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2465                                         struct kobj_attribute *attr, char *buf)
2466 {
2467         struct hstate *h;
2468         unsigned long nr_huge_pages;
2469         int nid;
2470
2471         h = kobj_to_hstate(kobj, &nid);
2472         if (nid == NUMA_NO_NODE)
2473                 nr_huge_pages = h->nr_huge_pages;
2474         else
2475                 nr_huge_pages = h->nr_huge_pages_node[nid];
2476
2477         return sprintf(buf, "%lu\n", nr_huge_pages);
2478 }
2479
2480 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2481                                            struct hstate *h, int nid,
2482                                            unsigned long count, size_t len)
2483 {
2484         int err;
2485         NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
2486
2487         if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
2488                 err = -EINVAL;
2489                 goto out;
2490         }
2491
2492         if (nid == NUMA_NO_NODE) {
2493                 /*
2494                  * global hstate attribute
2495                  */
2496                 if (!(obey_mempolicy &&
2497                                 init_nodemask_of_mempolicy(nodes_allowed))) {
2498                         NODEMASK_FREE(nodes_allowed);
2499                         nodes_allowed = &node_states[N_MEMORY];
2500                 }
2501         } else if (nodes_allowed) {
2502                 /*
2503                  * per node hstate attribute: adjust count to global,
2504                  * but restrict alloc/free to the specified node.
2505                  */
2506                 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2507                 init_nodemask_of_node(nodes_allowed, nid);
2508         } else
2509                 nodes_allowed = &node_states[N_MEMORY];
2510
2511         h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
2512
2513         if (nodes_allowed != &node_states[N_MEMORY])
2514                 NODEMASK_FREE(nodes_allowed);
2515
2516         return len;
2517 out:
2518         NODEMASK_FREE(nodes_allowed);
2519         return err;
2520 }
2521
2522 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2523                                          struct kobject *kobj, const char *buf,
2524                                          size_t len)
2525 {
2526         struct hstate *h;
2527         unsigned long count;
2528         int nid;
2529         int err;
2530
2531         err = kstrtoul(buf, 10, &count);
2532         if (err)
2533                 return err;
2534
2535         h = kobj_to_hstate(kobj, &nid);
2536         return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2537 }
2538
2539 static ssize_t nr_hugepages_show(struct kobject *kobj,
2540                                        struct kobj_attribute *attr, char *buf)
2541 {
2542         return nr_hugepages_show_common(kobj, attr, buf);
2543 }
2544
2545 static ssize_t nr_hugepages_store(struct kobject *kobj,
2546                struct kobj_attribute *attr, const char *buf, size_t len)
2547 {
2548         return nr_hugepages_store_common(false, kobj, buf, len);
2549 }
2550 HSTATE_ATTR(nr_hugepages);
2551
2552 #ifdef CONFIG_NUMA
2553
2554 /*
2555  * hstate attribute for optionally mempolicy-based constraint on persistent
2556  * huge page alloc/free.
2557  */
2558 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2559                                        struct kobj_attribute *attr, char *buf)
2560 {
2561         return nr_hugepages_show_common(kobj, attr, buf);
2562 }
2563
2564 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2565                struct kobj_attribute *attr, const char *buf, size_t len)
2566 {
2567         return nr_hugepages_store_common(true, kobj, buf, len);
2568 }
2569 HSTATE_ATTR(nr_hugepages_mempolicy);
2570 #endif
2571
2572
2573 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2574                                         struct kobj_attribute *attr, char *buf)
2575 {
2576         struct hstate *h = kobj_to_hstate(kobj, NULL);
2577         return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2578 }
2579
2580 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2581                 struct kobj_attribute *attr, const char *buf, size_t count)
2582 {
2583         int err;
2584         unsigned long input;
2585         struct hstate *h = kobj_to_hstate(kobj, NULL);
2586
2587         if (hstate_is_gigantic(h))
2588                 return -EINVAL;
2589
2590         err = kstrtoul(buf, 10, &input);
2591         if (err)
2592                 return err;
2593
2594         spin_lock(&hugetlb_lock);
2595         h->nr_overcommit_huge_pages = input;
2596         spin_unlock(&hugetlb_lock);
2597
2598         return count;
2599 }
2600 HSTATE_ATTR(nr_overcommit_hugepages);
2601
2602 static ssize_t free_hugepages_show(struct kobject *kobj,
2603                                         struct kobj_attribute *attr, char *buf)
2604 {
2605         struct hstate *h;
2606         unsigned long free_huge_pages;
2607         int nid;
2608
2609         h = kobj_to_hstate(kobj, &nid);
2610         if (nid == NUMA_NO_NODE)
2611                 free_huge_pages = h->free_huge_pages;
2612         else
2613                 free_huge_pages = h->free_huge_pages_node[nid];
2614
2615         return sprintf(buf, "%lu\n", free_huge_pages);
2616 }
2617 HSTATE_ATTR_RO(free_hugepages);
2618
2619 static ssize_t resv_hugepages_show(struct kobject *kobj,
2620                                         struct kobj_attribute *attr, char *buf)
2621 {
2622         struct hstate *h = kobj_to_hstate(kobj, NULL);
2623         return sprintf(buf, "%lu\n", h->resv_huge_pages);
2624 }
2625 HSTATE_ATTR_RO(resv_hugepages);
2626
2627 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2628                                         struct kobj_attribute *attr, char *buf)
2629 {
2630         struct hstate *h;
2631         unsigned long surplus_huge_pages;
2632         int nid;
2633
2634         h = kobj_to_hstate(kobj, &nid);
2635         if (nid == NUMA_NO_NODE)
2636                 surplus_huge_pages = h->surplus_huge_pages;
2637         else
2638                 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2639
2640         return sprintf(buf, "%lu\n", surplus_huge_pages);
2641 }
2642 HSTATE_ATTR_RO(surplus_hugepages);
2643
2644 static struct attribute *hstate_attrs[] = {
2645         &nr_hugepages_attr.attr,
2646         &nr_overcommit_hugepages_attr.attr,
2647         &free_hugepages_attr.attr,
2648         &resv_hugepages_attr.attr,
2649         &surplus_hugepages_attr.attr,
2650 #ifdef CONFIG_NUMA
2651         &nr_hugepages_mempolicy_attr.attr,
2652 #endif
2653         NULL,
2654 };
2655
2656 static struct attribute_group hstate_attr_group = {
2657         .attrs = hstate_attrs,
2658 };
2659
2660 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2661                                     struct kobject **hstate_kobjs,
2662                                     struct attribute_group *hstate_attr_group)
2663 {
2664         int retval;
2665         int hi = hstate_index(h);
2666
2667         hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2668         if (!hstate_kobjs[hi])
2669                 return -ENOMEM;
2670
2671         retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2672         if (retval)
2673                 kobject_put(hstate_kobjs[hi]);
2674
2675         return retval;
2676 }
2677
2678 static void __init hugetlb_sysfs_init(void)
2679 {
2680         struct hstate *h;
2681         int err;
2682
2683         hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2684         if (!hugepages_kobj)
2685                 return;
2686
2687         for_each_hstate(h) {
2688                 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2689                                          hstate_kobjs, &hstate_attr_group);
2690                 if (err)
2691                         pr_err("Hugetlb: Unable to add hstate %s", h->name);
2692         }
2693 }
2694
2695 #ifdef CONFIG_NUMA
2696
2697 /*
2698  * node_hstate/s - associate per node hstate attributes, via their kobjects,
2699  * with node devices in node_devices[] using a parallel array.  The array
2700  * index of a node device or _hstate == node id.
2701  * This is here to avoid any static dependency of the node device driver, in
2702  * the base kernel, on the hugetlb module.
2703  */
2704 struct node_hstate {
2705         struct kobject          *hugepages_kobj;
2706         struct kobject          *hstate_kobjs[HUGE_MAX_HSTATE];
2707 };
2708 static struct node_hstate node_hstates[MAX_NUMNODES];
2709
2710 /*
2711  * A subset of global hstate attributes for node devices
2712  */
2713 static struct attribute *per_node_hstate_attrs[] = {
2714         &nr_hugepages_attr.attr,
2715         &free_hugepages_attr.attr,
2716         &surplus_hugepages_attr.attr,
2717         NULL,
2718 };
2719
2720 static struct attribute_group per_node_hstate_attr_group = {
2721         .attrs = per_node_hstate_attrs,
2722 };
2723
2724 /*
2725  * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2726  * Returns node id via non-NULL nidp.
2727  */
2728 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2729 {
2730         int nid;
2731
2732         for (nid = 0; nid < nr_node_ids; nid++) {
2733                 struct node_hstate *nhs = &node_hstates[nid];
2734                 int i;
2735                 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2736                         if (nhs->hstate_kobjs[i] == kobj) {
2737                                 if (nidp)
2738                                         *nidp = nid;
2739                                 return &hstates[i];
2740                         }
2741         }
2742
2743         BUG();
2744         return NULL;
2745 }
2746
2747 /*
2748  * Unregister hstate attributes from a single node device.
2749  * No-op if no hstate attributes attached.
2750  */
2751 static void hugetlb_unregister_node(struct node *node)
2752 {
2753         struct hstate *h;
2754         struct node_hstate *nhs = &node_hstates[node->dev.id];
2755
2756         if (!nhs->hugepages_kobj)
2757                 return;         /* no hstate attributes */
2758
2759         for_each_hstate(h) {
2760                 int idx = hstate_index(h);
2761                 if (nhs->hstate_kobjs[idx]) {
2762                         kobject_put(nhs->hstate_kobjs[idx]);
2763                         nhs->hstate_kobjs[idx] = NULL;
2764                 }
2765         }
2766
2767         kobject_put(nhs->hugepages_kobj);
2768         nhs->hugepages_kobj = NULL;
2769 }
2770
2771
2772 /*
2773  * Register hstate attributes for a single node device.
2774  * No-op if attributes already registered.
2775  */
2776 static void hugetlb_register_node(struct node *node)
2777 {
2778         struct hstate *h;
2779         struct node_hstate *nhs = &node_hstates[node->dev.id];
2780         int err;
2781
2782         if (nhs->hugepages_kobj)
2783                 return;         /* already allocated */
2784
2785         nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2786                                                         &node->dev.kobj);
2787         if (!nhs->hugepages_kobj)
2788                 return;
2789
2790         for_each_hstate(h) {
2791                 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2792                                                 nhs->hstate_kobjs,
2793                                                 &per_node_hstate_attr_group);
2794                 if (err) {
2795                         pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2796                                 h->name, node->dev.id);
2797                         hugetlb_unregister_node(node);
2798                         break;
2799                 }
2800         }
2801 }
2802
2803 /*
2804  * hugetlb init time:  register hstate attributes for all registered node
2805  * devices of nodes that have memory.  All on-line nodes should have
2806  * registered their associated device by this time.
2807  */
2808 static void __init hugetlb_register_all_nodes(void)
2809 {
2810         int nid;
2811
2812         for_each_node_state(nid, N_MEMORY) {
2813                 struct node *node = node_devices[nid];
2814                 if (node->dev.id == nid)
2815                         hugetlb_register_node(node);
2816         }
2817
2818         /*
2819          * Let the node device driver know we're here so it can
2820          * [un]register hstate attributes on node hotplug.
2821          */
2822         register_hugetlbfs_with_node(hugetlb_register_node,
2823                                      hugetlb_unregister_node);
2824 }
2825 #else   /* !CONFIG_NUMA */
2826
2827 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2828 {
2829         BUG();
2830         if (nidp)
2831                 *nidp = -1;
2832         return NULL;
2833 }
2834
2835 static void hugetlb_register_all_nodes(void) { }
2836
2837 #endif
2838
2839 static int __init hugetlb_init(void)
2840 {
2841         int i;
2842
2843         if (!hugepages_supported())
2844                 return 0;
2845
2846         if (!size_to_hstate(default_hstate_size)) {
2847                 if (default_hstate_size != 0) {
2848                         pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n",
2849                                default_hstate_size, HPAGE_SIZE);
2850                 }
2851
2852                 default_hstate_size = HPAGE_SIZE;
2853                 if (!size_to_hstate(default_hstate_size))
2854                         hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2855         }
2856         default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2857         if (default_hstate_max_huge_pages) {
2858                 if (!default_hstate.max_huge_pages)
2859                         default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2860         }
2861
2862         hugetlb_init_hstates();
2863         gather_bootmem_prealloc();
2864         report_hugepages();
2865
2866         hugetlb_sysfs_init();
2867         hugetlb_register_all_nodes();
2868         hugetlb_cgroup_file_init();
2869
2870 #ifdef CONFIG_SMP
2871         num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2872 #else
2873         num_fault_mutexes = 1;
2874 #endif
2875         hugetlb_fault_mutex_table =
2876                 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
2877         BUG_ON(!hugetlb_fault_mutex_table);
2878
2879         for (i = 0; i < num_fault_mutexes; i++)
2880                 mutex_init(&hugetlb_fault_mutex_table[i]);
2881         return 0;
2882 }
2883 subsys_initcall(hugetlb_init);
2884
2885 /* Should be called on processing a hugepagesz=... option */
2886 void __init hugetlb_bad_size(void)
2887 {
2888         parsed_valid_hugepagesz = false;
2889 }
2890
2891 void __init hugetlb_add_hstate(unsigned int order)
2892 {
2893         struct hstate *h;
2894         unsigned long i;
2895
2896         if (size_to_hstate(PAGE_SIZE << order)) {
2897                 pr_warn("hugepagesz= specified twice, ignoring\n");
2898                 return;
2899         }
2900         BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2901         BUG_ON(order == 0);
2902         h = &hstates[hugetlb_max_hstate++];
2903         h->order = order;
2904         h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2905         h->nr_huge_pages = 0;
2906         h->free_huge_pages = 0;
2907         for (i = 0; i < MAX_NUMNODES; ++i)
2908                 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2909         INIT_LIST_HEAD(&h->hugepage_activelist);
2910         h->next_nid_to_alloc = first_memory_node;
2911         h->next_nid_to_free = first_memory_node;
2912         snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2913                                         huge_page_size(h)/1024);
2914
2915         parsed_hstate = h;
2916 }
2917
2918 static int __init hugetlb_nrpages_setup(char *s)
2919 {
2920         unsigned long *mhp;
2921         static unsigned long *last_mhp;
2922
2923         if (!parsed_valid_hugepagesz) {
2924                 pr_warn("hugepages = %s preceded by "
2925                         "an unsupported hugepagesz, ignoring\n", s);
2926                 parsed_valid_hugepagesz = true;
2927                 return 1;
2928         }
2929         /*
2930          * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2931          * so this hugepages= parameter goes to the "default hstate".
2932          */
2933         else if (!hugetlb_max_hstate)
2934                 mhp = &default_hstate_max_huge_pages;
2935         else
2936                 mhp = &parsed_hstate->max_huge_pages;
2937
2938         if (mhp == last_mhp) {
2939                 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
2940                 return 1;
2941         }
2942
2943         if (sscanf(s, "%lu", mhp) <= 0)
2944                 *mhp = 0;
2945
2946         /*
2947          * Global state is always initialized later in hugetlb_init.
2948          * But we need to allocate >= MAX_ORDER hstates here early to still
2949          * use the bootmem allocator.
2950          */
2951         if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2952                 hugetlb_hstate_alloc_pages(parsed_hstate);
2953
2954         last_mhp = mhp;
2955
2956         return 1;
2957 }
2958 __setup("hugepages=", hugetlb_nrpages_setup);
2959
2960 static int __init hugetlb_default_setup(char *s)
2961 {
2962         default_hstate_size = memparse(s, &s);
2963         return 1;
2964 }
2965 __setup("default_hugepagesz=", hugetlb_default_setup);
2966
2967 static unsigned int cpuset_mems_nr(unsigned int *array)
2968 {
2969         int node;
2970         unsigned int nr = 0;
2971
2972         for_each_node_mask(node, cpuset_current_mems_allowed)
2973                 nr += array[node];
2974
2975         return nr;
2976 }
2977
2978 #ifdef CONFIG_SYSCTL
2979 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2980                          struct ctl_table *table, int write,
2981                          void __user *buffer, size_t *length, loff_t *ppos)
2982 {
2983         struct hstate *h = &default_hstate;
2984         unsigned long tmp = h->max_huge_pages;
2985         int ret;
2986
2987         if (!hugepages_supported())
2988                 return -EOPNOTSUPP;
2989
2990         table->data = &tmp;
2991         table->maxlen = sizeof(unsigned long);
2992         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2993         if (ret)
2994                 goto out;
2995
2996         if (write)
2997                 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2998                                                   NUMA_NO_NODE, tmp, *length);
2999 out:
3000         return ret;
3001 }
3002
3003 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
3004                           void __user *buffer, size_t *length, loff_t *ppos)
3005 {
3006
3007         return hugetlb_sysctl_handler_common(false, table, write,
3008                                                         buffer, length, ppos);
3009 }
3010
3011 #ifdef CONFIG_NUMA
3012 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
3013                           void __user *buffer, size_t *length, loff_t *ppos)
3014 {
3015         return hugetlb_sysctl_handler_common(true, table, write,
3016                                                         buffer, length, ppos);
3017 }
3018 #endif /* CONFIG_NUMA */
3019
3020 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
3021                         void __user *buffer,
3022                         size_t *length, loff_t *ppos)
3023 {
3024         struct hstate *h = &default_hstate;
3025         unsigned long tmp;
3026         int ret;
3027
3028         if (!hugepages_supported())
3029                 return -EOPNOTSUPP;
3030
3031         tmp = h->nr_overcommit_huge_pages;
3032
3033         if (write && hstate_is_gigantic(h))
3034                 return -EINVAL;
3035
3036         table->data = &tmp;
3037         table->maxlen = sizeof(unsigned long);
3038         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
3039         if (ret)
3040                 goto out;
3041
3042         if (write) {
3043                 spin_lock(&hugetlb_lock);
3044                 h->nr_overcommit_huge_pages = tmp;
3045                 spin_unlock(&hugetlb_lock);
3046         }
3047 out:
3048         return ret;
3049 }
3050
3051 #endif /* CONFIG_SYSCTL */
3052
3053 void hugetlb_report_meminfo(struct seq_file *m)
3054 {
3055         struct hstate *h = &default_hstate;
3056         if (!hugepages_supported())
3057                 return;
3058         seq_printf(m,
3059                         "HugePages_Total:   %5lu\n"
3060                         "HugePages_Free:    %5lu\n"
3061                         "HugePages_Rsvd:    %5lu\n"
3062                         "HugePages_Surp:    %5lu\n"
3063                         "Hugepagesize:   %8lu kB\n",
3064                         h->nr_huge_pages,
3065                         h->free_huge_pages,
3066                         h->resv_huge_pages,
3067                         h->surplus_huge_pages,
3068                         1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3069 }
3070
3071 int hugetlb_report_node_meminfo(int nid, char *buf)
3072 {
3073         struct hstate *h = &default_hstate;
3074         if (!hugepages_supported())
3075                 return 0;
3076         return sprintf(buf,
3077                 "Node %d HugePages_Total: %5u\n"
3078                 "Node %d HugePages_Free:  %5u\n"
3079                 "Node %d HugePages_Surp:  %5u\n",
3080                 nid, h->nr_huge_pages_node[nid],
3081                 nid, h->free_huge_pages_node[nid],
3082                 nid, h->surplus_huge_pages_node[nid]);
3083 }
3084
3085 void hugetlb_show_meminfo(void)
3086 {
3087         struct hstate *h;
3088         int nid;
3089
3090         if (!hugepages_supported())
3091                 return;
3092
3093         for_each_node_state(nid, N_MEMORY)
3094                 for_each_hstate(h)
3095                         pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3096                                 nid,
3097                                 h->nr_huge_pages_node[nid],
3098                                 h->free_huge_pages_node[nid],
3099                                 h->surplus_huge_pages_node[nid],
3100                                 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3101 }
3102
3103 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3104 {
3105         seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3106                    atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3107 }
3108
3109 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3110 unsigned long hugetlb_total_pages(void)
3111 {
3112         struct hstate *h;
3113         unsigned long nr_total_pages = 0;
3114
3115         for_each_hstate(h)
3116                 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3117         return nr_total_pages;
3118 }
3119
3120 static int hugetlb_acct_memory(struct hstate *h, long delta)
3121 {
3122         int ret = -ENOMEM;
3123
3124         spin_lock(&hugetlb_lock);
3125         /*
3126          * When cpuset is configured, it breaks the strict hugetlb page
3127          * reservation as the accounting is done on a global variable. Such
3128          * reservation is completely rubbish in the presence of cpuset because
3129          * the reservation is not checked against page availability for the
3130          * current cpuset. Application can still potentially OOM'ed by kernel
3131          * with lack of free htlb page in cpuset that the task is in.
3132          * Attempt to enforce strict accounting with cpuset is almost
3133          * impossible (or too ugly) because cpuset is too fluid that
3134          * task or memory node can be dynamically moved between cpusets.
3135          *
3136          * The change of semantics for shared hugetlb mapping with cpuset is
3137          * undesirable. However, in order to preserve some of the semantics,
3138          * we fall back to check against current free page availability as
3139          * a best attempt and hopefully to minimize the impact of changing
3140          * semantics that cpuset has.
3141          */
3142         if (delta > 0) {
3143                 if (gather_surplus_pages(h, delta) < 0)
3144                         goto out;
3145
3146                 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
3147                         return_unused_surplus_pages(h, delta);
3148                         goto out;
3149                 }
3150         }
3151
3152         ret = 0;
3153         if (delta < 0)
3154                 return_unused_surplus_pages(h, (unsigned long) -delta);
3155
3156 out:
3157         spin_unlock(&hugetlb_lock);
3158         return ret;
3159 }
3160
3161 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3162 {
3163         struct resv_map *resv = vma_resv_map(vma);
3164
3165         /*
3166          * This new VMA should share its siblings reservation map if present.
3167          * The VMA will only ever have a valid reservation map pointer where
3168          * it is being copied for another still existing VMA.  As that VMA
3169          * has a reference to the reservation map it cannot disappear until
3170          * after this open call completes.  It is therefore safe to take a
3171          * new reference here without additional locking.
3172          */
3173         if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3174                 kref_get(&resv->refs);
3175 }
3176
3177 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3178 {
3179         struct hstate *h = hstate_vma(vma);
3180         struct resv_map *resv = vma_resv_map(vma);
3181         struct hugepage_subpool *spool = subpool_vma(vma);
3182         unsigned long reserve, start, end;
3183         long gbl_reserve;
3184
3185         if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3186                 return;
3187
3188         start = vma_hugecache_offset(h, vma, vma->vm_start);
3189         end = vma_hugecache_offset(h, vma, vma->vm_end);
3190
3191         reserve = (end - start) - region_count(resv, start, end);
3192
3193         kref_put(&resv->refs, resv_map_release);
3194
3195         if (reserve) {
3196                 /*
3197                  * Decrement reserve counts.  The global reserve count may be
3198                  * adjusted if the subpool has a minimum size.
3199                  */
3200                 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3201                 hugetlb_acct_memory(h, -gbl_reserve);
3202         }
3203 }
3204
3205 /*
3206  * We cannot handle pagefaults against hugetlb pages at all.  They cause
3207  * handle_mm_fault() to try to instantiate regular-sized pages in the
3208  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
3209  * this far.
3210  */
3211 static int hugetlb_vm_op_fault(struct vm_fault *vmf)
3212 {
3213         BUG();
3214         return 0;
3215 }
3216
3217 const struct vm_operations_struct hugetlb_vm_ops = {
3218         .fault = hugetlb_vm_op_fault,
3219         .open = hugetlb_vm_op_open,
3220         .close = hugetlb_vm_op_close,
3221 };
3222
3223 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3224                                 int writable)
3225 {
3226         pte_t entry;
3227
3228         if (writable) {
3229                 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3230                                          vma->vm_page_prot)));
3231         } else {
3232                 entry = huge_pte_wrprotect(mk_huge_pte(page,
3233                                            vma->vm_page_prot));
3234         }
3235         entry = pte_mkyoung(entry);
3236         entry = pte_mkhuge(entry);
3237         entry = arch_make_huge_pte(entry, vma, page, writable);
3238
3239         return entry;
3240 }
3241
3242 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3243                                    unsigned long address, pte_t *ptep)
3244 {
3245         pte_t entry;
3246
3247         entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3248         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3249                 update_mmu_cache(vma, address, ptep);
3250 }
3251
3252 bool is_hugetlb_entry_migration(pte_t pte)
3253 {
3254         swp_entry_t swp;
3255
3256         if (huge_pte_none(pte) || pte_present(pte))
3257                 return false;
3258         swp = pte_to_swp_entry(pte);
3259         if (non_swap_entry(swp) && is_migration_entry(swp))
3260                 return true;
3261         else
3262                 return false;
3263 }
3264
3265 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3266 {
3267         swp_entry_t swp;
3268
3269         if (huge_pte_none(pte) || pte_present(pte))
3270                 return 0;
3271         swp = pte_to_swp_entry(pte);
3272         if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3273                 return 1;
3274         else
3275                 return 0;
3276 }
3277
3278 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3279                             struct vm_area_struct *vma)
3280 {
3281         pte_t *src_pte, *dst_pte, entry;
3282         struct page *ptepage;
3283         unsigned long addr;
3284         int cow;
3285         struct hstate *h = hstate_vma(vma);
3286         unsigned long sz = huge_page_size(h);
3287         unsigned long mmun_start;       /* For mmu_notifiers */
3288         unsigned long mmun_end;         /* For mmu_notifiers */
3289         int ret = 0;
3290
3291         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3292
3293         mmun_start = vma->vm_start;
3294         mmun_end = vma->vm_end;
3295         if (cow)
3296                 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
3297
3298         for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3299                 spinlock_t *src_ptl, *dst_ptl;
3300                 src_pte = huge_pte_offset(src, addr, sz);
3301                 if (!src_pte)
3302                         continue;
3303                 dst_pte = huge_pte_alloc(dst, addr, sz);
3304                 if (!dst_pte) {
3305                         ret = -ENOMEM;
3306                         break;
3307                 }
3308
3309                 /* If the pagetables are shared don't copy or take references */
3310                 if (dst_pte == src_pte)
3311                         continue;
3312
3313                 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3314                 src_ptl = huge_pte_lockptr(h, src, src_pte);
3315                 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3316                 entry = huge_ptep_get(src_pte);
3317                 if (huge_pte_none(entry)) { /* skip none entry */
3318                         ;
3319                 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3320                                     is_hugetlb_entry_hwpoisoned(entry))) {
3321                         swp_entry_t swp_entry = pte_to_swp_entry(entry);
3322
3323                         if (is_write_migration_entry(swp_entry) && cow) {
3324                                 /*
3325                                  * COW mappings require pages in both
3326                                  * parent and child to be set to read.
3327                                  */
3328                                 make_migration_entry_read(&swp_entry);
3329                                 entry = swp_entry_to_pte(swp_entry);
3330                                 set_huge_swap_pte_at(src, addr, src_pte,
3331                                                      entry, sz);
3332                         }
3333                         set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3334                 } else {
3335                         if (cow) {
3336                                 huge_ptep_set_wrprotect(src, addr, src_pte);
3337                                 mmu_notifier_invalidate_range(src, mmun_start,
3338                                                                    mmun_end);
3339                         }
3340                         entry = huge_ptep_get(src_pte);
3341                         ptepage = pte_page(entry);
3342                         get_page(ptepage);
3343                         page_dup_rmap(ptepage, true);
3344                         set_huge_pte_at(dst, addr, dst_pte, entry);
3345                         hugetlb_count_add(pages_per_huge_page(h), dst);
3346                 }
3347                 spin_unlock(src_ptl);
3348                 spin_unlock(dst_ptl);
3349         }
3350
3351         if (cow)
3352                 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
3353
3354         return ret;
3355 }
3356
3357 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3358                             unsigned long start, unsigned long end,
3359                             struct page *ref_page)
3360 {
3361         struct mm_struct *mm = vma->vm_mm;
3362         unsigned long address;
3363         pte_t *ptep;
3364         pte_t pte;
3365         spinlock_t *ptl;
3366         struct page *page;
3367         struct hstate *h = hstate_vma(vma);
3368         unsigned long sz = huge_page_size(h);
3369         const unsigned long mmun_start = start; /* For mmu_notifiers */
3370         const unsigned long mmun_end   = end;   /* For mmu_notifiers */
3371
3372         WARN_ON(!is_vm_hugetlb_page(vma));
3373         BUG_ON(start & ~huge_page_mask(h));
3374         BUG_ON(end & ~huge_page_mask(h));
3375
3376         /*
3377          * This is a hugetlb vma, all the pte entries should point
3378          * to huge page.
3379          */
3380         tlb_remove_check_page_size_change(tlb, sz);
3381         tlb_start_vma(tlb, vma);
3382         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3383         address = start;
3384         for (; address < end; address += sz) {
3385                 ptep = huge_pte_offset(mm, address, sz);
3386                 if (!ptep)
3387                         continue;
3388
3389                 ptl = huge_pte_lock(h, mm, ptep);
3390                 if (huge_pmd_unshare(mm, &address, ptep)) {
3391                         spin_unlock(ptl);
3392                         continue;
3393                 }
3394
3395                 pte = huge_ptep_get(ptep);
3396                 if (huge_pte_none(pte)) {
3397                         spin_unlock(ptl);
3398                         continue;
3399                 }
3400
3401                 /*
3402                  * Migrating hugepage or HWPoisoned hugepage is already
3403                  * unmapped and its refcount is dropped, so just clear pte here.
3404                  */
3405                 if (unlikely(!pte_present(pte))) {
3406                         huge_pte_clear(mm, address, ptep, sz);
3407                         spin_unlock(ptl);
3408                         continue;
3409                 }
3410
3411                 page = pte_page(pte);
3412                 /*
3413                  * If a reference page is supplied, it is because a specific
3414                  * page is being unmapped, not a range. Ensure the page we
3415                  * are about to unmap is the actual page of interest.
3416                  */
3417                 if (ref_page) {
3418                         if (page != ref_page) {
3419                                 spin_unlock(ptl);
3420                                 continue;
3421                         }
3422                         /*
3423                          * Mark the VMA as having unmapped its page so that
3424                          * future faults in this VMA will fail rather than
3425                          * looking like data was lost
3426                          */
3427                         set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3428                 }
3429
3430                 pte = huge_ptep_get_and_clear(mm, address, ptep);
3431                 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
3432                 if (huge_pte_dirty(pte))
3433                         set_page_dirty(page);
3434
3435                 hugetlb_count_sub(pages_per_huge_page(h), mm);
3436                 page_remove_rmap(page, true);
3437
3438                 spin_unlock(ptl);
3439                 tlb_remove_page_size(tlb, page, huge_page_size(h));
3440                 /*
3441                  * Bail out after unmapping reference page if supplied
3442                  */
3443                 if (ref_page)
3444                         break;
3445         }
3446         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3447         tlb_end_vma(tlb, vma);
3448 }
3449
3450 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3451                           struct vm_area_struct *vma, unsigned long start,
3452                           unsigned long end, struct page *ref_page)
3453 {
3454         __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3455
3456         /*
3457          * Clear this flag so that x86's huge_pmd_share page_table_shareable
3458          * test will fail on a vma being torn down, and not grab a page table
3459          * on its way out.  We're lucky that the flag has such an appropriate
3460          * name, and can in fact be safely cleared here. We could clear it
3461          * before the __unmap_hugepage_range above, but all that's necessary
3462          * is to clear it before releasing the i_mmap_rwsem. This works
3463          * because in the context this is called, the VMA is about to be
3464          * destroyed and the i_mmap_rwsem is held.
3465          */
3466         vma->vm_flags &= ~VM_MAYSHARE;
3467 }
3468
3469 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3470                           unsigned long end, struct page *ref_page)
3471 {
3472         struct mm_struct *mm;
3473         struct mmu_gather tlb;
3474
3475         mm = vma->vm_mm;
3476
3477         tlb_gather_mmu(&tlb, mm, start, end);
3478         __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3479         tlb_finish_mmu(&tlb, start, end);
3480 }
3481
3482 /*
3483  * This is called when the original mapper is failing to COW a MAP_PRIVATE
3484  * mappping it owns the reserve page for. The intention is to unmap the page
3485  * from other VMAs and let the children be SIGKILLed if they are faulting the
3486  * same region.
3487  */
3488 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3489                               struct page *page, unsigned long address)
3490 {
3491         struct hstate *h = hstate_vma(vma);
3492         struct vm_area_struct *iter_vma;
3493         struct address_space *mapping;
3494         pgoff_t pgoff;
3495
3496         /*
3497          * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3498          * from page cache lookup which is in HPAGE_SIZE units.
3499          */
3500         address = address & huge_page_mask(h);
3501         pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3502                         vma->vm_pgoff;
3503         mapping = vma->vm_file->f_mapping;
3504
3505         /*
3506          * Take the mapping lock for the duration of the table walk. As
3507          * this mapping should be shared between all the VMAs,
3508          * __unmap_hugepage_range() is called as the lock is already held
3509          */
3510         i_mmap_lock_write(mapping);
3511         vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3512                 /* Do not unmap the current VMA */
3513                 if (iter_vma == vma)
3514                         continue;
3515
3516                 /*
3517                  * Shared VMAs have their own reserves and do not affect
3518                  * MAP_PRIVATE accounting but it is possible that a shared
3519                  * VMA is using the same page so check and skip such VMAs.
3520                  */
3521                 if (iter_vma->vm_flags & VM_MAYSHARE)
3522                         continue;
3523
3524                 /*
3525                  * Unmap the page from other VMAs without their own reserves.
3526                  * They get marked to be SIGKILLed if they fault in these
3527                  * areas. This is because a future no-page fault on this VMA
3528                  * could insert a zeroed page instead of the data existing
3529                  * from the time of fork. This would look like data corruption
3530                  */
3531                 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3532                         unmap_hugepage_range(iter_vma, address,
3533                                              address + huge_page_size(h), page);
3534         }
3535         i_mmap_unlock_write(mapping);
3536 }
3537
3538 /*
3539  * Hugetlb_cow() should be called with page lock of the original hugepage held.
3540  * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3541  * cannot race with other handlers or page migration.
3542  * Keep the pte_same checks anyway to make transition from the mutex easier.
3543  */
3544 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3545                        unsigned long address, pte_t *ptep,
3546                        struct page *pagecache_page, spinlock_t *ptl)
3547 {
3548         pte_t pte;
3549         struct hstate *h = hstate_vma(vma);
3550         struct page *old_page, *new_page;
3551         int ret = 0, outside_reserve = 0;
3552         unsigned long mmun_start;       /* For mmu_notifiers */
3553         unsigned long mmun_end;         /* For mmu_notifiers */
3554
3555         pte = huge_ptep_get(ptep);
3556         old_page = pte_page(pte);
3557
3558 retry_avoidcopy:
3559         /* If no-one else is actually using this page, avoid the copy
3560          * and just make the page writable */
3561         if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3562                 page_move_anon_rmap(old_page, vma);
3563                 set_huge_ptep_writable(vma, address, ptep);
3564                 return 0;
3565         }
3566
3567         /*
3568          * If the process that created a MAP_PRIVATE mapping is about to
3569          * perform a COW due to a shared page count, attempt to satisfy
3570          * the allocation without using the existing reserves. The pagecache
3571          * page is used to determine if the reserve at this address was
3572          * consumed or not. If reserves were used, a partial faulted mapping
3573          * at the time of fork() could consume its reserves on COW instead
3574          * of the full address range.
3575          */
3576         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3577                         old_page != pagecache_page)
3578                 outside_reserve = 1;
3579
3580         get_page(old_page);
3581
3582         /*
3583          * Drop page table lock as buddy allocator may be called. It will
3584          * be acquired again before returning to the caller, as expected.
3585          */
3586         spin_unlock(ptl);
3587         new_page = alloc_huge_page(vma, address, outside_reserve);
3588
3589         if (IS_ERR(new_page)) {
3590                 /*
3591                  * If a process owning a MAP_PRIVATE mapping fails to COW,
3592                  * it is due to references held by a child and an insufficient
3593                  * huge page pool. To guarantee the original mappers
3594                  * reliability, unmap the page from child processes. The child
3595                  * may get SIGKILLed if it later faults.
3596                  */
3597                 if (outside_reserve) {
3598                         put_page(old_page);
3599                         BUG_ON(huge_pte_none(pte));
3600                         unmap_ref_private(mm, vma, old_page, address);
3601                         BUG_ON(huge_pte_none(pte));
3602                         spin_lock(ptl);
3603                         ptep = huge_pte_offset(mm, address & huge_page_mask(h),
3604                                                huge_page_size(h));
3605                         if (likely(ptep &&
3606                                    pte_same(huge_ptep_get(ptep), pte)))
3607                                 goto retry_avoidcopy;
3608                         /*
3609                          * race occurs while re-acquiring page table
3610                          * lock, and our job is done.
3611                          */
3612                         return 0;
3613                 }
3614
3615                 ret = (PTR_ERR(new_page) == -ENOMEM) ?
3616                         VM_FAULT_OOM : VM_FAULT_SIGBUS;
3617                 goto out_release_old;
3618         }
3619
3620         /*
3621          * When the original hugepage is shared one, it does not have
3622          * anon_vma prepared.
3623          */
3624         if (unlikely(anon_vma_prepare(vma))) {
3625                 ret = VM_FAULT_OOM;
3626                 goto out_release_all;
3627         }
3628
3629         copy_user_huge_page(new_page, old_page, address, vma,
3630                             pages_per_huge_page(h));
3631         __SetPageUptodate(new_page);
3632         set_page_huge_active(new_page);
3633
3634         mmun_start = address & huge_page_mask(h);
3635         mmun_end = mmun_start + huge_page_size(h);
3636         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3637
3638         /*
3639          * Retake the page table lock to check for racing updates
3640          * before the page tables are altered
3641          */
3642         spin_lock(ptl);
3643         ptep = huge_pte_offset(mm, address & huge_page_mask(h),
3644                                huge_page_size(h));
3645         if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3646                 ClearPagePrivate(new_page);
3647
3648                 /* Break COW */
3649                 huge_ptep_clear_flush(vma, address, ptep);
3650                 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
3651                 set_huge_pte_at(mm, address, ptep,
3652                                 make_huge_pte(vma, new_page, 1));
3653                 page_remove_rmap(old_page, true);
3654                 hugepage_add_new_anon_rmap(new_page, vma, address);
3655                 /* Make the old page be freed below */
3656                 new_page = old_page;
3657         }
3658         spin_unlock(ptl);
3659         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3660 out_release_all:
3661         restore_reserve_on_error(h, vma, address, new_page);
3662         put_page(new_page);
3663 out_release_old:
3664         put_page(old_page);
3665
3666         spin_lock(ptl); /* Caller expects lock to be held */
3667         return ret;
3668 }
3669
3670 /* Return the pagecache page at a given address within a VMA */
3671 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3672                         struct vm_area_struct *vma, unsigned long address)
3673 {
3674         struct address_space *mapping;
3675         pgoff_t idx;
3676
3677         mapping = vma->vm_file->f_mapping;
3678         idx = vma_hugecache_offset(h, vma, address);
3679
3680         return find_lock_page(mapping, idx);
3681 }
3682
3683 /*
3684  * Return whether there is a pagecache page to back given address within VMA.
3685  * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3686  */
3687 static bool hugetlbfs_pagecache_present(struct hstate *h,
3688                         struct vm_area_struct *vma, unsigned long address)
3689 {
3690         struct address_space *mapping;
3691         pgoff_t idx;
3692         struct page *page;
3693
3694         mapping = vma->vm_file->f_mapping;
3695         idx = vma_hugecache_offset(h, vma, address);
3696
3697         page = find_get_page(mapping, idx);
3698         if (page)
3699                 put_page(page);
3700         return page != NULL;
3701 }
3702
3703 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3704                            pgoff_t idx)
3705 {
3706         struct inode *inode = mapping->host;
3707         struct hstate *h = hstate_inode(inode);
3708         int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3709
3710         if (err)
3711                 return err;
3712         ClearPagePrivate(page);
3713
3714         spin_lock(&inode->i_lock);
3715         inode->i_blocks += blocks_per_huge_page(h);
3716         spin_unlock(&inode->i_lock);
3717         return 0;
3718 }
3719
3720 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
3721                            struct address_space *mapping, pgoff_t idx,
3722                            unsigned long address, pte_t *ptep, unsigned int flags)
3723 {
3724         struct hstate *h = hstate_vma(vma);
3725         int ret = VM_FAULT_SIGBUS;
3726         int anon_rmap = 0;
3727         unsigned long size;
3728         struct page *page;
3729         pte_t new_pte;
3730         spinlock_t *ptl;
3731
3732         /*
3733          * Currently, we are forced to kill the process in the event the
3734          * original mapper has unmapped pages from the child due to a failed
3735          * COW. Warn that such a situation has occurred as it may not be obvious
3736          */
3737         if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3738                 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
3739                            current->pid);
3740                 return ret;
3741         }
3742
3743         /*
3744          * Use page lock to guard against racing truncation
3745          * before we get page_table_lock.
3746          */
3747 retry:
3748         page = find_lock_page(mapping, idx);
3749         if (!page) {
3750                 size = i_size_read(mapping->host) >> huge_page_shift(h);
3751                 if (idx >= size)
3752                         goto out;
3753
3754                 /*
3755                  * Check for page in userfault range
3756                  */
3757                 if (userfaultfd_missing(vma)) {
3758                         u32 hash;
3759                         struct vm_fault vmf = {
3760                                 .vma = vma,
3761                                 .address = address,
3762                                 .flags = flags,
3763                                 /*
3764                                  * Hard to debug if it ends up being
3765                                  * used by a callee that assumes
3766                                  * something about the other
3767                                  * uninitialized fields... same as in
3768                                  * memory.c
3769                                  */
3770                         };
3771
3772                         /*
3773                          * hugetlb_fault_mutex must be dropped before
3774                          * handling userfault.  Reacquire after handling
3775                          * fault to make calling code simpler.
3776                          */
3777                         hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping,
3778                                                         idx, address);
3779                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3780                         ret = handle_userfault(&vmf, VM_UFFD_MISSING);
3781                         mutex_lock(&hugetlb_fault_mutex_table[hash]);
3782                         goto out;
3783                 }
3784
3785                 page = alloc_huge_page(vma, address, 0);
3786                 if (IS_ERR(page)) {
3787                         ret = PTR_ERR(page);
3788                         if (ret == -ENOMEM)
3789                                 ret = VM_FAULT_OOM;
3790                         else
3791                                 ret = VM_FAULT_SIGBUS;
3792                         goto out;
3793                 }
3794                 clear_huge_page(page, address, pages_per_huge_page(h));
3795                 __SetPageUptodate(page);
3796                 set_page_huge_active(page);
3797
3798                 if (vma->vm_flags & VM_MAYSHARE) {
3799                         int err = huge_add_to_page_cache(page, mapping, idx);
3800                         if (err) {
3801                                 put_page(page);
3802                                 if (err == -EEXIST)
3803                                         goto retry;
3804                                 goto out;
3805                         }
3806                 } else {
3807                         lock_page(page);
3808                         if (unlikely(anon_vma_prepare(vma))) {
3809                                 ret = VM_FAULT_OOM;
3810                                 goto backout_unlocked;
3811                         }
3812                         anon_rmap = 1;
3813                 }
3814         } else {
3815                 /*
3816                  * If memory error occurs between mmap() and fault, some process
3817                  * don't have hwpoisoned swap entry for errored virtual address.
3818                  * So we need to block hugepage fault by PG_hwpoison bit check.
3819                  */
3820                 if (unlikely(PageHWPoison(page))) {
3821                         ret = VM_FAULT_HWPOISON |
3822                                 VM_FAULT_SET_HINDEX(hstate_index(h));
3823                         goto backout_unlocked;
3824                 }
3825         }
3826
3827         /*
3828          * If we are going to COW a private mapping later, we examine the
3829          * pending reservations for this page now. This will ensure that
3830          * any allocations necessary to record that reservation occur outside
3831          * the spinlock.
3832          */
3833         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3834                 if (vma_needs_reservation(h, vma, address) < 0) {
3835                         ret = VM_FAULT_OOM;
3836                         goto backout_unlocked;
3837                 }
3838                 /* Just decrements count, does not deallocate */
3839                 vma_end_reservation(h, vma, address);
3840         }
3841
3842         ptl = huge_pte_lock(h, mm, ptep);
3843         size = i_size_read(mapping->host) >> huge_page_shift(h);
3844         if (idx >= size)
3845                 goto backout;
3846
3847         ret = 0;
3848         if (!huge_pte_none(huge_ptep_get(ptep)))
3849                 goto backout;
3850
3851         if (anon_rmap) {
3852                 ClearPagePrivate(page);
3853                 hugepage_add_new_anon_rmap(page, vma, address);
3854         } else
3855                 page_dup_rmap(page, true);
3856         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3857                                 && (vma->vm_flags & VM_SHARED)));
3858         set_huge_pte_at(mm, address, ptep, new_pte);
3859
3860         hugetlb_count_add(pages_per_huge_page(h), mm);
3861         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3862                 /* Optimization, do the COW without a second fault */
3863                 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
3864         }
3865
3866         spin_unlock(ptl);
3867         unlock_page(page);
3868 out:
3869         return ret;
3870
3871 backout:
3872         spin_unlock(ptl);