~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/mm/hugetlb.c

Version: ~ [ linux-5.13-rc5 ] ~ [ linux-5.12.9 ] ~ [ linux-5.11.22 ] ~ [ linux-5.10.42 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.124 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.193 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.235 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.271 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.271 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.18.140 ] ~ [ linux-3.16.85 ] ~ [ linux-3.14.79 ] ~ [ linux-3.12.74 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 /*
  2  * Generic hugetlb support.
  3  * (C) Nadia Yvette Chambers, April 2004
  4  */
  5 #include <linux/list.h>
  6 #include <linux/init.h>
  7 #include <linux/mm.h>
  8 #include <linux/seq_file.h>
  9 #include <linux/sysctl.h>
 10 #include <linux/highmem.h>
 11 #include <linux/mmu_notifier.h>
 12 #include <linux/nodemask.h>
 13 #include <linux/pagemap.h>
 14 #include <linux/mempolicy.h>
 15 #include <linux/compiler.h>
 16 #include <linux/cpuset.h>
 17 #include <linux/mutex.h>
 18 #include <linux/bootmem.h>
 19 #include <linux/sysfs.h>
 20 #include <linux/slab.h>
 21 #include <linux/mmdebug.h>
 22 #include <linux/sched/signal.h>
 23 #include <linux/rmap.h>
 24 #include <linux/string_helpers.h>
 25 #include <linux/swap.h>
 26 #include <linux/swapops.h>
 27 #include <linux/jhash.h>
 28 
 29 #include <asm/page.h>
 30 #include <asm/pgtable.h>
 31 #include <asm/tlb.h>
 32 
 33 #include <linux/io.h>
 34 #include <linux/hugetlb.h>
 35 #include <linux/hugetlb_cgroup.h>
 36 #include <linux/node.h>
 37 #include <linux/userfaultfd_k.h>
 38 #include "internal.h"
 39 
 40 int hugepages_treat_as_movable;
 41 
 42 int hugetlb_max_hstate __read_mostly;
 43 unsigned int default_hstate_idx;
 44 struct hstate hstates[HUGE_MAX_HSTATE];
 45 /*
 46  * Minimum page order among possible hugepage sizes, set to a proper value
 47  * at boot time.
 48  */
 49 static unsigned int minimum_order __read_mostly = UINT_MAX;
 50 
 51 __initdata LIST_HEAD(huge_boot_pages);
 52 
 53 /* for command line parsing */
 54 static struct hstate * __initdata parsed_hstate;
 55 static unsigned long __initdata default_hstate_max_huge_pages;
 56 static unsigned long __initdata default_hstate_size;
 57 static bool __initdata parsed_valid_hugepagesz = true;
 58 
 59 /*
 60  * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
 61  * free_huge_pages, and surplus_huge_pages.
 62  */
 63 DEFINE_SPINLOCK(hugetlb_lock);
 64 
 65 /*
 66  * Serializes faults on the same logical page.  This is used to
 67  * prevent spurious OOMs when the hugepage pool is fully utilized.
 68  */
 69 static int num_fault_mutexes;
 70 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
 71 
 72 /* Forward declaration */
 73 static int hugetlb_acct_memory(struct hstate *h, long delta);
 74 
 75 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
 76 {
 77         bool free = (spool->count == 0) && (spool->used_hpages == 0);
 78 
 79         spin_unlock(&spool->lock);
 80 
 81         /* If no pages are used, and no other handles to the subpool
 82          * remain, give up any reservations mased on minimum size and
 83          * free the subpool */
 84         if (free) {
 85                 if (spool->min_hpages != -1)
 86                         hugetlb_acct_memory(spool->hstate,
 87                                                 -spool->min_hpages);
 88                 kfree(spool);
 89         }
 90 }
 91 
 92 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
 93                                                 long min_hpages)
 94 {
 95         struct hugepage_subpool *spool;
 96 
 97         spool = kzalloc(sizeof(*spool), GFP_KERNEL);
 98         if (!spool)
 99                 return NULL;
100 
101         spin_lock_init(&spool->lock);
102         spool->count = 1;
103         spool->max_hpages = max_hpages;
104         spool->hstate = h;
105         spool->min_hpages = min_hpages;
106 
107         if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
108                 kfree(spool);
109                 return NULL;
110         }
111         spool->rsv_hpages = min_hpages;
112 
113         return spool;
114 }
115 
116 void hugepage_put_subpool(struct hugepage_subpool *spool)
117 {
118         spin_lock(&spool->lock);
119         BUG_ON(!spool->count);
120         spool->count--;
121         unlock_or_release_subpool(spool);
122 }
123 
124 /*
125  * Subpool accounting for allocating and reserving pages.
126  * Return -ENOMEM if there are not enough resources to satisfy the
127  * the request.  Otherwise, return the number of pages by which the
128  * global pools must be adjusted (upward).  The returned value may
129  * only be different than the passed value (delta) in the case where
130  * a subpool minimum size must be manitained.
131  */
132 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
133                                       long delta)
134 {
135         long ret = delta;
136 
137         if (!spool)
138                 return ret;
139 
140         spin_lock(&spool->lock);
141 
142         if (spool->max_hpages != -1) {          /* maximum size accounting */
143                 if ((spool->used_hpages + delta) <= spool->max_hpages)
144                         spool->used_hpages += delta;
145                 else {
146                         ret = -ENOMEM;
147                         goto unlock_ret;
148                 }
149         }
150 
151         /* minimum size accounting */
152         if (spool->min_hpages != -1 && spool->rsv_hpages) {
153                 if (delta > spool->rsv_hpages) {
154                         /*
155                          * Asking for more reserves than those already taken on
156                          * behalf of subpool.  Return difference.
157                          */
158                         ret = delta - spool->rsv_hpages;
159                         spool->rsv_hpages = 0;
160                 } else {
161                         ret = 0;        /* reserves already accounted for */
162                         spool->rsv_hpages -= delta;
163                 }
164         }
165 
166 unlock_ret:
167         spin_unlock(&spool->lock);
168         return ret;
169 }
170 
171 /*
172  * Subpool accounting for freeing and unreserving pages.
173  * Return the number of global page reservations that must be dropped.
174  * The return value may only be different than the passed value (delta)
175  * in the case where a subpool minimum size must be maintained.
176  */
177 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
178                                        long delta)
179 {
180         long ret = delta;
181 
182         if (!spool)
183                 return delta;
184 
185         spin_lock(&spool->lock);
186 
187         if (spool->max_hpages != -1)            /* maximum size accounting */
188                 spool->used_hpages -= delta;
189 
190          /* minimum size accounting */
191         if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
192                 if (spool->rsv_hpages + delta <= spool->min_hpages)
193                         ret = 0;
194                 else
195                         ret = spool->rsv_hpages + delta - spool->min_hpages;
196 
197                 spool->rsv_hpages += delta;
198                 if (spool->rsv_hpages > spool->min_hpages)
199                         spool->rsv_hpages = spool->min_hpages;
200         }
201 
202         /*
203          * If hugetlbfs_put_super couldn't free spool due to an outstanding
204          * quota reference, free it now.
205          */
206         unlock_or_release_subpool(spool);
207 
208         return ret;
209 }
210 
211 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
212 {
213         return HUGETLBFS_SB(inode->i_sb)->spool;
214 }
215 
216 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
217 {
218         return subpool_inode(file_inode(vma->vm_file));
219 }
220 
221 /*
222  * Region tracking -- allows tracking of reservations and instantiated pages
223  *                    across the pages in a mapping.
224  *
225  * The region data structures are embedded into a resv_map and protected
226  * by a resv_map's lock.  The set of regions within the resv_map represent
227  * reservations for huge pages, or huge pages that have already been
228  * instantiated within the map.  The from and to elements are huge page
229  * indicies into the associated mapping.  from indicates the starting index
230  * of the region.  to represents the first index past the end of  the region.
231  *
232  * For example, a file region structure with from == 0 and to == 4 represents
233  * four huge pages in a mapping.  It is important to note that the to element
234  * represents the first element past the end of the region. This is used in
235  * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
236  *
237  * Interval notation of the form [from, to) will be used to indicate that
238  * the endpoint from is inclusive and to is exclusive.
239  */
240 struct file_region {
241         struct list_head link;
242         long from;
243         long to;
244 };
245 
246 /*
247  * Add the huge page range represented by [f, t) to the reserve
248  * map.  In the normal case, existing regions will be expanded
249  * to accommodate the specified range.  Sufficient regions should
250  * exist for expansion due to the previous call to region_chg
251  * with the same range.  However, it is possible that region_del
252  * could have been called after region_chg and modifed the map
253  * in such a way that no region exists to be expanded.  In this
254  * case, pull a region descriptor from the cache associated with
255  * the map and use that for the new range.
256  *
257  * Return the number of new huge pages added to the map.  This
258  * number is greater than or equal to zero.
259  */
260 static long region_add(struct resv_map *resv, long f, long t)
261 {
262         struct list_head *head = &resv->regions;
263         struct file_region *rg, *nrg, *trg;
264         long add = 0;
265 
266         spin_lock(&resv->lock);
267         /* Locate the region we are either in or before. */
268         list_for_each_entry(rg, head, link)
269                 if (f <= rg->to)
270                         break;
271 
272         /*
273          * If no region exists which can be expanded to include the
274          * specified range, the list must have been modified by an
275          * interleving call to region_del().  Pull a region descriptor
276          * from the cache and use it for this range.
277          */
278         if (&rg->link == head || t < rg->from) {
279                 VM_BUG_ON(resv->region_cache_count <= 0);
280 
281                 resv->region_cache_count--;
282                 nrg = list_first_entry(&resv->region_cache, struct file_region,
283                                         link);
284                 list_del(&nrg->link);
285 
286                 nrg->from = f;
287                 nrg->to = t;
288                 list_add(&nrg->link, rg->link.prev);
289 
290                 add += t - f;
291                 goto out_locked;
292         }
293 
294         /* Round our left edge to the current segment if it encloses us. */
295         if (f > rg->from)
296                 f = rg->from;
297 
298         /* Check for and consume any regions we now overlap with. */
299         nrg = rg;
300         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
301                 if (&rg->link == head)
302                         break;
303                 if (rg->from > t)
304                         break;
305 
306                 /* If this area reaches higher then extend our area to
307                  * include it completely.  If this is not the first area
308                  * which we intend to reuse, free it. */
309                 if (rg->to > t)
310                         t = rg->to;
311                 if (rg != nrg) {
312                         /* Decrement return value by the deleted range.
313                          * Another range will span this area so that by
314                          * end of routine add will be >= zero
315                          */
316                         add -= (rg->to - rg->from);
317                         list_del(&rg->link);
318                         kfree(rg);
319                 }
320         }
321 
322         add += (nrg->from - f);         /* Added to beginning of region */
323         nrg->from = f;
324         add += t - nrg->to;             /* Added to end of region */
325         nrg->to = t;
326 
327 out_locked:
328         resv->adds_in_progress--;
329         spin_unlock(&resv->lock);
330         VM_BUG_ON(add < 0);
331         return add;
332 }
333 
334 /*
335  * Examine the existing reserve map and determine how many
336  * huge pages in the specified range [f, t) are NOT currently
337  * represented.  This routine is called before a subsequent
338  * call to region_add that will actually modify the reserve
339  * map to add the specified range [f, t).  region_chg does
340  * not change the number of huge pages represented by the
341  * map.  However, if the existing regions in the map can not
342  * be expanded to represent the new range, a new file_region
343  * structure is added to the map as a placeholder.  This is
344  * so that the subsequent region_add call will have all the
345  * regions it needs and will not fail.
346  *
347  * Upon entry, region_chg will also examine the cache of region descriptors
348  * associated with the map.  If there are not enough descriptors cached, one
349  * will be allocated for the in progress add operation.
350  *
351  * Returns the number of huge pages that need to be added to the existing
352  * reservation map for the range [f, t).  This number is greater or equal to
353  * zero.  -ENOMEM is returned if a new file_region structure or cache entry
354  * is needed and can not be allocated.
355  */
356 static long region_chg(struct resv_map *resv, long f, long t)
357 {
358         struct list_head *head = &resv->regions;
359         struct file_region *rg, *nrg = NULL;
360         long chg = 0;
361 
362 retry:
363         spin_lock(&resv->lock);
364 retry_locked:
365         resv->adds_in_progress++;
366 
367         /*
368          * Check for sufficient descriptors in the cache to accommodate
369          * the number of in progress add operations.
370          */
371         if (resv->adds_in_progress > resv->region_cache_count) {
372                 struct file_region *trg;
373 
374                 VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1);
375                 /* Must drop lock to allocate a new descriptor. */
376                 resv->adds_in_progress--;
377                 spin_unlock(&resv->lock);
378 
379                 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
380                 if (!trg) {
381                         kfree(nrg);
382                         return -ENOMEM;
383                 }
384 
385                 spin_lock(&resv->lock);
386                 list_add(&trg->link, &resv->region_cache);
387                 resv->region_cache_count++;
388                 goto retry_locked;
389         }
390 
391         /* Locate the region we are before or in. */
392         list_for_each_entry(rg, head, link)
393                 if (f <= rg->to)
394                         break;
395 
396         /* If we are below the current region then a new region is required.
397          * Subtle, allocate a new region at the position but make it zero
398          * size such that we can guarantee to record the reservation. */
399         if (&rg->link == head || t < rg->from) {
400                 if (!nrg) {
401                         resv->adds_in_progress--;
402                         spin_unlock(&resv->lock);
403                         nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
404                         if (!nrg)
405                                 return -ENOMEM;
406 
407                         nrg->from = f;
408                         nrg->to   = f;
409                         INIT_LIST_HEAD(&nrg->link);
410                         goto retry;
411                 }
412 
413                 list_add(&nrg->link, rg->link.prev);
414                 chg = t - f;
415                 goto out_nrg;
416         }
417 
418         /* Round our left edge to the current segment if it encloses us. */
419         if (f > rg->from)
420                 f = rg->from;
421         chg = t - f;
422 
423         /* Check for and consume any regions we now overlap with. */
424         list_for_each_entry(rg, rg->link.prev, link) {
425                 if (&rg->link == head)
426                         break;
427                 if (rg->from > t)
428                         goto out;
429 
430                 /* We overlap with this area, if it extends further than
431                  * us then we must extend ourselves.  Account for its
432                  * existing reservation. */
433                 if (rg->to > t) {
434                         chg += rg->to - t;
435                         t = rg->to;
436                 }
437                 chg -= rg->to - rg->from;
438         }
439 
440 out:
441         spin_unlock(&resv->lock);
442         /*  We already know we raced and no longer need the new region */
443         kfree(nrg);
444         return chg;
445 out_nrg:
446         spin_unlock(&resv->lock);
447         return chg;
448 }
449 
450 /*
451  * Abort the in progress add operation.  The adds_in_progress field
452  * of the resv_map keeps track of the operations in progress between
453  * calls to region_chg and region_add.  Operations are sometimes
454  * aborted after the call to region_chg.  In such cases, region_abort
455  * is called to decrement the adds_in_progress counter.
456  *
457  * NOTE: The range arguments [f, t) are not needed or used in this
458  * routine.  They are kept to make reading the calling code easier as
459  * arguments will match the associated region_chg call.
460  */
461 static void region_abort(struct resv_map *resv, long f, long t)
462 {
463         spin_lock(&resv->lock);
464         VM_BUG_ON(!resv->region_cache_count);
465         resv->adds_in_progress--;
466         spin_unlock(&resv->lock);
467 }
468 
469 /*
470  * Delete the specified range [f, t) from the reserve map.  If the
471  * t parameter is LONG_MAX, this indicates that ALL regions after f
472  * should be deleted.  Locate the regions which intersect [f, t)
473  * and either trim, delete or split the existing regions.
474  *
475  * Returns the number of huge pages deleted from the reserve map.
476  * In the normal case, the return value is zero or more.  In the
477  * case where a region must be split, a new region descriptor must
478  * be allocated.  If the allocation fails, -ENOMEM will be returned.
479  * NOTE: If the parameter t == LONG_MAX, then we will never split
480  * a region and possibly return -ENOMEM.  Callers specifying
481  * t == LONG_MAX do not need to check for -ENOMEM error.
482  */
483 static long region_del(struct resv_map *resv, long f, long t)
484 {
485         struct list_head *head = &resv->regions;
486         struct file_region *rg, *trg;
487         struct file_region *nrg = NULL;
488         long del = 0;
489 
490 retry:
491         spin_lock(&resv->lock);
492         list_for_each_entry_safe(rg, trg, head, link) {
493                 /*
494                  * Skip regions before the range to be deleted.  file_region
495                  * ranges are normally of the form [from, to).  However, there
496                  * may be a "placeholder" entry in the map which is of the form
497                  * (from, to) with from == to.  Check for placeholder entries
498                  * at the beginning of the range to be deleted.
499                  */
500                 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
501                         continue;
502 
503                 if (rg->from >= t)
504                         break;
505 
506                 if (f > rg->from && t < rg->to) { /* Must split region */
507                         /*
508                          * Check for an entry in the cache before dropping
509                          * lock and attempting allocation.
510                          */
511                         if (!nrg &&
512                             resv->region_cache_count > resv->adds_in_progress) {
513                                 nrg = list_first_entry(&resv->region_cache,
514                                                         struct file_region,
515                                                         link);
516                                 list_del(&nrg->link);
517                                 resv->region_cache_count--;
518                         }
519 
520                         if (!nrg) {
521                                 spin_unlock(&resv->lock);
522                                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
523                                 if (!nrg)
524                                         return -ENOMEM;
525                                 goto retry;
526                         }
527 
528                         del += t - f;
529 
530                         /* New entry for end of split region */
531                         nrg->from = t;
532                         nrg->to = rg->to;
533                         INIT_LIST_HEAD(&nrg->link);
534 
535                         /* Original entry is trimmed */
536                         rg->to = f;
537 
538                         list_add(&nrg->link, &rg->link);
539                         nrg = NULL;
540                         break;
541                 }
542 
543                 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
544                         del += rg->to - rg->from;
545                         list_del(&rg->link);
546                         kfree(rg);
547                         continue;
548                 }
549 
550                 if (f <= rg->from) {    /* Trim beginning of region */
551                         del += t - rg->from;
552                         rg->from = t;
553                 } else {                /* Trim end of region */
554                         del += rg->to - f;
555                         rg->to = f;
556                 }
557         }
558 
559         spin_unlock(&resv->lock);
560         kfree(nrg);
561         return del;
562 }
563 
564 /*
565  * A rare out of memory error was encountered which prevented removal of
566  * the reserve map region for a page.  The huge page itself was free'ed
567  * and removed from the page cache.  This routine will adjust the subpool
568  * usage count, and the global reserve count if needed.  By incrementing
569  * these counts, the reserve map entry which could not be deleted will
570  * appear as a "reserved" entry instead of simply dangling with incorrect
571  * counts.
572  */
573 void hugetlb_fix_reserve_counts(struct inode *inode)
574 {
575         struct hugepage_subpool *spool = subpool_inode(inode);
576         long rsv_adjust;
577 
578         rsv_adjust = hugepage_subpool_get_pages(spool, 1);
579         if (rsv_adjust) {
580                 struct hstate *h = hstate_inode(inode);
581 
582                 hugetlb_acct_memory(h, 1);
583         }
584 }
585 
586 /*
587  * Count and return the number of huge pages in the reserve map
588  * that intersect with the range [f, t).
589  */
590 static long region_count(struct resv_map *resv, long f, long t)
591 {
592         struct list_head *head = &resv->regions;
593         struct file_region *rg;
594         long chg = 0;
595 
596         spin_lock(&resv->lock);
597         /* Locate each segment we overlap with, and count that overlap. */
598         list_for_each_entry(rg, head, link) {
599                 long seg_from;
600                 long seg_to;
601 
602                 if (rg->to <= f)
603                         continue;
604                 if (rg->from >= t)
605                         break;
606 
607                 seg_from = max(rg->from, f);
608                 seg_to = min(rg->to, t);
609 
610                 chg += seg_to - seg_from;
611         }
612         spin_unlock(&resv->lock);
613 
614         return chg;
615 }
616 
617 /*
618  * Convert the address within this vma to the page offset within
619  * the mapping, in pagecache page units; huge pages here.
620  */
621 static pgoff_t vma_hugecache_offset(struct hstate *h,
622                         struct vm_area_struct *vma, unsigned long address)
623 {
624         return ((address - vma->vm_start) >> huge_page_shift(h)) +
625                         (vma->vm_pgoff >> huge_page_order(h));
626 }
627 
628 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
629                                      unsigned long address)
630 {
631         return vma_hugecache_offset(hstate_vma(vma), vma, address);
632 }
633 EXPORT_SYMBOL_GPL(linear_hugepage_index);
634 
635 /*
636  * Return the size of the pages allocated when backing a VMA. In the majority
637  * cases this will be same size as used by the page table entries.
638  */
639 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
640 {
641         struct hstate *hstate;
642 
643         if (!is_vm_hugetlb_page(vma))
644                 return PAGE_SIZE;
645 
646         hstate = hstate_vma(vma);
647 
648         return 1UL << huge_page_shift(hstate);
649 }
650 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
651 
652 /*
653  * Return the page size being used by the MMU to back a VMA. In the majority
654  * of cases, the page size used by the kernel matches the MMU size. On
655  * architectures where it differs, an architecture-specific version of this
656  * function is required.
657  */
658 #ifndef vma_mmu_pagesize
659 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
660 {
661         return vma_kernel_pagesize(vma);
662 }
663 #endif
664 
665 /*
666  * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
667  * bits of the reservation map pointer, which are always clear due to
668  * alignment.
669  */
670 #define HPAGE_RESV_OWNER    (1UL << 0)
671 #define HPAGE_RESV_UNMAPPED (1UL << 1)
672 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
673 
674 /*
675  * These helpers are used to track how many pages are reserved for
676  * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
677  * is guaranteed to have their future faults succeed.
678  *
679  * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
680  * the reserve counters are updated with the hugetlb_lock held. It is safe
681  * to reset the VMA at fork() time as it is not in use yet and there is no
682  * chance of the global counters getting corrupted as a result of the values.
683  *
684  * The private mapping reservation is represented in a subtly different
685  * manner to a shared mapping.  A shared mapping has a region map associated
686  * with the underlying file, this region map represents the backing file
687  * pages which have ever had a reservation assigned which this persists even
688  * after the page is instantiated.  A private mapping has a region map
689  * associated with the original mmap which is attached to all VMAs which
690  * reference it, this region map represents those offsets which have consumed
691  * reservation ie. where pages have been instantiated.
692  */
693 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
694 {
695         return (unsigned long)vma->vm_private_data;
696 }
697 
698 static void set_vma_private_data(struct vm_area_struct *vma,
699                                                         unsigned long value)
700 {
701         vma->vm_private_data = (void *)value;
702 }
703 
704 struct resv_map *resv_map_alloc(void)
705 {
706         struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
707         struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
708 
709         if (!resv_map || !rg) {
710                 kfree(resv_map);
711                 kfree(rg);
712                 return NULL;
713         }
714 
715         kref_init(&resv_map->refs);
716         spin_lock_init(&resv_map->lock);
717         INIT_LIST_HEAD(&resv_map->regions);
718 
719         resv_map->adds_in_progress = 0;
720 
721         INIT_LIST_HEAD(&resv_map->region_cache);
722         list_add(&rg->link, &resv_map->region_cache);
723         resv_map->region_cache_count = 1;
724 
725         return resv_map;
726 }
727 
728 void resv_map_release(struct kref *ref)
729 {
730         struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
731         struct list_head *head = &resv_map->region_cache;
732         struct file_region *rg, *trg;
733 
734         /* Clear out any active regions before we release the map. */
735         region_del(resv_map, 0, LONG_MAX);
736 
737         /* ... and any entries left in the cache */
738         list_for_each_entry_safe(rg, trg, head, link) {
739                 list_del(&rg->link);
740                 kfree(rg);
741         }
742 
743         VM_BUG_ON(resv_map->adds_in_progress);
744 
745         kfree(resv_map);
746 }
747 
748 static inline struct resv_map *inode_resv_map(struct inode *inode)
749 {
750         return inode->i_mapping->private_data;
751 }
752 
753 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
754 {
755         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
756         if (vma->vm_flags & VM_MAYSHARE) {
757                 struct address_space *mapping = vma->vm_file->f_mapping;
758                 struct inode *inode = mapping->host;
759 
760                 return inode_resv_map(inode);
761 
762         } else {
763                 return (struct resv_map *)(get_vma_private_data(vma) &
764                                                         ~HPAGE_RESV_MASK);
765         }
766 }
767 
768 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
769 {
770         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
771         VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
772 
773         set_vma_private_data(vma, (get_vma_private_data(vma) &
774                                 HPAGE_RESV_MASK) | (unsigned long)map);
775 }
776 
777 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
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) | flags);
783 }
784 
785 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
786 {
787         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
788 
789         return (get_vma_private_data(vma) & flag) != 0;
790 }
791 
792 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
793 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
794 {
795         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
796         if (!(vma->vm_flags & VM_MAYSHARE))
797                 vma->vm_private_data = (void *)0;
798 }
799 
800 /* Returns true if the VMA has associated reserve pages */
801 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
802 {
803         if (vma->vm_flags & VM_NORESERVE) {
804                 /*
805                  * This address is already reserved by other process(chg == 0),
806                  * so, we should decrement reserved count. Without decrementing,
807                  * reserve count remains after releasing inode, because this
808                  * allocated page will go into page cache and is regarded as
809                  * coming from reserved pool in releasing step.  Currently, we
810                  * don't have any other solution to deal with this situation
811                  * properly, so add work-around here.
812                  */
813                 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
814                         return true;
815                 else
816                         return false;
817         }
818 
819         /* Shared mappings always use reserves */
820         if (vma->vm_flags & VM_MAYSHARE) {
821                 /*
822                  * We know VM_NORESERVE is not set.  Therefore, there SHOULD
823                  * be a region map for all pages.  The only situation where
824                  * there is no region map is if a hole was punched via
825                  * fallocate.  In this case, there really are no reverves to
826                  * use.  This situation is indicated if chg != 0.
827                  */
828                 if (chg)
829                         return false;
830                 else
831                         return true;
832         }
833 
834         /*
835          * Only the process that called mmap() has reserves for
836          * private mappings.
837          */
838         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
839                 /*
840                  * Like the shared case above, a hole punch or truncate
841                  * could have been performed on the private mapping.
842                  * Examine the value of chg to determine if reserves
843                  * actually exist or were previously consumed.
844                  * Very Subtle - The value of chg comes from a previous
845                  * call to vma_needs_reserves().  The reserve map for
846                  * private mappings has different (opposite) semantics
847                  * than that of shared mappings.  vma_needs_reserves()
848                  * has already taken this difference in semantics into
849                  * account.  Therefore, the meaning of chg is the same
850                  * as in the shared case above.  Code could easily be
851                  * combined, but keeping it separate draws attention to
852                  * subtle differences.
853                  */
854                 if (chg)
855                         return false;
856                 else
857                         return true;
858         }
859 
860         return false;
861 }
862 
863 static void enqueue_huge_page(struct hstate *h, struct page *page)
864 {
865         int nid = page_to_nid(page);
866         list_move(&page->lru, &h->hugepage_freelists[nid]);
867         h->free_huge_pages++;
868         h->free_huge_pages_node[nid]++;
869 }
870 
871 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
872 {
873         struct page *page;
874 
875         list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
876                 if (!PageHWPoison(page))
877                         break;
878         /*
879          * if 'non-isolated free hugepage' not found on the list,
880          * the allocation fails.
881          */
882         if (&h->hugepage_freelists[nid] == &page->lru)
883                 return NULL;
884         list_move(&page->lru, &h->hugepage_activelist);
885         set_page_refcounted(page);
886         h->free_huge_pages--;
887         h->free_huge_pages_node[nid]--;
888         return page;
889 }
890 
891 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
892                 nodemask_t *nmask)
893 {
894         unsigned int cpuset_mems_cookie;
895         struct zonelist *zonelist;
896         struct zone *zone;
897         struct zoneref *z;
898         int node = -1;
899 
900         zonelist = node_zonelist(nid, gfp_mask);
901 
902 retry_cpuset:
903         cpuset_mems_cookie = read_mems_allowed_begin();
904         for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
905                 struct page *page;
906 
907                 if (!cpuset_zone_allowed(zone, gfp_mask))
908                         continue;
909                 /*
910                  * no need to ask again on the same node. Pool is node rather than
911                  * zone aware
912                  */
913                 if (zone_to_nid(zone) == node)
914                         continue;
915                 node = zone_to_nid(zone);
916 
917                 page = dequeue_huge_page_node_exact(h, node);
918                 if (page)
919                         return page;
920         }
921         if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
922                 goto retry_cpuset;
923 
924         return NULL;
925 }
926 
927 /* Movability of hugepages depends on migration support. */
928 static inline gfp_t htlb_alloc_mask(struct hstate *h)
929 {
930         if (hugepages_treat_as_movable || hugepage_migration_supported(h))
931                 return GFP_HIGHUSER_MOVABLE;
932         else
933                 return GFP_HIGHUSER;
934 }
935 
936 static struct page *dequeue_huge_page_vma(struct hstate *h,
937                                 struct vm_area_struct *vma,
938                                 unsigned long address, int avoid_reserve,
939                                 long chg)
940 {
941         struct page *page;
942         struct mempolicy *mpol;
943         gfp_t gfp_mask;
944         nodemask_t *nodemask;
945         int nid;
946 
947         /*
948          * A child process with MAP_PRIVATE mappings created by their parent
949          * have no page reserves. This check ensures that reservations are
950          * not "stolen". The child may still get SIGKILLed
951          */
952         if (!vma_has_reserves(vma, chg) &&
953                         h->free_huge_pages - h->resv_huge_pages == 0)
954                 goto err;
955 
956         /* If reserves cannot be used, ensure enough pages are in the pool */
957         if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
958                 goto err;
959 
960         gfp_mask = htlb_alloc_mask(h);
961         nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
962         page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
963         if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
964                 SetPagePrivate(page);
965                 h->resv_huge_pages--;
966         }
967 
968         mpol_cond_put(mpol);
969         return page;
970 
971 err:
972         return NULL;
973 }
974 
975 /*
976  * common helper functions for hstate_next_node_to_{alloc|free}.
977  * We may have allocated or freed a huge page based on a different
978  * nodes_allowed previously, so h->next_node_to_{alloc|free} might
979  * be outside of *nodes_allowed.  Ensure that we use an allowed
980  * node for alloc or free.
981  */
982 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
983 {
984         nid = next_node_in(nid, *nodes_allowed);
985         VM_BUG_ON(nid >= MAX_NUMNODES);
986 
987         return nid;
988 }
989 
990 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
991 {
992         if (!node_isset(nid, *nodes_allowed))
993                 nid = next_node_allowed(nid, nodes_allowed);
994         return nid;
995 }
996 
997 /*
998  * returns the previously saved node ["this node"] from which to
999  * allocate a persistent huge page for the pool and advance the
1000  * next node from which to allocate, handling wrap at end of node
1001  * mask.
1002  */
1003 static int hstate_next_node_to_alloc(struct hstate *h,
1004                                         nodemask_t *nodes_allowed)
1005 {
1006         int nid;
1007 
1008         VM_BUG_ON(!nodes_allowed);
1009 
1010         nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1011         h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1012 
1013         return nid;
1014 }
1015 
1016 /*
1017  * helper for free_pool_huge_page() - return the previously saved
1018  * node ["this node"] from which to free a huge page.  Advance the
1019  * next node id whether or not we find a free huge page to free so
1020  * that the next attempt to free addresses the next node.
1021  */
1022 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1023 {
1024         int nid;
1025 
1026         VM_BUG_ON(!nodes_allowed);
1027 
1028         nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1029         h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1030 
1031         return nid;
1032 }
1033 
1034 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask)           \
1035         for (nr_nodes = nodes_weight(*mask);                            \
1036                 nr_nodes > 0 &&                                         \
1037                 ((node = hstate_next_node_to_alloc(hs, mask)) || 1);    \
1038                 nr_nodes--)
1039 
1040 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask)            \
1041         for (nr_nodes = nodes_weight(*mask);                            \
1042                 nr_nodes > 0 &&                                         \
1043                 ((node = hstate_next_node_to_free(hs, mask)) || 1);     \
1044                 nr_nodes--)
1045 
1046 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1047 static void destroy_compound_gigantic_page(struct page *page,
1048                                         unsigned int order)
1049 {
1050         int i;
1051         int nr_pages = 1 << order;
1052         struct page *p = page + 1;
1053 
1054         atomic_set(compound_mapcount_ptr(page), 0);
1055         for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1056                 clear_compound_head(p);
1057                 set_page_refcounted(p);
1058         }
1059 
1060         set_compound_order(page, 0);
1061         __ClearPageHead(page);
1062 }
1063 
1064 static void free_gigantic_page(struct page *page, unsigned int order)
1065 {
1066         free_contig_range(page_to_pfn(page), 1 << order);
1067 }
1068 
1069 static int __alloc_gigantic_page(unsigned long start_pfn,
1070                                 unsigned long nr_pages, gfp_t gfp_mask)
1071 {
1072         unsigned long end_pfn = start_pfn + nr_pages;
1073         return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
1074                                   gfp_mask);
1075 }
1076 
1077 static bool pfn_range_valid_gigantic(struct zone *z,
1078                         unsigned long start_pfn, unsigned long nr_pages)
1079 {
1080         unsigned long i, end_pfn = start_pfn + nr_pages;
1081         struct page *page;
1082 
1083         for (i = start_pfn; i < end_pfn; i++) {
1084                 if (!pfn_valid(i))
1085                         return false;
1086 
1087                 page = pfn_to_page(i);
1088 
1089                 if (page_zone(page) != z)
1090                         return false;
1091 
1092                 if (PageReserved(page))
1093                         return false;
1094 
1095                 if (page_count(page) > 0)
1096                         return false;
1097 
1098                 if (PageHuge(page))
1099                         return false;
1100         }
1101 
1102         return true;
1103 }
1104 
1105 static bool zone_spans_last_pfn(const struct zone *zone,
1106                         unsigned long start_pfn, unsigned long nr_pages)
1107 {
1108         unsigned long last_pfn = start_pfn + nr_pages - 1;
1109         return zone_spans_pfn(zone, last_pfn);
1110 }
1111 
1112 static struct page *alloc_gigantic_page(int nid, struct hstate *h)
1113 {
1114         unsigned int order = huge_page_order(h);
1115         unsigned long nr_pages = 1 << order;
1116         unsigned long ret, pfn, flags;
1117         struct zonelist *zonelist;
1118         struct zone *zone;
1119         struct zoneref *z;
1120         gfp_t gfp_mask;
1121 
1122         gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1123         zonelist = node_zonelist(nid, gfp_mask);
1124         for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), NULL) {
1125                 spin_lock_irqsave(&zone->lock, flags);
1126 
1127                 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
1128                 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
1129                         if (pfn_range_valid_gigantic(zone, pfn, nr_pages)) {
1130                                 /*
1131                                  * We release the zone lock here because
1132                                  * alloc_contig_range() will also lock the zone
1133                                  * at some point. If there's an allocation
1134                                  * spinning on this lock, it may win the race
1135                                  * and cause alloc_contig_range() to fail...
1136                                  */
1137                                 spin_unlock_irqrestore(&zone->lock, flags);
1138                                 ret = __alloc_gigantic_page(pfn, nr_pages, gfp_mask);
1139                                 if (!ret)
1140                                         return pfn_to_page(pfn);
1141                                 spin_lock_irqsave(&zone->lock, flags);
1142                         }
1143                         pfn += nr_pages;
1144                 }
1145 
1146                 spin_unlock_irqrestore(&zone->lock, flags);
1147         }
1148 
1149         return NULL;
1150 }
1151 
1152 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1153 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1154 
1155 static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid)
1156 {
1157         struct page *page;
1158 
1159         page = alloc_gigantic_page(nid, h);
1160         if (page) {
1161                 prep_compound_gigantic_page(page, huge_page_order(h));
1162                 prep_new_huge_page(h, page, nid);
1163         }
1164 
1165         return page;
1166 }
1167 
1168 static int alloc_fresh_gigantic_page(struct hstate *h,
1169                                 nodemask_t *nodes_allowed)
1170 {
1171         struct page *page = NULL;
1172         int nr_nodes, node;
1173 
1174         for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1175                 page = alloc_fresh_gigantic_page_node(h, node);
1176                 if (page)
1177                         return 1;
1178         }
1179 
1180         return 0;
1181 }
1182 
1183 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1184 static inline bool gigantic_page_supported(void) { return false; }
1185 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1186 static inline void destroy_compound_gigantic_page(struct page *page,
1187                                                 unsigned int order) { }
1188 static inline int alloc_fresh_gigantic_page(struct hstate *h,
1189                                         nodemask_t *nodes_allowed) { return 0; }
1190 #endif
1191 
1192 static void update_and_free_page(struct hstate *h, struct page *page)
1193 {
1194         int i;
1195 
1196         if (hstate_is_gigantic(h) && !gigantic_page_supported())
1197                 return;
1198 
1199         h->nr_huge_pages--;
1200         h->nr_huge_pages_node[page_to_nid(page)]--;
1201         for (i = 0; i < pages_per_huge_page(h); i++) {
1202                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1203                                 1 << PG_referenced | 1 << PG_dirty |
1204                                 1 << PG_active | 1 << PG_private |
1205                                 1 << PG_writeback);
1206         }
1207         VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1208         set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1209         set_page_refcounted(page);
1210         if (hstate_is_gigantic(h)) {
1211                 destroy_compound_gigantic_page(page, huge_page_order(h));
1212                 free_gigantic_page(page, huge_page_order(h));
1213         } else {
1214                 __free_pages(page, huge_page_order(h));
1215         }
1216 }
1217 
1218 struct hstate *size_to_hstate(unsigned long size)
1219 {
1220         struct hstate *h;
1221 
1222         for_each_hstate(h) {
1223                 if (huge_page_size(h) == size)
1224                         return h;
1225         }
1226         return NULL;
1227 }
1228 
1229 /*
1230  * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1231  * to hstate->hugepage_activelist.)
1232  *
1233  * This function can be called for tail pages, but never returns true for them.
1234  */
1235 bool page_huge_active(struct page *page)
1236 {
1237         VM_BUG_ON_PAGE(!PageHuge(page), page);
1238         return PageHead(page) && PagePrivate(&page[1]);
1239 }
1240 
1241 /* never called for tail page */
1242 static void set_page_huge_active(struct page *page)
1243 {
1244         VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1245         SetPagePrivate(&page[1]);
1246 }
1247 
1248 static void clear_page_huge_active(struct page *page)
1249 {
1250         VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1251         ClearPagePrivate(&page[1]);
1252 }
1253 
1254 void free_huge_page(struct page *page)
1255 {
1256         /*
1257          * Can't pass hstate in here because it is called from the
1258          * compound page destructor.
1259          */
1260         struct hstate *h = page_hstate(page);
1261         int nid = page_to_nid(page);
1262         struct hugepage_subpool *spool =
1263                 (struct hugepage_subpool *)page_private(page);
1264         bool restore_reserve;
1265 
1266         set_page_private(page, 0);
1267         page->mapping = NULL;
1268         VM_BUG_ON_PAGE(page_count(page), page);
1269         VM_BUG_ON_PAGE(page_mapcount(page), page);
1270         restore_reserve = PagePrivate(page);
1271         ClearPagePrivate(page);
1272 
1273         /*
1274          * A return code of zero implies that the subpool will be under its
1275          * minimum size if the reservation is not restored after page is free.
1276          * Therefore, force restore_reserve operation.
1277          */
1278         if (hugepage_subpool_put_pages(spool, 1) == 0)
1279                 restore_reserve = true;
1280 
1281         spin_lock(&hugetlb_lock);
1282         clear_page_huge_active(page);
1283         hugetlb_cgroup_uncharge_page(hstate_index(h),
1284                                      pages_per_huge_page(h), page);
1285         if (restore_reserve)
1286                 h->resv_huge_pages++;
1287 
1288         if (h->surplus_huge_pages_node[nid]) {
1289                 /* remove the page from active list */
1290                 list_del(&page->lru);
1291                 update_and_free_page(h, page);
1292                 h->surplus_huge_pages--;
1293                 h->surplus_huge_pages_node[nid]--;
1294         } else {
1295                 arch_clear_hugepage_flags(page);
1296                 enqueue_huge_page(h, page);
1297         }
1298         spin_unlock(&hugetlb_lock);
1299 }
1300 
1301 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1302 {
1303         INIT_LIST_HEAD(&page->lru);
1304         set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1305         spin_lock(&hugetlb_lock);
1306         set_hugetlb_cgroup(page, NULL);
1307         h->nr_huge_pages++;
1308         h->nr_huge_pages_node[nid]++;
1309         spin_unlock(&hugetlb_lock);
1310         put_page(page); /* free it into the hugepage allocator */
1311 }
1312 
1313 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1314 {
1315         int i;
1316         int nr_pages = 1 << order;
1317         struct page *p = page + 1;
1318 
1319         /* we rely on prep_new_huge_page to set the destructor */
1320         set_compound_order(page, order);
1321         __ClearPageReserved(page);
1322         __SetPageHead(page);
1323         for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1324                 /*
1325                  * For gigantic hugepages allocated through bootmem at
1326                  * boot, it's safer to be consistent with the not-gigantic
1327                  * hugepages and clear the PG_reserved bit from all tail pages
1328                  * too.  Otherwse drivers using get_user_pages() to access tail
1329                  * pages may get the reference counting wrong if they see
1330                  * PG_reserved set on a tail page (despite the head page not
1331                  * having PG_reserved set).  Enforcing this consistency between
1332                  * head and tail pages allows drivers to optimize away a check
1333                  * on the head page when they need know if put_page() is needed
1334                  * after get_user_pages().
1335                  */
1336                 __ClearPageReserved(p);
1337                 set_page_count(p, 0);
1338                 set_compound_head(p, page);
1339         }
1340         atomic_set(compound_mapcount_ptr(page), -1);
1341 }
1342 
1343 /*
1344  * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1345  * transparent huge pages.  See the PageTransHuge() documentation for more
1346  * details.
1347  */
1348 int PageHuge(struct page *page)
1349 {
1350         if (!PageCompound(page))
1351                 return 0;
1352 
1353         page = compound_head(page);
1354         return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1355 }
1356 EXPORT_SYMBOL_GPL(PageHuge);
1357 
1358 /*
1359  * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1360  * normal or transparent huge pages.
1361  */
1362 int PageHeadHuge(struct page *page_head)
1363 {
1364         if (!PageHead(page_head))
1365                 return 0;
1366 
1367         return get_compound_page_dtor(page_head) == free_huge_page;
1368 }
1369 
1370 pgoff_t __basepage_index(struct page *page)
1371 {
1372         struct page *page_head = compound_head(page);
1373         pgoff_t index = page_index(page_head);
1374         unsigned long compound_idx;
1375 
1376         if (!PageHuge(page_head))
1377                 return page_index(page);
1378 
1379         if (compound_order(page_head) >= MAX_ORDER)
1380                 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1381         else
1382                 compound_idx = page - page_head;
1383 
1384         return (index << compound_order(page_head)) + compound_idx;
1385 }
1386 
1387 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
1388 {
1389         struct page *page;
1390 
1391         page = __alloc_pages_node(nid,
1392                 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
1393                                                 __GFP_RETRY_MAYFAIL|__GFP_NOWARN,
1394                 huge_page_order(h));
1395         if (page) {
1396                 prep_new_huge_page(h, page, nid);
1397         }
1398 
1399         return page;
1400 }
1401 
1402 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1403 {
1404         struct page *page;
1405         int nr_nodes, node;
1406         int ret = 0;
1407 
1408         for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1409                 page = alloc_fresh_huge_page_node(h, node);
1410                 if (page) {
1411                         ret = 1;
1412                         break;
1413                 }
1414         }
1415 
1416         if (ret)
1417                 count_vm_event(HTLB_BUDDY_PGALLOC);
1418         else
1419                 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1420 
1421         return ret;
1422 }
1423 
1424 /*
1425  * Free huge page from pool from next node to free.
1426  * Attempt to keep persistent huge pages more or less
1427  * balanced over allowed nodes.
1428  * Called with hugetlb_lock locked.
1429  */
1430 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1431                                                          bool acct_surplus)
1432 {
1433         int nr_nodes, node;
1434         int ret = 0;
1435 
1436         for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1437                 /*
1438                  * If we're returning unused surplus pages, only examine
1439                  * nodes with surplus pages.
1440                  */
1441                 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1442                     !list_empty(&h->hugepage_freelists[node])) {
1443                         struct page *page =
1444                                 list_entry(h->hugepage_freelists[node].next,
1445                                           struct page, lru);
1446                         list_del(&page->lru);
1447                         h->free_huge_pages--;
1448                         h->free_huge_pages_node[node]--;
1449                         if (acct_surplus) {
1450                                 h->surplus_huge_pages--;
1451                                 h->surplus_huge_pages_node[node]--;
1452                         }
1453                         update_and_free_page(h, page);
1454                         ret = 1;
1455                         break;
1456                 }
1457         }
1458 
1459         return ret;
1460 }
1461 
1462 /*
1463  * Dissolve a given free hugepage into free buddy pages. This function does
1464  * nothing for in-use (including surplus) hugepages. Returns -EBUSY if the
1465  * number of free hugepages would be reduced below the number of reserved
1466  * hugepages.
1467  */
1468 int dissolve_free_huge_page(struct page *page)
1469 {
1470         int rc = 0;
1471 
1472         spin_lock(&hugetlb_lock);
1473         if (PageHuge(page) && !page_count(page)) {
1474                 struct page *head = compound_head(page);
1475                 struct hstate *h = page_hstate(head);
1476                 int nid = page_to_nid(head);
1477                 if (h->free_huge_pages - h->resv_huge_pages == 0) {
1478                         rc = -EBUSY;
1479                         goto out;
1480                 }
1481                 /*
1482                  * Move PageHWPoison flag from head page to the raw error page,
1483                  * which makes any subpages rather than the error page reusable.
1484                  */
1485                 if (PageHWPoison(head) && page != head) {
1486                         SetPageHWPoison(page);
1487                         ClearPageHWPoison(head);
1488                 }
1489                 list_del(&head->lru);
1490                 h->free_huge_pages--;
1491                 h->free_huge_pages_node[nid]--;
1492                 h->max_huge_pages--;
1493                 update_and_free_page(h, head);
1494         }
1495 out:
1496         spin_unlock(&hugetlb_lock);
1497         return rc;
1498 }
1499 
1500 /*
1501  * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1502  * make specified memory blocks removable from the system.
1503  * Note that this will dissolve a free gigantic hugepage completely, if any
1504  * part of it lies within the given range.
1505  * Also note that if dissolve_free_huge_page() returns with an error, all
1506  * free hugepages that were dissolved before that error are lost.
1507  */
1508 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1509 {
1510         unsigned long pfn;
1511         struct page *page;
1512         int rc = 0;
1513 
1514         if (!hugepages_supported())
1515                 return rc;
1516 
1517         for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1518                 page = pfn_to_page(pfn);
1519                 if (PageHuge(page) && !page_count(page)) {
1520                         rc = dissolve_free_huge_page(page);
1521                         if (rc)
1522                                 break;
1523                 }
1524         }
1525 
1526         return rc;
1527 }
1528 
1529 static struct page *__hugetlb_alloc_buddy_huge_page(struct hstate *h,
1530                 gfp_t gfp_mask, int nid, nodemask_t *nmask)
1531 {
1532         int order = huge_page_order(h);
1533 
1534         gfp_mask |= __GFP_COMP|__GFP_RETRY_MAYFAIL|__GFP_NOWARN;
1535         if (nid == NUMA_NO_NODE)
1536                 nid = numa_mem_id();
1537         return __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
1538 }
1539 
1540 static struct page *__alloc_buddy_huge_page(struct hstate *h, gfp_t gfp_mask,
1541                 int nid, nodemask_t *nmask)
1542 {
1543         struct page *page;
1544         unsigned int r_nid;
1545 
1546         if (hstate_is_gigantic(h))
1547                 return NULL;
1548 
1549         /*
1550          * Assume we will successfully allocate the surplus page to
1551          * prevent racing processes from causing the surplus to exceed
1552          * overcommit
1553          *
1554          * This however introduces a different race, where a process B
1555          * tries to grow the static hugepage pool while alloc_pages() is
1556          * called by process A. B will only examine the per-node
1557          * counters in determining if surplus huge pages can be
1558          * converted to normal huge pages in adjust_pool_surplus(). A
1559          * won't be able to increment the per-node counter, until the
1560          * lock is dropped by B, but B doesn't drop hugetlb_lock until
1561          * no more huge pages can be converted from surplus to normal
1562          * state (and doesn't try to convert again). Thus, we have a
1563          * case where a surplus huge page exists, the pool is grown, and
1564          * the surplus huge page still exists after, even though it
1565          * should just have been converted to a normal huge page. This
1566          * does not leak memory, though, as the hugepage will be freed
1567          * once it is out of use. It also does not allow the counters to
1568          * go out of whack in adjust_pool_surplus() as we don't modify
1569          * the node values until we've gotten the hugepage and only the
1570          * per-node value is checked there.
1571          */
1572         spin_lock(&hugetlb_lock);
1573         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1574                 spin_unlock(&hugetlb_lock);
1575                 return NULL;
1576         } else {
1577                 h->nr_huge_pages++;
1578                 h->surplus_huge_pages++;
1579         }
1580         spin_unlock(&hugetlb_lock);
1581 
1582         page = __hugetlb_alloc_buddy_huge_page(h, gfp_mask, nid, nmask);
1583 
1584         spin_lock(&hugetlb_lock);
1585         if (page) {
1586                 INIT_LIST_HEAD(&page->lru);
1587                 r_nid = page_to_nid(page);
1588                 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1589                 set_hugetlb_cgroup(page, NULL);
1590                 /*
1591                  * We incremented the global counters already
1592                  */
1593                 h->nr_huge_pages_node[r_nid]++;
1594                 h->surplus_huge_pages_node[r_nid]++;
1595                 __count_vm_event(HTLB_BUDDY_PGALLOC);
1596         } else {
1597                 h->nr_huge_pages--;
1598                 h->surplus_huge_pages--;
1599                 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1600         }
1601         spin_unlock(&hugetlb_lock);
1602 
1603         return page;
1604 }
1605 
1606 /*
1607  * Use the VMA's mpolicy to allocate a huge page from the buddy.
1608  */
1609 static
1610 struct page *__alloc_buddy_huge_page_with_mpol(struct hstate *h,
1611                 struct vm_area_struct *vma, unsigned long addr)
1612 {
1613         struct page *page;
1614         struct mempolicy *mpol;
1615         gfp_t gfp_mask = htlb_alloc_mask(h);
1616         int nid;
1617         nodemask_t *nodemask;
1618 
1619         nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1620         page = __alloc_buddy_huge_page(h, gfp_mask, nid, nodemask);
1621         mpol_cond_put(mpol);
1622 
1623         return page;
1624 }
1625 
1626 /*
1627  * This allocation function is useful in the context where vma is irrelevant.
1628  * E.g. soft-offlining uses this function because it only cares physical
1629  * address of error page.
1630  */
1631 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1632 {
1633         gfp_t gfp_mask = htlb_alloc_mask(h);
1634         struct page *page = NULL;
1635 
1636         if (nid != NUMA_NO_NODE)
1637                 gfp_mask |= __GFP_THISNODE;
1638 
1639         spin_lock(&hugetlb_lock);
1640         if (h->free_huge_pages - h->resv_huge_pages > 0)
1641                 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL);
1642         spin_unlock(&hugetlb_lock);
1643 
1644         if (!page)
1645                 page = __alloc_buddy_huge_page(h, gfp_mask, nid, NULL);
1646 
1647         return page;
1648 }
1649 
1650 
1651 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
1652                 nodemask_t *nmask)
1653 {
1654         gfp_t gfp_mask = htlb_alloc_mask(h);
1655 
1656         spin_lock(&hugetlb_lock);
1657         if (h->free_huge_pages - h->resv_huge_pages > 0) {
1658                 struct page *page;
1659 
1660                 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1661                 if (page) {
1662                         spin_unlock(&hugetlb_lock);
1663                         return page;
1664                 }
1665         }
1666         spin_unlock(&hugetlb_lock);
1667 
1668         /* No reservations, try to overcommit */
1669 
1670         return __alloc_buddy_huge_page(h, gfp_mask, preferred_nid, nmask);
1671 }
1672 
1673 /*
1674  * Increase the hugetlb pool such that it can accommodate a reservation
1675  * of size 'delta'.
1676  */
1677 static int gather_surplus_pages(struct hstate *h, int delta)
1678 {
1679         struct list_head surplus_list;
1680         struct page *page, *tmp;
1681         int ret, i;
1682         int needed, allocated;
1683         bool alloc_ok = true;
1684 
1685         needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1686         if (needed <= 0) {
1687                 h->resv_huge_pages += delta;
1688                 return 0;
1689         }
1690 
1691         allocated = 0;
1692         INIT_LIST_HEAD(&surplus_list);
1693 
1694         ret = -ENOMEM;
1695 retry:
1696         spin_unlock(&hugetlb_lock);
1697         for (i = 0; i < needed; i++) {
1698                 page = __alloc_buddy_huge_page(h, htlb_alloc_mask(h),
1699                                 NUMA_NO_NODE, NULL);
1700                 if (!page) {
1701                         alloc_ok = false;
1702                         break;
1703                 }
1704                 list_add(&page->lru, &surplus_list);
1705                 cond_resched();
1706         }
1707         allocated += i;
1708 
1709         /*
1710          * After retaking hugetlb_lock, we need to recalculate 'needed'
1711          * because either resv_huge_pages or free_huge_pages may have changed.
1712          */
1713         spin_lock(&hugetlb_lock);
1714         needed = (h->resv_huge_pages + delta) -
1715                         (h->free_huge_pages + allocated);
1716         if (needed > 0) {
1717                 if (alloc_ok)
1718                         goto retry;
1719                 /*
1720                  * We were not able to allocate enough pages to
1721                  * satisfy the entire reservation so we free what
1722                  * we've allocated so far.
1723                  */
1724                 goto free;
1725         }
1726         /*
1727          * The surplus_list now contains _at_least_ the number of extra pages
1728          * needed to accommodate the reservation.  Add the appropriate number
1729          * of pages to the hugetlb pool and free the extras back to the buddy
1730          * allocator.  Commit the entire reservation here to prevent another
1731          * process from stealing the pages as they are added to the pool but
1732          * before they are reserved.
1733          */
1734         needed += allocated;
1735         h->resv_huge_pages += delta;
1736         ret = 0;
1737 
1738         /* Free the needed pages to the hugetlb pool */
1739         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1740                 if ((--needed) < 0)
1741                         break;
1742                 /*
1743                  * This page is now managed by the hugetlb allocator and has
1744                  * no users -- drop the buddy allocator's reference.
1745                  */
1746                 put_page_testzero(page);
1747                 VM_BUG_ON_PAGE(page_count(page), page);
1748                 enqueue_huge_page(h, page);
1749         }
1750 free:
1751         spin_unlock(&hugetlb_lock);
1752 
1753         /* Free unnecessary surplus pages to the buddy allocator */
1754         list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1755                 put_page(page);
1756         spin_lock(&hugetlb_lock);
1757 
1758         return ret;
1759 }
1760 
1761 /*
1762  * This routine has two main purposes:
1763  * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1764  *    in unused_resv_pages.  This corresponds to the prior adjustments made
1765  *    to the associated reservation map.
1766  * 2) Free any unused surplus pages that may have been allocated to satisfy
1767  *    the reservation.  As many as unused_resv_pages may be freed.
1768  *
1769  * Called with hugetlb_lock held.  However, the lock could be dropped (and
1770  * reacquired) during calls to cond_resched_lock.  Whenever dropping the lock,
1771  * we must make sure nobody else can claim pages we are in the process of
1772  * freeing.  Do this by ensuring resv_huge_page always is greater than the
1773  * number of huge pages we plan to free when dropping the lock.
1774  */
1775 static void return_unused_surplus_pages(struct hstate *h,
1776                                         unsigned long unused_resv_pages)
1777 {
1778         unsigned long nr_pages;
1779 
1780         /* Cannot return gigantic pages currently */
1781         if (hstate_is_gigantic(h))
1782                 goto out;
1783 
1784         /*
1785          * Part (or even all) of the reservation could have been backed
1786          * by pre-allocated pages. Only free surplus pages.
1787          */
1788         nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1789 
1790         /*
1791          * We want to release as many surplus pages as possible, spread
1792          * evenly across all nodes with memory. Iterate across these nodes
1793          * until we can no longer free unreserved surplus pages. This occurs
1794          * when the nodes with surplus pages have no free pages.
1795          * free_pool_huge_page() will balance the the freed pages across the
1796          * on-line nodes with memory and will handle the hstate accounting.
1797          *
1798          * Note that we decrement resv_huge_pages as we free the pages.  If
1799          * we drop the lock, resv_huge_pages will still be sufficiently large
1800          * to cover subsequent pages we may free.
1801          */
1802         while (nr_pages--) {
1803                 h->resv_huge_pages--;
1804                 unused_resv_pages--;
1805                 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1806                         goto out;
1807                 cond_resched_lock(&hugetlb_lock);
1808         }
1809 
1810 out:
1811         /* Fully uncommit the reservation */
1812         h->resv_huge_pages -= unused_resv_pages;
1813 }
1814 
1815 
1816 /*
1817  * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1818  * are used by the huge page allocation routines to manage reservations.
1819  *
1820  * vma_needs_reservation is called to determine if the huge page at addr
1821  * within the vma has an associated reservation.  If a reservation is
1822  * needed, the value 1 is returned.  The caller is then responsible for
1823  * managing the global reservation and subpool usage counts.  After
1824  * the huge page has been allocated, vma_commit_reservation is called
1825  * to add the page to the reservation map.  If the page allocation fails,
1826  * the reservation must be ended instead of committed.  vma_end_reservation
1827  * is called in such cases.
1828  *
1829  * In the normal case, vma_commit_reservation returns the same value
1830  * as the preceding vma_needs_reservation call.  The only time this
1831  * is not the case is if a reserve map was changed between calls.  It
1832  * is the responsibility of the caller to notice the difference and
1833  * take appropriate action.
1834  *
1835  * vma_add_reservation is used in error paths where a reservation must
1836  * be restored when a newly allocated huge page must be freed.  It is
1837  * to be called after calling vma_needs_reservation to determine if a
1838  * reservation exists.
1839  */
1840 enum vma_resv_mode {
1841         VMA_NEEDS_RESV,
1842         VMA_COMMIT_RESV,
1843         VMA_END_RESV,
1844         VMA_ADD_RESV,
1845 };
1846 static long __vma_reservation_common(struct hstate *h,
1847                                 struct vm_area_struct *vma, unsigned long addr,
1848                                 enum vma_resv_mode mode)
1849 {
1850         struct resv_map *resv;
1851         pgoff_t idx;
1852         long ret;
1853 
1854         resv = vma_resv_map(vma);
1855         if (!resv)
1856                 return 1;
1857 
1858         idx = vma_hugecache_offset(h, vma, addr);
1859         switch (mode) {
1860         case VMA_NEEDS_RESV:
1861                 ret = region_chg(resv, idx, idx + 1);
1862                 break;
1863         case VMA_COMMIT_RESV:
1864                 ret = region_add(resv, idx, idx + 1);
1865                 break;
1866         case VMA_END_RESV:
1867                 region_abort(resv, idx, idx + 1);
1868                 ret = 0;
1869                 break;
1870         case VMA_ADD_RESV:
1871                 if (vma->vm_flags & VM_MAYSHARE)
1872                         ret = region_add(resv, idx, idx + 1);
1873                 else {
1874                         region_abort(resv, idx, idx + 1);
1875                         ret = region_del(resv, idx, idx + 1);
1876                 }
1877                 break;
1878         default:
1879                 BUG();
1880         }
1881 
1882         if (vma->vm_flags & VM_MAYSHARE)
1883                 return ret;
1884         else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
1885                 /*
1886                  * In most cases, reserves always exist for private mappings.
1887                  * However, a file associated with mapping could have been
1888                  * hole punched or truncated after reserves were consumed.
1889                  * As subsequent fault on such a range will not use reserves.
1890                  * Subtle - The reserve map for private mappings has the
1891                  * opposite meaning than that of shared mappings.  If NO
1892                  * entry is in the reserve map, it means a reservation exists.
1893                  * If an entry exists in the reserve map, it means the
1894                  * reservation has already been consumed.  As a result, the
1895                  * return value of this routine is the opposite of the
1896                  * value returned from reserve map manipulation routines above.
1897                  */
1898                 if (ret)
1899                         return 0;
1900                 else
1901                         return 1;
1902         }
1903         else
1904                 return ret < 0 ? ret : 0;
1905 }
1906 
1907 static long vma_needs_reservation(struct hstate *h,
1908                         struct vm_area_struct *vma, unsigned long addr)
1909 {
1910         return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
1911 }
1912 
1913 static long vma_commit_reservation(struct hstate *h,
1914                         struct vm_area_struct *vma, unsigned long addr)
1915 {
1916         return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1917 }
1918 
1919 static void vma_end_reservation(struct hstate *h,
1920                         struct vm_area_struct *vma, unsigned long addr)
1921 {
1922         (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
1923 }
1924 
1925 static long vma_add_reservation(struct hstate *h,
1926                         struct vm_area_struct *vma, unsigned long addr)
1927 {
1928         return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
1929 }
1930 
1931 /*
1932  * This routine is called to restore a reservation on error paths.  In the
1933  * specific error paths, a huge page was allocated (via alloc_huge_page)
1934  * and is about to be freed.  If a reservation for the page existed,
1935  * alloc_huge_page would have consumed the reservation and set PagePrivate
1936  * in the newly allocated page.  When the page is freed via free_huge_page,
1937  * the global reservation count will be incremented if PagePrivate is set.
1938  * However, free_huge_page can not adjust the reserve map.  Adjust the
1939  * reserve map here to be consistent with global reserve count adjustments
1940  * to be made by free_huge_page.
1941  */
1942 static void restore_reserve_on_error(struct hstate *h,
1943                         struct vm_area_struct *vma, unsigned long address,
1944                         struct page *page)
1945 {
1946         if (unlikely(PagePrivate(page))) {
1947                 long rc = vma_needs_reservation(h, vma, address);
1948 
1949                 if (unlikely(rc < 0)) {
1950                         /*
1951                          * Rare out of memory condition in reserve map
1952                          * manipulation.  Clear PagePrivate so that
1953                          * global reserve count will not be incremented
1954                          * by free_huge_page.  This will make it appear
1955                          * as though the reservation for this page was
1956                          * consumed.  This may prevent the task from
1957                          * faulting in the page at a later time.  This
1958                          * is better than inconsistent global huge page
1959                          * accounting of reserve counts.
1960                          */
1961                         ClearPagePrivate(page);
1962                 } else if (rc) {
1963                         rc = vma_add_reservation(h, vma, address);
1964                         if (unlikely(rc < 0))
1965                                 /*
1966                                  * See above comment about rare out of
1967                                  * memory condition.
1968                                  */
1969                                 ClearPagePrivate(page);
1970                 } else
1971                         vma_end_reservation(h, vma, address);
1972         }
1973 }
1974 
1975 struct page *alloc_huge_page(struct vm_area_struct *vma,
1976                                     unsigned long addr, int avoid_reserve)
1977 {
1978         struct hugepage_subpool *spool = subpool_vma(vma);
1979         struct hstate *h = hstate_vma(vma);
1980         struct page *page;
1981         long map_chg, map_commit;
1982         long gbl_chg;
1983         int ret, idx;
1984         struct hugetlb_cgroup *h_cg;
1985 
1986         idx = hstate_index(h);
1987         /*
1988          * Examine the region/reserve map to determine if the process
1989          * has a reservation for the page to be allocated.  A return
1990          * code of zero indicates a reservation exists (no change).
1991          */
1992         map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
1993         if (map_chg < 0)
1994                 return ERR_PTR(-ENOMEM);
1995 
1996         /*
1997          * Processes that did not create the mapping will have no
1998          * reserves as indicated by the region/reserve map. Check
1999          * that the allocation will not exceed the subpool limit.
2000          * Allocations for MAP_NORESERVE mappings also need to be
2001          * checked against any subpool limit.
2002          */
2003         if (map_chg || avoid_reserve) {
2004                 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2005                 if (gbl_chg < 0) {
2006                         vma_end_reservation(h, vma, addr);
2007                         return ERR_PTR(-ENOSPC);
2008                 }
2009 
2010                 /*
2011                  * Even though there was no reservation in the region/reserve
2012                  * map, there could be reservations associated with the
2013                  * subpool that can be used.  This would be indicated if the
2014                  * return value of hugepage_subpool_get_pages() is zero.
2015                  * However, if avoid_reserve is specified we still avoid even
2016                  * the subpool reservations.
2017                  */
2018                 if (avoid_reserve)
2019                         gbl_chg = 1;
2020         }
2021 
2022         ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2023         if (ret)
2024                 goto out_subpool_put;
2025 
2026         spin_lock(&hugetlb_lock);
2027         /*
2028          * glb_chg is passed to indicate whether or not a page must be taken
2029          * from the global free pool (global change).  gbl_chg == 0 indicates
2030          * a reservation exists for the allocation.
2031          */
2032         page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2033         if (!page) {
2034                 spin_unlock(&hugetlb_lock);
2035                 page = __alloc_buddy_huge_page_with_mpol(h, vma, addr);
2036                 if (!page)
2037                         goto out_uncharge_cgroup;
2038                 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2039                         SetPagePrivate(page);
2040                         h->resv_huge_pages--;
2041                 }
2042                 spin_lock(&hugetlb_lock);
2043                 list_move(&page->lru, &h->hugepage_activelist);
2044                 /* Fall through */
2045         }
2046         hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2047         spin_unlock(&hugetlb_lock);
2048 
2049         set_page_private(page, (unsigned long)spool);
2050 
2051         map_commit = vma_commit_reservation(h, vma, addr);
2052         if (unlikely(map_chg > map_commit)) {
2053                 /*
2054                  * The page was added to the reservation map between
2055                  * vma_needs_reservation and vma_commit_reservation.
2056                  * This indicates a race with hugetlb_reserve_pages.
2057                  * Adjust for the subpool count incremented above AND
2058                  * in hugetlb_reserve_pages for the same page.  Also,
2059                  * the reservation count added in hugetlb_reserve_pages
2060                  * no longer applies.
2061                  */
2062                 long rsv_adjust;
2063 
2064                 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2065                 hugetlb_acct_memory(h, -rsv_adjust);
2066         }
2067         return page;
2068 
2069 out_uncharge_cgroup:
2070         hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2071 out_subpool_put:
2072         if (map_chg || avoid_reserve)
2073                 hugepage_subpool_put_pages(spool, 1);
2074         vma_end_reservation(h, vma, addr);
2075         return ERR_PTR(-ENOSPC);
2076 }
2077 
2078 /*
2079  * alloc_huge_page()'s wrapper which simply returns the page if allocation
2080  * succeeds, otherwise NULL. This function is called from new_vma_page(),
2081  * where no ERR_VALUE is expected to be returned.
2082  */
2083 struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
2084                                 unsigned long addr, int avoid_reserve)
2085 {
2086         struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
2087         if (IS_ERR(page))
2088                 page = NULL;
2089         return page;
2090 }
2091 
2092 int alloc_bootmem_huge_page(struct hstate *h)
2093         __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2094 int __alloc_bootmem_huge_page(struct hstate *h)
2095 {
2096         struct huge_bootmem_page *m;
2097         int nr_nodes, node;
2098 
2099         for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2100                 void *addr;
2101 
2102                 addr = memblock_virt_alloc_try_nid_nopanic(
2103                                 huge_page_size(h), huge_page_size(h),
2104                                 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
2105                 if (addr) {
2106                         /*
2107                          * Use the beginning of the huge page to store the
2108                          * huge_bootmem_page struct (until gather_bootmem
2109                          * puts them into the mem_map).
2110                          */
2111                         m = addr;
2112                         goto found;
2113                 }
2114         }
2115         return 0;
2116 
2117 found:
2118         BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2119         /* Put them into a private list first because mem_map is not up yet */
2120         list_add(&m->list, &huge_boot_pages);
2121         m->hstate = h;
2122         return 1;
2123 }
2124 
2125 static void __init prep_compound_huge_page(struct page *page,
2126                 unsigned int order)
2127 {
2128         if (unlikely(order > (MAX_ORDER - 1)))
2129                 prep_compound_gigantic_page(page, order);
2130         else
2131                 prep_compound_page(page, order);
2132 }
2133 
2134 /* Put bootmem huge pages into the standard lists after mem_map is up */
2135 static void __init gather_bootmem_prealloc(void)
2136 {
2137         struct huge_bootmem_page *m;
2138 
2139         list_for_each_entry(m, &huge_boot_pages, list) {
2140                 struct hstate *h = m->hstate;
2141                 struct page *page;
2142 
2143 #ifdef CONFIG_HIGHMEM
2144                 page = pfn_to_page(m->phys >> PAGE_SHIFT);
2145                 memblock_free_late(__pa(m),
2146                                    sizeof(struct huge_bootmem_page));
2147 #else
2148                 page = virt_to_page(m);
2149 #endif
2150                 WARN_ON(page_count(page) != 1);
2151                 prep_compound_huge_page(page, h->order);
2152                 WARN_ON(PageReserved(page));
2153                 prep_new_huge_page(h, page, page_to_nid(page));
2154                 /*
2155                  * If we had gigantic hugepages allocated at boot time, we need
2156                  * to restore the 'stolen' pages to totalram_pages in order to
2157                  * fix confusing memory reports from free(1) and another
2158                  * side-effects, like CommitLimit going negative.
2159                  */
2160                 if (hstate_is_gigantic(h))
2161                         adjust_managed_page_count(page, 1 << h->order);
2162         }
2163 }
2164 
2165 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2166 {
2167         unsigned long i;
2168 
2169         for (i = 0; i < h->max_huge_pages; ++i) {
2170                 if (hstate_is_gigantic(h)) {
2171                         if (!alloc_bootmem_huge_page(h))
2172                                 break;
2173                 } else if (!alloc_fresh_huge_page(h,
2174                                          &node_states[N_MEMORY]))
2175                         break;
2176                 cond_resched();
2177         }
2178         if (i < h->max_huge_pages) {
2179                 char buf[32];
2180 
2181                 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2182                 pr_warn("HugeTLB: allocating %lu of page size %s failed.  Only allocated %lu hugepages.\n",
2183                         h->max_huge_pages, buf, i);
2184                 h->max_huge_pages = i;
2185         }
2186 }
2187 
2188 static void __init hugetlb_init_hstates(void)
2189 {
2190         struct hstate *h;
2191 
2192         for_each_hstate(h) {
2193                 if (minimum_order > huge_page_order(h))
2194                         minimum_order = huge_page_order(h);
2195 
2196                 /* oversize hugepages were init'ed in early boot */
2197                 if (!hstate_is_gigantic(h))
2198                         hugetlb_hstate_alloc_pages(h);
2199         }
2200         VM_BUG_ON(minimum_order == UINT_MAX);
2201 }
2202 
2203 static void __init report_hugepages(void)
2204 {
2205         struct hstate *h;
2206 
2207         for_each_hstate(h) {
2208                 char buf[32];
2209 
2210                 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2211                 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2212                         buf, h->free_huge_pages);
2213         }
2214 }
2215 
2216 #ifdef CONFIG_HIGHMEM
2217 static void try_to_free_low(struct hstate *h, unsigned long count,
2218                                                 nodemask_t *nodes_allowed)
2219 {
2220         int i;
2221 
2222         if (hstate_is_gigantic(h))
2223                 return;
2224 
2225         for_each_node_mask(i, *nodes_allowed) {
2226                 struct page *page, *next;
2227                 struct list_head *freel = &h->hugepage_freelists[i];
2228                 list_for_each_entry_safe(page, next, freel, lru) {
2229                         if (count >= h->nr_huge_pages)
2230                                 return;
2231                         if (PageHighMem(page))
2232                                 continue;
2233                         list_del(&page->lru);
2234                         update_and_free_page(h, page);
2235                         h->free_huge_pages--;
2236                         h->free_huge_pages_node[page_to_nid(page)]--;
2237                 }
2238         }
2239 }
2240 #else
2241 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2242                                                 nodemask_t *nodes_allowed)
2243 {
2244 }
2245 #endif
2246 
2247 /*
2248  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
2249  * balanced by operating on them in a round-robin fashion.
2250  * Returns 1 if an adjustment was made.
2251  */
2252 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2253                                 int delta)
2254 {
2255         int nr_nodes, node;
2256 
2257         VM_BUG_ON(delta != -1 && delta != 1);
2258 
2259         if (delta < 0) {
2260                 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2261                         if (h->surplus_huge_pages_node[node])
2262                                 goto found;
2263                 }
2264         } else {
2265                 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2266                         if (h->surplus_huge_pages_node[node] <
2267                                         h->nr_huge_pages_node[node])
2268                                 goto found;
2269                 }
2270         }
2271         return 0;
2272 
2273 found:
2274         h->surplus_huge_pages += delta;
2275         h->surplus_huge_pages_node[node] += delta;
2276         return 1;
2277 }
2278 
2279 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2280 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
2281                                                 nodemask_t *nodes_allowed)
2282 {
2283         unsigned long min_count, ret;
2284 
2285         if (hstate_is_gigantic(h) && !gigantic_page_supported())
2286                 return h->max_huge_pages;
2287 
2288         /*
2289          * Increase the pool size
2290          * First take pages out of surplus state.  Then make up the
2291          * remaining difference by allocating fresh huge pages.
2292          *
2293          * We might race with __alloc_buddy_huge_page() here and be unable
2294          * to convert a surplus huge page to a normal huge page. That is
2295          * not critical, though, it just means the overall size of the
2296          * pool might be one hugepage larger than it needs to be, but
2297          * within all the constraints specified by the sysctls.
2298          */
2299         spin_lock(&hugetlb_lock);
2300         while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2301                 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2302                         break;
2303         }
2304 
2305         while (count > persistent_huge_pages(h)) {
2306                 /*
2307                  * If this allocation races such that we no longer need the
2308                  * page, free_huge_page will handle it by freeing the page
2309                  * and reducing the surplus.
2310                  */
2311                 spin_unlock(&hugetlb_lock);
2312 
2313                 /* yield cpu to avoid soft lockup */
2314                 cond_resched();
2315 
2316                 if (hstate_is_gigantic(h))
2317                         ret = alloc_fresh_gigantic_page(h, nodes_allowed);
2318                 else
2319                         ret = alloc_fresh_huge_page(h, nodes_allowed);
2320                 spin_lock(&hugetlb_lock);
2321                 if (!ret)
2322                         goto out;
2323 
2324                 /* Bail for signals. Probably ctrl-c from user */
2325                 if (signal_pending(current))
2326                         goto out;
2327         }
2328 
2329         /*
2330          * Decrease the pool size
2331          * First return free pages to the buddy allocator (being careful
2332          * to keep enough around to satisfy reservations).  Then place
2333          * pages into surplus state as needed so the pool will shrink
2334          * to the desired size as pages become free.
2335          *
2336          * By placing pages into the surplus state independent of the
2337          * overcommit value, we are allowing the surplus pool size to
2338          * exceed overcommit. There are few sane options here. Since
2339          * __alloc_buddy_huge_page() is checking the global counter,
2340          * though, we'll note that we're not allowed to exceed surplus
2341          * and won't grow the pool anywhere else. Not until one of the
2342          * sysctls are changed, or the surplus pages go out of use.
2343          */
2344         min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2345         min_count = max(count, min_count);
2346         try_to_free_low(h, min_count, nodes_allowed);
2347         while (min_count < persistent_huge_pages(h)) {
2348                 if (!free_pool_huge_page(h, nodes_allowed, 0))
2349                         break;
2350                 cond_resched_lock(&hugetlb_lock);
2351         }
2352         while (count < persistent_huge_pages(h)) {
2353                 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2354                         break;
2355         }
2356 out:
2357         ret = persistent_huge_pages(h);
2358         spin_unlock(&hugetlb_lock);
2359         return ret;
2360 }
2361 
2362 #define HSTATE_ATTR_RO(_name) \
2363         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2364 
2365 #define HSTATE_ATTR(_name) \
2366         static struct kobj_attribute _name##_attr = \
2367                 __ATTR(_name, 0644, _name##_show, _name##_store)
2368 
2369 static struct kobject *hugepages_kobj;
2370 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2371 
2372 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2373 
2374 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2375 {
2376         int i;
2377 
2378         for (i = 0; i < HUGE_MAX_HSTATE; i++)
2379                 if (hstate_kobjs[i] == kobj) {
2380                         if (nidp)
2381                                 *nidp = NUMA_NO_NODE;
2382                         return &hstates[i];
2383                 }
2384 
2385         return kobj_to_node_hstate(kobj, nidp);
2386 }
2387 
2388 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2389                                         struct kobj_attribute *attr, char *buf)
2390 {
2391         struct hstate *h;
2392         unsigned long nr_huge_pages;
2393         int nid;
2394 
2395         h = kobj_to_hstate(kobj, &nid);
2396         if (nid == NUMA_NO_NODE)
2397                 nr_huge_pages = h->nr_huge_pages;
2398         else
2399                 nr_huge_pages = h->nr_huge_pages_node[nid];
2400 
2401         return sprintf(buf, "%lu\n", nr_huge_pages);
2402 }
2403 
2404 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2405                                            struct hstate *h, int nid,
2406                                            unsigned long count, size_t len)
2407 {
2408         int err;
2409         NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
2410 
2411         if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
2412                 err = -EINVAL;
2413                 goto out;
2414         }
2415 
2416         if (nid == NUMA_NO_NODE) {
2417                 /*
2418                  * global hstate attribute
2419                  */
2420                 if (!(obey_mempolicy &&
2421                                 init_nodemask_of_mempolicy(nodes_allowed))) {
2422                         NODEMASK_FREE(nodes_allowed);
2423                         nodes_allowed = &node_states[N_MEMORY];
2424                 }
2425         } else if (nodes_allowed) {
2426                 /*
2427                  * per node hstate attribute: adjust count to global,
2428                  * but restrict alloc/free to the specified node.
2429                  */
2430                 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2431                 init_nodemask_of_node(nodes_allowed, nid);
2432         } else
2433                 nodes_allowed = &node_states[N_MEMORY];
2434 
2435         h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
2436 
2437         if (nodes_allowed != &node_states[N_MEMORY])
2438                 NODEMASK_FREE(nodes_allowed);
2439 
2440         return len;
2441 out:
2442         NODEMASK_FREE(nodes_allowed);
2443         return err;
2444 }
2445 
2446 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2447                                          struct kobject *kobj, const char *buf,
2448                                          size_t len)
2449 {
2450         struct hstate *h;
2451         unsigned long count;
2452         int nid;
2453         int err;
2454 
2455         err = kstrtoul(buf, 10, &count);
2456         if (err)
2457                 return err;
2458 
2459         h = kobj_to_hstate(kobj, &nid);
2460         return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2461 }
2462 
2463 static ssize_t nr_hugepages_show(struct kobject *kobj,
2464                                        struct kobj_attribute *attr, char *buf)
2465 {
2466         return nr_hugepages_show_common(kobj, attr, buf);
2467 }
2468 
2469 static ssize_t nr_hugepages_store(struct kobject *kobj,
2470                struct kobj_attribute *attr, const char *buf, size_t len)
2471 {
2472         return nr_hugepages_store_common(false, kobj, buf, len);
2473 }
2474 HSTATE_ATTR(nr_hugepages);
2475 
2476 #ifdef CONFIG_NUMA
2477 
2478 /*
2479  * hstate attribute for optionally mempolicy-based constraint on persistent
2480  * huge page alloc/free.
2481  */
2482 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2483                                        struct kobj_attribute *attr, char *buf)
2484 {
2485         return nr_hugepages_show_common(kobj, attr, buf);
2486 }
2487 
2488 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2489                struct kobj_attribute *attr, const char *buf, size_t len)
2490 {
2491         return nr_hugepages_store_common(true, kobj, buf, len);
2492 }
2493 HSTATE_ATTR(nr_hugepages_mempolicy);
2494 #endif
2495 
2496 
2497 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2498                                         struct kobj_attribute *attr, char *buf)
2499 {
2500         struct hstate *h = kobj_to_hstate(kobj, NULL);
2501         return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2502 }
2503 
2504 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2505                 struct kobj_attribute *attr, const char *buf, size_t count)
2506 {
2507         int err;
2508         unsigned long input;
2509         struct hstate *h = kobj_to_hstate(kobj, NULL);
2510 
2511         if (hstate_is_gigantic(h))
2512                 return -EINVAL;
2513 
2514         err = kstrtoul(buf, 10, &input);
2515         if (err)
2516                 return err;
2517 
2518         spin_lock(&hugetlb_lock);
2519         h->nr_overcommit_huge_pages = input;
2520         spin_unlock(&hugetlb_lock);
2521 
2522         return count;
2523 }
2524 HSTATE_ATTR(nr_overcommit_hugepages);
2525 
2526 static ssize_t free_hugepages_show(struct kobject *kobj,
2527                                         struct kobj_attribute *attr, char *buf)
2528 {
2529         struct hstate *h;
2530         unsigned long free_huge_pages;
2531         int nid;
2532 
2533         h = kobj_to_hstate(kobj, &nid);
2534         if (nid == NUMA_NO_NODE)
2535                 free_huge_pages = h->free_huge_pages;
2536         else
2537                 free_huge_pages = h->free_huge_pages_node[nid];
2538 
2539         return sprintf(buf, "%lu\n", free_huge_pages);
2540 }
2541 HSTATE_ATTR_RO(free_hugepages);
2542 
2543 static ssize_t resv_hugepages_show(struct kobject *kobj,
2544                                         struct kobj_attribute *attr, char *buf)
2545 {
2546         struct hstate *h = kobj_to_hstate(kobj, NULL);
2547         return sprintf(buf, "%lu\n", h->resv_huge_pages);
2548 }
2549 HSTATE_ATTR_RO(resv_hugepages);
2550 
2551 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2552                                         struct kobj_attribute *attr, char *buf)
2553 {
2554         struct hstate *h;
2555         unsigned long surplus_huge_pages;
2556         int nid;
2557 
2558         h = kobj_to_hstate(kobj, &nid);
2559         if (nid == NUMA_NO_NODE)
2560                 surplus_huge_pages = h->surplus_huge_pages;
2561         else
2562                 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2563 
2564         return sprintf(buf, "%lu\n", surplus_huge_pages);
2565 }
2566 HSTATE_ATTR_RO(surplus_hugepages);
2567 
2568 static struct attribute *hstate_attrs[] = {
2569         &nr_hugepages_attr.attr,
2570         &nr_overcommit_hugepages_attr.attr,
2571         &free_hugepages_attr.attr,
2572         &resv_hugepages_attr.attr,
2573         &surplus_hugepages_attr.attr,
2574 #ifdef CONFIG_NUMA
2575         &nr_hugepages_mempolicy_attr.attr,
2576 #endif
2577         NULL,
2578 };
2579 
2580 static const struct attribute_group hstate_attr_group = {
2581         .attrs = hstate_attrs,
2582 };
2583 
2584 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2585                                     struct kobject **hstate_kobjs,
2586                                     const struct attribute_group *hstate_attr_group)
2587 {
2588         int retval;
2589         int hi = hstate_index(h);
2590 
2591         hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2592         if (!hstate_kobjs[hi])
2593                 return -ENOMEM;
2594 
2595         retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2596         if (retval)
2597                 kobject_put(hstate_kobjs[hi]);
2598 
2599         return retval;
2600 }
2601 
2602 static void __init hugetlb_sysfs_init(void)
2603 {
2604         struct hstate *h;
2605         int err;
2606 
2607         hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2608         if (!hugepages_kobj)
2609                 return;
2610 
2611         for_each_hstate(h) {
2612                 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2613                                          hstate_kobjs, &hstate_attr_group);
2614                 if (err)
2615                         pr_err("Hugetlb: Unable to add hstate %s", h->name);
2616         }
2617 }
2618 
2619 #ifdef CONFIG_NUMA
2620 
2621 /*
2622  * node_hstate/s - associate per node hstate attributes, via their kobjects,
2623  * with node devices in node_devices[] using a parallel array.  The array
2624  * index of a node device or _hstate == node id.
2625  * This is here to avoid any static dependency of the node device driver, in
2626  * the base kernel, on the hugetlb module.
2627  */
2628 struct node_hstate {
2629         struct kobject          *hugepages_kobj;
2630         struct kobject          *hstate_kobjs[HUGE_MAX_HSTATE];
2631 };
2632 static struct node_hstate node_hstates[MAX_NUMNODES];
2633 
2634 /*
2635  * A subset of global hstate attributes for node devices
2636  */
2637 static struct attribute *per_node_hstate_attrs[] = {
2638         &nr_hugepages_attr.attr,
2639         &free_hugepages_attr.attr,
2640         &surplus_hugepages_attr.attr,
2641         NULL,
2642 };
2643 
2644 static const struct attribute_group per_node_hstate_attr_group = {
2645         .attrs = per_node_hstate_attrs,
2646 };
2647 
2648 /*
2649  * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2650  * Returns node id via non-NULL nidp.
2651  */
2652 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2653 {
2654         int nid;
2655 
2656         for (nid = 0; nid < nr_node_ids; nid++) {
2657                 struct node_hstate *nhs = &node_hstates[nid];
2658                 int i;
2659                 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2660                         if (nhs->hstate_kobjs[i] == kobj) {
2661                                 if (nidp)
2662                                         *nidp = nid;
2663                                 return &hstates[i];
2664                         }
2665         }
2666 
2667         BUG();
2668         return NULL;
2669 }
2670 
2671 /*
2672  * Unregister hstate attributes from a single node device.
2673  * No-op if no hstate attributes attached.
2674  */
2675 static void hugetlb_unregister_node(struct node *node)
2676 {
2677         struct hstate *h;
2678         struct node_hstate *nhs = &node_hstates[node->dev.id];
2679 
2680         if (!nhs->hugepages_kobj)
2681                 return;         /* no hstate attributes */
2682 
2683         for_each_hstate(h) {
2684                 int idx = hstate_index(h);
2685                 if (nhs->hstate_kobjs[idx]) {
2686                         kobject_put(nhs->hstate_kobjs[idx]);
2687                         nhs->hstate_kobjs[idx] = NULL;
2688                 }
2689         }
2690 
2691         kobject_put(nhs->hugepages_kobj);
2692         nhs->hugepages_kobj = NULL;
2693 }
2694 
2695 
2696 /*
2697  * Register hstate attributes for a single node device.
2698  * No-op if attributes already registered.
2699  */
2700 static void hugetlb_register_node(struct node *node)
2701 {
2702         struct hstate *h;
2703         struct node_hstate *nhs = &node_hstates[node->dev.id];
2704         int err;
2705 
2706         if (nhs->hugepages_kobj)
2707                 return;         /* already allocated */
2708 
2709         nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2710                                                         &node->dev.kobj);
2711         if (!nhs->hugepages_kobj)
2712                 return;
2713 
2714         for_each_hstate(h) {
2715                 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2716                                                 nhs->hstate_kobjs,
2717                                                 &per_node_hstate_attr_group);
2718                 if (err) {
2719                         pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2720                                 h->name, node->dev.id);
2721                         hugetlb_unregister_node(node);
2722                         break;
2723                 }
2724         }
2725 }
2726 
2727 /*
2728  * hugetlb init time:  register hstate attributes for all registered node
2729  * devices of nodes that have memory.  All on-line nodes should have
2730  * registered their associated device by this time.
2731  */
2732 static void __init hugetlb_register_all_nodes(void)
2733 {
2734         int nid;
2735 
2736         for_each_node_state(nid, N_MEMORY) {
2737                 struct node *node = node_devices[nid];
2738                 if (node->dev.id == nid)
2739                         hugetlb_register_node(node);
2740         }
2741 
2742         /*
2743          * Let the node device driver know we're here so it can
2744          * [un]register hstate attributes on node hotplug.
2745          */
2746         register_hugetlbfs_with_node(hugetlb_register_node,
2747                                      hugetlb_unregister_node);
2748 }
2749 #else   /* !CONFIG_NUMA */
2750 
2751 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2752 {
2753         BUG();
2754         if (nidp)
2755                 *nidp = -1;
2756         return NULL;
2757 }
2758 
2759 static void hugetlb_register_all_nodes(void) { }
2760 
2761 #endif
2762 
2763 static int __init hugetlb_init(void)
2764 {
2765         int i;
2766 
2767         if (!hugepages_supported())
2768                 return 0;
2769 
2770         if (!size_to_hstate(default_hstate_size)) {
2771                 if (default_hstate_size != 0) {
2772                         pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n",
2773                                default_hstate_size, HPAGE_SIZE);
2774                 }
2775 
2776                 default_hstate_size = HPAGE_SIZE;
2777                 if (!size_to_hstate(default_hstate_size))
2778                         hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2779         }
2780         default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2781         if (default_hstate_max_huge_pages) {
2782                 if (!default_hstate.max_huge_pages)
2783                         default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2784         }
2785 
2786         hugetlb_init_hstates();
2787         gather_bootmem_prealloc();
2788         report_hugepages();
2789 
2790         hugetlb_sysfs_init();
2791         hugetlb_register_all_nodes();
2792         hugetlb_cgroup_file_init();
2793 
2794 #ifdef CONFIG_SMP
2795         num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2796 #else
2797         num_fault_mutexes = 1;
2798 #endif
2799         hugetlb_fault_mutex_table =
2800                 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
2801         BUG_ON(!hugetlb_fault_mutex_table);
2802 
2803         for (i = 0; i < num_fault_mutexes; i++)
2804                 mutex_init(&hugetlb_fault_mutex_table[i]);
2805         return 0;
2806 }
2807 subsys_initcall(hugetlb_init);
2808 
2809 /* Should be called on processing a hugepagesz=... option */
2810 void __init hugetlb_bad_size(void)
2811 {
2812         parsed_valid_hugepagesz = false;
2813 }
2814 
2815 void __init hugetlb_add_hstate(unsigned int order)
2816 {
2817         struct hstate *h;
2818         unsigned long i;
2819 
2820         if (size_to_hstate(PAGE_SIZE << order)) {
2821                 pr_warn("hugepagesz= specified twice, ignoring\n");
2822                 return;
2823         }
2824         BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2825         BUG_ON(order == 0);
2826         h = &hstates[hugetlb_max_hstate++];
2827         h->order = order;
2828         h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2829         h->nr_huge_pages = 0;
2830         h->free_huge_pages = 0;
2831         for (i = 0; i < MAX_NUMNODES; ++i)
2832                 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2833         INIT_LIST_HEAD(&h->hugepage_activelist);
2834         h->next_nid_to_alloc = first_memory_node;
2835         h->next_nid_to_free = first_memory_node;
2836         snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2837                                         huge_page_size(h)/1024);
2838 
2839         parsed_hstate = h;
2840 }
2841 
2842 static int __init hugetlb_nrpages_setup(char *s)
2843 {
2844         unsigned long *mhp;
2845         static unsigned long *last_mhp;
2846 
2847         if (!parsed_valid_hugepagesz) {
2848                 pr_warn("hugepages = %s preceded by "
2849                         "an unsupported hugepagesz, ignoring\n", s);
2850                 parsed_valid_hugepagesz = true;
2851                 return 1;
2852         }
2853         /*
2854          * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2855          * so this hugepages= parameter goes to the "default hstate".
2856          */
2857         else if (!hugetlb_max_hstate)
2858                 mhp = &default_hstate_max_huge_pages;
2859         else
2860                 mhp = &parsed_hstate->max_huge_pages;
2861 
2862         if (mhp == last_mhp) {
2863                 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
2864                 return 1;
2865         }
2866 
2867         if (sscanf(s, "%lu", mhp) <= 0)
2868                 *mhp = 0;
2869 
2870         /*
2871          * Global state is always initialized later in hugetlb_init.
2872          * But we need to allocate >= MAX_ORDER hstates here early to still
2873          * use the bootmem allocator.
2874          */
2875         if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2876                 hugetlb_hstate_alloc_pages(parsed_hstate);
2877 
2878         last_mhp = mhp;
2879 
2880         return 1;
2881 }
2882 __setup("hugepages=", hugetlb_nrpages_setup);
2883 
2884 static int __init hugetlb_default_setup(char *s)
2885 {
2886         default_hstate_size = memparse(s, &s);
2887         return 1;
2888 }
2889 __setup("default_hugepagesz=", hugetlb_default_setup);
2890 
2891 static unsigned int cpuset_mems_nr(unsigned int *array)
2892 {
2893         int node;
2894         unsigned int nr = 0;
2895 
2896         for_each_node_mask(node, cpuset_current_mems_allowed)
2897                 nr += array[node];
2898 
2899         return nr;
2900 }
2901 
2902 #ifdef CONFIG_SYSCTL
2903 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2904                          struct ctl_table *table, int write,
2905                          void __user *buffer, size_t *length, loff_t *ppos)
2906 {
2907         struct hstate *h = &default_hstate;
2908         unsigned long tmp = h->max_huge_pages;
2909         int ret;
2910 
2911         if (!hugepages_supported())
2912                 return -EOPNOTSUPP;
2913 
2914         table->data = &tmp;
2915         table->maxlen = sizeof(unsigned long);
2916         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2917         if (ret)
2918                 goto out;
2919 
2920         if (write)
2921                 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2922                                                   NUMA_NO_NODE, tmp, *length);
2923 out:
2924         return ret;
2925 }
2926 
2927 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2928                           void __user *buffer, size_t *length, loff_t *ppos)
2929 {
2930 
2931         return hugetlb_sysctl_handler_common(false, table, write,
2932                                                         buffer, length, ppos);
2933 }
2934 
2935 #ifdef CONFIG_NUMA
2936 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2937                           void __user *buffer, size_t *length, loff_t *ppos)
2938 {
2939         return hugetlb_sysctl_handler_common(true, table, write,
2940                                                         buffer, length, ppos);
2941 }
2942 #endif /* CONFIG_NUMA */
2943 
2944 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2945                         void __user *buffer,
2946                         size_t *length, loff_t *ppos)
2947 {
2948         struct hstate *h = &default_hstate;
2949         unsigned long tmp;
2950         int ret;
2951 
2952         if (!hugepages_supported())
2953                 return -EOPNOTSUPP;
2954 
2955         tmp = h->nr_overcommit_huge_pages;
2956 
2957         if (write && hstate_is_gigantic(h))
2958                 return -EINVAL;
2959 
2960         table->data = &tmp;
2961         table->maxlen = sizeof(unsigned long);
2962         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2963         if (ret)
2964                 goto out;
2965 
2966         if (write) {
2967                 spin_lock(&hugetlb_lock);
2968                 h->nr_overcommit_huge_pages = tmp;
2969                 spin_unlock(&hugetlb_lock);
2970         }
2971 out:
2972         return ret;
2973 }
2974 
2975 #endif /* CONFIG_SYSCTL */
2976 
2977 void hugetlb_report_meminfo(struct seq_file *m)
2978 {
2979         struct hstate *h = &default_hstate;
2980         if (!hugepages_supported())
2981                 return;
2982         seq_printf(m,
2983                         "HugePages_Total:   %5lu\n"
2984                         "HugePages_Free:    %5lu\n"
2985                         "HugePages_Rsvd:    %5lu\n"
2986                         "HugePages_Surp:    %5lu\n"
2987                         "Hugepagesize:   %8lu kB\n",
2988                         h->nr_huge_pages,
2989                         h->free_huge_pages,
2990                         h->resv_huge_pages,
2991                         h->surplus_huge_pages,
2992                         1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2993 }
2994 
2995 int hugetlb_report_node_meminfo(int nid, char *buf)
2996 {
2997         struct hstate *h = &default_hstate;
2998         if (!hugepages_supported())
2999                 return 0;
3000         return sprintf(buf,
3001                 "Node %d HugePages_Total: %5u\n"
3002                 "Node %d HugePages_Free:  %5u\n"
3003                 "Node %d HugePages_Surp:  %5u\n",
3004                 nid, h->nr_huge_pages_node[nid],
3005                 nid, h->free_huge_pages_node[nid],
3006                 nid, h->surplus_huge_pages_node[nid]);
3007 }
3008 
3009 void hugetlb_show_meminfo(void)
3010 {
3011         struct hstate *h;
3012         int nid;
3013 
3014         if (!hugepages_supported())
3015                 return;
3016 
3017         for_each_node_state(nid, N_MEMORY)
3018                 for_each_hstate(h)
3019                         pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3020                                 nid,
3021                                 h->nr_huge_pages_node[nid],
3022                                 h->free_huge_pages_node[nid],
3023                                 h->surplus_huge_pages_node[nid],
3024                                 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3025 }
3026 
3027 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3028 {
3029         seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3030                    atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3031 }
3032 
3033 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3034 unsigned long hugetlb_total_pages(void)
3035 {
3036         struct hstate *h;
3037         unsigned long nr_total_pages = 0;
3038 
3039         for_each_hstate(h)
3040                 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3041         return nr_total_pages;
3042 }
3043 
3044 static int hugetlb_acct_memory(struct hstate *h, long delta)
3045 {
3046         int ret = -ENOMEM;
3047 
3048         spin_lock(&hugetlb_lock);
3049         /*
3050          * When cpuset is configured, it breaks the strict hugetlb page
3051          * reservation as the accounting is done on a global variable. Such
3052          * reservation is completely rubbish in the presence of cpuset because
3053          * the reservation is not checked against page availability for the
3054          * current cpuset. Application can still potentially OOM'ed by kernel
3055          * with lack of free htlb page in cpuset that the task is in.
3056          * Attempt to enforce strict accounting with cpuset is almost
3057          * impossible (or too ugly) because cpuset is too fluid that
3058          * task or memory node can be dynamically moved between cpusets.
3059          *
3060          * The change of semantics for shared hugetlb mapping with cpuset is
3061          * undesirable. However, in order to preserve some of the semantics,
3062          * we fall back to check against current free page availability as
3063          * a best attempt and hopefully to minimize the impact of changing
3064          * semantics that cpuset has.
3065          */
3066         if (delta > 0) {
3067                 if (gather_surplus_pages(h, delta) < 0)
3068                         goto out;
3069 
3070                 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
3071                         return_unused_surplus_pages(h, delta);
3072                         goto out;
3073                 }
3074         }
3075 
3076         ret = 0;
3077         if (delta < 0)
3078                 return_unused_surplus_pages(h, (unsigned long) -delta);
3079 
3080 out:
3081         spin_unlock(&hugetlb_lock);
3082         return ret;
3083 }
3084 
3085 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3086 {
3087         struct resv_map *resv = vma_resv_map(vma);
3088 
3089         /*
3090          * This new VMA should share its siblings reservation map if present.
3091          * The VMA will only ever have a valid reservation map pointer where
3092          * it is being copied for another still existing VMA.  As that VMA
3093          * has a reference to the reservation map it cannot disappear until
3094          * after this open call completes.  It is therefore safe to take a
3095          * new reference here without additional locking.
3096          */
3097         if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3098                 kref_get(&resv->refs);
3099 }
3100 
3101 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3102 {
3103         struct hstate *h = hstate_vma(vma);
3104         struct resv_map *resv = vma_resv_map(vma);
3105         struct hugepage_subpool *spool = subpool_vma(vma);
3106         unsigned long reserve, start, end;
3107         long gbl_reserve;
3108 
3109         if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3110                 return;
3111 
3112         start = vma_hugecache_offset(h, vma, vma->vm_start);
3113         end = vma_hugecache_offset(h, vma, vma->vm_end);
3114 
3115         reserve = (end - start) - region_count(resv, start, end);
3116 
3117         kref_put(&resv->refs, resv_map_release);
3118 
3119         if (reserve) {
3120                 /*
3121                  * Decrement reserve counts.  The global reserve count may be
3122                  * adjusted if the subpool has a minimum size.
3123                  */
3124                 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3125                 hugetlb_acct_memory(h, -gbl_reserve);
3126         }
3127 }
3128 
3129 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3130 {
3131         if (addr & ~(huge_page_mask(hstate_vma(vma))))
3132                 return -EINVAL;
3133         return 0;
3134 }
3135 
3136 /*
3137  * We cannot handle pagefaults against hugetlb pages at all.  They cause
3138  * handle_mm_fault() to try to instantiate regular-sized pages in the
3139  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
3140  * this far.
3141  */
3142 static int hugetlb_vm_op_fault(struct vm_fault *vmf)
3143 {
3144         BUG();
3145         return 0;
3146 }
3147 
3148 const struct vm_operations_struct hugetlb_vm_ops = {
3149         .fault = hugetlb_vm_op_fault,
3150         .open = hugetlb_vm_op_open,
3151         .close = hugetlb_vm_op_close,
3152         .split = hugetlb_vm_op_split,
3153 };
3154 
3155 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3156                                 int writable)
3157 {
3158         pte_t entry;
3159 
3160         if (writable) {
3161                 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3162                                          vma->vm_page_prot)));
3163         } else {
3164                 entry = huge_pte_wrprotect(mk_huge_pte(page,
3165                                            vma->vm_page_prot));
3166         }
3167         entry = pte_mkyoung(entry);
3168         entry = pte_mkhuge(entry);
3169         entry = arch_make_huge_pte(entry, vma, page, writable);
3170 
3171         return entry;
3172 }
3173 
3174 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3175                                    unsigned long address, pte_t *ptep)
3176 {
3177         pte_t entry;
3178 
3179         entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3180         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3181                 update_mmu_cache(vma, address, ptep);
3182 }
3183 
3184 bool is_hugetlb_entry_migration(pte_t pte)
3185 {
3186         swp_entry_t swp;
3187 
3188         if (huge_pte_none(pte) || pte_present(pte))
3189                 return false;
3190         swp = pte_to_swp_entry(pte);
3191         if (non_swap_entry(swp) && is_migration_entry(swp))
3192                 return true;
3193         else
3194                 return false;
3195 }
3196 
3197 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3198 {
3199         swp_entry_t swp;
3200 
3201         if (huge_pte_none(pte) || pte_present(pte))
3202                 return 0;
3203         swp = pte_to_swp_entry(pte);
3204         if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3205                 return 1;
3206         else
3207                 return 0;
3208 }
3209 
3210 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3211                             struct vm_area_struct *vma)
3212 {
3213         pte_t *src_pte, *dst_pte, entry;
3214         struct page *ptepage;
3215         unsigned long addr;
3216         int cow;
3217         struct hstate *h = hstate_vma(vma);
3218         unsigned long sz = huge_page_size(h);
3219         unsigned long mmun_start;       /* For mmu_notifiers */
3220         unsigned long mmun_end;         /* For mmu_notifiers */
3221         int ret = 0;
3222 
3223         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3224 
3225         mmun_start = vma->vm_start;
3226         mmun_end = vma->vm_end;
3227         if (cow)
3228                 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
3229 
3230         for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3231                 spinlock_t *src_ptl, *dst_ptl;
3232                 src_pte = huge_pte_offset(src, addr, sz);
3233                 if (!src_pte)
3234                         continue;
3235                 dst_pte = huge_pte_alloc(dst, addr, sz);
3236                 if (!dst_pte) {
3237                         ret = -ENOMEM;
3238                         break;
3239                 }
3240 
3241                 /* If the pagetables are shared don't copy or take references */
3242                 if (dst_pte == src_pte)
3243                         continue;
3244 
3245                 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3246                 src_ptl = huge_pte_lockptr(h, src, src_pte);
3247                 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3248                 entry = huge_ptep_get(src_pte);
3249                 if (huge_pte_none(entry)) { /* skip none entry */
3250                         ;
3251                 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3252                                     is_hugetlb_entry_hwpoisoned(entry))) {
3253                         swp_entry_t swp_entry = pte_to_swp_entry(entry);
3254 
3255                         if (is_write_migration_entry(swp_entry) && cow) {
3256                                 /*
3257                                  * COW mappings require pages in both
3258                                  * parent and child to be set to read.
3259                                  */
3260                                 make_migration_entry_read(&swp_entry);
3261                                 entry = swp_entry_to_pte(swp_entry);
3262                                 set_huge_swap_pte_at(src, addr, src_pte,
3263                                                      entry, sz);
3264                         }
3265                         set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3266                 } else {
3267                         if (cow) {
3268                                 /*
3269                                  * No need to notify as we are downgrading page
3270                                  * table protection not changing it to point
3271                                  * to a new page.
3272                                  *
3273                                  * See Documentation/vm/mmu_notifier.txt
3274                                  */
3275                                 huge_ptep_set_wrprotect(src, addr, src_pte);
3276                         }
3277                         entry = huge_ptep_get(src_pte);
3278                         ptepage = pte_page(entry);
3279                         get_page(ptepage);
3280                         page_dup_rmap(ptepage, true);
3281                         set_huge_pte_at(dst, addr, dst_pte, entry);
3282                         hugetlb_count_add(pages_per_huge_page(h), dst);
3283                 }
3284                 spin_unlock(src_ptl);
3285                 spin_unlock(dst_ptl);
3286         }
3287 
3288         if (cow)
3289                 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
3290 
3291         return ret;
3292 }
3293 
3294 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3295                             unsigned long start, unsigned long end,
3296                             struct page *ref_page)
3297 {
3298         struct mm_struct *mm = vma->vm_mm;
3299         unsigned long address;
3300         pte_t *ptep;
3301         pte_t pte;
3302         spinlock_t *ptl;
3303         struct page *page;
3304         struct hstate *h = hstate_vma(vma);
3305         unsigned long sz = huge_page_size(h);
3306         const unsigned long mmun_start = start; /* For mmu_notifiers */
3307         const unsigned long mmun_end   = end;   /* For mmu_notifiers */
3308 
3309         WARN_ON(!is_vm_hugetlb_page(vma));
3310         BUG_ON(start & ~huge_page_mask(h));
3311         BUG_ON(end & ~huge_page_mask(h));
3312 
3313         /*
3314          * This is a hugetlb vma, all the pte entries should point
3315          * to huge page.
3316          */
3317         tlb_remove_check_page_size_change(tlb, sz);
3318         tlb_start_vma(tlb, vma);
3319         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3320         address = start;
3321         for (; address < end; address += sz) {
3322                 ptep = huge_pte_offset(mm, address, sz);
3323                 if (!ptep)
3324                         continue;
3325 
3326                 ptl = huge_pte_lock(h, mm, ptep);
3327                 if (huge_pmd_unshare(mm, &address, ptep)) {
3328                         spin_unlock(ptl);
3329                         continue;
3330                 }
3331 
3332                 pte = huge_ptep_get(ptep);
3333                 if (huge_pte_none(pte)) {
3334                         spin_unlock(ptl);
3335                         continue;
3336                 }
3337 
3338                 /*
3339                  * Migrating hugepage or HWPoisoned hugepage is already
3340                  * unmapped and its refcount is dropped, so just clear pte here.
3341                  */
3342                 if (unlikely(!pte_present(pte))) {
3343                         huge_pte_clear(mm, address, ptep, sz);
3344                         spin_unlock(ptl);
3345                         continue;
3346                 }
3347 
3348                 page = pte_page(pte);
3349                 /*
3350                  * If a reference page is supplied, it is because a specific
3351                  * page is being unmapped, not a range. Ensure the page we
3352                  * are about to unmap is the actual page of interest.
3353                  */
3354                 if (ref_page) {
3355                         if (page != ref_page) {
3356                                 spin_unlock(ptl);
3357                                 continue;
3358                         }
3359                         /*
3360                          * Mark the VMA as having unmapped its page so that
3361                          * future faults in this VMA will fail rather than
3362                          * looking like data was lost
3363                          */
3364                         set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3365                 }
3366 
3367                 pte = huge_ptep_get_and_clear(mm, address, ptep);
3368                 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
3369                 if (huge_pte_dirty(pte))
3370                         set_page_dirty(page);
3371 
3372                 hugetlb_count_sub(pages_per_huge_page(h), mm);
3373                 page_remove_rmap(page, true);
3374 
3375                 spin_unlock(ptl);
3376                 tlb_remove_page_size(tlb, page, huge_page_size(h));
3377                 /*
3378                  * Bail out after unmapping reference page if supplied
3379                  */
3380                 if (ref_page)
3381                         break;
3382         }
3383         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3384         tlb_end_vma(tlb, vma);
3385 }
3386 
3387 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3388                           struct vm_area_struct *vma, unsigned long start,
3389                           unsigned long end, struct page *ref_page)
3390 {
3391         __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3392 
3393         /*
3394          * Clear this flag so that x86's huge_pmd_share page_table_shareable
3395          * test will fail on a vma being torn down, and not grab a page table
3396          * on its way out.  We're lucky that the flag has such an appropriate
3397          * name, and can in fact be safely cleared here. We could clear it
3398          * before the __unmap_hugepage_range above, but all that's necessary
3399          * is to clear it before releasing the i_mmap_rwsem. This works
3400          * because in the context this is called, the VMA is about to be
3401          * destroyed and the i_mmap_rwsem is held.
3402          */
3403         vma->vm_flags &= ~VM_MAYSHARE;
3404 }
3405 
3406 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3407                           unsigned long end, struct page *ref_page)
3408 {
3409         struct mm_struct *mm;
3410         struct mmu_gather tlb;
3411 
3412         mm = vma->vm_mm;
3413 
3414         tlb_gather_mmu(&tlb, mm, start, end);
3415         __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3416         tlb_finish_mmu(&tlb, start, end);
3417 }
3418 
3419 /*
3420  * This is called when the original mapper is failing to COW a MAP_PRIVATE
3421  * mappping it owns the reserve page for. The intention is to unmap the page
3422  * from other VMAs and let the children be SIGKILLed if they are faulting the
3423  * same region.
3424  */
3425 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3426                               struct page *page, unsigned long address)
3427 {
3428         struct hstate *h = hstate_vma(vma);
3429         struct vm_area_struct *iter_vma;
3430         struct address_space *mapping;
3431         pgoff_t pgoff;
3432 
3433         /*
3434          * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3435          * from page cache lookup which is in HPAGE_SIZE units.
3436          */
3437         address = address & huge_page_mask(h);
3438         pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3439                         vma->vm_pgoff;
3440         mapping = vma->vm_file->f_mapping;
3441 
3442         /*
3443          * Take the mapping lock for the duration of the table walk. As
3444          * this mapping should be shared between all the VMAs,
3445          * __unmap_hugepage_range() is called as the lock is already held
3446          */
3447         i_mmap_lock_write(mapping);
3448         vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3449                 /* Do not unmap the current VMA */
3450                 if (iter_vma == vma)
3451                         continue;
3452 
3453                 /*
3454                  * Shared VMAs have their own reserves and do not affect
3455                  * MAP_PRIVATE accounting but it is possible that a shared
3456                  * VMA is using the same page so check and skip such VMAs.
3457                  */
3458                 if (iter_vma->vm_flags & VM_MAYSHARE)
3459                         continue;
3460 
3461                 /*
3462                  * Unmap the page from other VMAs without their own reserves.
3463                  * They get marked to be SIGKILLed if they fault in these
3464                  * areas. This is because a future no-page fault on this VMA
3465                  * could insert a zeroed page instead of the data existing
3466                  * from the time of fork. This would look like data corruption
3467                  */
3468                 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3469                         unmap_hugepage_range(iter_vma, address,
3470                                              address + huge_page_size(h), page);
3471         }
3472         i_mmap_unlock_write(mapping);
3473 }
3474 
3475 /*
3476  * Hugetlb_cow() should be called with page lock of the original hugepage held.
3477  * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3478  * cannot race with other handlers or page migration.
3479  * Keep the pte_same checks anyway to make transition from the mutex easier.
3480  */
3481 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3482                        unsigned long address, pte_t *ptep,
3483                        struct page *pagecache_page, spinlock_t *ptl)
3484 {
3485         pte_t pte;
3486         struct hstate *h = hstate_vma(vma);
3487         struct page *old_page, *new_page;
3488         int ret = 0, outside_reserve = 0;
3489         unsigned long mmun_start;       /* For mmu_notifiers */
3490         unsigned long mmun_end;         /* For mmu_notifiers */
3491 
3492         pte = huge_ptep_get(ptep);
3493         old_page = pte_page(pte);
3494 
3495 retry_avoidcopy:
3496         /* If no-one else is actually using this page, avoid the copy
3497          * and just make the page writable */
3498         if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3499                 page_move_anon_rmap(old_page, vma);
3500                 set_huge_ptep_writable(vma, address, ptep);
3501                 return 0;
3502         }
3503 
3504         /*
3505          * If the process that created a MAP_PRIVATE mapping is about to
3506          * perform a COW due to a shared page count, attempt to satisfy
3507          * the allocation without using the existing reserves. The pagecache
3508          * page is used to determine if the reserve at this address was
3509          * consumed or not. If reserves were used, a partial faulted mapping
3510          * at the time of fork() could consume its reserves on COW instead
3511          * of the full address range.
3512          */
3513         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3514                         old_page != pagecache_page)
3515                 outside_reserve = 1;
3516 
3517         get_page(old_page);
3518 
3519         /*
3520          * Drop page table lock as buddy allocator may be called. It will
3521          * be acquired again before returning to the caller, as expected.
3522          */
3523         spin_unlock(ptl);
3524         new_page = alloc_huge_page(vma, address, outside_reserve);
3525 
3526         if (IS_ERR(new_page)) {
3527                 /*
3528                  * If a process owning a MAP_PRIVATE mapping fails to COW,
3529                  * it is due to references held by a child and an insufficient
3530                  * huge page pool. To guarantee the original mappers
3531                  * reliability, unmap the page from child processes. The child
3532                  * may get SIGKILLed if it later faults.
3533                  */
3534                 if (outside_reserve) {
3535                         put_page(old_page);
3536                         BUG_ON(huge_pte_none(pte));
3537                         unmap_ref_private(mm, vma, old_page, address);
3538                         BUG_ON(huge_pte_none(pte));
3539                         spin_lock(ptl);
3540                         ptep = huge_pte_offset(mm, address & huge_page_mask(h),
3541                                                huge_page_size(h));
3542                         if (likely(ptep &&
3543                                    pte_same(huge_ptep_get(ptep), pte)))
3544                                 goto retry_avoidcopy;
3545                         /*
3546                          * race occurs while re-acquiring page table
3547                          * lock, and our job is done.
3548                          */
3549                         return 0;
3550                 }
3551 
3552                 ret = (PTR_ERR(new_page) == -ENOMEM) ?
3553                         VM_FAULT_OOM : VM_FAULT_SIGBUS;
3554                 goto out_release_old;
3555         }
3556 
3557         /*
3558          * When the original hugepage is shared one, it does not have
3559          * anon_vma prepared.
3560          */
3561         if (unlikely(anon_vma_prepare(vma))) {
3562                 ret = VM_FAULT_OOM;
3563                 goto out_release_all;
3564         }
3565 
3566         copy_user_huge_page(new_page, old_page, address, vma,
3567                             pages_per_huge_page(h));
3568         __SetPageUptodate(new_page);
3569         set_page_huge_active(new_page);
3570 
3571         mmun_start = address & huge_page_mask(h);
3572         mmun_end = mmun_start + huge_page_size(h);
3573         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3574 
3575         /*
3576          * Retake the page table lock to check for racing updates
3577          * before the page tables are altered
3578          */
3579         spin_lock(ptl);
3580         ptep = huge_pte_offset(mm, address & huge_page_mask(h),
3581                                huge_page_size(h));
3582         if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3583                 ClearPagePrivate(new_page);
3584 
3585                 /* Break COW */
3586                 huge_ptep_clear_flush(vma, address, ptep);
3587                 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
3588                 set_huge_pte_at(mm, address, ptep,
3589                                 make_huge_pte(vma, new_page, 1));
3590                 page_remove_rmap(old_page, true);
3591                 hugepage_add_new_anon_rmap(new_page, vma, address);
3592                 /* Make the old page be freed below */
3593                 new_page = old_page;
3594         }
3595         spin_unlock(ptl);
3596         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3597 out_release_all:
3598         restore_reserve_on_error(h, vma, address, new_page);
3599         put_page(new_page);
3600 out_release_old:
3601         put_page(old_page);
3602 
3603         spin_lock(ptl); /* Caller expects lock to be held */
3604         return ret;
3605 }
3606 
3607 /* Return the pagecache page at a given address within a VMA */
3608 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3609                         struct vm_area_struct *vma, unsigned long address)
3610 {
3611         struct address_space *mapping;
3612         pgoff_t idx;
3613 
3614         mapping = vma->vm_file->f_mapping;
3615         idx = vma_hugecache_offset(h, vma, address);
3616 
3617         return find_lock_page(mapping, idx);
3618 }
3619 
3620 /*
3621  * Return whether there is a pagecache page to back given address within VMA.
3622  * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3623  */
3624 static bool hugetlbfs_pagecache_present(struct hstate *h,
3625                         struct vm_area_struct *vma, unsigned long address)
3626 {
3627         struct address_space *mapping;
3628         pgoff_t idx;
3629         struct page *page;
3630 
3631         mapping = vma->vm_file->f_mapping;
3632         idx = vma_hugecache_offset(h, vma, address);
3633 
3634         page = find_get_page(mapping, idx);
3635         if (page)
3636                 put_page(page);
3637         return page != NULL;
3638 }
3639 
3640 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3641                            pgoff_t idx)
3642 {
3643         struct inode *inode = mapping->host;
3644         struct hstate *h = hstate_inode(inode);
3645         int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3646 
3647         if (err)
3648                 return err;
3649         ClearPagePrivate(page);
3650 
3651         spin_lock(&inode->i_lock);
3652         inode->i_blocks += blocks_per_huge_page(h);
3653         spin_unlock(&inode->i_lock);
3654         return 0;
3655 }
3656 
3657 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
3658                            struct address_space *mapping, pgoff_t idx,
3659                            unsigned long address, pte_t *ptep, unsigned int flags)
3660 {
3661         struct hstate *h = hstate_vma(vma);
3662         int ret = VM_FAULT_SIGBUS;
3663         int anon_rmap = 0;
3664         unsigned long size;
3665         struct page *page;
3666         pte_t new_pte;
3667         spinlock_t *ptl;
3668 
3669         /*
3670          * Currently, we are forced to kill the process in the event the
3671          * original mapper has unmapped pages from the child due to a failed
3672          * COW. Warn that such a situation has occurred as it may not be obvious
3673          */
3674         if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3675                 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
3676                            current->pid);
3677                 return ret;
3678         }
3679 
3680         /*
3681          * Use page lock to guard against racing truncation
3682          * before we get page_table_lock.
3683          */
3684 retry:
3685         page = find_lock_page(mapping, idx);
3686         if (!page) {
3687                 size = i_size_read(mapping->host) >> huge_page_shift(h);
3688                 if (idx >= size)
3689                         goto out;
3690 
3691                 /*
3692                  * Check for page in userfault range
3693                  */
3694                 if (userfaultfd_missing(vma)) {
3695                         u32 hash;
3696                         struct vm_fault vmf = {
3697                                 .vma = vma,
3698                                 .address = address,
3699                                 .flags = flags,
3700                                 /*
3701                                  * Hard to debug if it ends up being
3702                                  * used by a callee that assumes
3703                                  * something about the other
3704                                  * uninitialized fields... same as in
3705                                  * memory.c
3706                                  */
3707                         };
3708 
3709                         /*
3710                          * hugetlb_fault_mutex must be dropped before
3711                          * handling userfault.  Reacquire after handling
3712                          * fault to make calling code simpler.
3713                          */
3714                         hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping,
3715                                                         idx, address);
3716                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3717                         ret = handle_userfault(&vmf, VM_UFFD_MISSING);
3718                         mutex_lock(&hugetlb_fault_mutex_table[hash]);
3719                         goto out;
3720                 }
3721 
3722                 page = alloc_huge_page(vma, address, 0);
3723                 if (IS_ERR(page)) {
3724                         ret = PTR_ERR(page);
3725                         if (ret == -ENOMEM)
3726                                 ret = VM_FAULT_OOM;
3727                         else
3728                                 ret = VM_FAULT_SIGBUS;
3729                         goto out;
3730                 }
3731                 clear_huge_page(page, address, pages_per_huge_page(h));
3732                 __SetPageUptodate(page);
3733                 set_page_huge_active(page);
3734 
3735                 if (vma->vm_flags & VM_MAYSHARE) {
3736                         int err = huge_add_to_page_cache(page, mapping, idx);
3737                         if (err) {
3738                                 put_page(page);
3739                                 if (err == -EEXIST)
3740                                         goto retry;
3741                                 goto out;
3742                         }
3743                 } else {
3744                         lock_page(page);
3745                         if (unlikely(anon_vma_prepare(vma))) {
3746                                 ret = VM_FAULT_OOM;
3747                                 goto backout_unlocked;
3748                         }
3749                         anon_rmap = 1;
3750                 }
3751         } else {
3752                 /*
3753                  * If memory error occurs between mmap() and fault, some process
3754                  * don't have hwpoisoned swap entry for errored virtual address.
3755                  * So we need to block hugepage fault by PG_hwpoison bit check.
3756                  */
3757                 if (unlikely(PageHWPoison(page))) {
3758                         ret = VM_FAULT_HWPOISON |
3759                                 VM_FAULT_SET_HINDEX(hstate_index(h));
3760                         goto backout_unlocked;
3761                 }
3762         }
3763 
3764         /*
3765          * If we are going to COW a private mapping later, we examine the
3766          * pending reservations for this page now. This will ensure that
3767          * any allocations necessary to record that reservation occur outside
3768          * the spinlock.
3769          */
3770         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3771                 if (vma_needs_reservation(h, vma, address) < 0) {
3772                         ret = VM_FAULT_OOM;
3773                         goto backout_unlocked;
3774                 }
3775                 /* Just decrements count, does not deallocate */
3776                 vma_end_reservation(h, vma, address);
3777         }
3778 
3779         ptl = huge_pte_lock(h, mm, ptep);
3780         size = i_size_read(mapping->host) >> huge_page_shift(h);
3781         if (idx >= size)
3782                 goto backout;
3783 
3784         ret = 0;
3785         if (!huge_pte_none(huge_ptep_get(ptep)))
3786                 goto backout;
3787 
3788         if (anon_rmap) {
3789                 ClearPagePrivate(page);
3790                 hugepage_add_new_anon_rmap(page, vma, address);
3791         } else
3792                 page_dup_rmap(page, true);
3793         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3794                                 && (vma->vm_flags & VM_SHARED)));
3795         set_huge_pte_at(mm, address, ptep, new_pte);
3796 
3797         hugetlb_count_add(pages_per_huge_page(h), mm);
3798         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3799                 /* Optimization, do the COW without a second fault */
3800                 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
3801         }
3802 
3803         spin_unlock(ptl);
3804         unlock_page(page);
3805 out:
3806         return ret;
3807 
3808 backout:
3809         spin_unlock(ptl);
3810 backout_unlocked:
3811         unlock_page(page);
3812         restore_reserve_on_error(h, vma, address, page);
3813         put_page(page);
3814         goto out;
3815 }
3816 
3817 #ifdef CONFIG_SMP
3818 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3819                             struct vm_area_struct *vma,
3820                             struct address_space *mapping,
3821                             pgoff_t idx, unsigned long address)
3822 {
3823         unsigned long key[2];
3824         u32 hash;
3825 
3826         if (vma->vm_flags & VM_SHARED) {
3827                 key[0] = (unsigned long) mapping;
3828                 key[1] = idx;
3829         } else {
3830                 key[0] = (unsigned long) mm;
3831                 key[1] = address >> huge_page_shift(h);
3832         }
3833 
3834         hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3835 
3836         return hash & (num_fault_mutexes - 1);
3837 }
3838 #else
3839 /*
3840  * For uniprocesor systems we always use a single mutex, so just
3841  * return 0 and avoid the hashing overhead.
3842  */
3843 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3844                             struct vm_area_struct *vma,
3845                             struct address_space *mapping,
3846                             pgoff_t idx, unsigned long address)
3847 {
3848         return 0;
3849 }
3850 #endif
3851 
3852 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3853                         unsigned long address, unsigned int flags)
3854 {
3855         pte_t *ptep, entry;
3856         spinlock_t *ptl;
3857         int ret;
3858         u32 hash;
3859         pgoff_t idx;
3860         struct page *page = NULL;
3861         struct page *pagecache_page = NULL;
3862         struct hstate *h = hstate_vma(vma);
3863         struct address_space *mapping;
3864         int need_wait_lock = 0;
3865 
3866         address &= huge_page_mask(h);
3867 
3868         ptep = huge_pte_offset(mm, address, huge_page_size(h));
3869         if (ptep) {
3870                 entry = huge_ptep_get(ptep);
3871                 if (unlikely(is_hugetlb_entry_migration(entry))) {
3872                         migration_entry_wait_huge(vma, mm, ptep);
3873                         return 0;
3874                 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
3875                         return VM_FAULT_HWPOISON_LARGE |
3876                                 VM_FAULT_SET_HINDEX(hstate_index(h));
3877         } else {
3878                 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
3879                 if (!ptep)
3880                         return VM_FAULT_OOM;
3881         }
3882 
3883         mapping = vma->vm_file->f_mapping;
3884         idx = vma_hugecache_offset(h, vma, address);
3885 
3886         /*
3887          * Serialize hugepage allocation and instantiation, so that we don't
3888          * get spurious allocation failures if two CPUs race to instantiate
3889          * the same page in the page cache.
3890          */
3891         hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address);
3892         mutex_lock(&hugetlb_fault_mutex_table[hash]);
3893 
3894         entry = huge_ptep_get(ptep);
3895         if (huge_pte_none(entry)) {
3896                 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
3897                 goto out_mutex;
3898         }
3899 
3900         ret = 0;
3901 
3902         /*
3903          * entry could be a migration/hwpoison entry at this point, so this
3904          * check prevents the kernel from going below assuming that we have
3905          * a active hugepage in pagecache. This goto expects the 2nd page fault,
3906          * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3907          * handle it.
3908          */
3909         if (!pte_present(entry))
3910                 goto out_mutex;
3911 
3912         /*
3913          * If we are going to COW the mapping later, we examine the pending
3914          * reservations for this page now. This will ensure that any
3915          * allocations necessary to record that reservation occur outside the
3916          * spinlock. For private mappings, we also lookup the pagecache
3917          * page now as it is used to determine if a reservation has been
3918          * consumed.
3919          */
3920         if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
3921                 if (vma_needs_reservation(h, vma, address) < 0) {
3922                         ret = VM_FAULT_OOM;
3923                         goto out_mutex;
3924                 }
3925                 /* Just decrements count, does not deallocate */
3926                 vma_end_reservation(h, vma, address);
3927 
3928                 if (!(vma->vm_flags & VM_MAYSHARE))
3929                         pagecache_page = hugetlbfs_pagecache_page(h,
3930                                                                 vma, address);
3931         }
3932 
3933         ptl = huge_pte_lock(h, mm, ptep);
3934 
3935         /* Check for a racing update before calling hugetlb_cow */
3936         if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
3937                 goto out_ptl;
3938 
3939         /*
3940          * hugetlb_cow() requires page locks of pte_page(entry) and
3941          * pagecache_page, so here we need take the former one
3942          * when page != pagecache_page or !pagecache_page.
3943          */
3944         page = pte_page(entry);
3945         if (page != pagecache_page)
3946                 if (!trylock_page(page)) {
3947                         need_wait_lock = 1;
3948                         goto out_ptl;
3949                 }
3950 
3951         get_page(page);
3952 
3953         if (flags & FAULT_FLAG_WRITE) {
3954                 if (!huge_pte_write(entry)) {
3955                         ret = hugetlb_cow(mm, vma, address, ptep,
3956                                           pagecache_page, ptl);
3957                         goto out_put_page;
3958                 }
3959                 entry = huge_pte_mkdirty(entry);
3960         }
3961         entry = pte_mkyoung(entry);
3962         if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3963                                                 flags & FAULT_FLAG_WRITE))
3964                 update_mmu_cache(vma, address, ptep);
3965 out_put_page:
3966         if (page != pagecache_page)
3967                 unlock_page(page);
3968         put_page(page);
3969 out_ptl:
3970         spin_unlock(ptl);
3971 
3972         if (pagecache_page) {
3973                 unlock_page(pagecache_page);
3974                 put_page(pagecache_page);
3975         }
3976 out_mutex:
3977         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3978         /*
3979          * Generally it's safe to hold refcount during waiting page lock. But
3980          * here we just wait to defer the next page fault to avoid busy loop and
3981          * the page is not used after unlocked before returning from the current
3982          * page fault. So we are safe from accessing freed page, even if we wait
3983          * here without taking refcount.
3984          */
3985         if (need_wait_lock)
3986                 wait_on_page_locked(page);
3987         return ret;
3988 }
3989 
3990 /*
3991  * Used by userfaultfd UFFDIO_COPY.  Based on mcopy_atomic_pte with
3992  * modifications for huge pages.
3993  */
3994 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
3995                             pte_t *dst_pte,
3996                             struct vm_area_struct *dst_vma,
3997                             unsigned long dst_addr,
3998                             unsigned long src_addr,
3999                             struct page **pagep)
4000 {
4001         struct address_space *mapping;
4002         pgoff_t idx;
4003         unsigned long size;
4004         int vm_shared = dst_vma->vm_flags & VM_SHARED;
4005         struct hstate *h = hstate_vma(dst_vma);
4006         pte_t _dst_pte;
4007         spinlock_t *ptl;
4008         int ret;
4009         struct page *page;
4010 
4011         if (!*pagep) {
4012                 ret = -ENOMEM;
4013                 page = alloc_huge_page(dst_vma, dst_addr, 0);
4014                 if (IS_ERR(page))
4015                         goto out;
4016 
4017                 ret = copy_huge_page_from_user(page,
4018                                                 (const void __user *) src_addr,
4019                                                 pages_per_huge_page(h), false);
4020 
4021                 /* fallback to copy_from_user outside mmap_sem */
4022                 if (unlikely(ret)) {
4023                         ret = -EFAULT;
4024                         *pagep = page;
4025                         /* don't free the page */
4026                         goto out;
4027                 }
4028         } else {
4029                 page = *pagep;
4030                 *pagep = NULL;
4031         }
4032 
4033         /*
4034          * The memory barrier inside __SetPageUptodate makes sure that
4035          * preceding stores to the page contents become visible before
4036          * the set_pte_at() write.
4037          */
4038         __SetPageUptodate(page);
4039         set_page_huge_active(page);
4040 
4041         mapping = dst_vma->vm_file->f_mapping;
4042         idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4043 
4044         /*
4045          * If shared, add to page cache
4046          */
4047         if (vm_shared) {
4048                 size = i_size_read(mapping->host) >> huge_page_shift(h);
4049                 ret = -EFAULT;
4050                 if (idx >= size)
4051                         goto out_release_nounlock;
4052 
4053                 /*
4054                  * Serialization between remove_inode_hugepages() and
4055                  * huge_add_to_page_cache() below happens through the
4056                  * hugetlb_fault_mutex_table that here must be hold by
4057                  * the caller.
4058                  */
4059                 ret = huge_add_to_page_cache(page, mapping, idx);
4060                 if (ret)
4061                         goto out_release_nounlock;
4062         }
4063 
4064         ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4065         spin_lock(ptl);
4066 
4067         /*
4068          * Recheck the i_size after holding PT lock to make sure not
4069          * to leave any page mapped (as page_mapped()) beyond the end
4070          * of the i_size (remove_inode_hugepages() is strict about
4071          * enforcing that). If we bail out here, we'll also leave a
4072          * page in the radix tree in the vm_shared case beyond the end
4073          * of the i_size, but remove_inode_hugepages() will take care
4074          * of it as soon as we drop the hugetlb_fault_mutex_table.
4075          */
4076         size = i_size_read(mapping->host) >> huge_page_shift(h);
4077         ret = -EFAULT;
4078         if (idx >= size)
4079                 goto out_release_unlock;
4080 
4081         ret = -EEXIST;
4082         if (!huge_pte_none(huge_ptep_get(dst_pte)))
4083                 goto out_release_unlock;
4084 
4085         if (vm_shared) {
4086                 page_dup_rmap(page, true);
4087         } else {
4088                 ClearPagePrivate(page);
4089                 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4090         }
4091 
4092         _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4093         if (dst_vma->vm_flags & VM_WRITE)
4094                 _dst_pte = huge_pte_mkdirty(_dst_pte);
4095         _dst_pte = pte_mkyoung(_dst_pte);
4096 
4097         set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4098 
4099         (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4100                                         dst_vma->vm_flags & VM_WRITE);
4101         hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4102 
4103         /* No need to invalidate - it was non-present before */
4104         update_mmu_cache(dst_vma, dst_addr, dst_pte);
4105 
4106         spin_unlock(ptl);
4107         if (vm_shared)
4108                 unlock_page(page);
4109         ret = 0;
4110 out:
4111         return ret;
4112 out_release_unlock:
4113         spin_unlock(ptl);
4114         if (vm_shared)
4115                 unlock_page(page);
4116 out_release_nounlock:
4117         put_page(page);
4118         goto out;
4119 }
4120 
4121 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4122                          struct page **pages, struct vm_area_struct **vmas,
4123                          unsigned long *position, unsigned long *nr_pages,
4124                          long i, unsigned int flags, int *nonblocking)
4125 {
4126         unsigned long pfn_offset;
4127         unsigned long vaddr = *position;
4128         unsigned long remainder = *nr_pages;
4129         struct hstate *h = hstate_vma(vma);
4130         int err = -EFAULT;
4131 
4132         while (vaddr < vma->vm_end && remainder) {
4133                 pte_t *pte;
4134                 spinlock_t *ptl = NULL;
4135                 int absent;
4136                 struct page *page;
4137 
4138                 /*
4139                  * If we have a pending SIGKILL, don't keep faulting pages and
4140                  * potentially allocating memory.
4141                  */
4142                 if (unlikely(fatal_signal_pending(current))) {
4143                         remainder = 0;
4144                         break;
4145                 }
4146 
4147                 /*
4148                  * Some archs (sparc64, sh*) have multiple pte_ts to
4149                  * each hugepage.  We have to make sure we get the
4150                  * first, for the page indexing below to work.
4151                  *
4152                  * Note that page table lock is not held when pte is null.
4153                  */
4154                 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4155                                       huge_page_size(h));
4156                 if (pte)
4157                         ptl = huge_pte_lock(h, mm, pte);
4158                 absent = !pte || huge_pte_none(huge_ptep_get(pte));
4159 
4160                 /*
4161                  * When coredumping, it suits get_dump_page if we just return
4162                  * an error where there's an empty slot with no huge pagecache
4163                  * to back it.  This way, we avoid allocating a hugepage, and
4164                  * the sparse dumpfile avoids allocating disk blocks, but its
4165                  * huge holes still show up with zeroes where they need to be.
4166                  */
4167                 if (absent && (flags & FOLL_DUMP) &&
4168                     !hugetlbfs_pagecache_present(h, vma, vaddr)) {
4169                         if (pte)
4170                                 spin_unlock(ptl);
4171                         remainder = 0;
4172                         break;
4173                 }
4174 
4175                 /*
4176                  * We need call hugetlb_fault for both hugepages under migration
4177                  * (in which case hugetlb_fault waits for the migration,) and
4178                  * hwpoisoned hugepages (in which case we need to prevent the
4179                  * caller from accessing to them.) In order to do this, we use
4180                  * here is_swap_pte instead of is_hugetlb_entry_migration and
4181                  * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4182                  * both cases, and because we can't follow correct pages
4183                  * directly from any kind of swap entries.
4184                  */
4185                 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4186                     ((flags & FOLL_WRITE) &&
4187                       !huge_pte_write(huge_ptep_get(pte)))) {
4188                         int ret;
4189                         unsigned int fault_flags = 0;
4190 
4191                         if (pte)
4192                                 spin_unlock(ptl);
4193                         if (flags & FOLL_WRITE)
4194                                 fault_flags |= FAULT_FLAG_WRITE;
4195                         if (nonblocking)
4196                                 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
4197                         if (flags & FOLL_NOWAIT)
4198                                 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4199                                         FAULT_FLAG_RETRY_NOWAIT;
4200                         if (flags & FOLL_TRIED) {
4201                                 VM_WARN_ON_ONCE(fault_flags &
4202                                                 FAULT_FLAG_ALLOW_RETRY);
4203                                 fault_flags |= FAULT_FLAG_TRIED;
4204                         }
4205                         ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4206                         if (ret & VM_FAULT_ERROR) {
4207                                 err = vm_fault_to_errno(ret, flags);
4208                                 remainder = 0;
4209                                 break;
4210                         }
4211                         if (ret & VM_FAULT_RETRY) {
4212                                 if (nonblocking)
4213                                         *nonblocking = 0;
4214                                 *nr_pages = 0;
4215                                 /*
4216                                  * VM_FAULT_RETRY must not return an
4217                                  * error, it will return zero
4218                                  * instead.
4219                                  *
4220                                  * No need to update "position" as the
4221                                  * caller will not check it after
4222                                  * *nr_pages is set to 0.
4223                                  */
4224                                 return i;
4225                         }
4226                         continue;
4227                 }
4228 
4229                 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4230                 page = pte_page(huge_ptep_get(pte));
4231 same_page:
4232                 if (pages) {
4233                         pages[i] = mem_map_offset(page, pfn_offset);
4234                         get_page(pages[i]);
4235                 }
4236 
4237                 if (vmas)
4238                         vmas[i] = vma;
4239 
4240                 vaddr += PAGE_SIZE;
4241                 ++pfn_offset;
4242                 --remainder;
4243                 ++i;
4244                 if (vaddr < vma->vm_end && remainder &&
4245                                 pfn_offset < pages_per_huge_page(h)) {
4246                         /*
4247                          * We use pfn_offset to avoid touching the pageframes
4248                          * of this compound page.
4249                          */
4250                         goto same_page;
4251                 }
4252                 spin_unlock(ptl);
4253         }
4254         *nr_pages = remainder;
4255         /*
4256          * setting position is actually required only if remainder is
4257          * not zero but it's faster not to add a "if (remainder)"
4258          * branch.
4259          */
4260         *position = vaddr;
4261 
4262         return i ? i : err;
4263 }
4264 
4265 #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4266 /*
4267  * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4268  * implement this.
4269  */
4270 #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4271 #endif
4272 
4273 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
4274                 unsigned long address, unsigned long end, pgprot_t newprot)
4275 {
4276         struct mm_struct *mm = vma->vm_mm;
4277         unsigned long start = address;
4278         pte_t *ptep;
4279         pte_t pte;
4280         struct hstate *h = hstate_vma(vma);
4281         unsigned long pages = 0;
4282 
4283         BUG_ON(address >= end);
4284         flush_cache_range(vma, address, end);
4285 
4286         mmu_notifier_invalidate_range_start(mm, start, end);
4287         i_mmap_lock_write(vma->vm_file->f_mapping);
4288         for (; address < end; address += huge_page_size(h)) {
4289                 spinlock_t *ptl;
4290                 ptep = huge_pte_offset(mm, address, huge_page_size(h));
4291                 if (!ptep)
4292                         continue;
4293                 ptl = huge_pte_lock(h, mm, ptep);
4294                 if (huge_pmd_unshare(mm, &address, ptep)) {
4295                         pages++;
4296                         spin_unlock(ptl);
4297                         continue;
4298                 }
4299                 pte = huge_ptep_get(ptep);
4300                 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
4301                         spin_unlock(ptl);
4302                         continue;
4303                 }
4304                 if (unlikely(is_hugetlb_entry_migration(pte))) {
4305                         swp_entry_t entry = pte_to_swp_entry(pte);
4306 
4307                         if (is_write_migration_entry(entry)) {
4308                                 pte_t newpte;
4309 
4310                                 make_migration_entry_read(&entry);
4311                                 newpte = swp_entry_to_pte(entry);
4312                                 set_huge_swap_pte_at(mm, address, ptep,
4313                                                      newpte, huge_page_size(h));
4314                                 pages++;
4315                         }
4316                         spin_unlock(ptl);
4317                         continue;
4318                 }
4319                 if (!huge_pte_none(pte)) {
4320                         pte = huge_ptep_get_and_clear(mm, address, ptep);
4321                         pte = pte_mkhuge(huge_pte_modify(pte, newprot));
4322                         pte = arch_make_huge_pte(pte, vma, NULL, 0);
4323                         set_huge_pte_at(mm, address, ptep, pte);
4324                         pages++;
4325                 }
4326                 spin_unlock(ptl);
4327         }
4328         /*
4329          * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
4330          * may have cleared our pud entry and done put_page on the page table:
4331          * once we release i_mmap_rwsem, another task can do the final put_page
4332          * and that page table be reused and filled with junk.
4333          */
4334         flush_hugetlb_tlb_range(vma, start, end);
4335         /*
4336          * No need to call mmu_notifier_invalidate_range() we are downgrading
4337          * page table protection not changing it to point to a new page.
4338          *
4339          * See Documentation/vm/mmu_notifier.txt
4340          */
4341         i_mmap_unlock_write(vma->vm_file->f_mapping);
4342         mmu_notifier_invalidate_range_end(mm, start, end);
4343 
4344         return pages << h->order;
4345 }
4346 
4347 int hugetlb_reserve_pages(struct inode *inode,
4348                                         long from, long to,
4349                                         struct vm_area_struct *vma,
4350                                         vm_flags_t vm_flags)
4351 {
4352         long ret, chg;
4353         struct hstate *h = hstate_inode(inode);
4354         struct hugepage_subpool *spool = subpool_inode(inode);
4355         struct resv_map *resv_map;
4356         long gbl_reserve;
4357 
4358         /* This should never happen */
4359         if (from > to) {
4360                 VM_WARN(1, "%s called with a negative range\n", __func__);
4361                 return -EINVAL;
4362         }
4363 
4364         /*
4365          * Only apply hugepage reservation if asked. At fault time, an
4366          * attempt will be made for VM_NORESERVE to allocate a page
4367          * without using reserves
4368          */
4369         if (vm_flags & VM_NORESERVE)
4370                 return 0;
4371 
4372         /*
4373          * Shared mappings base their reservation on the number of pages that
4374          * are already allocated on behalf of the file. Private mappings need
4375          * to reserve the full area even if read-only as mprotect() may be
4376          * called to make the mapping read-write. Assume !vma is a shm mapping
4377          */
4378         if (!vma || vma->vm_flags & VM_MAYSHARE) {
4379                 resv_map = inode_resv_map(inode);
4380 
4381                 chg = region_chg(resv_map, from, to);
4382 
4383         } else {
4384                 resv_map = resv_map_alloc();
4385                 if (!resv_map)
4386                         return -ENOMEM;
4387 
4388                 chg = to - from;
4389 
4390                 set_vma_resv_map(vma, resv_map);
4391                 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4392         }
4393 
4394         if (chg < 0) {
4395                 ret = chg;
4396                 goto out_err;
4397         }
4398 
4399         /*
4400          * There must be enough pages in the subpool for the mapping. If
4401          * the subpool has a minimum size, there may be some global
4402          * reservations already in place (gbl_reserve).
4403          */
4404         gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4405         if (gbl_reserve < 0) {
4406                 ret = -ENOSPC;
4407                 goto out_err;
4408         }
4409 
4410         /*
4411          * Check enough hugepages are available for the reservation.
4412          * Hand the pages back to the subpool if there are not
4413          */
4414         ret = hugetlb_acct_memory(h, gbl_reserve);
4415         if (ret < 0) {
4416                 /* put back original number of pages, chg */
4417                 (void)hugepage_subpool_put_pages(spool, chg);
4418                 goto out_err;
4419         }
4420 
4421         /*
4422          * Account for the reservations made. Shared mappings record regions
4423          * that have reservations as they are shared by multiple VMAs.
4424          * When the last VMA disappears, the region map says how much
4425          * the reservation was and the page cache tells how much of
4426          * the reservation was consumed. Private mappings are per-VMA and
4427          * only the consumed reservations are tracked. When the VMA
4428          * disappears, the original reservation is the VMA size and the
4429          * consumed reservations are stored in the map. Hence, nothing
4430          * else has to be done for private mappings here
4431          */
4432         if (!vma || vma->vm_flags & VM_MAYSHARE) {
4433                 long add = region_add(resv_map, from, to);
4434 
4435                 if (unlikely(chg > add)) {
4436                         /*
4437                          * pages in this range were added to the reserve
4438                          * map between region_chg and region_add.  This
4439                          * indicates a race with alloc_huge_page.  Adjust
4440                          * the subpool and reserve counts modified above
4441                          * based on the difference.
4442                          */
4443                         long rsv_adjust;
4444 
4445                         rsv_adjust = hugepage_subpool_put_pages(spool,
4446                                                                 chg - add);
4447                         hugetlb_acct_memory(h, -rsv_adjust);
4448                 }
4449         }
4450         return 0;
4451 out_err:
4452         if (!vma || vma->vm_flags & VM_MAYSHARE)
4453                 /* Don't call region_abort if region_chg failed */
4454                 if (chg >= 0)
4455                         region_abort(resv_map, from, to);
4456         if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4457                 kref_put(&resv_map->refs, resv_map_release);
4458         return ret;
4459 }
4460 
4461 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4462                                                                 long freed)
4463 {
4464         struct hstate *h = hstate_inode(inode);
4465         struct resv_map *resv_map = inode_resv_map(inode);
4466         long chg = 0;
4467         struct hugepage_subpool *spool = subpool_inode(inode);
4468         long gbl_reserve;
4469 
4470         if (resv_map) {
4471                 chg = region_del(resv_map, start, end);
4472                 /*
4473                  * region_del() can fail in the rare case where a region
4474                  * must be split and another region descriptor can not be
4475                  * allocated.  If end == LONG_MAX, it will not fail.
4476                  */
4477                 if (chg < 0)
4478                         return chg;
4479         }
4480 
4481         spin_lock(&inode->i_lock);
4482         inode->i_blocks -= (blocks_per_huge_page(h) * freed);
4483         spin_unlock(&inode->i_lock);
4484 
4485         /*
4486          * If the subpool has a minimum size, the number of global
4487          * reservations to be released may be adjusted.
4488          */
4489         gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4490         hugetlb_acct_memory(h, -gbl_reserve);
4491 
4492         return 0;
4493 }
4494 
4495 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4496 static unsigned long page_table_shareable(struct vm_area_struct *svma,
4497                                 struct vm_area_struct *vma,
4498                                 unsigned long addr, pgoff_t idx)
4499 {
4500         unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4501                                 svma->vm_start;
4502         unsigned long sbase = saddr & PUD_MASK;
4503         unsigned long s_end = sbase + PUD_SIZE;
4504 
4505         /* Allow segments to share if only one is marked locked */
4506         unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4507         unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
4508 
4509         /*
4510          * match the virtual addresses, permission and the alignment of the
4511          * page table page.
4512          */
4513         if (pmd_index(addr) != pmd_index(saddr) ||
4514             vm_flags != svm_flags ||
4515             sbase < svma->vm_start || svma->vm_end < s_end)
4516                 return 0;
4517 
4518         return saddr;
4519 }
4520 
4521 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
4522 {
4523         unsigned long base = addr & PUD_MASK;
4524         unsigned long end = base + PUD_SIZE;
4525 
4526         /*
4527          * check on proper vm_flags and page table alignment
4528          */
4529         if (vma->vm_flags & VM_MAYSHARE &&
4530             vma->vm_start <= base && end <= vma->vm_end)
4531                 return true;
4532         return false;
4533 }
4534 
4535 /*
4536  * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4537  * and returns the corresponding pte. While this is not necessary for the
4538  * !shared pmd case because we can allocate the pmd later as well, it makes the
4539  * code much cleaner. pmd allocation is essential for the shared case because
4540  * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4541  * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4542  * bad pmd for sharing.
4543  */
4544 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4545 {
4546         struct vm_area_struct *vma = find_vma(mm, addr);
4547         struct address_space *mapping = vma->vm_file->f_mapping;
4548         pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4549                         vma->vm_pgoff;
4550         struct vm_area_struct *svma;
4551         unsigned long saddr;
4552         pte_t *spte = NULL;
4553         pte_t *pte;
4554         spinlock_t *ptl;
4555 
4556         if (!vma_shareable(vma, addr))
4557                 return (pte_t *)pmd_alloc(mm, pud, addr);
4558 
4559         i_mmap_lock_write(mapping);
4560         vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4561                 if (svma == vma)
4562                         continue;
4563 
4564                 saddr = page_table_shareable(svma, vma, addr, idx);
4565                 if (saddr) {
4566                         spte = huge_pte_offset(svma->vm_mm, saddr,
4567                                                vma_mmu_pagesize(svma));
4568                         if (spte) {
4569                                 get_page(virt_to_page(spte));
4570                                 break;
4571                         }
4572                 }
4573         }
4574 
4575         if (!spte)
4576                 goto out;
4577 
4578         ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
4579         if (pud_none(*pud)) {
4580                 pud_populate(mm, pud,
4581                                 (pmd_t *)((unsigned long)spte & PAGE_MASK));
4582                 mm_inc_nr_pmds(mm);
4583         } else {
4584                 put_page(virt_to_page(spte));
4585         }
4586         spin_unlock(ptl);
4587 out:
4588         pte = (pte_t *)pmd_alloc(mm, pud, addr);
4589         i_mmap_unlock_write(mapping);
4590         return pte;
4591 }
4592 
4593 /*
4594  * unmap huge page backed by shared pte.
4595  *
4596  * Hugetlb pte page is ref counted at the time of mapping.  If pte is shared
4597  * indicated by page_count > 1, unmap is achieved by clearing pud and
4598  * decrementing the ref count. If count == 1, the pte page is not shared.
4599  *
4600  * called with page table lock held.
4601  *
4602  * returns: 1 successfully unmapped a shared pte page
4603  *          0 the underlying pte page is not shared, or it is the last user
4604  */
4605 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4606 {
4607         pgd_t *pgd = pgd_offset(mm, *addr);
4608         p4d_t *p4d = p4d_offset(pgd, *addr);
4609         pud_t *pud = pud_offset(p4d, *addr);
4610 
4611         BUG_ON(page_count(virt_to_page(ptep)) == 0);
4612         if (page_count(virt_to_page(ptep)) == 1)
4613                 return 0;
4614 
4615         pud_clear(pud);
4616         put_page(virt_to_page(ptep));
4617         mm_dec_nr_pmds(mm);
4618         *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4619         return 1;
4620 }
4621 #define want_pmd_share()        (1)
4622 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4623 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4624 {
4625         return NULL;
4626 }
4627 
4628 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4629 {
4630         return 0;
4631 }
4632 #define want_pmd_share()        (0)
4633 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4634 
4635 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4636 pte_t *huge_pte_alloc(struct mm_struct *mm,
4637                         unsigned long addr, unsigned long sz)
4638 {
4639         pgd_t *pgd;
4640         p4d_t *p4d;
4641         pud_t *pud;
4642         pte_t *pte = NULL;
4643 
4644         pgd = pgd_offset(mm, addr);
4645         p4d = p4d_alloc(mm, pgd, addr);
4646         if (!p4d)
4647                 return NULL;
4648         pud = pud_alloc(mm, p4d, addr);
4649         if (pud) {
4650                 if (sz == PUD_SIZE) {
4651                         pte = (pte_t *)pud;
4652                 } else {
4653                         BUG_ON(sz != PMD_SIZE);
4654                         if (want_pmd_share() && pud_none(*pud))
4655                                 pte = huge_pmd_share(mm, addr, pud);
4656                         else
4657                                 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4658                 }
4659         }
4660         BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
4661 
4662         return pte;
4663 }
4664 
4665 /*
4666  * huge_pte_offset() - Walk the page table to resolve the hugepage
4667  * entry at address @addr
4668  *
4669  * Return: Pointer to page table or swap entry (PUD or PMD) for
4670  * address @addr, or NULL if a p*d_none() entry is encountered and the
4671  * size @sz doesn't match the hugepage size at this level of the page
4672  * table.
4673  */
4674 pte_t *huge_pte_offset(struct mm_struct *mm,
4675                        unsigned long addr, unsigned long sz)
4676 {
4677         pgd_t *pgd;
4678         p4d_t *p4d;
4679         pud_t *pud;
4680         pmd_t *pmd;
4681 
4682         pgd = pgd_offset(mm, addr);
4683         if (!pgd_present(*pgd))
4684                 return NULL;
4685         p4d = p4d_offset(pgd, addr);
4686         if (!p4d_present(*p4d))
4687                 return NULL;
4688 
4689         pud = pud_offset(p4d, addr);
4690         if (sz != PUD_SIZE && pud_none(*pud))
4691                 return NULL;
4692         /* hugepage or swap? */
4693         if (pud_huge(*pud) || !pud_present(*pud))
4694                 return (pte_t *)pud;
4695 
4696         pmd = pmd_offset(pud, addr);
4697         if (sz != PMD_SIZE && pmd_none(*pmd))
4698                 return NULL;
4699         /* hugepage or swap? */
4700         if (pmd_huge(*pmd) || !pmd_present(*pmd))
4701                 return (pte_t *)pmd;
4702 
4703         return NULL;
4704 }
4705 
4706 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4707 
4708 /*
4709  * These functions are overwritable if your architecture needs its own
4710  * behavior.
4711  */
4712 struct page * __weak
4713 follow_huge_addr(struct mm_struct *mm, unsigned long address,
4714                               int write)
4715 {
4716         return ERR_PTR(-EINVAL);
4717 }
4718 
4719 struct page * __weak
4720 follow_huge_pd(struct vm_area_struct *vma,
4721                unsigned long address, hugepd_t hpd, int flags, int pdshift)
4722 {
4723         WARN(1, "hugepd follow called with no support for hugepage directory format\n");
4724         return NULL;
4725 }
4726 
4727 struct page * __weak
4728 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
4729                 pmd_t *pmd, int flags)
4730 {
4731         struct page *page = NULL;
4732         spinlock_t *ptl;
4733         pte_t pte;
4734 retry:
4735         ptl = pmd_lockptr(mm, pmd);
4736         spin_lock(ptl);
4737         /*
4738          * make sure that the address range covered by this pmd is not
4739          * unmapped from other threads.
4740          */
4741         if (!pmd_huge(*pmd))
4742                 goto out;
4743         pte = huge_ptep_get((pte_t *)pmd);
4744         if (pte_present(pte)) {
4745                 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
4746                 if (flags & FOLL_GET)
4747                         get_page(page);
4748         } else {
4749                 if (is_hugetlb_entry_migration(pte)) {
4750                         spin_unlock(ptl);
4751                         __migration_entry_wait(mm, (pte_t *)pmd, ptl);
4752                         goto retry;
4753                 }
4754                 /*
4755                  * hwpoisoned entry is treated as no_page_table in
4756                  * follow_page_mask().
4757                  */
4758         }
4759 out:
4760         spin_unlock(ptl);
4761         return page;
4762 }
4763 
4764 struct page * __weak
4765 follow_huge_pud(struct mm_struct *mm, unsigned long address,
4766                 pud_t *pud, int flags)
4767 {
4768         if (flags & FOLL_GET)
4769                 return NULL;
4770 
4771         return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
4772 }
4773 
4774 struct page * __weak
4775 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
4776 {
4777         if (flags & FOLL_GET)
4778                 return NULL;
4779 
4780         return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
4781 }
4782 
4783 bool isolate_huge_page(struct page *page, struct list_head *list)
4784 {
4785         bool ret = true;
4786 
4787         VM_BUG_ON_PAGE(!PageHead(page), page);
4788         spin_lock(&hugetlb_lock);
4789         if (!page_huge_active(page) || !get_page_unless_zero(page)) {
4790                 ret = false;
4791                 goto unlock;
4792         }
4793         clear_page_huge_active(page);
4794         list_move_tail(&page->lru, list);
4795 unlock:
4796         spin_unlock(&hugetlb_lock);
4797         return ret;
4798 }
4799 
4800 void putback_active_hugepage(struct page *page)
4801 {
4802         VM_BUG_ON_PAGE(!PageHead(page), page);
4803         spin_lock(&hugetlb_lock);
4804         set_page_huge_active(page);
4805         list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
4806         spin_unlock(&hugetlb_lock);
4807         put_page(page);
4808 }
4809 

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp