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

TOMOYO Linux Cross Reference
Linux/mm/hugetlb.c

Version: ~ [ linux-5.3-rc5 ] ~ [ linux-5.2.9 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.67 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.139 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.189 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.189 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.140 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.72 ] ~ [ linux-3.15.10 ] ~ [ linux-3.14.79 ] ~ [ linux-3.13.11 ] ~ [ linux-3.12.74 ] ~ [ linux-3.11.10 ] ~ [ linux-3.10.108 ] ~ [ linux-3.9.11 ] ~ [ linux-3.8.13 ] ~ [ linux-3.7.10 ] ~ [ linux-3.6.11 ] ~ [ linux-3.5.7 ] ~ [ linux-3.4.113 ] ~ [ linux-3.3.8 ] ~ [ linux-3.2.102 ] ~ [ linux-3.1.10 ] ~ [ linux-3.0.101 ] ~ [ 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/sched/signal.h>
 22 #include <linux/rmap.h>
 23 #include <linux/swap.h>
 24 #include <linux/swapops.h>
 25 #include <linux/page-isolation.h>
 26 #include <linux/jhash.h>
 27 
 28 #include <asm/page.h>
 29 #include <asm/pgtable.h>
 30 #include <asm/tlb.h>
 31 
 32 #include <linux/io.h>
 33 #include <linux/hugetlb.h>
 34 #include <linux/hugetlb_cgroup.h>
 35 #include <linux/node.h>
 36 #include <linux/userfaultfd_k.h>
 37 #include "internal.h"
 38 
 39 int hugepages_treat_as_movable;
 40 
 41 int hugetlb_max_hstate __read_mostly;
 42 unsigned int default_hstate_idx;
 43 struct hstate hstates[HUGE_MAX_HSTATE];
 44 /*
 45  * Minimum page order among possible hugepage sizes, set to a proper value
 46  * at boot time.
 47  */
 48 static unsigned int minimum_order __read_mostly = UINT_MAX;
 49 
 50 __initdata LIST_HEAD(huge_boot_pages);
 51 
 52 /* for command line parsing */
 53 static struct hstate * __initdata parsed_hstate;
 54 static unsigned long __initdata default_hstate_max_huge_pages;
 55 static unsigned long __initdata default_hstate_size;
 56 static bool __initdata parsed_valid_hugepagesz = true;
 57 
 58 /*
 59  * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
 60  * free_huge_pages, and surplus_huge_pages.
 61  */
 62 DEFINE_SPINLOCK(hugetlb_lock);
 63 
 64 /*
 65  * Serializes faults on the same logical page.  This is used to
 66  * prevent spurious OOMs when the hugepage pool is fully utilized.
 67  */
 68 static int num_fault_mutexes;
 69 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
 70 
 71 /* Forward declaration */
 72 static int hugetlb_acct_memory(struct hstate *h, long delta);
 73 
 74 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
 75 {
 76         bool free = (spool->count == 0) && (spool->used_hpages == 0);
 77 
 78         spin_unlock(&spool->lock);
 79 
 80         /* If no pages are used, and no other handles to the subpool
 81          * remain, give up any reservations mased on minimum size and
 82          * free the subpool */
 83         if (free) {
 84                 if (spool->min_hpages != -1)
 85                         hugetlb_acct_memory(spool->hstate,
 86                                                 -spool->min_hpages);
 87                 kfree(spool);
 88         }
 89 }
 90 
 91 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
 92                                                 long min_hpages)
 93 {
 94         struct hugepage_subpool *spool;
 95 
 96         spool = kzalloc(sizeof(*spool), GFP_KERNEL);
 97         if (!spool)
 98                 return NULL;
 99 
100         spin_lock_init(&spool->lock);
101         spool->count = 1;
102         spool->max_hpages = max_hpages;
103         spool->hstate = h;
104         spool->min_hpages = min_hpages;
105 
106         if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
107                 kfree(spool);
108                 return NULL;
109         }
110         spool->rsv_hpages = min_hpages;
111 
112         return spool;
113 }
114 
115 void hugepage_put_subpool(struct hugepage_subpool *spool)
116 {
117         spin_lock(&spool->lock);
118         BUG_ON(!spool->count);
119         spool->count--;
120         unlock_or_release_subpool(spool);
121 }
122 
123 /*
124  * Subpool accounting for allocating and reserving pages.
125  * Return -ENOMEM if there are not enough resources to satisfy the
126  * the request.  Otherwise, return the number of pages by which the
127  * global pools must be adjusted (upward).  The returned value may
128  * only be different than the passed value (delta) in the case where
129  * a subpool minimum size must be manitained.
130  */
131 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
132                                       long delta)
133 {
134         long ret = delta;
135 
136         if (!spool)
137                 return ret;
138 
139         spin_lock(&spool->lock);
140 
141         if (spool->max_hpages != -1) {          /* maximum size accounting */
142                 if ((spool->used_hpages + delta) <= spool->max_hpages)
143                         spool->used_hpages += delta;
144                 else {
145                         ret = -ENOMEM;
146                         goto unlock_ret;
147                 }
148         }
149 
150         /* minimum size accounting */
151         if (spool->min_hpages != -1 && spool->rsv_hpages) {
152                 if (delta > spool->rsv_hpages) {
153                         /*
154                          * Asking for more reserves than those already taken on
155                          * behalf of subpool.  Return difference.
156                          */
157                         ret = delta - spool->rsv_hpages;
158                         spool->rsv_hpages = 0;
159                 } else {
160                         ret = 0;        /* reserves already accounted for */
161                         spool->rsv_hpages -= delta;
162                 }
163         }
164 
165 unlock_ret:
166         spin_unlock(&spool->lock);
167         return ret;
168 }
169 
170 /*
171  * Subpool accounting for freeing and unreserving pages.
172  * Return the number of global page reservations that must be dropped.
173  * The return value may only be different than the passed value (delta)
174  * in the case where a subpool minimum size must be maintained.
175  */
176 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
177                                        long delta)
178 {
179         long ret = delta;
180 
181         if (!spool)
182                 return delta;
183 
184         spin_lock(&spool->lock);
185 
186         if (spool->max_hpages != -1)            /* maximum size accounting */
187                 spool->used_hpages -= delta;
188 
189          /* minimum size accounting */
190         if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
191                 if (spool->rsv_hpages + delta <= spool->min_hpages)
192                         ret = 0;
193                 else
194                         ret = spool->rsv_hpages + delta - spool->min_hpages;
195 
196                 spool->rsv_hpages += delta;
197                 if (spool->rsv_hpages > spool->min_hpages)
198                         spool->rsv_hpages = spool->min_hpages;
199         }
200 
201         /*
202          * If hugetlbfs_put_super couldn't free spool due to an outstanding
203          * quota reference, free it now.
204          */
205         unlock_or_release_subpool(spool);
206 
207         return ret;
208 }
209 
210 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
211 {
212         return HUGETLBFS_SB(inode->i_sb)->spool;
213 }
214 
215 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
216 {
217         return subpool_inode(file_inode(vma->vm_file));
218 }
219 
220 /*
221  * Region tracking -- allows tracking of reservations and instantiated pages
222  *                    across the pages in a mapping.
223  *
224  * The region data structures are embedded into a resv_map and protected
225  * by a resv_map's lock.  The set of regions within the resv_map represent
226  * reservations for huge pages, or huge pages that have already been
227  * instantiated within the map.  The from and to elements are huge page
228  * indicies into the associated mapping.  from indicates the starting index
229  * of the region.  to represents the first index past the end of  the region.
230  *
231  * For example, a file region structure with from == 0 and to == 4 represents
232  * four huge pages in a mapping.  It is important to note that the to element
233  * represents the first element past the end of the region. This is used in
234  * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
235  *
236  * Interval notation of the form [from, to) will be used to indicate that
237  * the endpoint from is inclusive and to is exclusive.
238  */
239 struct file_region {
240         struct list_head link;
241         long from;
242         long to;
243 };
244 
245 /*
246  * Add the huge page range represented by [f, t) to the reserve
247  * map.  In the normal case, existing regions will be expanded
248  * to accommodate the specified range.  Sufficient regions should
249  * exist for expansion due to the previous call to region_chg
250  * with the same range.  However, it is possible that region_del
251  * could have been called after region_chg and modifed the map
252  * in such a way that no region exists to be expanded.  In this
253  * case, pull a region descriptor from the cache associated with
254  * the map and use that for the new range.
255  *
256  * Return the number of new huge pages added to the map.  This
257  * number is greater than or equal to zero.
258  */
259 static long region_add(struct resv_map *resv, long f, long t)
260 {
261         struct list_head *head = &resv->regions;
262         struct file_region *rg, *nrg, *trg;
263         long add = 0;
264 
265         spin_lock(&resv->lock);
266         /* Locate the region we are either in or before. */
267         list_for_each_entry(rg, head, link)
268                 if (f <= rg->to)
269                         break;
270 
271         /*
272          * If no region exists which can be expanded to include the
273          * specified range, the list must have been modified by an
274          * interleving call to region_del().  Pull a region descriptor
275          * from the cache and use it for this range.
276          */
277         if (&rg->link == head || t < rg->from) {
278                 VM_BUG_ON(resv->region_cache_count <= 0);
279 
280                 resv->region_cache_count--;
281                 nrg = list_first_entry(&resv->region_cache, struct file_region,
282                                         link);
283                 list_del(&nrg->link);
284 
285                 nrg->from = f;
286                 nrg->to = t;
287                 list_add(&nrg->link, rg->link.prev);
288 
289                 add += t - f;
290                 goto out_locked;
291         }
292 
293         /* Round our left edge to the current segment if it encloses us. */
294         if (f > rg->from)
295                 f = rg->from;
296 
297         /* Check for and consume any regions we now overlap with. */
298         nrg = rg;
299         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
300                 if (&rg->link == head)
301                         break;
302                 if (rg->from > t)
303                         break;
304 
305                 /* If this area reaches higher then extend our area to
306                  * include it completely.  If this is not the first area
307                  * which we intend to reuse, free it. */
308                 if (rg->to > t)
309                         t = rg->to;
310                 if (rg != nrg) {
311                         /* Decrement return value by the deleted range.
312                          * Another range will span this area so that by
313                          * end of routine add will be >= zero
314                          */
315                         add -= (rg->to - rg->from);
316                         list_del(&rg->link);
317                         kfree(rg);
318                 }
319         }
320 
321         add += (nrg->from - f);         /* Added to beginning of region */
322         nrg->from = f;
323         add += t - nrg->to;             /* Added to end of region */
324         nrg->to = t;
325 
326 out_locked:
327         resv->adds_in_progress--;
328         spin_unlock(&resv->lock);
329         VM_BUG_ON(add < 0);
330         return add;
331 }
332 
333 /*
334  * Examine the existing reserve map and determine how many
335  * huge pages in the specified range [f, t) are NOT currently
336  * represented.  This routine is called before a subsequent
337  * call to region_add that will actually modify the reserve
338  * map to add the specified range [f, t).  region_chg does
339  * not change the number of huge pages represented by the
340  * map.  However, if the existing regions in the map can not
341  * be expanded to represent the new range, a new file_region
342  * structure is added to the map as a placeholder.  This is
343  * so that the subsequent region_add call will have all the
344  * regions it needs and will not fail.
345  *
346  * Upon entry, region_chg will also examine the cache of region descriptors
347  * associated with the map.  If there are not enough descriptors cached, one
348  * will be allocated for the in progress add operation.
349  *
350  * Returns the number of huge pages that need to be added to the existing
351  * reservation map for the range [f, t).  This number is greater or equal to
352  * zero.  -ENOMEM is returned if a new file_region structure or cache entry
353  * is needed and can not be allocated.
354  */
355 static long region_chg(struct resv_map *resv, long f, long t)
356 {
357         struct list_head *head = &resv->regions;
358         struct file_region *rg, *nrg = NULL;
359         long chg = 0;
360 
361 retry:
362         spin_lock(&resv->lock);
363 retry_locked:
364         resv->adds_in_progress++;
365 
366         /*
367          * Check for sufficient descriptors in the cache to accommodate
368          * the number of in progress add operations.
369          */
370         if (resv->adds_in_progress > resv->region_cache_count) {
371                 struct file_region *trg;
372 
373                 VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1);
374                 /* Must drop lock to allocate a new descriptor. */
375                 resv->adds_in_progress--;
376                 spin_unlock(&resv->lock);
377 
378                 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
379                 if (!trg) {
380                         kfree(nrg);
381                         return -ENOMEM;
382                 }
383 
384                 spin_lock(&resv->lock);
385                 list_add(&trg->link, &resv->region_cache);
386                 resv->region_cache_count++;
387                 goto retry_locked;
388         }
389 
390         /* Locate the region we are before or in. */
391         list_for_each_entry(rg, head, link)
392                 if (f <= rg->to)
393                         break;
394 
395         /* If we are below the current region then a new region is required.
396          * Subtle, allocate a new region at the position but make it zero
397          * size such that we can guarantee to record the reservation. */
398         if (&rg->link == head || t < rg->from) {
399                 if (!nrg) {
400                         resv->adds_in_progress--;
401                         spin_unlock(&resv->lock);
402                         nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
403                         if (!nrg)
404                                 return -ENOMEM;
405 
406                         nrg->from = f;
407                         nrg->to   = f;
408                         INIT_LIST_HEAD(&nrg->link);
409                         goto retry;
410                 }
411 
412                 list_add(&nrg->link, rg->link.prev);
413                 chg = t - f;
414                 goto out_nrg;
415         }
416 
417         /* Round our left edge to the current segment if it encloses us. */
418         if (f > rg->from)
419                 f = rg->from;
420         chg = t - f;
421 
422         /* Check for and consume any regions we now overlap with. */
423         list_for_each_entry(rg, rg->link.prev, link) {
424                 if (&rg->link == head)
425                         break;
426                 if (rg->from > t)
427                         goto out;
428 
429                 /* We overlap with this area, if it extends further than
430                  * us then we must extend ourselves.  Account for its
431                  * existing reservation. */
432                 if (rg->to > t) {
433                         chg += rg->to - t;
434                         t = rg->to;
435                 }
436                 chg -= rg->to - rg->from;
437         }
438 
439 out:
440         spin_unlock(&resv->lock);
441         /*  We already know we raced and no longer need the new region */
442         kfree(nrg);
443         return chg;
444 out_nrg:
445         spin_unlock(&resv->lock);
446         return chg;
447 }
448 
449 /*
450  * Abort the in progress add operation.  The adds_in_progress field
451  * of the resv_map keeps track of the operations in progress between
452  * calls to region_chg and region_add.  Operations are sometimes
453  * aborted after the call to region_chg.  In such cases, region_abort
454  * is called to decrement the adds_in_progress counter.
455  *
456  * NOTE: The range arguments [f, t) are not needed or used in this
457  * routine.  They are kept to make reading the calling code easier as
458  * arguments will match the associated region_chg call.
459  */
460 static void region_abort(struct resv_map *resv, long f, long t)
461 {
462         spin_lock(&resv->lock);
463         VM_BUG_ON(!resv->region_cache_count);
464         resv->adds_in_progress--;
465         spin_unlock(&resv->lock);
466 }
467 
468 /*
469  * Delete the specified range [f, t) from the reserve map.  If the
470  * t parameter is LONG_MAX, this indicates that ALL regions after f
471  * should be deleted.  Locate the regions which intersect [f, t)
472  * and either trim, delete or split the existing regions.
473  *
474  * Returns the number of huge pages deleted from the reserve map.
475  * In the normal case, the return value is zero or more.  In the
476  * case where a region must be split, a new region descriptor must
477  * be allocated.  If the allocation fails, -ENOMEM will be returned.
478  * NOTE: If the parameter t == LONG_MAX, then we will never split
479  * a region and possibly return -ENOMEM.  Callers specifying
480  * t == LONG_MAX do not need to check for -ENOMEM error.
481  */
482 static long region_del(struct resv_map *resv, long f, long t)
483 {
484         struct list_head *head = &resv->regions;
485         struct file_region *rg, *trg;
486         struct file_region *nrg = NULL;
487         long del = 0;
488 
489 retry:
490         spin_lock(&resv->lock);
491         list_for_each_entry_safe(rg, trg, head, link) {
492                 /*
493                  * Skip regions before the range to be deleted.  file_region
494                  * ranges are normally of the form [from, to).  However, there
495                  * may be a "placeholder" entry in the map which is of the form
496                  * (from, to) with from == to.  Check for placeholder entries
497                  * at the beginning of the range to be deleted.
498                  */
499                 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
500                         continue;
501 
502                 if (rg->from >= t)
503                         break;
504 
505                 if (f > rg->from && t < rg->to) { /* Must split region */
506                         /*
507                          * Check for an entry in the cache before dropping
508                          * lock and attempting allocation.
509                          */
510                         if (!nrg &&
511                             resv->region_cache_count > resv->adds_in_progress) {
512                                 nrg = list_first_entry(&resv->region_cache,
513                                                         struct file_region,
514                                                         link);
515                                 list_del(&nrg->link);
516                                 resv->region_cache_count--;
517                         }
518 
519                         if (!nrg) {
520                                 spin_unlock(&resv->lock);
521                                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
522                                 if (!nrg)
523                                         return -ENOMEM;
524                                 goto retry;
525                         }
526 
527                         del += t - f;
528 
529                         /* New entry for end of split region */
530                         nrg->from = t;
531                         nrg->to = rg->to;
532                         INIT_LIST_HEAD(&nrg->link);
533 
534                         /* Original entry is trimmed */
535                         rg->to = f;
536 
537                         list_add(&nrg->link, &rg->link);
538                         nrg = NULL;
539                         break;
540                 }
541 
542                 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
543                         del += rg->to - rg->from;
544                         list_del(&rg->link);
545                         kfree(rg);
546                         continue;
547                 }
548 
549                 if (f <= rg->from) {    /* Trim beginning of region */
550                         del += t - rg->from;
551                         rg->from = t;
552                 } else {                /* Trim end of region */
553                         del += rg->to - f;
554                         rg->to = f;
555                 }
556         }
557 
558         spin_unlock(&resv->lock);
559         kfree(nrg);
560         return del;
561 }
562 
563 /*
564  * A rare out of memory error was encountered which prevented removal of
565  * the reserve map region for a page.  The huge page itself was free'ed
566  * and removed from the page cache.  This routine will adjust the subpool
567  * usage count, and the global reserve count if needed.  By incrementing
568  * these counts, the reserve map entry which could not be deleted will
569  * appear as a "reserved" entry instead of simply dangling with incorrect
570  * counts.
571  */
572 void hugetlb_fix_reserve_counts(struct inode *inode)
573 {
574         struct hugepage_subpool *spool = subpool_inode(inode);
575         long rsv_adjust;
576 
577         rsv_adjust = hugepage_subpool_get_pages(spool, 1);
578         if (rsv_adjust) {
579                 struct hstate *h = hstate_inode(inode);
580 
581                 hugetlb_acct_memory(h, 1);
582         }
583 }
584 
585 /*
586  * Count and return the number of huge pages in the reserve map
587  * that intersect with the range [f, t).
588  */
589 static long region_count(struct resv_map *resv, long f, long t)
590 {
591         struct list_head *head = &resv->regions;
592         struct file_region *rg;
593         long chg = 0;
594 
595         spin_lock(&resv->lock);
596         /* Locate each segment we overlap with, and count that overlap. */
597         list_for_each_entry(rg, head, link) {
598                 long seg_from;
599                 long seg_to;
600 
601                 if (rg->to <= f)
602                         continue;
603                 if (rg->from >= t)
604                         break;
605 
606                 seg_from = max(rg->from, f);
607                 seg_to = min(rg->to, t);
608 
609                 chg += seg_to - seg_from;
610         }
611         spin_unlock(&resv->lock);
612 
613         return chg;
614 }
615 
616 /*
617  * Convert the address within this vma to the page offset within
618  * the mapping, in pagecache page units; huge pages here.
619  */
620 static pgoff_t vma_hugecache_offset(struct hstate *h,
621                         struct vm_area_struct *vma, unsigned long address)
622 {
623         return ((address - vma->vm_start) >> huge_page_shift(h)) +
624                         (vma->vm_pgoff >> huge_page_order(h));
625 }
626 
627 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
628                                      unsigned long address)
629 {
630         return vma_hugecache_offset(hstate_vma(vma), vma, address);
631 }
632 EXPORT_SYMBOL_GPL(linear_hugepage_index);
633 
634 /*
635  * Return the size of the pages allocated when backing a VMA. In the majority
636  * cases this will be same size as used by the page table entries.
637  */
638 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
639 {
640         struct hstate *hstate;
641 
642         if (!is_vm_hugetlb_page(vma))
643                 return PAGE_SIZE;
644 
645         hstate = hstate_vma(vma);
646 
647         return 1UL << huge_page_shift(hstate);
648 }
649 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
650 
651 /*
652  * Return the page size being used by the MMU to back a VMA. In the majority
653  * of cases, the page size used by the kernel matches the MMU size. On
654  * architectures where it differs, an architecture-specific version of this
655  * function is required.
656  */
657 #ifndef vma_mmu_pagesize
658 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
659 {
660         return vma_kernel_pagesize(vma);
661 }
662 #endif
663 
664 /*
665  * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
666  * bits of the reservation map pointer, which are always clear due to
667  * alignment.
668  */
669 #define HPAGE_RESV_OWNER    (1UL << 0)
670 #define HPAGE_RESV_UNMAPPED (1UL << 1)
671 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
672 
673 /*
674  * These helpers are used to track how many pages are reserved for
675  * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
676  * is guaranteed to have their future faults succeed.
677  *
678  * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
679  * the reserve counters are updated with the hugetlb_lock held. It is safe
680  * to reset the VMA at fork() time as it is not in use yet and there is no
681  * chance of the global counters getting corrupted as a result of the values.
682  *
683  * The private mapping reservation is represented in a subtly different
684  * manner to a shared mapping.  A shared mapping has a region map associated
685  * with the underlying file, this region map represents the backing file
686  * pages which have ever had a reservation assigned which this persists even
687  * after the page is instantiated.  A private mapping has a region map
688  * associated with the original mmap which is attached to all VMAs which
689  * reference it, this region map represents those offsets which have consumed
690  * reservation ie. where pages have been instantiated.
691  */
692 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
693 {
694         return (unsigned long)vma->vm_private_data;
695 }
696 
697 static void set_vma_private_data(struct vm_area_struct *vma,
698                                                         unsigned long value)
699 {
700         vma->vm_private_data = (void *)value;
701 }
702 
703 struct resv_map *resv_map_alloc(void)
704 {
705         struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
706         struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
707 
708         if (!resv_map || !rg) {
709                 kfree(resv_map);
710                 kfree(rg);
711                 return NULL;
712         }
713 
714         kref_init(&resv_map->refs);
715         spin_lock_init(&resv_map->lock);
716         INIT_LIST_HEAD(&resv_map->regions);
717 
718         resv_map->adds_in_progress = 0;
719 
720         INIT_LIST_HEAD(&resv_map->region_cache);
721         list_add(&rg->link, &resv_map->region_cache);
722         resv_map->region_cache_count = 1;
723 
724         return resv_map;
725 }
726 
727 void resv_map_release(struct kref *ref)
728 {
729         struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
730         struct list_head *head = &resv_map->region_cache;
731         struct file_region *rg, *trg;
732 
733         /* Clear out any active regions before we release the map. */
734         region_del(resv_map, 0, LONG_MAX);
735 
736         /* ... and any entries left in the cache */
737         list_for_each_entry_safe(rg, trg, head, link) {
738                 list_del(&rg->link);
739                 kfree(rg);
740         }
741 
742         VM_BUG_ON(resv_map->adds_in_progress);
743 
744         kfree(resv_map);
745 }
746 
747 static inline struct resv_map *inode_resv_map(struct inode *inode)
748 {
749         return inode->i_mapping->private_data;
750 }
751 
752 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
753 {
754         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
755         if (vma->vm_flags & VM_MAYSHARE) {
756                 struct address_space *mapping = vma->vm_file->f_mapping;
757                 struct inode *inode = mapping->host;
758 
759                 return inode_resv_map(inode);
760 
761         } else {
762                 return (struct resv_map *)(get_vma_private_data(vma) &
763                                                         ~HPAGE_RESV_MASK);
764         }
765 }
766 
767 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
768 {
769         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
770         VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
771 
772         set_vma_private_data(vma, (get_vma_private_data(vma) &
773                                 HPAGE_RESV_MASK) | (unsigned long)map);
774 }
775 
776 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
777 {
778         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
779         VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
780 
781         set_vma_private_data(vma, get_vma_private_data(vma) | flags);
782 }
783 
784 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
785 {
786         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
787 
788         return (get_vma_private_data(vma) & flag) != 0;
789 }
790 
791 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
792 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
793 {
794         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
795         if (!(vma->vm_flags & VM_MAYSHARE))
796                 vma->vm_private_data = (void *)0;
797 }
798 
799 /* Returns true if the VMA has associated reserve pages */
800 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
801 {
802         if (vma->vm_flags & VM_NORESERVE) {
803                 /*
804                  * This address is already reserved by other process(chg == 0),
805                  * so, we should decrement reserved count. Without decrementing,
806                  * reserve count remains after releasing inode, because this
807                  * allocated page will go into page cache and is regarded as
808                  * coming from reserved pool in releasing step.  Currently, we
809                  * don't have any other solution to deal with this situation
810                  * properly, so add work-around here.
811                  */
812                 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
813                         return true;
814                 else
815                         return false;
816         }
817 
818         /* Shared mappings always use reserves */
819         if (vma->vm_flags & VM_MAYSHARE) {
820                 /*
821                  * We know VM_NORESERVE is not set.  Therefore, there SHOULD
822                  * be a region map for all pages.  The only situation where
823                  * there is no region map is if a hole was punched via
824                  * fallocate.  In this case, there really are no reverves to
825                  * use.  This situation is indicated if chg != 0.
826                  */
827                 if (chg)
828                         return false;
829                 else
830                         return true;
831         }
832 
833         /*
834          * Only the process that called mmap() has reserves for
835          * private mappings.
836          */
837         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
838                 /*
839                  * Like the shared case above, a hole punch or truncate
840                  * could have been performed on the private mapping.
841                  * Examine the value of chg to determine if reserves
842                  * actually exist or were previously consumed.
843                  * Very Subtle - The value of chg comes from a previous
844                  * call to vma_needs_reserves().  The reserve map for
845                  * private mappings has different (opposite) semantics
846                  * than that of shared mappings.  vma_needs_reserves()
847                  * has already taken this difference in semantics into
848                  * account.  Therefore, the meaning of chg is the same
849                  * as in the shared case above.  Code could easily be
850                  * combined, but keeping it separate draws attention to
851                  * subtle differences.
852                  */
853                 if (chg)
854                         return false;
855                 else
856                         return true;
857         }
858 
859         return false;
860 }
861 
862 static void enqueue_huge_page(struct hstate *h, struct page *page)
863 {
864         int nid = page_to_nid(page);
865         list_move(&page->lru, &h->hugepage_freelists[nid]);
866         h->free_huge_pages++;
867         h->free_huge_pages_node[nid]++;
868 }
869 
870 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
871 {
872         struct page *page;
873 
874         list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
875                 if (!is_migrate_isolate_page(page))
876                         break;
877         /*
878          * if 'non-isolated free hugepage' not found on the list,
879          * the allocation fails.
880          */
881         if (&h->hugepage_freelists[nid] == &page->lru)
882                 return NULL;
883         list_move(&page->lru, &h->hugepage_activelist);
884         set_page_refcounted(page);
885         h->free_huge_pages--;
886         h->free_huge_pages_node[nid]--;
887         return page;
888 }
889 
890 /* Movability of hugepages depends on migration support. */
891 static inline gfp_t htlb_alloc_mask(struct hstate *h)
892 {
893         if (hugepages_treat_as_movable || hugepage_migration_supported(h))
894                 return GFP_HIGHUSER_MOVABLE;
895         else
896                 return GFP_HIGHUSER;
897 }
898 
899 static struct page *dequeue_huge_page_vma(struct hstate *h,
900                                 struct vm_area_struct *vma,
901                                 unsigned long address, int avoid_reserve,
902                                 long chg)
903 {
904         struct page *page = NULL;
905         struct mempolicy *mpol;
906         nodemask_t *nodemask;
907         struct zonelist *zonelist;
908         struct zone *zone;
909         struct zoneref *z;
910         unsigned int cpuset_mems_cookie;
911 
912         /*
913          * A child process with MAP_PRIVATE mappings created by their parent
914          * have no page reserves. This check ensures that reservations are
915          * not "stolen". The child may still get SIGKILLed
916          */
917         if (!vma_has_reserves(vma, chg) &&
918                         h->free_huge_pages - h->resv_huge_pages == 0)
919                 goto err;
920 
921         /* If reserves cannot be used, ensure enough pages are in the pool */
922         if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
923                 goto err;
924 
925 retry_cpuset:
926         cpuset_mems_cookie = read_mems_allowed_begin();
927         zonelist = huge_zonelist(vma, address,
928                                         htlb_alloc_mask(h), &mpol, &nodemask);
929 
930         for_each_zone_zonelist_nodemask(zone, z, zonelist,
931                                                 MAX_NR_ZONES - 1, nodemask) {
932                 if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) {
933                         page = dequeue_huge_page_node(h, zone_to_nid(zone));
934                         if (page) {
935                                 if (avoid_reserve)
936                                         break;
937                                 if (!vma_has_reserves(vma, chg))
938                                         break;
939 
940                                 SetPagePrivate(page);
941                                 h->resv_huge_pages--;
942                                 break;
943                         }
944                 }
945         }
946 
947         mpol_cond_put(mpol);
948         if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
949                 goto retry_cpuset;
950         return page;
951 
952 err:
953         return NULL;
954 }
955 
956 /*
957  * common helper functions for hstate_next_node_to_{alloc|free}.
958  * We may have allocated or freed a huge page based on a different
959  * nodes_allowed previously, so h->next_node_to_{alloc|free} might
960  * be outside of *nodes_allowed.  Ensure that we use an allowed
961  * node for alloc or free.
962  */
963 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
964 {
965         nid = next_node_in(nid, *nodes_allowed);
966         VM_BUG_ON(nid >= MAX_NUMNODES);
967 
968         return nid;
969 }
970 
971 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
972 {
973         if (!node_isset(nid, *nodes_allowed))
974                 nid = next_node_allowed(nid, nodes_allowed);
975         return nid;
976 }
977 
978 /*
979  * returns the previously saved node ["this node"] from which to
980  * allocate a persistent huge page for the pool and advance the
981  * next node from which to allocate, handling wrap at end of node
982  * mask.
983  */
984 static int hstate_next_node_to_alloc(struct hstate *h,
985                                         nodemask_t *nodes_allowed)
986 {
987         int nid;
988 
989         VM_BUG_ON(!nodes_allowed);
990 
991         nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
992         h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
993 
994         return nid;
995 }
996 
997 /*
998  * helper for free_pool_huge_page() - return the previously saved
999  * node ["this node"] from which to free a huge page.  Advance the
1000  * next node id whether or not we find a free huge page to free so
1001  * that the next attempt to free addresses the next node.
1002  */
1003 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1004 {
1005         int nid;
1006 
1007         VM_BUG_ON(!nodes_allowed);
1008 
1009         nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1010         h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1011 
1012         return nid;
1013 }
1014 
1015 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask)           \
1016         for (nr_nodes = nodes_weight(*mask);                            \
1017                 nr_nodes > 0 &&                                         \
1018                 ((node = hstate_next_node_to_alloc(hs, mask)) || 1);    \
1019                 nr_nodes--)
1020 
1021 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask)            \
1022         for (nr_nodes = nodes_weight(*mask);                            \
1023                 nr_nodes > 0 &&                                         \
1024                 ((node = hstate_next_node_to_free(hs, mask)) || 1);     \
1025                 nr_nodes--)
1026 
1027 #if defined(CONFIG_ARCH_HAS_GIGANTIC_PAGE) && \
1028         ((defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || \
1029         defined(CONFIG_CMA))
1030 static void destroy_compound_gigantic_page(struct page *page,
1031                                         unsigned int order)
1032 {
1033         int i;
1034         int nr_pages = 1 << order;
1035         struct page *p = page + 1;
1036 
1037         atomic_set(compound_mapcount_ptr(page), 0);
1038         for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1039                 clear_compound_head(p);
1040                 set_page_refcounted(p);
1041         }
1042 
1043         set_compound_order(page, 0);
1044         __ClearPageHead(page);
1045 }
1046 
1047 static void free_gigantic_page(struct page *page, unsigned int order)
1048 {
1049         free_contig_range(page_to_pfn(page), 1 << order);
1050 }
1051 
1052 static int __alloc_gigantic_page(unsigned long start_pfn,
1053                                 unsigned long nr_pages)
1054 {
1055         unsigned long end_pfn = start_pfn + nr_pages;
1056         return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
1057                                   GFP_KERNEL);
1058 }
1059 
1060 static bool pfn_range_valid_gigantic(struct zone *z,
1061                         unsigned long start_pfn, unsigned long nr_pages)
1062 {
1063         unsigned long i, end_pfn = start_pfn + nr_pages;
1064         struct page *page;
1065 
1066         for (i = start_pfn; i < end_pfn; i++) {
1067                 if (!pfn_valid(i))
1068                         return false;
1069 
1070                 page = pfn_to_page(i);
1071 
1072                 if (page_zone(page) != z)
1073                         return false;
1074 
1075                 if (PageReserved(page))
1076                         return false;
1077 
1078                 if (page_count(page) > 0)
1079                         return false;
1080 
1081                 if (PageHuge(page))
1082                         return false;
1083         }
1084 
1085         return true;
1086 }
1087 
1088 static bool zone_spans_last_pfn(const struct zone *zone,
1089                         unsigned long start_pfn, unsigned long nr_pages)
1090 {
1091         unsigned long last_pfn = start_pfn + nr_pages - 1;
1092         return zone_spans_pfn(zone, last_pfn);
1093 }
1094 
1095 static struct page *alloc_gigantic_page(int nid, unsigned int order)
1096 {
1097         unsigned long nr_pages = 1 << order;
1098         unsigned long ret, pfn, flags;
1099         struct zone *z;
1100 
1101         z = NODE_DATA(nid)->node_zones;
1102         for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) {
1103                 spin_lock_irqsave(&z->lock, flags);
1104 
1105                 pfn = ALIGN(z->zone_start_pfn, nr_pages);
1106                 while (zone_spans_last_pfn(z, pfn, nr_pages)) {
1107                         if (pfn_range_valid_gigantic(z, pfn, nr_pages)) {
1108                                 /*
1109                                  * We release the zone lock here because
1110                                  * alloc_contig_range() will also lock the zone
1111                                  * at some point. If there's an allocation
1112                                  * spinning on this lock, it may win the race
1113                                  * and cause alloc_contig_range() to fail...
1114                                  */
1115                                 spin_unlock_irqrestore(&z->lock, flags);
1116                                 ret = __alloc_gigantic_page(pfn, nr_pages);
1117                                 if (!ret)
1118                                         return pfn_to_page(pfn);
1119                                 spin_lock_irqsave(&z->lock, flags);
1120                         }
1121                         pfn += nr_pages;
1122                 }
1123 
1124                 spin_unlock_irqrestore(&z->lock, flags);
1125         }
1126 
1127         return NULL;
1128 }
1129 
1130 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1131 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1132 
1133 static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid)
1134 {
1135         struct page *page;
1136 
1137         page = alloc_gigantic_page(nid, huge_page_order(h));
1138         if (page) {
1139                 prep_compound_gigantic_page(page, huge_page_order(h));
1140                 prep_new_huge_page(h, page, nid);
1141         }
1142 
1143         return page;
1144 }
1145 
1146 static int alloc_fresh_gigantic_page(struct hstate *h,
1147                                 nodemask_t *nodes_allowed)
1148 {
1149         struct page *page = NULL;
1150         int nr_nodes, node;
1151 
1152         for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1153                 page = alloc_fresh_gigantic_page_node(h, node);
1154                 if (page)
1155                         return 1;
1156         }
1157 
1158         return 0;
1159 }
1160 
1161 static inline bool gigantic_page_supported(void) { return true; }
1162 #else
1163 static inline bool gigantic_page_supported(void) { return false; }
1164 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1165 static inline void destroy_compound_gigantic_page(struct page *page,
1166                                                 unsigned int order) { }
1167 static inline int alloc_fresh_gigantic_page(struct hstate *h,
1168                                         nodemask_t *nodes_allowed) { return 0; }
1169 #endif
1170 
1171 static void update_and_free_page(struct hstate *h, struct page *page)
1172 {
1173         int i;
1174 
1175         if (hstate_is_gigantic(h) && !gigantic_page_supported())
1176                 return;
1177 
1178         h->nr_huge_pages--;
1179         h->nr_huge_pages_node[page_to_nid(page)]--;
1180         for (i = 0; i < pages_per_huge_page(h); i++) {
1181                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1182                                 1 << PG_referenced | 1 << PG_dirty |
1183                                 1 << PG_active | 1 << PG_private |
1184                                 1 << PG_writeback);
1185         }
1186         VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1187         set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1188         set_page_refcounted(page);
1189         if (hstate_is_gigantic(h)) {
1190                 destroy_compound_gigantic_page(page, huge_page_order(h));
1191                 free_gigantic_page(page, huge_page_order(h));
1192         } else {
1193                 __free_pages(page, huge_page_order(h));
1194         }
1195 }
1196 
1197 struct hstate *size_to_hstate(unsigned long size)
1198 {
1199         struct hstate *h;
1200 
1201         for_each_hstate(h) {
1202                 if (huge_page_size(h) == size)
1203                         return h;
1204         }
1205         return NULL;
1206 }
1207 
1208 /*
1209  * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1210  * to hstate->hugepage_activelist.)
1211  *
1212  * This function can be called for tail pages, but never returns true for them.
1213  */
1214 bool page_huge_active(struct page *page)
1215 {
1216         VM_BUG_ON_PAGE(!PageHuge(page), page);
1217         return PageHead(page) && PagePrivate(&page[1]);
1218 }
1219 
1220 /* never called for tail page */
1221 static void set_page_huge_active(struct page *page)
1222 {
1223         VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1224         SetPagePrivate(&page[1]);
1225 }
1226 
1227 static void clear_page_huge_active(struct page *page)
1228 {
1229         VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1230         ClearPagePrivate(&page[1]);
1231 }
1232 
1233 void free_huge_page(struct page *page)
1234 {
1235         /*
1236          * Can't pass hstate in here because it is called from the
1237          * compound page destructor.
1238          */
1239         struct hstate *h = page_hstate(page);
1240         int nid = page_to_nid(page);
1241         struct hugepage_subpool *spool =
1242                 (struct hugepage_subpool *)page_private(page);
1243         bool restore_reserve;
1244 
1245         set_page_private(page, 0);
1246         page->mapping = NULL;
1247         VM_BUG_ON_PAGE(page_count(page), page);
1248         VM_BUG_ON_PAGE(page_mapcount(page), page);
1249         restore_reserve = PagePrivate(page);
1250         ClearPagePrivate(page);
1251 
1252         /*
1253          * A return code of zero implies that the subpool will be under its
1254          * minimum size if the reservation is not restored after page is free.
1255          * Therefore, force restore_reserve operation.
1256          */
1257         if (hugepage_subpool_put_pages(spool, 1) == 0)
1258                 restore_reserve = true;
1259 
1260         spin_lock(&hugetlb_lock);
1261         clear_page_huge_active(page);
1262         hugetlb_cgroup_uncharge_page(hstate_index(h),
1263                                      pages_per_huge_page(h), page);
1264         if (restore_reserve)
1265                 h->resv_huge_pages++;
1266 
1267         if (h->surplus_huge_pages_node[nid]) {
1268                 /* remove the page from active list */
1269                 list_del(&page->lru);
1270                 update_and_free_page(h, page);
1271                 h->surplus_huge_pages--;
1272                 h->surplus_huge_pages_node[nid]--;
1273         } else {
1274                 arch_clear_hugepage_flags(page);
1275                 enqueue_huge_page(h, page);
1276         }
1277         spin_unlock(&hugetlb_lock);
1278 }
1279 
1280 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1281 {
1282         INIT_LIST_HEAD(&page->lru);
1283         set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1284         spin_lock(&hugetlb_lock);
1285         set_hugetlb_cgroup(page, NULL);
1286         h->nr_huge_pages++;
1287         h->nr_huge_pages_node[nid]++;
1288         spin_unlock(&hugetlb_lock);
1289         put_page(page); /* free it into the hugepage allocator */
1290 }
1291 
1292 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1293 {
1294         int i;
1295         int nr_pages = 1 << order;
1296         struct page *p = page + 1;
1297 
1298         /* we rely on prep_new_huge_page to set the destructor */
1299         set_compound_order(page, order);
1300         __ClearPageReserved(page);
1301         __SetPageHead(page);
1302         for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1303                 /*
1304                  * For gigantic hugepages allocated through bootmem at
1305                  * boot, it's safer to be consistent with the not-gigantic
1306                  * hugepages and clear the PG_reserved bit from all tail pages
1307                  * too.  Otherwse drivers using get_user_pages() to access tail
1308                  * pages may get the reference counting wrong if they see
1309                  * PG_reserved set on a tail page (despite the head page not
1310                  * having PG_reserved set).  Enforcing this consistency between
1311                  * head and tail pages allows drivers to optimize away a check
1312                  * on the head page when they need know if put_page() is needed
1313                  * after get_user_pages().
1314                  */
1315                 __ClearPageReserved(p);
1316                 set_page_count(p, 0);
1317                 set_compound_head(p, page);
1318         }
1319         atomic_set(compound_mapcount_ptr(page), -1);
1320 }
1321 
1322 /*
1323  * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1324  * transparent huge pages.  See the PageTransHuge() documentation for more
1325  * details.
1326  */
1327 int PageHuge(struct page *page)
1328 {
1329         if (!PageCompound(page))
1330                 return 0;
1331 
1332         page = compound_head(page);
1333         return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1334 }
1335 EXPORT_SYMBOL_GPL(PageHuge);
1336 
1337 /*
1338  * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1339  * normal or transparent huge pages.
1340  */
1341 int PageHeadHuge(struct page *page_head)
1342 {
1343         if (!PageHead(page_head))
1344                 return 0;
1345 
1346         return get_compound_page_dtor(page_head) == free_huge_page;
1347 }
1348 
1349 pgoff_t __basepage_index(struct page *page)
1350 {
1351         struct page *page_head = compound_head(page);
1352         pgoff_t index = page_index(page_head);
1353         unsigned long compound_idx;
1354 
1355         if (!PageHuge(page_head))
1356                 return page_index(page);
1357 
1358         if (compound_order(page_head) >= MAX_ORDER)
1359                 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1360         else
1361                 compound_idx = page - page_head;
1362 
1363         return (index << compound_order(page_head)) + compound_idx;
1364 }
1365 
1366 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
1367 {
1368         struct page *page;
1369 
1370         page = __alloc_pages_node(nid,
1371                 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
1372                                                 __GFP_REPEAT|__GFP_NOWARN,
1373                 huge_page_order(h));
1374         if (page) {
1375                 prep_new_huge_page(h, page, nid);
1376         }
1377 
1378         return page;
1379 }
1380 
1381 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1382 {
1383         struct page *page;
1384         int nr_nodes, node;
1385         int ret = 0;
1386 
1387         for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1388                 page = alloc_fresh_huge_page_node(h, node);
1389                 if (page) {
1390                         ret = 1;
1391                         break;
1392                 }
1393         }
1394 
1395         if (ret)
1396                 count_vm_event(HTLB_BUDDY_PGALLOC);
1397         else
1398                 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1399 
1400         return ret;
1401 }
1402 
1403 /*
1404  * Free huge page from pool from next node to free.
1405  * Attempt to keep persistent huge pages more or less
1406  * balanced over allowed nodes.
1407  * Called with hugetlb_lock locked.
1408  */
1409 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1410                                                          bool acct_surplus)
1411 {
1412         int nr_nodes, node;
1413         int ret = 0;
1414 
1415         for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1416                 /*
1417                  * If we're returning unused surplus pages, only examine
1418                  * nodes with surplus pages.
1419                  */
1420                 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1421                     !list_empty(&h->hugepage_freelists[node])) {
1422                         struct page *page =
1423                                 list_entry(h->hugepage_freelists[node].next,
1424                                           struct page, lru);
1425                         list_del(&page->lru);
1426                         h->free_huge_pages--;
1427                         h->free_huge_pages_node[node]--;
1428                         if (acct_surplus) {
1429                                 h->surplus_huge_pages--;
1430                                 h->surplus_huge_pages_node[node]--;
1431                         }
1432                         update_and_free_page(h, page);
1433                         ret = 1;
1434                         break;
1435                 }
1436         }
1437 
1438         return ret;
1439 }
1440 
1441 /*
1442  * Dissolve a given free hugepage into free buddy pages. This function does
1443  * nothing for in-use (including surplus) hugepages. Returns -EBUSY if the
1444  * number of free hugepages would be reduced below the number of reserved
1445  * hugepages.
1446  */
1447 static int dissolve_free_huge_page(struct page *page)
1448 {
1449         int rc = 0;
1450 
1451         spin_lock(&hugetlb_lock);
1452         if (PageHuge(page) && !page_count(page)) {
1453                 struct page *head = compound_head(page);
1454                 struct hstate *h = page_hstate(head);
1455                 int nid = page_to_nid(head);
1456                 if (h->free_huge_pages - h->resv_huge_pages == 0) {
1457                         rc = -EBUSY;
1458                         goto out;
1459                 }
1460                 list_del(&head->lru);
1461                 h->free_huge_pages--;
1462                 h->free_huge_pages_node[nid]--;
1463                 h->max_huge_pages--;
1464                 update_and_free_page(h, head);
1465         }
1466 out:
1467         spin_unlock(&hugetlb_lock);
1468         return rc;
1469 }
1470 
1471 /*
1472  * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1473  * make specified memory blocks removable from the system.
1474  * Note that this will dissolve a free gigantic hugepage completely, if any
1475  * part of it lies within the given range.
1476  * Also note that if dissolve_free_huge_page() returns with an error, all
1477  * free hugepages that were dissolved before that error are lost.
1478  */
1479 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1480 {
1481         unsigned long pfn;
1482         struct page *page;
1483         int rc = 0;
1484 
1485         if (!hugepages_supported())
1486                 return rc;
1487 
1488         for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1489                 page = pfn_to_page(pfn);
1490                 if (PageHuge(page) && !page_count(page)) {
1491                         rc = dissolve_free_huge_page(page);
1492                         if (rc)
1493                                 break;
1494                 }
1495         }
1496 
1497         return rc;
1498 }
1499 
1500 /*
1501  * There are 3 ways this can get called:
1502  * 1. With vma+addr: we use the VMA's memory policy
1503  * 2. With !vma, but nid=NUMA_NO_NODE:  We try to allocate a huge
1504  *    page from any node, and let the buddy allocator itself figure
1505  *    it out.
1506  * 3. With !vma, but nid!=NUMA_NO_NODE.  We allocate a huge page
1507  *    strictly from 'nid'
1508  */
1509 static struct page *__hugetlb_alloc_buddy_huge_page(struct hstate *h,
1510                 struct vm_area_struct *vma, unsigned long addr, int nid)
1511 {
1512         int order = huge_page_order(h);
1513         gfp_t gfp = htlb_alloc_mask(h)|__GFP_COMP|__GFP_REPEAT|__GFP_NOWARN;
1514         unsigned int cpuset_mems_cookie;
1515 
1516         /*
1517          * We need a VMA to get a memory policy.  If we do not
1518          * have one, we use the 'nid' argument.
1519          *
1520          * The mempolicy stuff below has some non-inlined bits
1521          * and calls ->vm_ops.  That makes it hard to optimize at
1522          * compile-time, even when NUMA is off and it does
1523          * nothing.  This helps the compiler optimize it out.
1524          */
1525         if (!IS_ENABLED(CONFIG_NUMA) || !vma) {
1526                 /*
1527                  * If a specific node is requested, make sure to
1528                  * get memory from there, but only when a node
1529                  * is explicitly specified.
1530                  */
1531                 if (nid != NUMA_NO_NODE)
1532                         gfp |= __GFP_THISNODE;
1533                 /*
1534                  * Make sure to call something that can handle
1535                  * nid=NUMA_NO_NODE
1536                  */
1537                 return alloc_pages_node(nid, gfp, order);
1538         }
1539 
1540         /*
1541          * OK, so we have a VMA.  Fetch the mempolicy and try to
1542          * allocate a huge page with it.  We will only reach this
1543          * when CONFIG_NUMA=y.
1544          */
1545         do {
1546                 struct page *page;
1547                 struct mempolicy *mpol;
1548                 struct zonelist *zl;
1549                 nodemask_t *nodemask;
1550 
1551                 cpuset_mems_cookie = read_mems_allowed_begin();
1552                 zl = huge_zonelist(vma, addr, gfp, &mpol, &nodemask);
1553                 mpol_cond_put(mpol);
1554                 page = __alloc_pages_nodemask(gfp, order, zl, nodemask);
1555                 if (page)
1556                         return page;
1557         } while (read_mems_allowed_retry(cpuset_mems_cookie));
1558 
1559         return NULL;
1560 }
1561 
1562 /*
1563  * There are two ways to allocate a huge page:
1564  * 1. When you have a VMA and an address (like a fault)
1565  * 2. When you have no VMA (like when setting /proc/.../nr_hugepages)
1566  *
1567  * 'vma' and 'addr' are only for (1).  'nid' is always NUMA_NO_NODE in
1568  * this case which signifies that the allocation should be done with
1569  * respect for the VMA's memory policy.
1570  *
1571  * For (2), we ignore 'vma' and 'addr' and use 'nid' exclusively. This
1572  * implies that memory policies will not be taken in to account.
1573  */
1574 static struct page *__alloc_buddy_huge_page(struct hstate *h,
1575                 struct vm_area_struct *vma, unsigned long addr, int nid)
1576 {
1577         struct page *page;
1578         unsigned int r_nid;
1579 
1580         if (hstate_is_gigantic(h))
1581                 return NULL;
1582 
1583         /*
1584          * Make sure that anyone specifying 'nid' is not also specifying a VMA.
1585          * This makes sure the caller is picking _one_ of the modes with which
1586          * we can call this function, not both.
1587          */
1588         if (vma || (addr != -1)) {
1589                 VM_WARN_ON_ONCE(addr == -1);
1590                 VM_WARN_ON_ONCE(nid != NUMA_NO_NODE);
1591         }
1592         /*
1593          * Assume we will successfully allocate the surplus page to
1594          * prevent racing processes from causing the surplus to exceed
1595          * overcommit
1596          *
1597          * This however introduces a different race, where a process B
1598          * tries to grow the static hugepage pool while alloc_pages() is
1599          * called by process A. B will only examine the per-node
1600          * counters in determining if surplus huge pages can be
1601          * converted to normal huge pages in adjust_pool_surplus(). A
1602          * won't be able to increment the per-node counter, until the
1603          * lock is dropped by B, but B doesn't drop hugetlb_lock until
1604          * no more huge pages can be converted from surplus to normal
1605          * state (and doesn't try to convert again). Thus, we have a
1606          * case where a surplus huge page exists, the pool is grown, and
1607          * the surplus huge page still exists after, even though it
1608          * should just have been converted to a normal huge page. This
1609          * does not leak memory, though, as the hugepage will be freed
1610          * once it is out of use. It also does not allow the counters to
1611          * go out of whack in adjust_pool_surplus() as we don't modify
1612          * the node values until we've gotten the hugepage and only the
1613          * per-node value is checked there.
1614          */
1615         spin_lock(&hugetlb_lock);
1616         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1617                 spin_unlock(&hugetlb_lock);
1618                 return NULL;
1619         } else {
1620                 h->nr_huge_pages++;
1621                 h->surplus_huge_pages++;
1622         }
1623         spin_unlock(&hugetlb_lock);
1624 
1625         page = __hugetlb_alloc_buddy_huge_page(h, vma, addr, nid);
1626 
1627         spin_lock(&hugetlb_lock);
1628         if (page) {
1629                 INIT_LIST_HEAD(&page->lru);
1630                 r_nid = page_to_nid(page);
1631                 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1632                 set_hugetlb_cgroup(page, NULL);
1633                 /*
1634                  * We incremented the global counters already
1635                  */
1636                 h->nr_huge_pages_node[r_nid]++;
1637                 h->surplus_huge_pages_node[r_nid]++;
1638                 __count_vm_event(HTLB_BUDDY_PGALLOC);
1639         } else {
1640                 h->nr_huge_pages--;
1641                 h->surplus_huge_pages--;
1642                 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1643         }
1644         spin_unlock(&hugetlb_lock);
1645 
1646         return page;
1647 }
1648 
1649 /*
1650  * Allocate a huge page from 'nid'.  Note, 'nid' may be
1651  * NUMA_NO_NODE, which means that it may be allocated
1652  * anywhere.
1653  */
1654 static
1655 struct page *__alloc_buddy_huge_page_no_mpol(struct hstate *h, int nid)
1656 {
1657         unsigned long addr = -1;
1658 
1659         return __alloc_buddy_huge_page(h, NULL, addr, nid);
1660 }
1661 
1662 /*
1663  * Use the VMA's mpolicy to allocate a huge page from the buddy.
1664  */
1665 static
1666 struct page *__alloc_buddy_huge_page_with_mpol(struct hstate *h,
1667                 struct vm_area_struct *vma, unsigned long addr)
1668 {
1669         return __alloc_buddy_huge_page(h, vma, addr, NUMA_NO_NODE);
1670 }
1671 
1672 /*
1673  * This allocation function is useful in the context where vma is irrelevant.
1674  * E.g. soft-offlining uses this function because it only cares physical
1675  * address of error page.
1676  */
1677 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1678 {
1679         struct page *page = NULL;
1680 
1681         spin_lock(&hugetlb_lock);
1682         if (h->free_huge_pages - h->resv_huge_pages > 0)
1683                 page = dequeue_huge_page_node(h, nid);
1684         spin_unlock(&hugetlb_lock);
1685 
1686         if (!page)
1687                 page = __alloc_buddy_huge_page_no_mpol(h, nid);
1688 
1689         return page;
1690 }
1691 
1692 /*
1693  * Increase the hugetlb pool such that it can accommodate a reservation
1694  * of size 'delta'.
1695  */
1696 static int gather_surplus_pages(struct hstate *h, int delta)
1697 {
1698         struct list_head surplus_list;
1699         struct page *page, *tmp;
1700         int ret, i;
1701         int needed, allocated;
1702         bool alloc_ok = true;
1703 
1704         needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1705         if (needed <= 0) {
1706                 h->resv_huge_pages += delta;
1707                 return 0;
1708         }
1709 
1710         allocated = 0;
1711         INIT_LIST_HEAD(&surplus_list);
1712 
1713         ret = -ENOMEM;
1714 retry:
1715         spin_unlock(&hugetlb_lock);
1716         for (i = 0; i < needed; i++) {
1717                 page = __alloc_buddy_huge_page_no_mpol(h, NUMA_NO_NODE);
1718                 if (!page) {
1719                         alloc_ok = false;
1720                         break;
1721                 }
1722                 list_add(&page->lru, &surplus_list);
1723         }
1724         allocated += i;
1725 
1726         /*
1727          * After retaking hugetlb_lock, we need to recalculate 'needed'
1728          * because either resv_huge_pages or free_huge_pages may have changed.
1729          */
1730         spin_lock(&hugetlb_lock);
1731         needed = (h->resv_huge_pages + delta) -
1732                         (h->free_huge_pages + allocated);
1733         if (needed > 0) {
1734                 if (alloc_ok)
1735                         goto retry;
1736                 /*
1737                  * We were not able to allocate enough pages to
1738                  * satisfy the entire reservation so we free what
1739                  * we've allocated so far.
1740                  */
1741                 goto free;
1742         }
1743         /*
1744          * The surplus_list now contains _at_least_ the number of extra pages
1745          * needed to accommodate the reservation.  Add the appropriate number
1746          * of pages to the hugetlb pool and free the extras back to the buddy
1747          * allocator.  Commit the entire reservation here to prevent another
1748          * process from stealing the pages as they are added to the pool but
1749          * before they are reserved.
1750          */
1751         needed += allocated;
1752         h->resv_huge_pages += delta;
1753         ret = 0;
1754 
1755         /* Free the needed pages to the hugetlb pool */
1756         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1757                 if ((--needed) < 0)
1758                         break;
1759                 /*
1760                  * This page is now managed by the hugetlb allocator and has
1761                  * no users -- drop the buddy allocator's reference.
1762                  */
1763                 put_page_testzero(page);
1764                 VM_BUG_ON_PAGE(page_count(page), page);
1765                 enqueue_huge_page(h, page);
1766         }
1767 free:
1768         spin_unlock(&hugetlb_lock);
1769 
1770         /* Free unnecessary surplus pages to the buddy allocator */
1771         list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1772                 put_page(page);
1773         spin_lock(&hugetlb_lock);
1774 
1775         return ret;
1776 }
1777 
1778 /*
1779  * This routine has two main purposes:
1780  * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1781  *    in unused_resv_pages.  This corresponds to the prior adjustments made
1782  *    to the associated reservation map.
1783  * 2) Free any unused surplus pages that may have been allocated to satisfy
1784  *    the reservation.  As many as unused_resv_pages may be freed.
1785  *
1786  * Called with hugetlb_lock held.  However, the lock could be dropped (and
1787  * reacquired) during calls to cond_resched_lock.  Whenever dropping the lock,
1788  * we must make sure nobody else can claim pages we are in the process of
1789  * freeing.  Do this by ensuring resv_huge_page always is greater than the
1790  * number of huge pages we plan to free when dropping the lock.
1791  */
1792 static void return_unused_surplus_pages(struct hstate *h,
1793                                         unsigned long unused_resv_pages)
1794 {
1795         unsigned long nr_pages;
1796 
1797         /* Cannot return gigantic pages currently */
1798         if (hstate_is_gigantic(h))
1799                 goto out;
1800 
1801         /*
1802          * Part (or even all) of the reservation could have been backed
1803          * by pre-allocated pages. Only free surplus pages.
1804          */
1805         nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1806 
1807         /*
1808          * We want to release as many surplus pages as possible, spread
1809          * evenly across all nodes with memory. Iterate across these nodes
1810          * until we can no longer free unreserved surplus pages. This occurs
1811          * when the nodes with surplus pages have no free pages.
1812          * free_pool_huge_page() will balance the the freed pages across the
1813          * on-line nodes with memory and will handle the hstate accounting.
1814          *
1815          * Note that we decrement resv_huge_pages as we free the pages.  If
1816          * we drop the lock, resv_huge_pages will still be sufficiently large
1817          * to cover subsequent pages we may free.
1818          */
1819         while (nr_pages--) {
1820                 h->resv_huge_pages--;
1821                 unused_resv_pages--;
1822                 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1823                         goto out;
1824                 cond_resched_lock(&hugetlb_lock);
1825         }
1826 
1827 out:
1828         /* Fully uncommit the reservation */
1829         h->resv_huge_pages -= unused_resv_pages;
1830 }
1831 
1832 
1833 /*
1834  * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1835  * are used by the huge page allocation routines to manage reservations.
1836  *
1837  * vma_needs_reservation is called to determine if the huge page at addr
1838  * within the vma has an associated reservation.  If a reservation is
1839  * needed, the value 1 is returned.  The caller is then responsible for
1840  * managing the global reservation and subpool usage counts.  After
1841  * the huge page has been allocated, vma_commit_reservation is called
1842  * to add the page to the reservation map.  If the page allocation fails,
1843  * the reservation must be ended instead of committed.  vma_end_reservation
1844  * is called in such cases.
1845  *
1846  * In the normal case, vma_commit_reservation returns the same value
1847  * as the preceding vma_needs_reservation call.  The only time this
1848  * is not the case is if a reserve map was changed between calls.  It
1849  * is the responsibility of the caller to notice the difference and
1850  * take appropriate action.
1851  *
1852  * vma_add_reservation is used in error paths where a reservation must
1853  * be restored when a newly allocated huge page must be freed.  It is
1854  * to be called after calling vma_needs_reservation to determine if a
1855  * reservation exists.
1856  */
1857 enum vma_resv_mode {
1858         VMA_NEEDS_RESV,
1859         VMA_COMMIT_RESV,
1860         VMA_END_RESV,
1861         VMA_ADD_RESV,
1862 };
1863 static long __vma_reservation_common(struct hstate *h,
1864                                 struct vm_area_struct *vma, unsigned long addr,
1865                                 enum vma_resv_mode mode)
1866 {
1867         struct resv_map *resv;
1868         pgoff_t idx;
1869         long ret;
1870 
1871         resv = vma_resv_map(vma);
1872         if (!resv)
1873                 return 1;
1874 
1875         idx = vma_hugecache_offset(h, vma, addr);
1876         switch (mode) {
1877         case VMA_NEEDS_RESV:
1878                 ret = region_chg(resv, idx, idx + 1);
1879                 break;
1880         case VMA_COMMIT_RESV:
1881                 ret = region_add(resv, idx, idx + 1);
1882                 break;
1883         case VMA_END_RESV:
1884                 region_abort(resv, idx, idx + 1);
1885                 ret = 0;
1886                 break;
1887         case VMA_ADD_RESV:
1888                 if (vma->vm_flags & VM_MAYSHARE)
1889                         ret = region_add(resv, idx, idx + 1);
1890                 else {
1891                         region_abort(resv, idx, idx + 1);
1892                         ret = region_del(resv, idx, idx + 1);
1893                 }
1894                 break;
1895         default:
1896                 BUG();
1897         }
1898 
1899         if (vma->vm_flags & VM_MAYSHARE)
1900                 return ret;
1901         else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
1902                 /*
1903                  * In most cases, reserves always exist for private mappings.
1904                  * However, a file associated with mapping could have been
1905                  * hole punched or truncated after reserves were consumed.
1906                  * As subsequent fault on such a range will not use reserves.
1907                  * Subtle - The reserve map for private mappings has the
1908                  * opposite meaning than that of shared mappings.  If NO
1909                  * entry is in the reserve map, it means a reservation exists.
1910                  * If an entry exists in the reserve map, it means the
1911                  * reservation has already been consumed.  As a result, the
1912                  * return value of this routine is the opposite of the
1913                  * value returned from reserve map manipulation routines above.
1914                  */
1915                 if (ret)
1916                         return 0;
1917                 else
1918                         return 1;
1919         }
1920         else
1921                 return ret < 0 ? ret : 0;
1922 }
1923 
1924 static long vma_needs_reservation(struct hstate *h,
1925                         struct vm_area_struct *vma, unsigned long addr)
1926 {
1927         return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
1928 }
1929 
1930 static long vma_commit_reservation(struct hstate *h,
1931                         struct vm_area_struct *vma, unsigned long addr)
1932 {
1933         return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1934 }
1935 
1936 static void vma_end_reservation(struct hstate *h,
1937                         struct vm_area_struct *vma, unsigned long addr)
1938 {
1939         (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
1940 }
1941 
1942 static long vma_add_reservation(struct hstate *h,
1943                         struct vm_area_struct *vma, unsigned long addr)
1944 {
1945         return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
1946 }
1947 
1948 /*
1949  * This routine is called to restore a reservation on error paths.  In the
1950  * specific error paths, a huge page was allocated (via alloc_huge_page)
1951  * and is about to be freed.  If a reservation for the page existed,
1952  * alloc_huge_page would have consumed the reservation and set PagePrivate
1953  * in the newly allocated page.  When the page is freed via free_huge_page,
1954  * the global reservation count will be incremented if PagePrivate is set.
1955  * However, free_huge_page can not adjust the reserve map.  Adjust the
1956  * reserve map here to be consistent with global reserve count adjustments
1957  * to be made by free_huge_page.
1958  */
1959 static void restore_reserve_on_error(struct hstate *h,
1960                         struct vm_area_struct *vma, unsigned long address,
1961                         struct page *page)
1962 {
1963         if (unlikely(PagePrivate(page))) {
1964                 long rc = vma_needs_reservation(h, vma, address);
1965 
1966                 if (unlikely(rc < 0)) {
1967                         /*
1968                          * Rare out of memory condition in reserve map
1969                          * manipulation.  Clear PagePrivate so that
1970                          * global reserve count will not be incremented
1971                          * by free_huge_page.  This will make it appear
1972                          * as though the reservation for this page was
1973                          * consumed.  This may prevent the task from
1974                          * faulting in the page at a later time.  This
1975                          * is better than inconsistent global huge page
1976                          * accounting of reserve counts.
1977                          */
1978                         ClearPagePrivate(page);
1979                 } else if (rc) {
1980                         rc = vma_add_reservation(h, vma, address);
1981                         if (unlikely(rc < 0))
1982                                 /*
1983                                  * See above comment about rare out of
1984                                  * memory condition.
1985                                  */
1986                                 ClearPagePrivate(page);
1987                 } else
1988                         vma_end_reservation(h, vma, address);
1989         }
1990 }
1991 
1992 struct page *alloc_huge_page(struct vm_area_struct *vma,
1993                                     unsigned long addr, int avoid_reserve)
1994 {
1995         struct hugepage_subpool *spool = subpool_vma(vma);
1996         struct hstate *h = hstate_vma(vma);
1997         struct page *page;
1998         long map_chg, map_commit;
1999         long gbl_chg;
2000         int ret, idx;
2001         struct hugetlb_cgroup *h_cg;
2002 
2003         idx = hstate_index(h);
2004         /*
2005          * Examine the region/reserve map to determine if the process
2006          * has a reservation for the page to be allocated.  A return
2007          * code of zero indicates a reservation exists (no change).
2008          */
2009         map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2010         if (map_chg < 0)
2011                 return ERR_PTR(-ENOMEM);
2012 
2013         /*
2014          * Processes that did not create the mapping will have no
2015          * reserves as indicated by the region/reserve map. Check
2016          * that the allocation will not exceed the subpool limit.
2017          * Allocations for MAP_NORESERVE mappings also need to be
2018          * checked against any subpool limit.
2019          */
2020         if (map_chg || avoid_reserve) {
2021                 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2022                 if (gbl_chg < 0) {
2023                         vma_end_reservation(h, vma, addr);
2024                         return ERR_PTR(-ENOSPC);
2025                 }
2026 
2027                 /*
2028                  * Even though there was no reservation in the region/reserve
2029                  * map, there could be reservations associated with the
2030                  * subpool that can be used.  This would be indicated if the
2031                  * return value of hugepage_subpool_get_pages() is zero.
2032                  * However, if avoid_reserve is specified we still avoid even
2033                  * the subpool reservations.
2034                  */
2035                 if (avoid_reserve)
2036                         gbl_chg = 1;
2037         }
2038 
2039         ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2040         if (ret)
2041                 goto out_subpool_put;
2042 
2043         spin_lock(&hugetlb_lock);
2044         /*
2045          * glb_chg is passed to indicate whether or not a page must be taken
2046          * from the global free pool (global change).  gbl_chg == 0 indicates
2047          * a reservation exists for the allocation.
2048          */
2049         page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2050         if (!page) {
2051                 spin_unlock(&hugetlb_lock);
2052                 page = __alloc_buddy_huge_page_with_mpol(h, vma, addr);
2053                 if (!page)
2054                         goto out_uncharge_cgroup;
2055                 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2056                         SetPagePrivate(page);
2057                         h->resv_huge_pages--;
2058                 }
2059                 spin_lock(&hugetlb_lock);
2060                 list_move(&page->lru, &h->hugepage_activelist);
2061                 /* Fall through */
2062         }
2063         hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2064         spin_unlock(&hugetlb_lock);
2065 
2066         set_page_private(page, (unsigned long)spool);
2067 
2068         map_commit = vma_commit_reservation(h, vma, addr);
2069         if (unlikely(map_chg > map_commit)) {
2070                 /*
2071                  * The page was added to the reservation map between
2072                  * vma_needs_reservation and vma_commit_reservation.
2073                  * This indicates a race with hugetlb_reserve_pages.
2074                  * Adjust for the subpool count incremented above AND
2075                  * in hugetlb_reserve_pages for the same page.  Also,
2076                  * the reservation count added in hugetlb_reserve_pages
2077                  * no longer applies.
2078                  */
2079                 long rsv_adjust;
2080 
2081                 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2082                 hugetlb_acct_memory(h, -rsv_adjust);
2083         }
2084         return page;
2085 
2086 out_uncharge_cgroup:
2087         hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2088 out_subpool_put:
2089         if (map_chg || avoid_reserve)
2090                 hugepage_subpool_put_pages(spool, 1);
2091         vma_end_reservation(h, vma, addr);
2092         return ERR_PTR(-ENOSPC);
2093 }
2094 
2095 /*
2096  * alloc_huge_page()'s wrapper which simply returns the page if allocation
2097  * succeeds, otherwise NULL. This function is called from new_vma_page(),
2098  * where no ERR_VALUE is expected to be returned.
2099  */
2100 struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
2101                                 unsigned long addr, int avoid_reserve)
2102 {
2103         struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
2104         if (IS_ERR(page))
2105                 page = NULL;
2106         return page;
2107 }
2108 
2109 int __weak alloc_bootmem_huge_page(struct hstate *h)
2110 {
2111         struct huge_bootmem_page *m;
2112         int nr_nodes, node;
2113 
2114         for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2115                 void *addr;
2116 
2117                 addr = memblock_virt_alloc_try_nid_nopanic(
2118                                 huge_page_size(h), huge_page_size(h),
2119                                 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
2120                 if (addr) {
2121                         /*
2122                          * Use the beginning of the huge page to store the
2123                          * huge_bootmem_page struct (until gather_bootmem
2124                          * puts them into the mem_map).
2125                          */
2126                         m = addr;
2127                         goto found;
2128                 }
2129         }
2130         return 0;
2131 
2132 found:
2133         BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2134         /* Put them into a private list first because mem_map is not up yet */
2135         list_add(&m->list, &huge_boot_pages);
2136         m->hstate = h;
2137         return 1;
2138 }
2139 
2140 static void __init prep_compound_huge_page(struct page *page,
2141                 unsigned int order)
2142 {
2143         if (unlikely(order > (MAX_ORDER - 1)))
2144                 prep_compound_gigantic_page(page, order);
2145         else
2146                 prep_compound_page(page, order);
2147 }
2148 
2149 /* Put bootmem huge pages into the standard lists after mem_map is up */
2150 static void __init gather_bootmem_prealloc(void)
2151 {
2152         struct huge_bootmem_page *m;
2153 
2154         list_for_each_entry(m, &huge_boot_pages, list) {
2155                 struct hstate *h = m->hstate;
2156                 struct page *page;
2157 
2158 #ifdef CONFIG_HIGHMEM
2159                 page = pfn_to_page(m->phys >> PAGE_SHIFT);
2160                 memblock_free_late(__pa(m),
2161                                    sizeof(struct huge_bootmem_page));
2162 #else
2163                 page = virt_to_page(m);
2164 #endif
2165                 WARN_ON(page_count(page) != 1);
2166                 prep_compound_huge_page(page, h->order);
2167                 WARN_ON(PageReserved(page));
2168                 prep_new_huge_page(h, page, page_to_nid(page));
2169                 /*
2170                  * If we had gigantic hugepages allocated at boot time, we need
2171                  * to restore the 'stolen' pages to totalram_pages in order to
2172                  * fix confusing memory reports from free(1) and another
2173                  * side-effects, like CommitLimit going negative.
2174                  */
2175                 if (hstate_is_gigantic(h))
2176                         adjust_managed_page_count(page, 1 << h->order);
2177         }
2178 }
2179 
2180 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2181 {
2182         unsigned long i;
2183 
2184         for (i = 0; i < h->max_huge_pages; ++i) {
2185                 if (hstate_is_gigantic(h)) {
2186                         if (!alloc_bootmem_huge_page(h))
2187                                 break;
2188                 } else if (!alloc_fresh_huge_page(h,
2189                                          &node_states[N_MEMORY]))
2190                         break;
2191         }
2192         h->max_huge_pages = i;
2193 }
2194 
2195 static void __init hugetlb_init_hstates(void)
2196 {
2197         struct hstate *h;
2198 
2199         for_each_hstate(h) {
2200                 if (minimum_order > huge_page_order(h))
2201                         minimum_order = huge_page_order(h);
2202 
2203                 /* oversize hugepages were init'ed in early boot */
2204                 if (!hstate_is_gigantic(h))
2205                         hugetlb_hstate_alloc_pages(h);
2206         }
2207         VM_BUG_ON(minimum_order == UINT_MAX);
2208 }
2209 
2210 static char * __init memfmt(char *buf, unsigned long n)
2211 {
2212         if (n >= (1UL << 30))
2213                 sprintf(buf, "%lu GB", n >> 30);
2214         else if (n >= (1UL << 20))
2215                 sprintf(buf, "%lu MB", n >> 20);
2216         else
2217                 sprintf(buf, "%lu KB", n >> 10);
2218         return buf;
2219 }
2220 
2221 static void __init report_hugepages(void)
2222 {
2223         struct hstate *h;
2224 
2225         for_each_hstate(h) {
2226                 char buf[32];
2227                 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2228                         memfmt(buf, huge_page_size(h)),
2229                         h->free_huge_pages);
2230         }
2231 }
2232 
2233 #ifdef CONFIG_HIGHMEM
2234 static void try_to_free_low(struct hstate *h, unsigned long count,
2235                                                 nodemask_t *nodes_allowed)
2236 {
2237         int i;
2238 
2239         if (hstate_is_gigantic(h))
2240                 return;
2241 
2242         for_each_node_mask(i, *nodes_allowed) {
2243                 struct page *page, *next;
2244                 struct list_head *freel = &h->hugepage_freelists[i];
2245                 list_for_each_entry_safe(page, next, freel, lru) {
2246                         if (count >= h->nr_huge_pages)
2247                                 return;
2248                         if (PageHighMem(page))
2249                                 continue;
2250                         list_del(&page->lru);
2251                         update_and_free_page(h, page);
2252                         h->free_huge_pages--;
2253                         h->free_huge_pages_node[page_to_nid(page)]--;
2254                 }
2255         }
2256 }
2257 #else
2258 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2259                                                 nodemask_t *nodes_allowed)
2260 {
2261 }
2262 #endif
2263 
2264 /*
2265  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
2266  * balanced by operating on them in a round-robin fashion.
2267  * Returns 1 if an adjustment was made.
2268  */
2269 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2270                                 int delta)
2271 {
2272         int nr_nodes, node;
2273 
2274         VM_BUG_ON(delta != -1 && delta != 1);
2275 
2276         if (delta < 0) {
2277                 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2278                         if (h->surplus_huge_pages_node[node])
2279                                 goto found;
2280                 }
2281         } else {
2282                 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2283                         if (h->surplus_huge_pages_node[node] <
2284                                         h->nr_huge_pages_node[node])
2285                                 goto found;
2286                 }
2287         }
2288         return 0;
2289 
2290 found:
2291         h->surplus_huge_pages += delta;
2292         h->surplus_huge_pages_node[node] += delta;
2293         return 1;
2294 }
2295 
2296 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2297 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
2298                                                 nodemask_t *nodes_allowed)
2299 {
2300         unsigned long min_count, ret;
2301 
2302         if (hstate_is_gigantic(h) && !gigantic_page_supported())
2303                 return h->max_huge_pages;
2304 
2305         /*
2306          * Increase the pool size
2307          * First take pages out of surplus state.  Then make up the
2308          * remaining difference by allocating fresh huge pages.
2309          *
2310          * We might race with __alloc_buddy_huge_page() here and be unable
2311          * to convert a surplus huge page to a normal huge page. That is
2312          * not critical, though, it just means the overall size of the
2313          * pool might be one hugepage larger than it needs to be, but
2314          * within all the constraints specified by the sysctls.
2315          */
2316         spin_lock(&hugetlb_lock);
2317         while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2318                 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2319                         break;
2320         }
2321 
2322         while (count > persistent_huge_pages(h)) {
2323                 /*
2324                  * If this allocation races such that we no longer need the
2325                  * page, free_huge_page will handle it by freeing the page
2326                  * and reducing the surplus.
2327                  */
2328                 spin_unlock(&hugetlb_lock);
2329 
2330                 /* yield cpu to avoid soft lockup */
2331                 cond_resched();
2332 
2333                 if (hstate_is_gigantic(h))
2334                         ret = alloc_fresh_gigantic_page(h, nodes_allowed);
2335                 else
2336                         ret = alloc_fresh_huge_page(h, nodes_allowed);
2337                 spin_lock(&hugetlb_lock);
2338                 if (!ret)
2339                         goto out;
2340 
2341                 /* Bail for signals. Probably ctrl-c from user */
2342                 if (signal_pending(current))
2343                         goto out;
2344         }
2345 
2346         /*
2347          * Decrease the pool size
2348          * First return free pages to the buddy allocator (being careful
2349          * to keep enough around to satisfy reservations).  Then place
2350          * pages into surplus state as needed so the pool will shrink
2351          * to the desired size as pages become free.
2352          *
2353          * By placing pages into the surplus state independent of the
2354          * overcommit value, we are allowing the surplus pool size to
2355          * exceed overcommit. There are few sane options here. Since
2356          * __alloc_buddy_huge_page() is checking the global counter,
2357          * though, we'll note that we're not allowed to exceed surplus
2358          * and won't grow the pool anywhere else. Not until one of the
2359          * sysctls are changed, or the surplus pages go out of use.
2360          */
2361         min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2362         min_count = max(count, min_count);
2363         try_to_free_low(h, min_count, nodes_allowed);
2364         while (min_count < persistent_huge_pages(h)) {
2365                 if (!free_pool_huge_page(h, nodes_allowed, 0))
2366                         break;
2367                 cond_resched_lock(&hugetlb_lock);
2368         }
2369         while (count < persistent_huge_pages(h)) {
2370                 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2371                         break;
2372         }
2373 out:
2374         ret = persistent_huge_pages(h);
2375         spin_unlock(&hugetlb_lock);
2376         return ret;
2377 }
2378 
2379 #define HSTATE_ATTR_RO(_name) \
2380         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2381 
2382 #define HSTATE_ATTR(_name) \
2383         static struct kobj_attribute _name##_attr = \
2384                 __ATTR(_name, 0644, _name##_show, _name##_store)
2385 
2386 static struct kobject *hugepages_kobj;
2387 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2388 
2389 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2390 
2391 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2392 {
2393         int i;
2394 
2395         for (i = 0; i < HUGE_MAX_HSTATE; i++)
2396                 if (hstate_kobjs[i] == kobj) {
2397                         if (nidp)
2398                                 *nidp = NUMA_NO_NODE;
2399                         return &hstates[i];
2400                 }
2401 
2402         return kobj_to_node_hstate(kobj, nidp);
2403 }
2404 
2405 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2406                                         struct kobj_attribute *attr, char *buf)
2407 {
2408         struct hstate *h;
2409         unsigned long nr_huge_pages;
2410         int nid;
2411 
2412         h = kobj_to_hstate(kobj, &nid);
2413         if (nid == NUMA_NO_NODE)
2414                 nr_huge_pages = h->nr_huge_pages;
2415         else
2416                 nr_huge_pages = h->nr_huge_pages_node[nid];
2417 
2418         return sprintf(buf, "%lu\n", nr_huge_pages);
2419 }
2420 
2421 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2422                                            struct hstate *h, int nid,
2423                                            unsigned long count, size_t len)
2424 {
2425         int err;
2426         NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
2427 
2428         if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
2429                 err = -EINVAL;
2430                 goto out;
2431         }
2432 
2433         if (nid == NUMA_NO_NODE) {
2434                 /*
2435                  * global hstate attribute
2436                  */
2437                 if (!(obey_mempolicy &&
2438                                 init_nodemask_of_mempolicy(nodes_allowed))) {
2439                         NODEMASK_FREE(nodes_allowed);
2440                         nodes_allowed = &node_states[N_MEMORY];
2441                 }
2442         } else if (nodes_allowed) {
2443                 /*
2444                  * per node hstate attribute: adjust count to global,
2445                  * but restrict alloc/free to the specified node.
2446                  */
2447                 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2448                 init_nodemask_of_node(nodes_allowed, nid);
2449         } else
2450                 nodes_allowed = &node_states[N_MEMORY];
2451 
2452         h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
2453 
2454         if (nodes_allowed != &node_states[N_MEMORY])
2455                 NODEMASK_FREE(nodes_allowed);
2456 
2457         return len;
2458 out:
2459         NODEMASK_FREE(nodes_allowed);
2460         return err;
2461 }
2462 
2463 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2464                                          struct kobject *kobj, const char *buf,
2465                                          size_t len)
2466 {
2467         struct hstate *h;
2468         unsigned long count;
2469         int nid;
2470         int err;
2471 
2472         err = kstrtoul(buf, 10, &count);
2473         if (err)
2474                 return err;
2475 
2476         h = kobj_to_hstate(kobj, &nid);
2477         return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2478 }
2479 
2480 static ssize_t nr_hugepages_show(struct kobject *kobj,
2481                                        struct kobj_attribute *attr, char *buf)
2482 {
2483         return nr_hugepages_show_common(kobj, attr, buf);
2484 }
2485 
2486 static ssize_t nr_hugepages_store(struct kobject *kobj,
2487                struct kobj_attribute *attr, const char *buf, size_t len)
2488 {
2489         return nr_hugepages_store_common(false, kobj, buf, len);
2490 }
2491 HSTATE_ATTR(nr_hugepages);
2492 
2493 #ifdef CONFIG_NUMA
2494 
2495 /*
2496  * hstate attribute for optionally mempolicy-based constraint on persistent
2497  * huge page alloc/free.
2498  */
2499 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2500                                        struct kobj_attribute *attr, char *buf)
2501 {
2502         return nr_hugepages_show_common(kobj, attr, buf);
2503 }
2504 
2505 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2506                struct kobj_attribute *attr, const char *buf, size_t len)
2507 {
2508         return nr_hugepages_store_common(true, kobj, buf, len);
2509 }
2510 HSTATE_ATTR(nr_hugepages_mempolicy);
2511 #endif
2512 
2513 
2514 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2515                                         struct kobj_attribute *attr, char *buf)
2516 {
2517         struct hstate *h = kobj_to_hstate(kobj, NULL);
2518         return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2519 }
2520 
2521 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2522                 struct kobj_attribute *attr, const char *buf, size_t count)
2523 {
2524         int err;
2525         unsigned long input;
2526         struct hstate *h = kobj_to_hstate(kobj, NULL);
2527 
2528         if (hstate_is_gigantic(h))
2529                 return -EINVAL;
2530 
2531         err = kstrtoul(buf, 10, &input);
2532         if (err)
2533                 return err;
2534 
2535         spin_lock(&hugetlb_lock);
2536         h->nr_overcommit_huge_pages = input;
2537         spin_unlock(&hugetlb_lock);
2538 
2539         return count;
2540 }
2541 HSTATE_ATTR(nr_overcommit_hugepages);
2542 
2543 static ssize_t free_hugepages_show(struct kobject *kobj,
2544                                         struct kobj_attribute *attr, char *buf)
2545 {
2546         struct hstate *h;
2547         unsigned long free_huge_pages;
2548         int nid;
2549 
2550         h = kobj_to_hstate(kobj, &nid);
2551         if (nid == NUMA_NO_NODE)
2552                 free_huge_pages = h->free_huge_pages;
2553         else
2554                 free_huge_pages = h->free_huge_pages_node[nid];
2555 
2556         return sprintf(buf, "%lu\n", free_huge_pages);
2557 }
2558 HSTATE_ATTR_RO(free_hugepages);
2559 
2560 static ssize_t resv_hugepages_show(struct kobject *kobj,
2561                                         struct kobj_attribute *attr, char *buf)
2562 {
2563         struct hstate *h = kobj_to_hstate(kobj, NULL);
2564         return sprintf(buf, "%lu\n", h->resv_huge_pages);
2565 }
2566 HSTATE_ATTR_RO(resv_hugepages);
2567 
2568 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2569                                         struct kobj_attribute *attr, char *buf)
2570 {
2571         struct hstate *h;
2572         unsigned long surplus_huge_pages;
2573         int nid;
2574 
2575         h = kobj_to_hstate(kobj, &nid);
2576         if (nid == NUMA_NO_NODE)
2577                 surplus_huge_pages = h->surplus_huge_pages;
2578         else
2579                 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2580 
2581         return sprintf(buf, "%lu\n", surplus_huge_pages);
2582 }
2583 HSTATE_ATTR_RO(surplus_hugepages);
2584 
2585 static struct attribute *hstate_attrs[] = {
2586         &nr_hugepages_attr.attr,
2587         &nr_overcommit_hugepages_attr.attr,
2588         &free_hugepages_attr.attr,
2589         &resv_hugepages_attr.attr,
2590         &surplus_hugepages_attr.attr,
2591 #ifdef CONFIG_NUMA
2592         &nr_hugepages_mempolicy_attr.attr,
2593 #endif
2594         NULL,
2595 };
2596 
2597 static struct attribute_group hstate_attr_group = {
2598         .attrs = hstate_attrs,
2599 };
2600 
2601 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2602                                     struct kobject **hstate_kobjs,
2603                                     struct attribute_group *hstate_attr_group)
2604 {
2605         int retval;
2606         int hi = hstate_index(h);
2607 
2608         hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2609         if (!hstate_kobjs[hi])
2610                 return -ENOMEM;
2611 
2612         retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2613         if (retval)
2614                 kobject_put(hstate_kobjs[hi]);
2615 
2616         return retval;
2617 }
2618 
2619 static void __init hugetlb_sysfs_init(void)
2620 {
2621         struct hstate *h;
2622         int err;
2623 
2624         hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2625         if (!hugepages_kobj)
2626                 return;
2627 
2628         for_each_hstate(h) {
2629                 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2630                                          hstate_kobjs, &hstate_attr_group);
2631                 if (err)
2632                         pr_err("Hugetlb: Unable to add hstate %s", h->name);
2633         }
2634 }
2635 
2636 #ifdef CONFIG_NUMA
2637 
2638 /*
2639  * node_hstate/s - associate per node hstate attributes, via their kobjects,
2640  * with node devices in node_devices[] using a parallel array.  The array
2641  * index of a node device or _hstate == node id.
2642  * This is here to avoid any static dependency of the node device driver, in
2643  * the base kernel, on the hugetlb module.
2644  */
2645 struct node_hstate {
2646         struct kobject          *hugepages_kobj;
2647         struct kobject          *hstate_kobjs[HUGE_MAX_HSTATE];
2648 };
2649 static struct node_hstate node_hstates[MAX_NUMNODES];
2650 
2651 /*
2652  * A subset of global hstate attributes for node devices
2653  */
2654 static struct attribute *per_node_hstate_attrs[] = {
2655         &nr_hugepages_attr.attr,
2656         &free_hugepages_attr.attr,
2657         &surplus_hugepages_attr.attr,
2658         NULL,
2659 };
2660 
2661 static struct attribute_group per_node_hstate_attr_group = {
2662         .attrs = per_node_hstate_attrs,
2663 };
2664 
2665 /*
2666  * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2667  * Returns node id via non-NULL nidp.
2668  */
2669 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2670 {
2671         int nid;
2672 
2673         for (nid = 0; nid < nr_node_ids; nid++) {
2674                 struct node_hstate *nhs = &node_hstates[nid];
2675                 int i;
2676                 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2677                         if (nhs->hstate_kobjs[i] == kobj) {
2678                                 if (nidp)
2679                                         *nidp = nid;
2680                                 return &hstates[i];
2681                         }
2682         }
2683 
2684         BUG();
2685         return NULL;
2686 }
2687 
2688 /*
2689  * Unregister hstate attributes from a single node device.
2690  * No-op if no hstate attributes attached.
2691  */
2692 static void hugetlb_unregister_node(struct node *node)
2693 {
2694         struct hstate *h;
2695         struct node_hstate *nhs = &node_hstates[node->dev.id];
2696 
2697         if (!nhs->hugepages_kobj)
2698                 return;         /* no hstate attributes */
2699 
2700         for_each_hstate(h) {
2701                 int idx = hstate_index(h);
2702                 if (nhs->hstate_kobjs[idx]) {
2703                         kobject_put(nhs->hstate_kobjs[idx]);
2704                         nhs->hstate_kobjs[idx] = NULL;
2705                 }
2706         }
2707 
2708         kobject_put(nhs->hugepages_kobj);
2709         nhs->hugepages_kobj = NULL;
2710 }
2711 
2712 
2713 /*
2714  * Register hstate attributes for a single node device.
2715  * No-op if attributes already registered.
2716  */
2717 static void hugetlb_register_node(struct node *node)
2718 {
2719         struct hstate *h;
2720         struct node_hstate *nhs = &node_hstates[node->dev.id];
2721         int err;
2722 
2723         if (nhs->hugepages_kobj)
2724                 return;         /* already allocated */
2725 
2726         nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2727                                                         &node->dev.kobj);
2728         if (!nhs->hugepages_kobj)
2729                 return;
2730 
2731         for_each_hstate(h) {
2732                 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2733                                                 nhs->hstate_kobjs,
2734                                                 &per_node_hstate_attr_group);
2735                 if (err) {
2736                         pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2737                                 h->name, node->dev.id);
2738                         hugetlb_unregister_node(node);
2739                         break;
2740                 }
2741         }
2742 }
2743 
2744 /*
2745  * hugetlb init time:  register hstate attributes for all registered node
2746  * devices of nodes that have memory.  All on-line nodes should have
2747  * registered their associated device by this time.
2748  */
2749 static void __init hugetlb_register_all_nodes(void)
2750 {
2751         int nid;
2752 
2753         for_each_node_state(nid, N_MEMORY) {
2754                 struct node *node = node_devices[nid];
2755                 if (node->dev.id == nid)
2756                         hugetlb_register_node(node);
2757         }
2758 
2759         /*
2760          * Let the node device driver know we're here so it can
2761          * [un]register hstate attributes on node hotplug.
2762          */
2763         register_hugetlbfs_with_node(hugetlb_register_node,
2764                                      hugetlb_unregister_node);
2765 }
2766 #else   /* !CONFIG_NUMA */
2767 
2768 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2769 {
2770         BUG();
2771         if (nidp)
2772                 *nidp = -1;
2773         return NULL;
2774 }
2775 
2776 static void hugetlb_register_all_nodes(void) { }
2777 
2778 #endif
2779 
2780 static int __init hugetlb_init(void)
2781 {
2782         int i;
2783 
2784         if (!hugepages_supported())
2785                 return 0;
2786 
2787         if (!size_to_hstate(default_hstate_size)) {
2788                 default_hstate_size = HPAGE_SIZE;
2789                 if (!size_to_hstate(default_hstate_size))
2790                         hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2791         }
2792         default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2793         if (default_hstate_max_huge_pages) {
2794                 if (!default_hstate.max_huge_pages)
2795                         default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2796         }
2797 
2798         hugetlb_init_hstates();
2799         gather_bootmem_prealloc();
2800         report_hugepages();
2801 
2802         hugetlb_sysfs_init();
2803         hugetlb_register_all_nodes();
2804         hugetlb_cgroup_file_init();
2805 
2806 #ifdef CONFIG_SMP
2807         num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2808 #else
2809         num_fault_mutexes = 1;
2810 #endif
2811         hugetlb_fault_mutex_table =
2812                 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
2813         BUG_ON(!hugetlb_fault_mutex_table);
2814 
2815         for (i = 0; i < num_fault_mutexes; i++)
2816                 mutex_init(&hugetlb_fault_mutex_table[i]);
2817         return 0;
2818 }
2819 subsys_initcall(hugetlb_init);
2820 
2821 /* Should be called on processing a hugepagesz=... option */
2822 void __init hugetlb_bad_size(void)
2823 {
2824         parsed_valid_hugepagesz = false;
2825 }
2826 
2827 void __init hugetlb_add_hstate(unsigned int order)
2828 {
2829         struct hstate *h;
2830         unsigned long i;
2831 
2832         if (size_to_hstate(PAGE_SIZE << order)) {
2833                 pr_warn("hugepagesz= specified twice, ignoring\n");
2834                 return;
2835         }
2836         BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2837         BUG_ON(order == 0);
2838         h = &hstates[hugetlb_max_hstate++];
2839         h->order = order;
2840         h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2841         h->nr_huge_pages = 0;
2842         h->free_huge_pages = 0;
2843         for (i = 0; i < MAX_NUMNODES; ++i)
2844                 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2845         INIT_LIST_HEAD(&h->hugepage_activelist);
2846         h->next_nid_to_alloc = first_memory_node;
2847         h->next_nid_to_free = first_memory_node;
2848         snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2849                                         huge_page_size(h)/1024);
2850 
2851         parsed_hstate = h;
2852 }
2853 
2854 static int __init hugetlb_nrpages_setup(char *s)
2855 {
2856         unsigned long *mhp;
2857         static unsigned long *last_mhp;
2858 
2859         if (!parsed_valid_hugepagesz) {
2860                 pr_warn("hugepages = %s preceded by "
2861                         "an unsupported hugepagesz, ignoring\n", s);
2862                 parsed_valid_hugepagesz = true;
2863                 return 1;
2864         }
2865         /*
2866          * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2867          * so this hugepages= parameter goes to the "default hstate".
2868          */
2869         else if (!hugetlb_max_hstate)
2870                 mhp = &default_hstate_max_huge_pages;
2871         else
2872                 mhp = &parsed_hstate->max_huge_pages;
2873 
2874         if (mhp == last_mhp) {
2875                 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
2876                 return 1;
2877         }
2878 
2879         if (sscanf(s, "%lu", mhp) <= 0)
2880                 *mhp = 0;
2881 
2882         /*
2883          * Global state is always initialized later in hugetlb_init.
2884          * But we need to allocate >= MAX_ORDER hstates here early to still
2885          * use the bootmem allocator.
2886          */
2887         if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2888                 hugetlb_hstate_alloc_pages(parsed_hstate);
2889 
2890         last_mhp = mhp;
2891 
2892         return 1;
2893 }
2894 __setup("hugepages=", hugetlb_nrpages_setup);
2895 
2896 static int __init hugetlb_default_setup(char *s)
2897 {
2898         default_hstate_size = memparse(s, &s);
2899         return 1;
2900 }
2901 __setup("default_hugepagesz=", hugetlb_default_setup);
2902 
2903 static unsigned int cpuset_mems_nr(unsigned int *array)
2904 {
2905         int node;
2906         unsigned int nr = 0;
2907 
2908         for_each_node_mask(node, cpuset_current_mems_allowed)
2909                 nr += array[node];
2910 
2911         return nr;
2912 }
2913 
2914 #ifdef CONFIG_SYSCTL
2915 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2916                          struct ctl_table *table, int write,
2917                          void __user *buffer, size_t *length, loff_t *ppos)
2918 {
2919         struct hstate *h = &default_hstate;
2920         unsigned long tmp = h->max_huge_pages;
2921         int ret;
2922 
2923         if (!hugepages_supported())
2924                 return -EOPNOTSUPP;
2925 
2926         table->data = &tmp;
2927         table->maxlen = sizeof(unsigned long);
2928         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2929         if (ret)
2930                 goto out;
2931 
2932         if (write)
2933                 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2934                                                   NUMA_NO_NODE, tmp, *length);
2935 out:
2936         return ret;
2937 }
2938 
2939 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2940                           void __user *buffer, size_t *length, loff_t *ppos)
2941 {
2942 
2943         return hugetlb_sysctl_handler_common(false, table, write,
2944                                                         buffer, length, ppos);
2945 }
2946 
2947 #ifdef CONFIG_NUMA
2948 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2949                           void __user *buffer, size_t *length, loff_t *ppos)
2950 {
2951         return hugetlb_sysctl_handler_common(true, table, write,
2952                                                         buffer, length, ppos);
2953 }
2954 #endif /* CONFIG_NUMA */
2955 
2956 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2957                         void __user *buffer,
2958                         size_t *length, loff_t *ppos)
2959 {
2960         struct hstate *h = &default_hstate;
2961         unsigned long tmp;
2962         int ret;
2963 
2964         if (!hugepages_supported())
2965                 return -EOPNOTSUPP;
2966 
2967         tmp = h->nr_overcommit_huge_pages;
2968 
2969         if (write && hstate_is_gigantic(h))
2970                 return -EINVAL;
2971 
2972         table->data = &tmp;
2973         table->maxlen = sizeof(unsigned long);
2974         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2975         if (ret)
2976                 goto out;
2977 
2978         if (write) {
2979                 spin_lock(&hugetlb_lock);
2980                 h->nr_overcommit_huge_pages = tmp;
2981                 spin_unlock(&hugetlb_lock);
2982         }
2983 out:
2984         return ret;
2985 }
2986 
2987 #endif /* CONFIG_SYSCTL */
2988 
2989 void hugetlb_report_meminfo(struct seq_file *m)
2990 {
2991         struct hstate *h = &default_hstate;
2992         if (!hugepages_supported())
2993                 return;
2994         seq_printf(m,
2995                         "HugePages_Total:   %5lu\n"
2996                         "HugePages_Free:    %5lu\n"
2997                         "HugePages_Rsvd:    %5lu\n"
2998                         "HugePages_Surp:    %5lu\n"
2999                         "Hugepagesize:   %8lu kB\n",
3000                         h->nr_huge_pages,
3001                         h->free_huge_pages,
3002                         h->resv_huge_pages,
3003                         h->surplus_huge_pages,
3004                         1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3005 }
3006 
3007 int hugetlb_report_node_meminfo(int nid, char *buf)
3008 {
3009         struct hstate *h = &default_hstate;
3010         if (!hugepages_supported())
3011                 return 0;
3012         return sprintf(buf,
3013                 "Node %d HugePages_Total: %5u\n"
3014                 "Node %d HugePages_Free:  %5u\n"
3015                 "Node %d HugePages_Surp:  %5u\n",
3016                 nid, h->nr_huge_pages_node[nid],
3017                 nid, h->free_huge_pages_node[nid],
3018                 nid, h->surplus_huge_pages_node[nid]);
3019 }
3020 
3021 void hugetlb_show_meminfo(void)
3022 {
3023         struct hstate *h;
3024         int nid;
3025 
3026         if (!hugepages_supported())
3027                 return;
3028 
3029         for_each_node_state(nid, N_MEMORY)
3030                 for_each_hstate(h)
3031                         pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3032                                 nid,
3033                                 h->nr_huge_pages_node[nid],
3034                                 h->free_huge_pages_node[nid],
3035                                 h->surplus_huge_pages_node[nid],
3036                                 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3037 }
3038 
3039 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3040 {
3041         seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3042                    atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3043 }
3044 
3045 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3046 unsigned long hugetlb_total_pages(void)
3047 {
3048         struct hstate *h;
3049         unsigned long nr_total_pages = 0;
3050 
3051         for_each_hstate(h)
3052                 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3053         return nr_total_pages;
3054 }
3055 
3056 static int hugetlb_acct_memory(struct hstate *h, long delta)
3057 {
3058         int ret = -ENOMEM;
3059 
3060         spin_lock(&hugetlb_lock);
3061         /*
3062          * When cpuset is configured, it breaks the strict hugetlb page
3063          * reservation as the accounting is done on a global variable. Such
3064          * reservation is completely rubbish in the presence of cpuset because
3065          * the reservation is not checked against page availability for the
3066          * current cpuset. Application can still potentially OOM'ed by kernel
3067          * with lack of free htlb page in cpuset that the task is in.
3068          * Attempt to enforce strict accounting with cpuset is almost
3069          * impossible (or too ugly) because cpuset is too fluid that
3070          * task or memory node can be dynamically moved between cpusets.
3071          *
3072          * The change of semantics for shared hugetlb mapping with cpuset is
3073          * undesirable. However, in order to preserve some of the semantics,
3074          * we fall back to check against current free page availability as
3075          * a best attempt and hopefully to minimize the impact of changing
3076          * semantics that cpuset has.
3077          */
3078         if (delta > 0) {
3079                 if (gather_surplus_pages(h, delta) < 0)
3080                         goto out;
3081 
3082                 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
3083                         return_unused_surplus_pages(h, delta);
3084                         goto out;
3085                 }
3086         }
3087 
3088         ret = 0;
3089         if (delta < 0)
3090                 return_unused_surplus_pages(h, (unsigned long) -delta);
3091 
3092 out:
3093         spin_unlock(&hugetlb_lock);
3094         return ret;
3095 }
3096 
3097 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3098 {
3099         struct resv_map *resv = vma_resv_map(vma);
3100 
3101         /*
3102          * This new VMA should share its siblings reservation map if present.
3103          * The VMA will only ever have a valid reservation map pointer where
3104          * it is being copied for another still existing VMA.  As that VMA
3105          * has a reference to the reservation map it cannot disappear until
3106          * after this open call completes.  It is therefore safe to take a
3107          * new reference here without additional locking.
3108          */
3109         if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3110                 kref_get(&resv->refs);
3111 }
3112 
3113 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3114 {
3115         struct hstate *h = hstate_vma(vma);
3116         struct resv_map *resv = vma_resv_map(vma);
3117         struct hugepage_subpool *spool = subpool_vma(vma);
3118         unsigned long reserve, start, end;
3119         long gbl_reserve;
3120 
3121         if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3122                 return;
3123 
3124         start = vma_hugecache_offset(h, vma, vma->vm_start);
3125         end = vma_hugecache_offset(h, vma, vma->vm_end);
3126 
3127         reserve = (end - start) - region_count(resv, start, end);
3128 
3129         kref_put(&resv->refs, resv_map_release);
3130 
3131         if (reserve) {
3132                 /*
3133                  * Decrement reserve counts.  The global reserve count may be
3134                  * adjusted if the subpool has a minimum size.
3135                  */
3136                 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3137                 hugetlb_acct_memory(h, -gbl_reserve);
3138         }
3139 }
3140 
3141 /*
3142  * We cannot handle pagefaults against hugetlb pages at all.  They cause
3143  * handle_mm_fault() to try to instantiate regular-sized pages in the
3144  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
3145  * this far.
3146  */
3147 static int hugetlb_vm_op_fault(struct vm_fault *vmf)
3148 {
3149         BUG();
3150         return 0;
3151 }
3152 
3153 const struct vm_operations_struct hugetlb_vm_ops = {
3154         .fault = hugetlb_vm_op_fault,
3155         .open = hugetlb_vm_op_open,
3156         .close = hugetlb_vm_op_close,
3157 };
3158 
3159 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3160                                 int writable)
3161 {
3162         pte_t entry;
3163 
3164         if (writable) {
3165                 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3166                                          vma->vm_page_prot)));
3167         } else {
3168                 entry = huge_pte_wrprotect(mk_huge_pte(page,
3169                                            vma->vm_page_prot));
3170         }
3171         entry = pte_mkyoung(entry);
3172         entry = pte_mkhuge(entry);
3173         entry = arch_make_huge_pte(entry, vma, page, writable);
3174 
3175         return entry;
3176 }
3177 
3178 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3179                                    unsigned long address, pte_t *ptep)
3180 {
3181         pte_t entry;
3182 
3183         entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3184         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3185                 update_mmu_cache(vma, address, ptep);
3186 }
3187 
3188 static int is_hugetlb_entry_migration(pte_t pte)
3189 {
3190         swp_entry_t swp;
3191 
3192         if (huge_pte_none(pte) || pte_present(pte))
3193                 return 0;
3194         swp = pte_to_swp_entry(pte);
3195         if (non_swap_entry(swp) && is_migration_entry(swp))
3196                 return 1;
3197         else
3198                 return 0;
3199 }
3200 
3201 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3202 {
3203         swp_entry_t swp;
3204 
3205         if (huge_pte_none(pte) || pte_present(pte))
3206                 return 0;
3207         swp = pte_to_swp_entry(pte);
3208         if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3209                 return 1;
3210         else
3211                 return 0;
3212 }
3213 
3214 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3215                             struct vm_area_struct *vma)
3216 {
3217         pte_t *src_pte, *dst_pte, entry;
3218         struct page *ptepage;
3219         unsigned long addr;
3220         int cow;
3221         struct hstate *h = hstate_vma(vma);
3222         unsigned long sz = huge_page_size(h);
3223         unsigned long mmun_start;       /* For mmu_notifiers */
3224         unsigned long mmun_end;         /* For mmu_notifiers */
3225         int ret = 0;
3226 
3227         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3228 
3229         mmun_start = vma->vm_start;
3230         mmun_end = vma->vm_end;
3231         if (cow)
3232                 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
3233 
3234         for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3235                 spinlock_t *src_ptl, *dst_ptl;
3236                 src_pte = huge_pte_offset(src, addr);
3237                 if (!src_pte)
3238                         continue;
3239                 dst_pte = huge_pte_alloc(dst, addr, sz);
3240                 if (!dst_pte) {
3241                         ret = -ENOMEM;
3242                         break;
3243                 }
3244 
3245                 /* If the pagetables are shared don't copy or take references */
3246                 if (dst_pte == src_pte)
3247                         continue;
3248 
3249                 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3250                 src_ptl = huge_pte_lockptr(h, src, src_pte);
3251                 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3252                 entry = huge_ptep_get(src_pte);
3253                 if (huge_pte_none(entry)) { /* skip none entry */
3254                         ;
3255                 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3256                                     is_hugetlb_entry_hwpoisoned(entry))) {
3257                         swp_entry_t swp_entry = pte_to_swp_entry(entry);
3258 
3259                         if (is_write_migration_entry(swp_entry) && cow) {
3260                                 /*
3261                                  * COW mappings require pages in both
3262                                  * parent and child to be set to read.
3263                                  */
3264                                 make_migration_entry_read(&swp_entry);
3265                                 entry = swp_entry_to_pte(swp_entry);
3266                                 set_huge_pte_at(src, addr, src_pte, entry);
3267                         }
3268                         set_huge_pte_at(dst, addr, dst_pte, entry);
3269                 } else {
3270                         if (cow) {
3271                                 huge_ptep_set_wrprotect(src, addr, src_pte);
3272                                 mmu_notifier_invalidate_range(src, mmun_start,
3273                                                                    mmun_end);
3274                         }
3275                         entry = huge_ptep_get(src_pte);
3276                         ptepage = pte_page(entry);
3277                         get_page(ptepage);
3278                         page_dup_rmap(ptepage, true);
3279                         set_huge_pte_at(dst, addr, dst_pte, entry);
3280                         hugetlb_count_add(pages_per_huge_page(h), dst);
3281                 }
3282                 spin_unlock(src_ptl);
3283                 spin_unlock(dst_ptl);
3284         }
3285 
3286         if (cow)
3287                 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
3288 
3289         return ret;
3290 }
3291 
3292 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3293                             unsigned long start, unsigned long end,
3294                             struct page *ref_page)
3295 {
3296         struct mm_struct *mm = vma->vm_mm;
3297         unsigned long address;
3298         pte_t *ptep;
3299         pte_t pte;
3300         spinlock_t *ptl;
3301         struct page *page;
3302         struct hstate *h = hstate_vma(vma);
3303         unsigned long sz = huge_page_size(h);
3304         const unsigned long mmun_start = start; /* For mmu_notifiers */
3305         const unsigned long mmun_end   = end;   /* For mmu_notifiers */
3306 
3307         WARN_ON(!is_vm_hugetlb_page(vma));
3308         BUG_ON(start & ~huge_page_mask(h));
3309         BUG_ON(end & ~huge_page_mask(h));
3310 
3311         /*
3312          * This is a hugetlb vma, all the pte entries should point
3313          * to huge page.
3314          */
3315         tlb_remove_check_page_size_change(tlb, sz);
3316         tlb_start_vma(tlb, vma);
3317         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3318         address = start;
3319         for (; address < end; address += sz) {
3320                 ptep = huge_pte_offset(mm, address);
3321                 if (!ptep)
3322                         continue;
3323 
3324                 ptl = huge_pte_lock(h, mm, ptep);
3325                 if (huge_pmd_unshare(mm, &address, ptep)) {
3326                         spin_unlock(ptl);
3327                         continue;
3328                 }
3329 
3330                 pte = huge_ptep_get(ptep);
3331                 if (huge_pte_none(pte)) {
3332                         spin_unlock(ptl);
3333                         continue;
3334                 }
3335 
3336                 /*
3337                  * Migrating hugepage or HWPoisoned hugepage is already
3338                  * unmapped and its refcount is dropped, so just clear pte here.
3339                  */
3340                 if (unlikely(!pte_present(pte))) {
3341                         huge_pte_clear(mm, address, ptep);
3342                         spin_unlock(ptl);
3343                         continue;
3344                 }
3345 
3346                 page = pte_page(pte);
3347                 /*
3348                  * If a reference page is supplied, it is because a specific
3349                  * page is being unmapped, not a range. Ensure the page we
3350                  * are about to unmap is the actual page of interest.
3351                  */
3352                 if (ref_page) {
3353                         if (page != ref_page) {
3354                                 spin_unlock(ptl);
3355                                 continue;
3356                         }
3357                         /*
3358                          * Mark the VMA as having unmapped its page so that
3359                          * future faults in this VMA will fail rather than
3360                          * looking like data was lost
3361                          */
3362                         set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3363                 }
3364 
3365                 pte = huge_ptep_get_and_clear(mm, address, ptep);
3366                 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
3367                 if (huge_pte_dirty(pte))
3368                         set_page_dirty(page);
3369 
3370                 hugetlb_count_sub(pages_per_huge_page(h), mm);
3371                 page_remove_rmap(page, true);
3372 
3373                 spin_unlock(ptl);
3374                 tlb_remove_page_size(tlb, page, huge_page_size(h));
3375                 /*
3376                  * Bail out after unmapping reference page if supplied
3377                  */
3378                 if (ref_page)
3379                         break;
3380         }
3381         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3382         tlb_end_vma(tlb, vma);
3383 }
3384 
3385 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3386                           struct vm_area_struct *vma, unsigned long start,
3387                           unsigned long end, struct page *ref_page)
3388 {
3389         __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3390 
3391         /*
3392          * Clear this flag so that x86's huge_pmd_share page_table_shareable
3393          * test will fail on a vma being torn down, and not grab a page table
3394          * on its way out.  We're lucky that the flag has such an appropriate
3395          * name, and can in fact be safely cleared here. We could clear it
3396          * before the __unmap_hugepage_range above, but all that's necessary
3397          * is to clear it before releasing the i_mmap_rwsem. This works
3398          * because in the context this is called, the VMA is about to be
3399          * destroyed and the i_mmap_rwsem is held.
3400          */
3401         vma->vm_flags &= ~VM_MAYSHARE;
3402 }
3403 
3404 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3405                           unsigned long end, struct page *ref_page)
3406 {
3407         struct mm_struct *mm;
3408         struct mmu_gather tlb;
3409 
3410         mm = vma->vm_mm;
3411 
3412         tlb_gather_mmu(&tlb, mm, start, end);
3413         __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3414         tlb_finish_mmu(&tlb, start, end);
3415 }
3416 
3417 /*
3418  * This is called when the original mapper is failing to COW a MAP_PRIVATE
3419  * mappping it owns the reserve page for. The intention is to unmap the page
3420  * from other VMAs and let the children be SIGKILLed if they are faulting the
3421  * same region.
3422  */
3423 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3424                               struct page *page, unsigned long address)
3425 {
3426         struct hstate *h = hstate_vma(vma);
3427         struct vm_area_struct *iter_vma;
3428         struct address_space *mapping;
3429         pgoff_t pgoff;
3430 
3431         /*
3432          * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3433          * from page cache lookup which is in HPAGE_SIZE units.
3434          */
3435         address = address & huge_page_mask(h);
3436         pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3437                         vma->vm_pgoff;
3438         mapping = vma->vm_file->f_mapping;
3439 
3440         /*
3441          * Take the mapping lock for the duration of the table walk. As
3442          * this mapping should be shared between all the VMAs,
3443          * __unmap_hugepage_range() is called as the lock is already held
3444          */
3445         i_mmap_lock_write(mapping);
3446         vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3447                 /* Do not unmap the current VMA */
3448                 if (iter_vma == vma)
3449                         continue;
3450 
3451                 /*
3452                  * Shared VMAs have their own reserves and do not affect
3453                  * MAP_PRIVATE accounting but it is possible that a shared
3454                  * VMA is using the same page so check and skip such VMAs.
3455                  */
3456                 if (iter_vma->vm_flags & VM_MAYSHARE)
3457                         continue;
3458 
3459                 /*
3460                  * Unmap the page from other VMAs without their own reserves.
3461                  * They get marked to be SIGKILLed if they fault in these
3462                  * areas. This is because a future no-page fault on this VMA
3463                  * could insert a zeroed page instead of the data existing
3464                  * from the time of fork. This would look like data corruption
3465                  */
3466                 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3467                         unmap_hugepage_range(iter_vma, address,
3468                                              address + huge_page_size(h), page);
3469         }
3470         i_mmap_unlock_write(mapping);
3471 }
3472 
3473 /*
3474  * Hugetlb_cow() should be called with page lock of the original hugepage held.
3475  * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3476  * cannot race with other handlers or page migration.
3477  * Keep the pte_same checks anyway to make transition from the mutex easier.
3478  */
3479 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3480                        unsigned long address, pte_t *ptep,
3481                        struct page *pagecache_page, spinlock_t *ptl)
3482 {
3483         pte_t pte;
3484         struct hstate *h = hstate_vma(vma);
3485         struct page *old_page, *new_page;
3486         int ret = 0, outside_reserve = 0;
3487         unsigned long mmun_start;       /* For mmu_notifiers */
3488         unsigned long mmun_end;         /* For mmu_notifiers */
3489 
3490         pte = huge_ptep_get(ptep);
3491         old_page = pte_page(pte);
3492 
3493 retry_avoidcopy:
3494         /* If no-one else is actually using this page, avoid the copy
3495          * and just make the page writable */
3496         if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3497                 page_move_anon_rmap(old_page, vma);
3498                 set_huge_ptep_writable(vma, address, ptep);
3499                 return 0;
3500         }
3501 
3502         /*
3503          * If the process that created a MAP_PRIVATE mapping is about to
3504          * perform a COW due to a shared page count, attempt to satisfy
3505          * the allocation without using the existing reserves. The pagecache
3506          * page is used to determine if the reserve at this address was
3507          * consumed or not. If reserves were used, a partial faulted mapping
3508          * at the time of fork() could consume its reserves on COW instead
3509          * of the full address range.
3510          */
3511         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3512                         old_page != pagecache_page)
3513                 outside_reserve = 1;
3514 
3515         get_page(old_page);
3516 
3517         /*
3518          * Drop page table lock as buddy allocator may be called. It will
3519          * be acquired again before returning to the caller, as expected.
3520          */
3521         spin_unlock(ptl);
3522         new_page = alloc_huge_page(vma, address, outside_reserve);
3523 
3524         if (IS_ERR(new_page)) {
3525                 /*
3526                  * If a process owning a MAP_PRIVATE mapping fails to COW,
3527                  * it is due to references held by a child and an insufficient
3528                  * huge page pool. To guarantee the original mappers
3529                  * reliability, unmap the page from child processes. The child
3530                  * may get SIGKILLed if it later faults.
3531                  */
3532                 if (outside_reserve) {
3533                         put_page(old_page);
3534                         BUG_ON(huge_pte_none(pte));
3535                         unmap_ref_private(mm, vma, old_page, address);
3536                         BUG_ON(huge_pte_none(pte));
3537                         spin_lock(ptl);
3538                         ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3539                         if (likely(ptep &&
3540                                    pte_same(huge_ptep_get(ptep), pte)))
3541                                 goto retry_avoidcopy;
3542                         /*
3543                          * race occurs while re-acquiring page table
3544                          * lock, and our job is done.
3545                          */
3546                         return 0;
3547                 }
3548 
3549                 ret = (PTR_ERR(new_page) == -ENOMEM) ?
3550                         VM_FAULT_OOM : VM_FAULT_SIGBUS;
3551                 goto out_release_old;
3552         }
3553 
3554         /*
3555          * When the original hugepage is shared one, it does not have
3556          * anon_vma prepared.
3557          */
3558         if (unlikely(anon_vma_prepare(vma))) {
3559                 ret = VM_FAULT_OOM;
3560                 goto out_release_all;
3561         }
3562 
3563         copy_user_huge_page(new_page, old_page, address, vma,
3564                             pages_per_huge_page(h));
3565         __SetPageUptodate(new_page);
3566         set_page_huge_active(new_page);
3567 
3568         mmun_start = address & huge_page_mask(h);
3569         mmun_end = mmun_start + huge_page_size(h);
3570         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3571 
3572         /*
3573          * Retake the page table lock to check for racing updates
3574          * before the page tables are altered
3575          */
3576         spin_lock(ptl);
3577         ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3578         if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3579                 ClearPagePrivate(new_page);
3580 
3581                 /* Break COW */
3582                 huge_ptep_clear_flush(vma, address, ptep);
3583                 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
3584                 set_huge_pte_at(mm, address, ptep,
3585                                 make_huge_pte(vma, new_page, 1));
3586                 page_remove_rmap(old_page, true);
3587                 hugepage_add_new_anon_rmap(new_page, vma, address);
3588                 /* Make the old page be freed below */
3589                 new_page = old_page;
3590         }
3591         spin_unlock(ptl);
3592         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3593 out_release_all:
3594         restore_reserve_on_error(h, vma, address, new_page);
3595         put_page(new_page);
3596 out_release_old:
3597         put_page(old_page);
3598 
3599         spin_lock(ptl); /* Caller expects lock to be held */
3600         return ret;
3601 }
3602 
3603 /* Return the pagecache page at a given address within a VMA */
3604 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3605                         struct vm_area_struct *vma, unsigned long address)
3606 {
3607         struct address_space *mapping;
3608         pgoff_t idx;
3609 
3610         mapping = vma->vm_file->f_mapping;
3611         idx = vma_hugecache_offset(h, vma, address);
3612 
3613         return find_lock_page(mapping, idx);
3614 }
3615 
3616 /*
3617  * Return whether there is a pagecache page to back given address within VMA.
3618  * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3619  */
3620 static bool hugetlbfs_pagecache_present(struct hstate *h,
3621                         struct vm_area_struct *vma, unsigned long address)
3622 {
3623         struct address_space *mapping;
3624         pgoff_t idx;
3625         struct page *page;
3626 
3627         mapping = vma->vm_file->f_mapping;
3628         idx = vma_hugecache_offset(h, vma, address);
3629 
3630         page = find_get_page(mapping, idx);
3631         if (page)
3632                 put_page(page);
3633         return page != NULL;
3634 }
3635 
3636 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3637                            pgoff_t idx)
3638 {
3639         struct inode *inode = mapping->host;
3640         struct hstate *h = hstate_inode(inode);
3641         int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3642 
3643         if (err)
3644                 return err;
3645         ClearPagePrivate(page);
3646 
3647         spin_lock(&inode->i_lock);
3648         inode->i_blocks += blocks_per_huge_page(h);
3649         spin_unlock(&inode->i_lock);
3650         return 0;
3651 }
3652 
3653 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
3654                            struct address_space *mapping, pgoff_t idx,
3655                            unsigned long address, pte_t *ptep, unsigned int flags)
3656 {
3657         struct hstate *h = hstate_vma(vma);
3658         int ret = VM_FAULT_SIGBUS;
3659         int anon_rmap = 0;
3660         unsigned long size;
3661         struct page *page;
3662         pte_t new_pte;
3663         spinlock_t *ptl;
3664 
3665         /*
3666          * Currently, we are forced to kill the process in the event the
3667          * original mapper has unmapped pages from the child due to a failed
3668          * COW. Warn that such a situation has occurred as it may not be obvious
3669          */
3670         if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3671                 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
3672                            current->pid);
3673                 return ret;
3674         }
3675 
3676         /*
3677          * Use page lock to guard against racing truncation
3678          * before we get page_table_lock.
3679          */
3680 retry:
3681         page = find_lock_page(mapping, idx);
3682         if (!page) {
3683                 size = i_size_read(mapping->host) >> huge_page_shift(h);
3684                 if (idx >= size)
3685                         goto out;
3686 
3687                 /*
3688                  * Check for page in userfault range
3689                  */
3690                 if (userfaultfd_missing(vma)) {
3691                         u32 hash;
3692                         struct vm_fault vmf = {
3693                                 .vma = vma,
3694                                 .address = address,
3695                                 .flags = flags,
3696                                 /*
3697                                  * Hard to debug if it ends up being
3698                                  * used by a callee that assumes
3699                                  * something about the other
3700                                  * uninitialized fields... same as in
3701                                  * memory.c
3702                                  */
3703                         };
3704 
3705                         /*
3706                          * hugetlb_fault_mutex must be dropped before
3707                          * handling userfault.  Reacquire after handling
3708                          * fault to make calling code simpler.
3709                          */
3710                         hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping,
3711                                                         idx, address);
3712                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3713                         ret = handle_userfault(&vmf, VM_UFFD_MISSING);
3714                         mutex_lock(&hugetlb_fault_mutex_table[hash]);
3715                         goto out;
3716                 }
3717 
3718                 page = alloc_huge_page(vma, address, 0);
3719                 if (IS_ERR(page)) {
3720                         ret = PTR_ERR(page);
3721                         if (ret == -ENOMEM)
3722                                 ret = VM_FAULT_OOM;
3723                         else
3724                                 ret = VM_FAULT_SIGBUS;
3725                         goto out;
3726                 }
3727                 clear_huge_page(page, address, pages_per_huge_page(h));
3728                 __SetPageUptodate(page);
3729                 set_page_huge_active(page);
3730 
3731                 if (vma->vm_flags & VM_MAYSHARE) {
3732                         int err = huge_add_to_page_cache(page, mapping, idx);
3733                         if (err) {
3734                                 put_page(page);
3735                                 if (err == -EEXIST)
3736                                         goto retry;
3737                                 goto out;
3738                         }
3739                 } else {
3740                         lock_page(page);
3741                         if (unlikely(anon_vma_prepare(vma))) {
3742                                 ret = VM_FAULT_OOM;
3743                                 goto backout_unlocked;
3744                         }
3745                         anon_rmap = 1;
3746                 }
3747         } else {
3748                 /*
3749                  * If memory error occurs between mmap() and fault, some process
3750                  * don't have hwpoisoned swap entry for errored virtual address.
3751                  * So we need to block hugepage fault by PG_hwpoison bit check.
3752                  */
3753                 if (unlikely(PageHWPoison(page))) {
3754                         ret = VM_FAULT_HWPOISON |
3755                                 VM_FAULT_SET_HINDEX(hstate_index(h));
3756                         goto backout_unlocked;
3757                 }
3758         }
3759 
3760         /*
3761          * If we are going to COW a private mapping later, we examine the
3762          * pending reservations for this page now. This will ensure that
3763          * any allocations necessary to record that reservation occur outside
3764          * the spinlock.
3765          */
3766         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3767                 if (vma_needs_reservation(h, vma, address) < 0) {
3768                         ret = VM_FAULT_OOM;
3769                         goto backout_unlocked;
3770                 }
3771                 /* Just decrements count, does not deallocate */
3772                 vma_end_reservation(h, vma, address);
3773         }
3774 
3775         ptl = huge_pte_lock(h, mm, ptep);
3776         size = i_size_read(mapping->host) >> huge_page_shift(h);
3777         if (idx >= size)
3778                 goto backout;
3779 
3780         ret = 0;
3781         if (!huge_pte_none(huge_ptep_get(ptep)))
3782                 goto backout;
3783 
3784         if (anon_rmap) {
3785                 ClearPagePrivate(page);
3786                 hugepage_add_new_anon_rmap(page, vma, address);
3787         } else
3788                 page_dup_rmap(page, true);
3789         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3790                                 && (vma->vm_flags & VM_SHARED)));
3791         set_huge_pte_at(mm, address, ptep, new_pte);
3792 
3793         hugetlb_count_add(pages_per_huge_page(h), mm);
3794         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3795                 /* Optimization, do the COW without a second fault */
3796                 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
3797         }
3798 
3799         spin_unlock(ptl);
3800         unlock_page(page);
3801 out:
3802         return ret;
3803 
3804 backout:
3805         spin_unlock(ptl);
3806 backout_unlocked:
3807         unlock_page(page);
3808         restore_reserve_on_error(h, vma, address, page);
3809         put_page(page);
3810         goto out;
3811 }
3812 
3813 #ifdef CONFIG_SMP
3814 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3815                             struct vm_area_struct *vma,
3816                             struct address_space *mapping,
3817                             pgoff_t idx, unsigned long address)
3818 {
3819         unsigned long key[2];
3820         u32 hash;
3821 
3822         if (vma->vm_flags & VM_SHARED) {
3823                 key[0] = (unsigned long) mapping;
3824                 key[1] = idx;
3825         } else {
3826                 key[0] = (unsigned long) mm;
3827                 key[1] = address >> huge_page_shift(h);
3828         }
3829 
3830         hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3831 
3832         return hash & (num_fault_mutexes - 1);
3833 }
3834 #else
3835 /*
3836  * For uniprocesor systems we always use a single mutex, so just
3837  * return 0 and avoid the hashing overhead.
3838  */
3839 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3840                             struct vm_area_struct *vma,
3841                             struct address_space *mapping,
3842                             pgoff_t idx, unsigned long address)
3843 {
3844         return 0;
3845 }
3846 #endif
3847 
3848 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3849                         unsigned long address, unsigned int flags)
3850 {
3851         pte_t *ptep, entry;
3852         spinlock_t *ptl;
3853         int ret;
3854         u32 hash;
3855         pgoff_t idx;
3856         struct page *page = NULL;
3857         struct page *pagecache_page = NULL;
3858         struct hstate *h = hstate_vma(vma);
3859         struct address_space *mapping;
3860         int need_wait_lock = 0;
3861 
3862         address &= huge_page_mask(h);
3863 
3864         ptep = huge_pte_offset(mm, address);
3865         if (ptep) {
3866                 entry = huge_ptep_get(ptep);
3867                 if (unlikely(is_hugetlb_entry_migration(entry))) {
3868                         migration_entry_wait_huge(vma, mm, ptep);
3869                         return 0;
3870                 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
3871                         return VM_FAULT_HWPOISON_LARGE |
3872                                 VM_FAULT_SET_HINDEX(hstate_index(h));
3873         } else {
3874                 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
3875                 if (!ptep)
3876                         return VM_FAULT_OOM;
3877         }
3878 
3879         mapping = vma->vm_file->f_mapping;
3880         idx = vma_hugecache_offset(h, vma, address);
3881 
3882         /*
3883          * Serialize hugepage allocation and instantiation, so that we don't
3884          * get spurious allocation failures if two CPUs race to instantiate
3885          * the same page in the page cache.
3886          */
3887         hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address);
3888         mutex_lock(&hugetlb_fault_mutex_table[hash]);
3889 
3890         entry = huge_ptep_get(ptep);
3891         if (huge_pte_none(entry)) {
3892                 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
3893                 goto out_mutex;
3894         }
3895 
3896         ret = 0;
3897 
3898         /*
3899          * entry could be a migration/hwpoison entry at this point, so this
3900          * check prevents the kernel from going below assuming that we have
3901          * a active hugepage in pagecache. This goto expects the 2nd page fault,
3902          * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3903          * handle it.
3904          */
3905         if (!pte_present(entry))
3906                 goto out_mutex;
3907 
3908         /*
3909          * If we are going to COW the mapping later, we examine the pending
3910          * reservations for this page now. This will ensure that any
3911          * allocations necessary to record that reservation occur outside the
3912          * spinlock. For private mappings, we also lookup the pagecache
3913          * page now as it is used to determine if a reservation has been
3914          * consumed.
3915          */
3916         if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
3917                 if (vma_needs_reservation(h, vma, address) < 0) {
3918                         ret = VM_FAULT_OOM;
3919                         goto out_mutex;
3920                 }
3921                 /* Just decrements count, does not deallocate */
3922                 vma_end_reservation(h, vma, address);
3923 
3924                 if (!(vma->vm_flags & VM_MAYSHARE))
3925                         pagecache_page = hugetlbfs_pagecache_page(h,
3926                                                                 vma, address);
3927         }
3928 
3929         ptl = huge_pte_lock(h, mm, ptep);
3930 
3931         /* Check for a racing update before calling hugetlb_cow */
3932         if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
3933                 goto out_ptl;
3934 
3935         /*
3936          * hugetlb_cow() requires page locks of pte_page(entry) and
3937          * pagecache_page, so here we need take the former one
3938          * when page != pagecache_page or !pagecache_page.
3939          */
3940         page = pte_page(entry);
3941         if (page != pagecache_page)
3942                 if (!trylock_page(page)) {
3943                         need_wait_lock = 1;
3944                         goto out_ptl;
3945                 }
3946 
3947         get_page(page);
3948 
3949         if (flags & FAULT_FLAG_WRITE) {
3950                 if (!huge_pte_write(entry)) {
3951                         ret = hugetlb_cow(mm, vma, address, ptep,
3952                                           pagecache_page, ptl);
3953                         goto out_put_page;
3954                 }
3955                 entry = huge_pte_mkdirty(entry);
3956         }
3957         entry = pte_mkyoung(entry);
3958         if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3959                                                 flags & FAULT_FLAG_WRITE))
3960                 update_mmu_cache(vma, address, ptep);
3961 out_put_page:
3962         if (page != pagecache_page)
3963                 unlock_page(page);
3964         put_page(page);
3965 out_ptl:
3966         spin_unlock(ptl);
3967 
3968         if (pagecache_page) {
3969                 unlock_page(pagecache_page);
3970                 put_page(pagecache_page);
3971         }
3972 out_mutex:
3973         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3974         /*
3975          * Generally it's safe to hold refcount during waiting page lock. But
3976          * here we just wait to defer the next page fault to avoid busy loop and
3977          * the page is not used after unlocked before returning from the current
3978          * page fault. So we are safe from accessing freed page, even if we wait
3979          * here without taking refcount.
3980          */
3981         if (need_wait_lock)
3982                 wait_on_page_locked(page);
3983         return ret;
3984 }
3985 
3986 /*
3987  * Used by userfaultfd UFFDIO_COPY.  Based on mcopy_atomic_pte with
3988  * modifications for huge pages.
3989  */
3990 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
3991                             pte_t *dst_pte,
3992                             struct vm_area_struct *dst_vma,
3993                             unsigned long dst_addr,
3994                             unsigned long src_addr,
3995                             struct page **pagep)
3996 {
3997         int vm_shared = dst_vma->vm_flags & VM_SHARED;
3998         struct hstate *h = hstate_vma(dst_vma);
3999         pte_t _dst_pte;
4000         spinlock_t *ptl;
4001         int ret;
4002         struct page *page;
4003 
4004         if (!*pagep) {
4005                 ret = -ENOMEM;
4006                 page = alloc_huge_page(dst_vma, dst_addr, 0);
4007                 if (IS_ERR(page))
4008                         goto out;
4009 
4010                 ret = copy_huge_page_from_user(page,
4011                                                 (const void __user *) src_addr,
4012                                                 pages_per_huge_page(h), false);
4013 
4014                 /* fallback to copy_from_user outside mmap_sem */
4015                 if (unlikely(ret)) {
4016                         ret = -EFAULT;
4017                         *pagep = page;
4018                         /* don't free the page */
4019                         goto out;
4020                 }
4021         } else {
4022                 page = *pagep;
4023                 *pagep = NULL;
4024         }
4025 
4026         /*
4027          * The memory barrier inside __SetPageUptodate makes sure that
4028          * preceding stores to the page contents become visible before
4029          * the set_pte_at() write.
4030          */
4031         __SetPageUptodate(page);
4032         set_page_huge_active(page);
4033 
4034         /*
4035          * If shared, add to page cache
4036          */
4037         if (vm_shared) {
4038                 struct address_space *mapping = dst_vma->vm_file->f_mapping;
4039                 pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4040 
4041                 ret = huge_add_to_page_cache(page, mapping, idx);
4042                 if (ret)
4043                         goto out_release_nounlock;
4044         }
4045 
4046         ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4047         spin_lock(ptl);
4048 
4049         ret = -EEXIST;
4050         if (!huge_pte_none(huge_ptep_get(dst_pte)))
4051                 goto out_release_unlock;
4052 
4053         if (vm_shared) {
4054                 page_dup_rmap(page, true);
4055         } else {
4056                 ClearPagePrivate(page);
4057                 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4058         }
4059 
4060         _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4061         if (dst_vma->vm_flags & VM_WRITE)
4062                 _dst_pte = huge_pte_mkdirty(_dst_pte);
4063         _dst_pte = pte_mkyoung(_dst_pte);
4064 
4065         set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4066 
4067         (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4068                                         dst_vma->vm_flags & VM_WRITE);
4069         hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4070 
4071         /* No need to invalidate - it was non-present before */
4072         update_mmu_cache(dst_vma, dst_addr, dst_pte);
4073 
4074         spin_unlock(ptl);
4075         if (vm_shared)
4076                 unlock_page(page);
4077         ret = 0;
4078 out:
4079         return ret;
4080 out_release_unlock:
4081         spin_unlock(ptl);
4082 out_release_nounlock:
4083         if (vm_shared)
4084                 unlock_page(page);
4085         put_page(page);
4086         goto out;
4087 }
4088 
4089 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4090                          struct page **pages, struct vm_area_struct **vmas,
4091                          unsigned long *position, unsigned long *nr_pages,
4092                          long i, unsigned int flags, int *nonblocking)
4093 {
4094         unsigned long pfn_offset;
4095         unsigned long vaddr = *position;
4096         unsigned long remainder = *nr_pages;
4097         struct hstate *h = hstate_vma(vma);
4098 
4099         while (vaddr < vma->vm_end && remainder) {
4100                 pte_t *pte;
4101                 spinlock_t *ptl = NULL;
4102                 int absent;
4103                 struct page *page;
4104 
4105                 /*
4106                  * If we have a pending SIGKILL, don't keep faulting pages and
4107                  * potentially allocating memory.
4108                  */
4109                 if (unlikely(fatal_signal_pending(current))) {
4110                         remainder = 0;
4111                         break;
4112                 }
4113 
4114                 /*
4115                  * Some archs (sparc64, sh*) have multiple pte_ts to
4116                  * each hugepage.  We have to make sure we get the
4117                  * first, for the page indexing below to work.
4118                  *
4119                  * Note that page table lock is not held when pte is null.
4120                  */
4121                 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
4122                 if (pte)
4123                         ptl = huge_pte_lock(h, mm, pte);
4124                 absent = !pte || huge_pte_none(huge_ptep_get(pte));
4125 
4126                 /*
4127                  * When coredumping, it suits get_dump_page if we just return
4128                  * an error where there's an empty slot with no huge pagecache
4129                  * to back it.  This way, we avoid allocating a hugepage, and
4130                  * the sparse dumpfile avoids allocating disk blocks, but its
4131                  * huge holes still show up with zeroes where they need to be.
4132                  */
4133                 if (absent && (flags & FOLL_DUMP) &&
4134                     !hugetlbfs_pagecache_present(h, vma, vaddr)) {
4135                         if (pte)
4136                                 spin_unlock(ptl);
4137                         remainder = 0;
4138                         break;
4139                 }
4140 
4141                 /*
4142                  * We need call hugetlb_fault for both hugepages under migration
4143                  * (in which case hugetlb_fault waits for the migration,) and
4144                  * hwpoisoned hugepages (in which case we need to prevent the
4145                  * caller from accessing to them.) In order to do this, we use
4146                  * here is_swap_pte instead of is_hugetlb_entry_migration and
4147                  * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4148                  * both cases, and because we can't follow correct pages
4149                  * directly from any kind of swap entries.
4150                  */
4151                 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4152                     ((flags & FOLL_WRITE) &&
4153                       !huge_pte_write(huge_ptep_get(pte)))) {
4154                         int ret;
4155                         unsigned int fault_flags = 0;
4156 
4157                         if (pte)
4158                                 spin_unlock(ptl);
4159                         if (flags & FOLL_WRITE)
4160                                 fault_flags |= FAULT_FLAG_WRITE;
4161                         if (nonblocking)
4162                                 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
4163                         if (flags & FOLL_NOWAIT)
4164                                 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4165                                         FAULT_FLAG_RETRY_NOWAIT;
4166                         if (flags & FOLL_TRIED) {
4167                                 VM_WARN_ON_ONCE(fault_flags &
4168                                                 FAULT_FLAG_ALLOW_RETRY);
4169                                 fault_flags |= FAULT_FLAG_TRIED;
4170                         }
4171                         ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4172                         if (ret & VM_FAULT_ERROR) {
4173                                 int err = vm_fault_to_errno(ret, flags);
4174 
4175                                 if (err)
4176                                         return err;
4177 
4178                                 remainder = 0;
4179                                 break;
4180                         }
4181                         if (ret & VM_FAULT_RETRY) {
4182                                 if (nonblocking)
4183                                         *nonblocking = 0;
4184                                 *nr_pages = 0;
4185                                 /*
4186                                  * VM_FAULT_RETRY must not return an
4187                                  * error, it will return zero
4188                                  * instead.
4189                                  *
4190                                  * No need to update "position" as the
4191                                  * caller will not check it after
4192                                  * *nr_pages is set to 0.
4193                                  */
4194                                 return i;
4195                         }
4196                         continue;
4197                 }
4198 
4199                 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4200                 page = pte_page(huge_ptep_get(pte));
4201 same_page:
4202                 if (pages) {
4203                         pages[i] = mem_map_offset(page, pfn_offset);
4204                         get_page(pages[i]);
4205                 }
4206 
4207                 if (vmas)
4208                         vmas[i] = vma;
4209 
4210                 vaddr += PAGE_SIZE;
4211                 ++pfn_offset;
4212                 --remainder;
4213                 ++i;
4214                 if (vaddr < vma->vm_end && remainder &&
4215                                 pfn_offset < pages_per_huge_page(h)) {
4216                         /*
4217                          * We use pfn_offset to avoid touching the pageframes
4218                          * of this compound page.
4219                          */
4220                         goto same_page;
4221                 }
4222                 spin_unlock(ptl);
4223         }
4224         *nr_pages = remainder;
4225         /*
4226          * setting position is actually required only if remainder is
4227          * not zero but it's faster not to add a "if (remainder)"
4228          * branch.
4229          */
4230         *position = vaddr;
4231 
4232         return i ? i : -EFAULT;
4233 }
4234 
4235 #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4236 /*
4237  * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4238  * implement this.
4239  */
4240 #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4241 #endif
4242 
4243 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
4244                 unsigned long address, unsigned long end, pgprot_t newprot)
4245 {
4246         struct mm_struct *mm = vma->vm_mm;
4247         unsigned long start = address;
4248         pte_t *ptep;
4249         pte_t pte;
4250         struct hstate *h = hstate_vma(vma);
4251         unsigned long pages = 0;
4252 
4253         BUG_ON(address >= end);
4254         flush_cache_range(vma, address, end);
4255 
4256         mmu_notifier_invalidate_range_start(mm, start, end);
4257         i_mmap_lock_write(vma->vm_file->f_mapping);
4258         for (; address < end; address += huge_page_size(h)) {
4259                 spinlock_t *ptl;
4260                 ptep = huge_pte_offset(mm, address);
4261                 if (!ptep)
4262                         continue;
4263                 ptl = huge_pte_lock(h, mm, ptep);
4264                 if (huge_pmd_unshare(mm, &address, ptep)) {
4265                         pages++;
4266                         spin_unlock(ptl);
4267                         continue;
4268                 }
4269                 pte = huge_ptep_get(ptep);
4270                 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
4271                         spin_unlock(ptl);
4272                         continue;
4273                 }
4274                 if (unlikely(is_hugetlb_entry_migration(pte))) {
4275                         swp_entry_t entry = pte_to_swp_entry(pte);
4276 
4277                         if (is_write_migration_entry(entry)) {
4278                                 pte_t newpte;
4279 
4280                                 make_migration_entry_read(&entry);
4281                                 newpte = swp_entry_to_pte(entry);
4282                                 set_huge_pte_at(mm, address, ptep, newpte);
4283                                 pages++;
4284                         }
4285                         spin_unlock(ptl);
4286                         continue;
4287                 }
4288                 if (!huge_pte_none(pte)) {
4289                         pte = huge_ptep_get_and_clear(mm, address, ptep);
4290                         pte = pte_mkhuge(huge_pte_modify(pte, newprot));
4291                         pte = arch_make_huge_pte(pte, vma, NULL, 0);
4292                         set_huge_pte_at(mm, address, ptep, pte);
4293                         pages++;
4294                 }
4295                 spin_unlock(ptl);
4296         }
4297         /*
4298          * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
4299          * may have cleared our pud entry and done put_page on the page table:
4300          * once we release i_mmap_rwsem, another task can do the final put_page
4301          * and that page table be reused and filled with junk.
4302          */
4303         flush_hugetlb_tlb_range(vma, start, end);
4304         mmu_notifier_invalidate_range(mm, start, end);
4305         i_mmap_unlock_write(vma->vm_file->f_mapping);
4306         mmu_notifier_invalidate_range_end(mm, start, end);
4307 
4308         return pages << h->order;
4309 }
4310 
4311 int hugetlb_reserve_pages(struct inode *inode,
4312                                         long from, long to,
4313                                         struct vm_area_struct *vma,
4314                                         vm_flags_t vm_flags)
4315 {
4316         long ret, chg;
4317         struct hstate *h = hstate_inode(inode);
4318         struct hugepage_subpool *spool = subpool_inode(inode);
4319         struct resv_map *resv_map;
4320         long gbl_reserve;
4321 
4322         /*
4323          * Only apply hugepage reservation if asked. At fault time, an
4324          * attempt will be made for VM_NORESERVE to allocate a page
4325          * without using reserves
4326          */
4327         if (vm_flags & VM_NORESERVE)
4328                 return 0;
4329 
4330         /*
4331          * Shared mappings base their reservation on the number of pages that
4332          * are already allocated on behalf of the file. Private mappings need
4333          * to reserve the full area even if read-only as mprotect() may be
4334          * called to make the mapping read-write. Assume !vma is a shm mapping
4335          */
4336         if (!vma || vma->vm_flags & VM_MAYSHARE) {
4337                 resv_map = inode_resv_map(inode);
4338 
4339                 chg = region_chg(resv_map, from, to);
4340 
4341         } else {
4342                 resv_map = resv_map_alloc();
4343                 if (!resv_map)
4344                         return -ENOMEM;
4345 
4346                 chg = to - from;
4347 
4348                 set_vma_resv_map(vma, resv_map);
4349                 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4350         }
4351 
4352         if (chg < 0) {
4353                 ret = chg;
4354                 goto out_err;
4355         }
4356 
4357         /*
4358          * There must be enough pages in the subpool for the mapping. If
4359          * the subpool has a minimum size, there may be some global
4360          * reservations already in place (gbl_reserve).
4361          */
4362         gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4363         if (gbl_reserve < 0) {
4364                 ret = -ENOSPC;
4365                 goto out_err;
4366         }
4367 
4368         /*
4369          * Check enough hugepages are available for the reservation.
4370          * Hand the pages back to the subpool if there are not
4371          */
4372         ret = hugetlb_acct_memory(h, gbl_reserve);
4373         if (ret < 0) {
4374                 /* put back original number of pages, chg */
4375                 (void)hugepage_subpool_put_pages(spool, chg);
4376                 goto out_err;
4377         }
4378 
4379         /*
4380          * Account for the reservations made. Shared mappings record regions
4381          * that have reservations as they are shared by multiple VMAs.
4382          * When the last VMA disappears, the region map says how much
4383          * the reservation was and the page cache tells how much of
4384          * the reservation was consumed. Private mappings are per-VMA and
4385          * only the consumed reservations are tracked. When the VMA
4386          * disappears, the original reservation is the VMA size and the
4387          * consumed reservations are stored in the map. Hence, nothing
4388          * else has to be done for private mappings here
4389          */
4390         if (!vma || vma->vm_flags & VM_MAYSHARE) {
4391                 long add = region_add(resv_map, from, to);
4392 
4393                 if (unlikely(chg > add)) {
4394                         /*
4395                          * pages in this range were added to the reserve
4396                          * map between region_chg and region_add.  This
4397                          * indicates a race with alloc_huge_page.  Adjust
4398                          * the subpool and reserve counts modified above
4399                          * based on the difference.
4400                          */
4401                         long rsv_adjust;
4402 
4403                         rsv_adjust = hugepage_subpool_put_pages(spool,
4404                                                                 chg - add);
4405                         hugetlb_acct_memory(h, -rsv_adjust);
4406                 }
4407         }
4408         return 0;
4409 out_err:
4410         if (!vma || vma->vm_flags & VM_MAYSHARE)
4411                 /* Don't call region_abort if region_chg failed */
4412                 if (chg >= 0)
4413                         region_abort(resv_map, from, to);
4414         if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4415                 kref_put(&resv_map->refs, resv_map_release);
4416         return ret;
4417 }
4418 
4419 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4420                                                                 long freed)
4421 {
4422         struct hstate *h = hstate_inode(inode);
4423         struct resv_map *resv_map = inode_resv_map(inode);
4424         long chg = 0;
4425         struct hugepage_subpool *spool = subpool_inode(inode);
4426         long gbl_reserve;
4427 
4428         if (resv_map) {
4429                 chg = region_del(resv_map, start, end);
4430                 /*
4431                  * region_del() can fail in the rare case where a region
4432                  * must be split and another region descriptor can not be
4433                  * allocated.  If end == LONG_MAX, it will not fail.
4434                  */
4435                 if (chg < 0)
4436                         return chg;
4437         }
4438 
4439         spin_lock(&inode->i_lock);
4440         inode->i_blocks -= (blocks_per_huge_page(h) * freed);
4441         spin_unlock(&inode->i_lock);
4442 
4443         /*
4444          * If the subpool has a minimum size, the number of global
4445          * reservations to be released may be adjusted.
4446          */
4447         gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4448         hugetlb_acct_memory(h, -gbl_reserve);
4449 
4450         return 0;
4451 }
4452 
4453 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4454 static unsigned long page_table_shareable(struct vm_area_struct *svma,
4455                                 struct vm_area_struct *vma,
4456                                 unsigned long addr, pgoff_t idx)
4457 {
4458         unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4459                                 svma->vm_start;
4460         unsigned long sbase = saddr & PUD_MASK;
4461         unsigned long s_end = sbase + PUD_SIZE;
4462 
4463         /* Allow segments to share if only one is marked locked */
4464         unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4465         unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
4466 
4467         /*
4468          * match the virtual addresses, permission and the alignment of the
4469          * page table page.
4470          */
4471         if (pmd_index(addr) != pmd_index(saddr) ||
4472             vm_flags != svm_flags ||
4473             sbase < svma->vm_start || svma->vm_end < s_end)
4474                 return 0;
4475 
4476         return saddr;
4477 }
4478 
4479 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
4480 {
4481         unsigned long base = addr & PUD_MASK;
4482         unsigned long end = base + PUD_SIZE;
4483 
4484         /*
4485          * check on proper vm_flags and page table alignment
4486          */
4487         if (vma->vm_flags & VM_MAYSHARE &&
4488             vma->vm_start <= base && end <= vma->vm_end)
4489                 return true;
4490         return false;
4491 }
4492 
4493 /*
4494  * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4495  * and returns the corresponding pte. While this is not necessary for the
4496  * !shared pmd case because we can allocate the pmd later as well, it makes the
4497  * code much cleaner. pmd allocation is essential for the shared case because
4498  * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4499  * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4500  * bad pmd for sharing.
4501  */
4502 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4503 {
4504         struct vm_area_struct *vma = find_vma(mm, addr);
4505         struct address_space *mapping = vma->vm_file->f_mapping;
4506         pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4507                         vma->vm_pgoff;
4508         struct vm_area_struct *svma;
4509         unsigned long saddr;
4510         pte_t *spte = NULL;
4511         pte_t *pte;
4512         spinlock_t *ptl;
4513 
4514         if (!vma_shareable(vma, addr))
4515                 return (pte_t *)pmd_alloc(mm, pud, addr);
4516 
4517         i_mmap_lock_write(mapping);
4518         vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4519                 if (svma == vma)
4520                         continue;
4521 
4522                 saddr = page_table_shareable(svma, vma, addr, idx);
4523                 if (saddr) {
4524                         spte = huge_pte_offset(svma->vm_mm, saddr);
4525                         if (spte) {
4526                                 get_page(virt_to_page(spte));
4527                                 break;
4528                         }
4529                 }
4530         }
4531 
4532         if (!spte)
4533                 goto out;
4534 
4535         ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
4536         if (pud_none(*pud)) {
4537                 pud_populate(mm, pud,
4538                                 (pmd_t *)((unsigned long)spte & PAGE_MASK));
4539                 mm_inc_nr_pmds(mm);
4540         } else {
4541                 put_page(virt_to_page(spte));
4542         }
4543         spin_unlock(ptl);
4544 out:
4545         pte = (pte_t *)pmd_alloc(mm, pud, addr);
4546         i_mmap_unlock_write(mapping);
4547         return pte;
4548 }
4549 
4550 /*
4551  * unmap huge page backed by shared pte.
4552  *
4553  * Hugetlb pte page is ref counted at the time of mapping.  If pte is shared
4554  * indicated by page_count > 1, unmap is achieved by clearing pud and
4555  * decrementing the ref count. If count == 1, the pte page is not shared.
4556  *
4557  * called with page table lock held.
4558  *
4559  * returns: 1 successfully unmapped a shared pte page
4560  *          0 the underlying pte page is not shared, or it is the last user
4561  */
4562 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4563 {
4564         pgd_t *pgd = pgd_offset(mm, *addr);
4565         p4d_t *p4d = p4d_offset(pgd, *addr);
4566         pud_t *pud = pud_offset(p4d, *addr);
4567 
4568         BUG_ON(page_count(virt_to_page(ptep)) == 0);
4569         if (page_count(virt_to_page(ptep)) == 1)
4570                 return 0;
4571 
4572         pud_clear(pud);
4573         put_page(virt_to_page(ptep));
4574         mm_dec_nr_pmds(mm);
4575         *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4576         return 1;
4577 }
4578 #define want_pmd_share()        (1)
4579 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4580 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4581 {
4582         return NULL;
4583 }
4584 
4585 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4586 {
4587         return 0;
4588 }
4589 #define want_pmd_share()        (0)
4590 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4591 
4592 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4593 pte_t *huge_pte_alloc(struct mm_struct *mm,
4594                         unsigned long addr, unsigned long sz)
4595 {
4596         pgd_t *pgd;
4597         p4d_t *p4d;
4598         pud_t *pud;
4599         pte_t *pte = NULL;
4600 
4601         pgd = pgd_offset(mm, addr);
4602         p4d = p4d_offset(pgd, addr);
4603         pud = pud_alloc(mm, p4d, addr);
4604         if (pud) {
4605                 if (sz == PUD_SIZE) {
4606                         pte = (pte_t *)pud;
4607                 } else {
4608                         BUG_ON(sz != PMD_SIZE);
4609                         if (want_pmd_share() && pud_none(*pud))
4610                                 pte = huge_pmd_share(mm, addr, pud);
4611                         else
4612                                 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4613                 }
4614         }
4615         BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
4616 
4617         return pte;
4618 }
4619 
4620 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
4621 {
4622         pgd_t *pgd;
4623         p4d_t *p4d;
4624         pud_t *pud;
4625         pmd_t *pmd;
4626 
4627         pgd = pgd_offset(mm, addr);
4628         if (!pgd_present(*pgd))
4629                 return NULL;
4630         p4d = p4d_offset(pgd, addr);
4631         if (!p4d_present(*p4d))
4632                 return NULL;
4633         pud = pud_offset(p4d, addr);
4634         if (!pud_present(*pud))
4635                 return NULL;
4636         if (pud_huge(*pud))
4637                 return (pte_t *)pud;
4638         pmd = pmd_offset(pud, addr);
4639         return (pte_t *) pmd;
4640 }
4641 
4642 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4643 
4644 /*
4645  * These functions are overwritable if your architecture needs its own
4646  * behavior.
4647  */
4648 struct page * __weak
4649 follow_huge_addr(struct mm_struct *mm, unsigned long address,
4650                               int write)
4651 {
4652         return ERR_PTR(-EINVAL);
4653 }
4654 
4655 struct page * __weak
4656 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
4657                 pmd_t *pmd, int flags)
4658 {
4659         struct page *page = NULL;
4660         spinlock_t *ptl;
4661         pte_t pte;
4662 retry:
4663         ptl = pmd_lockptr(mm, pmd);
4664         spin_lock(ptl);
4665         /*
4666          * make sure that the address range covered by this pmd is not
4667          * unmapped from other threads.
4668          */
4669         if (!pmd_huge(*pmd))
4670                 goto out;
4671         pte = huge_ptep_get((pte_t *)pmd);
4672         if (pte_present(pte)) {
4673                 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
4674                 if (flags & FOLL_GET)
4675                         get_page(page);
4676         } else {
4677                 if (is_hugetlb_entry_migration(pte)) {
4678                         spin_unlock(ptl);
4679                         __migration_entry_wait(mm, (pte_t *)pmd, ptl);
4680                         goto retry;
4681                 }
4682                 /*
4683                  * hwpoisoned entry is treated as no_page_table in
4684                  * follow_page_mask().
4685                  */
4686         }
4687 out:
4688         spin_unlock(ptl);
4689         return page;
4690 }
4691 
4692 struct page * __weak
4693 follow_huge_pud(struct mm_struct *mm, unsigned long address,
4694                 pud_t *pud, int flags)
4695 {
4696         if (flags & FOLL_GET)
4697                 return NULL;
4698 
4699         return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
4700 }
4701 
4702 #ifdef CONFIG_MEMORY_FAILURE
4703 
4704 /*
4705  * This function is called from memory failure code.
4706  */
4707 int dequeue_hwpoisoned_huge_page(struct page *hpage)
4708 {
4709         struct hstate *h = page_hstate(hpage);
4710         int nid = page_to_nid(hpage);
4711         int ret = -EBUSY;
4712 
4713         spin_lock(&hugetlb_lock);
4714         /*
4715          * Just checking !page_huge_active is not enough, because that could be
4716          * an isolated/hwpoisoned hugepage (which have >0 refcount).
4717          */
4718         if (!page_huge_active(hpage) && !page_count(hpage)) {
4719                 /*
4720                  * Hwpoisoned hugepage isn't linked to activelist or freelist,
4721                  * but dangling hpage->lru can trigger list-debug warnings
4722                  * (this happens when we call unpoison_memory() on it),
4723                  * so let it point to itself with list_del_init().
4724                  */
4725                 list_del_init(&hpage->lru);
4726                 set_page_refcounted(hpage);
4727                 h->free_huge_pages--;
4728                 h->free_huge_pages_node[nid]--;
4729                 ret = 0;
4730         }
4731         spin_unlock(&hugetlb_lock);
4732         return ret;
4733 }
4734 #endif
4735 
4736 bool isolate_huge_page(struct page *page, struct list_head *list)
4737 {
4738         bool ret = true;
4739 
4740         VM_BUG_ON_PAGE(!PageHead(page), page);
4741         spin_lock(&hugetlb_lock);
4742         if (!page_huge_active(page) || !get_page_unless_zero(page)) {
4743                 ret = false;
4744                 goto unlock;
4745         }
4746         clear_page_huge_active(page);
4747         list_move_tail(&page->lru, list);
4748 unlock:
4749         spin_unlock(&hugetlb_lock);
4750         return ret;
4751 }
4752 
4753 void putback_active_hugepage(struct page *page)
4754 {
4755         VM_BUG_ON_PAGE(!PageHead(page), page);
4756         spin_lock(&hugetlb_lock);
4757         set_page_huge_active(page);
4758         list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
4759         spin_unlock(&hugetlb_lock);
4760         put_page(page);
4761 }
4762 

~ [ 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