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

TOMOYO Linux Cross Reference
Linux/mm/hugetlb.c

Version: ~ [ linux-5.15-rc5 ] ~ [ linux-5.14.11 ] ~ [ linux-5.13.19 ] ~ [ linux-5.12.19 ] ~ [ linux-5.11.22 ] ~ [ linux-5.10.72 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.152 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.210 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.250 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.286 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.288 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.18.140 ] ~ [ linux-3.16.85 ] ~ [ linux-3.14.79 ] ~ [ linux-3.12.74 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

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

~ [ 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