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

TOMOYO Linux Cross Reference
Linux/mm/hugetlb.c

Version: ~ [ linux-5.16-rc3 ] ~ [ linux-5.15.5 ] ~ [ linux-5.14.21 ] ~ [ linux-5.13.19 ] ~ [ linux-5.12.19 ] ~ [ linux-5.11.22 ] ~ [ linux-5.10.82 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.162 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.218 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.256 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.291 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.293 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.18.140 ] ~ [ linux-3.16.85 ] ~ [ linux-3.14.79 ] ~ [ linux-3.12.74 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 /*
  2  * Generic hugetlb support.
  3  * (C) Nadia Yvette Chambers, April 2004
  4  */
  5 #include <linux/list.h>
  6 #include <linux/init.h>
  7 #include <linux/mm.h>
  8 #include <linux/seq_file.h>
  9 #include <linux/sysctl.h>
 10 #include <linux/highmem.h>
 11 #include <linux/mmu_notifier.h>
 12 #include <linux/nodemask.h>
 13 #include <linux/pagemap.h>
 14 #include <linux/mempolicy.h>
 15 #include <linux/compiler.h>
 16 #include <linux/cpuset.h>
 17 #include <linux/mutex.h>
 18 #include <linux/bootmem.h>
 19 #include <linux/sysfs.h>
 20 #include <linux/slab.h>
 21 #include <linux/mmdebug.h>
 22 #include <linux/sched/signal.h>
 23 #include <linux/rmap.h>
 24 #include <linux/string_helpers.h>
 25 #include <linux/swap.h>
 26 #include <linux/swapops.h>
 27 #include <linux/jhash.h>
 28 
 29 #include <asm/page.h>
 30 #include <asm/pgtable.h>
 31 #include <asm/tlb.h>
 32 
 33 #include <linux/io.h>
 34 #include <linux/hugetlb.h>
 35 #include <linux/hugetlb_cgroup.h>
 36 #include <linux/node.h>
 37 #include <linux/userfaultfd_k.h>
 38 #include <linux/page_owner.h>
 39 #include "internal.h"
 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         if (vma->vm_ops && vma->vm_ops->pagesize)
641                 return vma->vm_ops->pagesize(vma);
642         return PAGE_SIZE;
643 }
644 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
645 
646 /*
647  * Return the page size being used by the MMU to back a VMA. In the majority
648  * of cases, the page size used by the kernel matches the MMU size. On
649  * architectures where it differs, an architecture-specific 'strong'
650  * version of this symbol is required.
651  */
652 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
653 {
654         return vma_kernel_pagesize(vma);
655 }
656 
657 /*
658  * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
659  * bits of the reservation map pointer, which are always clear due to
660  * alignment.
661  */
662 #define HPAGE_RESV_OWNER    (1UL << 0)
663 #define HPAGE_RESV_UNMAPPED (1UL << 1)
664 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
665 
666 /*
667  * These helpers are used to track how many pages are reserved for
668  * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
669  * is guaranteed to have their future faults succeed.
670  *
671  * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
672  * the reserve counters are updated with the hugetlb_lock held. It is safe
673  * to reset the VMA at fork() time as it is not in use yet and there is no
674  * chance of the global counters getting corrupted as a result of the values.
675  *
676  * The private mapping reservation is represented in a subtly different
677  * manner to a shared mapping.  A shared mapping has a region map associated
678  * with the underlying file, this region map represents the backing file
679  * pages which have ever had a reservation assigned which this persists even
680  * after the page is instantiated.  A private mapping has a region map
681  * associated with the original mmap which is attached to all VMAs which
682  * reference it, this region map represents those offsets which have consumed
683  * reservation ie. where pages have been instantiated.
684  */
685 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
686 {
687         return (unsigned long)vma->vm_private_data;
688 }
689 
690 static void set_vma_private_data(struct vm_area_struct *vma,
691                                                         unsigned long value)
692 {
693         vma->vm_private_data = (void *)value;
694 }
695 
696 struct resv_map *resv_map_alloc(void)
697 {
698         struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
699         struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
700 
701         if (!resv_map || !rg) {
702                 kfree(resv_map);
703                 kfree(rg);
704                 return NULL;
705         }
706 
707         kref_init(&resv_map->refs);
708         spin_lock_init(&resv_map->lock);
709         INIT_LIST_HEAD(&resv_map->regions);
710 
711         resv_map->adds_in_progress = 0;
712 
713         INIT_LIST_HEAD(&resv_map->region_cache);
714         list_add(&rg->link, &resv_map->region_cache);
715         resv_map->region_cache_count = 1;
716 
717         return resv_map;
718 }
719 
720 void resv_map_release(struct kref *ref)
721 {
722         struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
723         struct list_head *head = &resv_map->region_cache;
724         struct file_region *rg, *trg;
725 
726         /* Clear out any active regions before we release the map. */
727         region_del(resv_map, 0, LONG_MAX);
728 
729         /* ... and any entries left in the cache */
730         list_for_each_entry_safe(rg, trg, head, link) {
731                 list_del(&rg->link);
732                 kfree(rg);
733         }
734 
735         VM_BUG_ON(resv_map->adds_in_progress);
736 
737         kfree(resv_map);
738 }
739 
740 static inline struct resv_map *inode_resv_map(struct inode *inode)
741 {
742         return inode->i_mapping->private_data;
743 }
744 
745 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
746 {
747         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
748         if (vma->vm_flags & VM_MAYSHARE) {
749                 struct address_space *mapping = vma->vm_file->f_mapping;
750                 struct inode *inode = mapping->host;
751 
752                 return inode_resv_map(inode);
753 
754         } else {
755                 return (struct resv_map *)(get_vma_private_data(vma) &
756                                                         ~HPAGE_RESV_MASK);
757         }
758 }
759 
760 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
761 {
762         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
763         VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
764 
765         set_vma_private_data(vma, (get_vma_private_data(vma) &
766                                 HPAGE_RESV_MASK) | (unsigned long)map);
767 }
768 
769 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
770 {
771         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
772         VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
773 
774         set_vma_private_data(vma, get_vma_private_data(vma) | flags);
775 }
776 
777 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
778 {
779         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
780 
781         return (get_vma_private_data(vma) & flag) != 0;
782 }
783 
784 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
785 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
786 {
787         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
788         if (!(vma->vm_flags & VM_MAYSHARE))
789                 vma->vm_private_data = (void *)0;
790 }
791 
792 /* Returns true if the VMA has associated reserve pages */
793 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
794 {
795         if (vma->vm_flags & VM_NORESERVE) {
796                 /*
797                  * This address is already reserved by other process(chg == 0),
798                  * so, we should decrement reserved count. Without decrementing,
799                  * reserve count remains after releasing inode, because this
800                  * allocated page will go into page cache and is regarded as
801                  * coming from reserved pool in releasing step.  Currently, we
802                  * don't have any other solution to deal with this situation
803                  * properly, so add work-around here.
804                  */
805                 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
806                         return true;
807                 else
808                         return false;
809         }
810 
811         /* Shared mappings always use reserves */
812         if (vma->vm_flags & VM_MAYSHARE) {
813                 /*
814                  * We know VM_NORESERVE is not set.  Therefore, there SHOULD
815                  * be a region map for all pages.  The only situation where
816                  * there is no region map is if a hole was punched via
817                  * fallocate.  In this case, there really are no reverves to
818                  * use.  This situation is indicated if chg != 0.
819                  */
820                 if (chg)
821                         return false;
822                 else
823                         return true;
824         }
825 
826         /*
827          * Only the process that called mmap() has reserves for
828          * private mappings.
829          */
830         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
831                 /*
832                  * Like the shared case above, a hole punch or truncate
833                  * could have been performed on the private mapping.
834                  * Examine the value of chg to determine if reserves
835                  * actually exist or were previously consumed.
836                  * Very Subtle - The value of chg comes from a previous
837                  * call to vma_needs_reserves().  The reserve map for
838                  * private mappings has different (opposite) semantics
839                  * than that of shared mappings.  vma_needs_reserves()
840                  * has already taken this difference in semantics into
841                  * account.  Therefore, the meaning of chg is the same
842                  * as in the shared case above.  Code could easily be
843                  * combined, but keeping it separate draws attention to
844                  * subtle differences.
845                  */
846                 if (chg)
847                         return false;
848                 else
849                         return true;
850         }
851 
852         return false;
853 }
854 
855 static void enqueue_huge_page(struct hstate *h, struct page *page)
856 {
857         int nid = page_to_nid(page);
858         list_move(&page->lru, &h->hugepage_freelists[nid]);
859         h->free_huge_pages++;
860         h->free_huge_pages_node[nid]++;
861 }
862 
863 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
864 {
865         struct page *page;
866 
867         list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
868                 if (!PageHWPoison(page))
869                         break;
870         /*
871          * if 'non-isolated free hugepage' not found on the list,
872          * the allocation fails.
873          */
874         if (&h->hugepage_freelists[nid] == &page->lru)
875                 return NULL;
876         list_move(&page->lru, &h->hugepage_activelist);
877         set_page_refcounted(page);
878         h->free_huge_pages--;
879         h->free_huge_pages_node[nid]--;
880         return page;
881 }
882 
883 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
884                 nodemask_t *nmask)
885 {
886         unsigned int cpuset_mems_cookie;
887         struct zonelist *zonelist;
888         struct zone *zone;
889         struct zoneref *z;
890         int node = -1;
891 
892         zonelist = node_zonelist(nid, gfp_mask);
893 
894 retry_cpuset:
895         cpuset_mems_cookie = read_mems_allowed_begin();
896         for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
897                 struct page *page;
898 
899                 if (!cpuset_zone_allowed(zone, gfp_mask))
900                         continue;
901                 /*
902                  * no need to ask again on the same node. Pool is node rather than
903                  * zone aware
904                  */
905                 if (zone_to_nid(zone) == node)
906                         continue;
907                 node = zone_to_nid(zone);
908 
909                 page = dequeue_huge_page_node_exact(h, node);
910                 if (page)
911                         return page;
912         }
913         if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
914                 goto retry_cpuset;
915 
916         return NULL;
917 }
918 
919 /* Movability of hugepages depends on migration support. */
920 static inline gfp_t htlb_alloc_mask(struct hstate *h)
921 {
922         if (hugepage_migration_supported(h))
923                 return GFP_HIGHUSER_MOVABLE;
924         else
925                 return GFP_HIGHUSER;
926 }
927 
928 static struct page *dequeue_huge_page_vma(struct hstate *h,
929                                 struct vm_area_struct *vma,
930                                 unsigned long address, int avoid_reserve,
931                                 long chg)
932 {
933         struct page *page;
934         struct mempolicy *mpol;
935         gfp_t gfp_mask;
936         nodemask_t *nodemask;
937         int nid;
938 
939         /*
940          * A child process with MAP_PRIVATE mappings created by their parent
941          * have no page reserves. This check ensures that reservations are
942          * not "stolen". The child may still get SIGKILLed
943          */
944         if (!vma_has_reserves(vma, chg) &&
945                         h->free_huge_pages - h->resv_huge_pages == 0)
946                 goto err;
947 
948         /* If reserves cannot be used, ensure enough pages are in the pool */
949         if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
950                 goto err;
951 
952         gfp_mask = htlb_alloc_mask(h);
953         nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
954         page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
955         if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
956                 SetPagePrivate(page);
957                 h->resv_huge_pages--;
958         }
959 
960         mpol_cond_put(mpol);
961         return page;
962 
963 err:
964         return NULL;
965 }
966 
967 /*
968  * common helper functions for hstate_next_node_to_{alloc|free}.
969  * We may have allocated or freed a huge page based on a different
970  * nodes_allowed previously, so h->next_node_to_{alloc|free} might
971  * be outside of *nodes_allowed.  Ensure that we use an allowed
972  * node for alloc or free.
973  */
974 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
975 {
976         nid = next_node_in(nid, *nodes_allowed);
977         VM_BUG_ON(nid >= MAX_NUMNODES);
978 
979         return nid;
980 }
981 
982 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
983 {
984         if (!node_isset(nid, *nodes_allowed))
985                 nid = next_node_allowed(nid, nodes_allowed);
986         return nid;
987 }
988 
989 /*
990  * returns the previously saved node ["this node"] from which to
991  * allocate a persistent huge page for the pool and advance the
992  * next node from which to allocate, handling wrap at end of node
993  * mask.
994  */
995 static int hstate_next_node_to_alloc(struct hstate *h,
996                                         nodemask_t *nodes_allowed)
997 {
998         int nid;
999 
1000         VM_BUG_ON(!nodes_allowed);
1001 
1002         nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1003         h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1004 
1005         return nid;
1006 }
1007 
1008 /*
1009  * helper for free_pool_huge_page() - return the previously saved
1010  * node ["this node"] from which to free a huge page.  Advance the
1011  * next node id whether or not we find a free huge page to free so
1012  * that the next attempt to free addresses the next node.
1013  */
1014 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1015 {
1016         int nid;
1017 
1018         VM_BUG_ON(!nodes_allowed);
1019 
1020         nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1021         h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1022 
1023         return nid;
1024 }
1025 
1026 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask)           \
1027         for (nr_nodes = nodes_weight(*mask);                            \
1028                 nr_nodes > 0 &&                                         \
1029                 ((node = hstate_next_node_to_alloc(hs, mask)) || 1);    \
1030                 nr_nodes--)
1031 
1032 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask)            \
1033         for (nr_nodes = nodes_weight(*mask);                            \
1034                 nr_nodes > 0 &&                                         \
1035                 ((node = hstate_next_node_to_free(hs, mask)) || 1);     \
1036                 nr_nodes--)
1037 
1038 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1039 static void destroy_compound_gigantic_page(struct page *page,
1040                                         unsigned int order)
1041 {
1042         int i;
1043         int nr_pages = 1 << order;
1044         struct page *p = page + 1;
1045 
1046         atomic_set(compound_mapcount_ptr(page), 0);
1047         for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1048                 clear_compound_head(p);
1049                 set_page_refcounted(p);
1050         }
1051 
1052         set_compound_order(page, 0);
1053         __ClearPageHead(page);
1054 }
1055 
1056 static void free_gigantic_page(struct page *page, unsigned int order)
1057 {
1058         free_contig_range(page_to_pfn(page), 1 << order);
1059 }
1060 
1061 static int __alloc_gigantic_page(unsigned long start_pfn,
1062                                 unsigned long nr_pages, gfp_t gfp_mask)
1063 {
1064         unsigned long end_pfn = start_pfn + nr_pages;
1065         return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
1066                                   gfp_mask);
1067 }
1068 
1069 static bool pfn_range_valid_gigantic(struct zone *z,
1070                         unsigned long start_pfn, unsigned long nr_pages)
1071 {
1072         unsigned long i, end_pfn = start_pfn + nr_pages;
1073         struct page *page;
1074 
1075         for (i = start_pfn; i < end_pfn; i++) {
1076                 if (!pfn_valid(i))
1077                         return false;
1078 
1079                 page = pfn_to_page(i);
1080 
1081                 if (page_zone(page) != z)
1082                         return false;
1083 
1084                 if (PageReserved(page))
1085                         return false;
1086 
1087                 if (page_count(page) > 0)
1088                         return false;
1089 
1090                 if (PageHuge(page))
1091                         return false;
1092         }
1093 
1094         return true;
1095 }
1096 
1097 static bool zone_spans_last_pfn(const struct zone *zone,
1098                         unsigned long start_pfn, unsigned long nr_pages)
1099 {
1100         unsigned long last_pfn = start_pfn + nr_pages - 1;
1101         return zone_spans_pfn(zone, last_pfn);
1102 }
1103 
1104 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1105                 int nid, nodemask_t *nodemask)
1106 {
1107         unsigned int order = huge_page_order(h);
1108         unsigned long nr_pages = 1 << order;
1109         unsigned long ret, pfn, flags;
1110         struct zonelist *zonelist;
1111         struct zone *zone;
1112         struct zoneref *z;
1113 
1114         zonelist = node_zonelist(nid, gfp_mask);
1115         for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nodemask) {
1116                 spin_lock_irqsave(&zone->lock, flags);
1117 
1118                 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
1119                 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
1120                         if (pfn_range_valid_gigantic(zone, pfn, nr_pages)) {
1121                                 /*
1122                                  * We release the zone lock here because
1123                                  * alloc_contig_range() will also lock the zone
1124                                  * at some point. If there's an allocation
1125                                  * spinning on this lock, it may win the race
1126                                  * and cause alloc_contig_range() to fail...
1127                                  */
1128                                 spin_unlock_irqrestore(&zone->lock, flags);
1129                                 ret = __alloc_gigantic_page(pfn, nr_pages, gfp_mask);
1130                                 if (!ret)
1131                                         return pfn_to_page(pfn);
1132                                 spin_lock_irqsave(&zone->lock, flags);
1133                         }
1134                         pfn += nr_pages;
1135                 }
1136 
1137                 spin_unlock_irqrestore(&zone->lock, flags);
1138         }
1139 
1140         return NULL;
1141 }
1142 
1143 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1144 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1145 
1146 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1147 static inline bool gigantic_page_supported(void) { return false; }
1148 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1149                 int nid, nodemask_t *nodemask) { return NULL; }
1150 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1151 static inline void destroy_compound_gigantic_page(struct page *page,
1152                                                 unsigned int order) { }
1153 #endif
1154 
1155 static void update_and_free_page(struct hstate *h, struct page *page)
1156 {
1157         int i;
1158 
1159         if (hstate_is_gigantic(h) && !gigantic_page_supported())
1160                 return;
1161 
1162         h->nr_huge_pages--;
1163         h->nr_huge_pages_node[page_to_nid(page)]--;
1164         for (i = 0; i < pages_per_huge_page(h); i++) {
1165                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1166                                 1 << PG_referenced | 1 << PG_dirty |
1167                                 1 << PG_active | 1 << PG_private |
1168                                 1 << PG_writeback);
1169         }
1170         VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1171         set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1172         set_page_refcounted(page);
1173         if (hstate_is_gigantic(h)) {
1174                 destroy_compound_gigantic_page(page, huge_page_order(h));
1175                 free_gigantic_page(page, huge_page_order(h));
1176         } else {
1177                 __free_pages(page, huge_page_order(h));
1178         }
1179 }
1180 
1181 struct hstate *size_to_hstate(unsigned long size)
1182 {
1183         struct hstate *h;
1184 
1185         for_each_hstate(h) {
1186                 if (huge_page_size(h) == size)
1187                         return h;
1188         }
1189         return NULL;
1190 }
1191 
1192 /*
1193  * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1194  * to hstate->hugepage_activelist.)
1195  *
1196  * This function can be called for tail pages, but never returns true for them.
1197  */
1198 bool page_huge_active(struct page *page)
1199 {
1200         VM_BUG_ON_PAGE(!PageHuge(page), page);
1201         return PageHead(page) && PagePrivate(&page[1]);
1202 }
1203 
1204 /* never called for tail page */
1205 static void set_page_huge_active(struct page *page)
1206 {
1207         VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1208         SetPagePrivate(&page[1]);
1209 }
1210 
1211 static void clear_page_huge_active(struct page *page)
1212 {
1213         VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1214         ClearPagePrivate(&page[1]);
1215 }
1216 
1217 /*
1218  * Internal hugetlb specific page flag. Do not use outside of the hugetlb
1219  * code
1220  */
1221 static inline bool PageHugeTemporary(struct page *page)
1222 {
1223         if (!PageHuge(page))
1224                 return false;
1225 
1226         return (unsigned long)page[2].mapping == -1U;
1227 }
1228 
1229 static inline void SetPageHugeTemporary(struct page *page)
1230 {
1231         page[2].mapping = (void *)-1U;
1232 }
1233 
1234 static inline void ClearPageHugeTemporary(struct page *page)
1235 {
1236         page[2].mapping = NULL;
1237 }
1238 
1239 void free_huge_page(struct page *page)
1240 {
1241         /*
1242          * Can't pass hstate in here because it is called from the
1243          * compound page destructor.
1244          */
1245         struct hstate *h = page_hstate(page);
1246         int nid = page_to_nid(page);
1247         struct hugepage_subpool *spool =
1248                 (struct hugepage_subpool *)page_private(page);
1249         bool restore_reserve;
1250 
1251         set_page_private(page, 0);
1252         page->mapping = NULL;
1253         VM_BUG_ON_PAGE(page_count(page), page);
1254         VM_BUG_ON_PAGE(page_mapcount(page), page);
1255         restore_reserve = PagePrivate(page);
1256         ClearPagePrivate(page);
1257 
1258         /*
1259          * A return code of zero implies that the subpool will be under its
1260          * minimum size if the reservation is not restored after page is free.
1261          * Therefore, force restore_reserve operation.
1262          */
1263         if (hugepage_subpool_put_pages(spool, 1) == 0)
1264                 restore_reserve = true;
1265 
1266         spin_lock(&hugetlb_lock);
1267         clear_page_huge_active(page);
1268         hugetlb_cgroup_uncharge_page(hstate_index(h),
1269                                      pages_per_huge_page(h), page);
1270         if (restore_reserve)
1271                 h->resv_huge_pages++;
1272 
1273         if (PageHugeTemporary(page)) {
1274                 list_del(&page->lru);
1275                 ClearPageHugeTemporary(page);
1276                 update_and_free_page(h, page);
1277         } else if (h->surplus_huge_pages_node[nid]) {
1278                 /* remove the page from active list */
1279                 list_del(&page->lru);
1280                 update_and_free_page(h, page);
1281                 h->surplus_huge_pages--;
1282                 h->surplus_huge_pages_node[nid]--;
1283         } else {
1284                 arch_clear_hugepage_flags(page);
1285                 enqueue_huge_page(h, page);
1286         }
1287         spin_unlock(&hugetlb_lock);
1288 }
1289 
1290 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1291 {
1292         INIT_LIST_HEAD(&page->lru);
1293         set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1294         spin_lock(&hugetlb_lock);
1295         set_hugetlb_cgroup(page, NULL);
1296         h->nr_huge_pages++;
1297         h->nr_huge_pages_node[nid]++;
1298         spin_unlock(&hugetlb_lock);
1299 }
1300 
1301 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1302 {
1303         int i;
1304         int nr_pages = 1 << order;
1305         struct page *p = page + 1;
1306 
1307         /* we rely on prep_new_huge_page to set the destructor */
1308         set_compound_order(page, order);
1309         __ClearPageReserved(page);
1310         __SetPageHead(page);
1311         for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1312                 /*
1313                  * For gigantic hugepages allocated through bootmem at
1314                  * boot, it's safer to be consistent with the not-gigantic
1315                  * hugepages and clear the PG_reserved bit from all tail pages
1316                  * too.  Otherwse drivers using get_user_pages() to access tail
1317                  * pages may get the reference counting wrong if they see
1318                  * PG_reserved set on a tail page (despite the head page not
1319                  * having PG_reserved set).  Enforcing this consistency between
1320                  * head and tail pages allows drivers to optimize away a check
1321                  * on the head page when they need know if put_page() is needed
1322                  * after get_user_pages().
1323                  */
1324                 __ClearPageReserved(p);
1325                 set_page_count(p, 0);
1326                 set_compound_head(p, page);
1327         }
1328         atomic_set(compound_mapcount_ptr(page), -1);
1329 }
1330 
1331 /*
1332  * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1333  * transparent huge pages.  See the PageTransHuge() documentation for more
1334  * details.
1335  */
1336 int PageHuge(struct page *page)
1337 {
1338         if (!PageCompound(page))
1339                 return 0;
1340 
1341         page = compound_head(page);
1342         return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1343 }
1344 EXPORT_SYMBOL_GPL(PageHuge);
1345 
1346 /*
1347  * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1348  * normal or transparent huge pages.
1349  */
1350 int PageHeadHuge(struct page *page_head)
1351 {
1352         if (!PageHead(page_head))
1353                 return 0;
1354 
1355         return get_compound_page_dtor(page_head) == free_huge_page;
1356 }
1357 
1358 pgoff_t __basepage_index(struct page *page)
1359 {
1360         struct page *page_head = compound_head(page);
1361         pgoff_t index = page_index(page_head);
1362         unsigned long compound_idx;
1363 
1364         if (!PageHuge(page_head))
1365                 return page_index(page);
1366 
1367         if (compound_order(page_head) >= MAX_ORDER)
1368                 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1369         else
1370                 compound_idx = page - page_head;
1371 
1372         return (index << compound_order(page_head)) + compound_idx;
1373 }
1374 
1375 static struct page *alloc_buddy_huge_page(struct hstate *h,
1376                 gfp_t gfp_mask, int nid, nodemask_t *nmask)
1377 {
1378         int order = huge_page_order(h);
1379         struct page *page;
1380 
1381         gfp_mask |= __GFP_COMP|__GFP_RETRY_MAYFAIL|__GFP_NOWARN;
1382         if (nid == NUMA_NO_NODE)
1383                 nid = numa_mem_id();
1384         page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
1385         if (page)
1386                 __count_vm_event(HTLB_BUDDY_PGALLOC);
1387         else
1388                 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1389 
1390         return page;
1391 }
1392 
1393 /*
1394  * Common helper to allocate a fresh hugetlb page. All specific allocators
1395  * should use this function to get new hugetlb pages
1396  */
1397 static struct page *alloc_fresh_huge_page(struct hstate *h,
1398                 gfp_t gfp_mask, int nid, nodemask_t *nmask)
1399 {
1400         struct page *page;
1401 
1402         if (hstate_is_gigantic(h))
1403                 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1404         else
1405                 page = alloc_buddy_huge_page(h, gfp_mask,
1406                                 nid, nmask);
1407         if (!page)
1408                 return NULL;
1409 
1410         if (hstate_is_gigantic(h))
1411                 prep_compound_gigantic_page(page, huge_page_order(h));
1412         prep_new_huge_page(h, page, page_to_nid(page));
1413 
1414         return page;
1415 }
1416 
1417 /*
1418  * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1419  * manner.
1420  */
1421 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1422 {
1423         struct page *page;
1424         int nr_nodes, node;
1425         gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1426 
1427         for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1428                 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed);
1429                 if (page)
1430                         break;
1431         }
1432 
1433         if (!page)
1434                 return 0;
1435 
1436         put_page(page); /* free it into the hugepage allocator */
1437 
1438         return 1;
1439 }
1440 
1441 /*
1442  * Free huge page from pool from next node to free.
1443  * Attempt to keep persistent huge pages more or less
1444  * balanced over allowed nodes.
1445  * Called with hugetlb_lock locked.
1446  */
1447 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1448                                                          bool acct_surplus)
1449 {
1450         int nr_nodes, node;
1451         int ret = 0;
1452 
1453         for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1454                 /*
1455                  * If we're returning unused surplus pages, only examine
1456                  * nodes with surplus pages.
1457                  */
1458                 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1459                     !list_empty(&h->hugepage_freelists[node])) {
1460                         struct page *page =
1461                                 list_entry(h->hugepage_freelists[node].next,
1462                                           struct page, lru);
1463                         list_del(&page->lru);
1464                         h->free_huge_pages--;
1465                         h->free_huge_pages_node[node]--;
1466                         if (acct_surplus) {
1467                                 h->surplus_huge_pages--;
1468                                 h->surplus_huge_pages_node[node]--;
1469                         }
1470                         update_and_free_page(h, page);
1471                         ret = 1;
1472                         break;
1473                 }
1474         }
1475 
1476         return ret;
1477 }
1478 
1479 /*
1480  * Dissolve a given free hugepage into free buddy pages. This function does
1481  * nothing for in-use (including surplus) hugepages. Returns -EBUSY if the
1482  * number of free hugepages would be reduced below the number of reserved
1483  * hugepages.
1484  */
1485 int dissolve_free_huge_page(struct page *page)
1486 {
1487         int rc = 0;
1488 
1489         spin_lock(&hugetlb_lock);
1490         if (PageHuge(page) && !page_count(page)) {
1491                 struct page *head = compound_head(page);
1492                 struct hstate *h = page_hstate(head);
1493                 int nid = page_to_nid(head);
1494                 if (h->free_huge_pages - h->resv_huge_pages == 0) {
1495                         rc = -EBUSY;
1496                         goto out;
1497                 }
1498                 /*
1499                  * Move PageHWPoison flag from head page to the raw error page,
1500                  * which makes any subpages rather than the error page reusable.
1501                  */
1502                 if (PageHWPoison(head) && page != head) {
1503                         SetPageHWPoison(page);
1504                         ClearPageHWPoison(head);
1505                 }
1506                 list_del(&head->lru);
1507                 h->free_huge_pages--;
1508                 h->free_huge_pages_node[nid]--;
1509                 h->max_huge_pages--;
1510                 update_and_free_page(h, head);
1511         }
1512 out:
1513         spin_unlock(&hugetlb_lock);
1514         return rc;
1515 }
1516 
1517 /*
1518  * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1519  * make specified memory blocks removable from the system.
1520  * Note that this will dissolve a free gigantic hugepage completely, if any
1521  * part of it lies within the given range.
1522  * Also note that if dissolve_free_huge_page() returns with an error, all
1523  * free hugepages that were dissolved before that error are lost.
1524  */
1525 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1526 {
1527         unsigned long pfn;
1528         struct page *page;
1529         int rc = 0;
1530 
1531         if (!hugepages_supported())
1532                 return rc;
1533 
1534         for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1535                 page = pfn_to_page(pfn);
1536                 if (PageHuge(page) && !page_count(page)) {
1537                         rc = dissolve_free_huge_page(page);
1538                         if (rc)
1539                                 break;
1540                 }
1541         }
1542 
1543         return rc;
1544 }
1545 
1546 /*
1547  * Allocates a fresh surplus page from the page allocator.
1548  */
1549 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
1550                 int nid, nodemask_t *nmask)
1551 {
1552         struct page *page = NULL;
1553 
1554         if (hstate_is_gigantic(h))
1555                 return NULL;
1556 
1557         spin_lock(&hugetlb_lock);
1558         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1559                 goto out_unlock;
1560         spin_unlock(&hugetlb_lock);
1561 
1562         page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask);
1563         if (!page)
1564                 return NULL;
1565 
1566         spin_lock(&hugetlb_lock);
1567         /*
1568          * We could have raced with the pool size change.
1569          * Double check that and simply deallocate the new page
1570          * if we would end up overcommiting the surpluses. Abuse
1571          * temporary page to workaround the nasty free_huge_page
1572          * codeflow
1573          */
1574         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1575                 SetPageHugeTemporary(page);
1576                 put_page(page);
1577                 page = NULL;
1578         } else {
1579                 h->surplus_huge_pages++;
1580                 h->surplus_huge_pages_node[page_to_nid(page)]++;
1581         }
1582 
1583 out_unlock:
1584         spin_unlock(&hugetlb_lock);
1585 
1586         return page;
1587 }
1588 
1589 static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
1590                 int nid, nodemask_t *nmask)
1591 {
1592         struct page *page;
1593 
1594         if (hstate_is_gigantic(h))
1595                 return NULL;
1596 
1597         page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask);
1598         if (!page)
1599                 return NULL;
1600 
1601         /*
1602          * We do not account these pages as surplus because they are only
1603          * temporary and will be released properly on the last reference
1604          */
1605         SetPageHugeTemporary(page);
1606 
1607         return page;
1608 }
1609 
1610 /*
1611  * Use the VMA's mpolicy to allocate a huge page from the buddy.
1612  */
1613 static
1614 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
1615                 struct vm_area_struct *vma, unsigned long addr)
1616 {
1617         struct page *page;
1618         struct mempolicy *mpol;
1619         gfp_t gfp_mask = htlb_alloc_mask(h);
1620         int nid;
1621         nodemask_t *nodemask;
1622 
1623         nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1624         page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
1625         mpol_cond_put(mpol);
1626 
1627         return page;
1628 }
1629 
1630 /* page migration callback function */
1631 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1632 {
1633         gfp_t gfp_mask = htlb_alloc_mask(h);
1634         struct page *page = NULL;
1635 
1636         if (nid != NUMA_NO_NODE)
1637                 gfp_mask |= __GFP_THISNODE;
1638 
1639         spin_lock(&hugetlb_lock);
1640         if (h->free_huge_pages - h->resv_huge_pages > 0)
1641                 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL);
1642         spin_unlock(&hugetlb_lock);
1643 
1644         if (!page)
1645                 page = alloc_migrate_huge_page(h, gfp_mask, nid, NULL);
1646 
1647         return page;
1648 }
1649 
1650 /* page migration callback function */
1651 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
1652                 nodemask_t *nmask)
1653 {
1654         gfp_t gfp_mask = htlb_alloc_mask(h);
1655 
1656         spin_lock(&hugetlb_lock);
1657         if (h->free_huge_pages - h->resv_huge_pages > 0) {
1658                 struct page *page;
1659 
1660                 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1661                 if (page) {
1662                         spin_unlock(&hugetlb_lock);
1663                         return page;
1664                 }
1665         }
1666         spin_unlock(&hugetlb_lock);
1667 
1668         return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
1669 }
1670 
1671 /* mempolicy aware migration callback */
1672 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
1673                 unsigned long address)
1674 {
1675         struct mempolicy *mpol;
1676         nodemask_t *nodemask;
1677         struct page *page;
1678         gfp_t gfp_mask;
1679         int node;
1680 
1681         gfp_mask = htlb_alloc_mask(h);
1682         node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1683         page = alloc_huge_page_nodemask(h, node, nodemask);
1684         mpol_cond_put(mpol);
1685 
1686         return page;
1687 }
1688 
1689 /*
1690  * Increase the hugetlb pool such that it can accommodate a reservation
1691  * of size 'delta'.
1692  */
1693 static int gather_surplus_pages(struct hstate *h, int delta)
1694 {
1695         struct list_head surplus_list;
1696         struct page *page, *tmp;
1697         int ret, i;
1698         int needed, allocated;
1699         bool alloc_ok = true;
1700 
1701         needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1702         if (needed <= 0) {
1703                 h->resv_huge_pages += delta;
1704                 return 0;
1705         }
1706 
1707         allocated = 0;
1708         INIT_LIST_HEAD(&surplus_list);
1709 
1710         ret = -ENOMEM;
1711 retry:
1712         spin_unlock(&hugetlb_lock);
1713         for (i = 0; i < needed; i++) {
1714                 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
1715                                 NUMA_NO_NODE, NULL);
1716                 if (!page) {
1717                         alloc_ok = false;
1718                         break;
1719                 }
1720                 list_add(&page->lru, &surplus_list);
1721                 cond_resched();
1722         }
1723         allocated += i;
1724 
1725         /*
1726          * After retaking hugetlb_lock, we need to recalculate 'needed'
1727          * because either resv_huge_pages or free_huge_pages may have changed.
1728          */
1729         spin_lock(&hugetlb_lock);
1730         needed = (h->resv_huge_pages + delta) -
1731                         (h->free_huge_pages + allocated);
1732         if (needed > 0) {
1733                 if (alloc_ok)
1734                         goto retry;
1735                 /*
1736                  * We were not able to allocate enough pages to
1737                  * satisfy the entire reservation so we free what
1738                  * we've allocated so far.
1739                  */
1740                 goto free;
1741         }
1742         /*
1743          * The surplus_list now contains _at_least_ the number of extra pages
1744          * needed to accommodate the reservation.  Add the appropriate number
1745          * of pages to the hugetlb pool and free the extras back to the buddy
1746          * allocator.  Commit the entire reservation here to prevent another
1747          * process from stealing the pages as they are added to the pool but
1748          * before they are reserved.
1749          */
1750         needed += allocated;
1751         h->resv_huge_pages += delta;
1752         ret = 0;
1753 
1754         /* Free the needed pages to the hugetlb pool */
1755         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1756                 if ((--needed) < 0)
1757                         break;
1758                 /*
1759                  * This page is now managed by the hugetlb allocator and has
1760                  * no users -- drop the buddy allocator's reference.
1761                  */
1762                 put_page_testzero(page);
1763                 VM_BUG_ON_PAGE(page_count(page), page);
1764                 enqueue_huge_page(h, page);
1765         }
1766 free:
1767         spin_unlock(&hugetlb_lock);
1768 
1769         /* Free unnecessary surplus pages to the buddy allocator */
1770         list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1771                 put_page(page);
1772         spin_lock(&hugetlb_lock);
1773 
1774         return ret;
1775 }
1776 
1777 /*
1778  * This routine has two main purposes:
1779  * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1780  *    in unused_resv_pages.  This corresponds to the prior adjustments made
1781  *    to the associated reservation map.
1782  * 2) Free any unused surplus pages that may have been allocated to satisfy
1783  *    the reservation.  As many as unused_resv_pages may be freed.
1784  *
1785  * Called with hugetlb_lock held.  However, the lock could be dropped (and
1786  * reacquired) during calls to cond_resched_lock.  Whenever dropping the lock,
1787  * we must make sure nobody else can claim pages we are in the process of
1788  * freeing.  Do this by ensuring resv_huge_page always is greater than the
1789  * number of huge pages we plan to free when dropping the lock.
1790  */
1791 static void return_unused_surplus_pages(struct hstate *h,
1792                                         unsigned long unused_resv_pages)
1793 {
1794         unsigned long nr_pages;
1795 
1796         /* Cannot return gigantic pages currently */
1797         if (hstate_is_gigantic(h))
1798                 goto out;
1799 
1800         /*
1801          * Part (or even all) of the reservation could have been backed
1802          * by pre-allocated pages. Only free surplus pages.
1803          */
1804         nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1805 
1806         /*
1807          * We want to release as many surplus pages as possible, spread
1808          * evenly across all nodes with memory. Iterate across these nodes
1809          * until we can no longer free unreserved surplus pages. This occurs
1810          * when the nodes with surplus pages have no free pages.
1811          * free_pool_huge_page() will balance the the freed pages across the
1812          * on-line nodes with memory and will handle the hstate accounting.
1813          *
1814          * Note that we decrement resv_huge_pages as we free the pages.  If
1815          * we drop the lock, resv_huge_pages will still be sufficiently large
1816          * to cover subsequent pages we may free.
1817          */
1818         while (nr_pages--) {
1819                 h->resv_huge_pages--;
1820                 unused_resv_pages--;
1821                 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1822                         goto out;
1823                 cond_resched_lock(&hugetlb_lock);
1824         }
1825 
1826 out:
1827         /* Fully uncommit the reservation */
1828         h->resv_huge_pages -= unused_resv_pages;
1829 }
1830 
1831 
1832 /*
1833  * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1834  * are used by the huge page allocation routines to manage reservations.
1835  *
1836  * vma_needs_reservation is called to determine if the huge page at addr
1837  * within the vma has an associated reservation.  If a reservation is
1838  * needed, the value 1 is returned.  The caller is then responsible for
1839  * managing the global reservation and subpool usage counts.  After
1840  * the huge page has been allocated, vma_commit_reservation is called
1841  * to add the page to the reservation map.  If the page allocation fails,
1842  * the reservation must be ended instead of committed.  vma_end_reservation
1843  * is called in such cases.
1844  *
1845  * In the normal case, vma_commit_reservation returns the same value
1846  * as the preceding vma_needs_reservation call.  The only time this
1847  * is not the case is if a reserve map was changed between calls.  It
1848  * is the responsibility of the caller to notice the difference and
1849  * take appropriate action.
1850  *
1851  * vma_add_reservation is used in error paths where a reservation must
1852  * be restored when a newly allocated huge page must be freed.  It is
1853  * to be called after calling vma_needs_reservation to determine if a
1854  * reservation exists.
1855  */
1856 enum vma_resv_mode {
1857         VMA_NEEDS_RESV,
1858         VMA_COMMIT_RESV,
1859         VMA_END_RESV,
1860         VMA_ADD_RESV,
1861 };
1862 static long __vma_reservation_common(struct hstate *h,
1863                                 struct vm_area_struct *vma, unsigned long addr,
1864                                 enum vma_resv_mode mode)
1865 {
1866         struct resv_map *resv;
1867         pgoff_t idx;
1868         long ret;
1869 
1870         resv = vma_resv_map(vma);
1871         if (!resv)
1872                 return 1;
1873 
1874         idx = vma_hugecache_offset(h, vma, addr);
1875         switch (mode) {
1876         case VMA_NEEDS_RESV:
1877                 ret = region_chg(resv, idx, idx + 1);
1878                 break;
1879         case VMA_COMMIT_RESV:
1880                 ret = region_add(resv, idx, idx + 1);
1881                 break;
1882         case VMA_END_RESV:
1883                 region_abort(resv, idx, idx + 1);
1884                 ret = 0;
1885                 break;
1886         case VMA_ADD_RESV:
1887                 if (vma->vm_flags & VM_MAYSHARE)
1888                         ret = region_add(resv, idx, idx + 1);
1889                 else {
1890                         region_abort(resv, idx, idx + 1);
1891                         ret = region_del(resv, idx, idx + 1);
1892                 }
1893                 break;
1894         default:
1895                 BUG();
1896         }
1897 
1898         if (vma->vm_flags & VM_MAYSHARE)
1899                 return ret;
1900         else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
1901                 /*
1902                  * In most cases, reserves always exist for private mappings.
1903                  * However, a file associated with mapping could have been
1904                  * hole punched or truncated after reserves were consumed.
1905                  * As subsequent fault on such a range will not use reserves.
1906                  * Subtle - The reserve map for private mappings has the
1907                  * opposite meaning than that of shared mappings.  If NO
1908                  * entry is in the reserve map, it means a reservation exists.
1909                  * If an entry exists in the reserve map, it means the
1910                  * reservation has already been consumed.  As a result, the
1911                  * return value of this routine is the opposite of the
1912                  * value returned from reserve map manipulation routines above.
1913                  */
1914                 if (ret)
1915                         return 0;
1916                 else
1917                         return 1;
1918         }
1919         else
1920                 return ret < 0 ? ret : 0;
1921 }
1922 
1923 static long vma_needs_reservation(struct hstate *h,
1924                         struct vm_area_struct *vma, unsigned long addr)
1925 {
1926         return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
1927 }
1928 
1929 static long vma_commit_reservation(struct hstate *h,
1930                         struct vm_area_struct *vma, unsigned long addr)
1931 {
1932         return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1933 }
1934 
1935 static void vma_end_reservation(struct hstate *h,
1936                         struct vm_area_struct *vma, unsigned long addr)
1937 {
1938         (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
1939 }
1940 
1941 static long vma_add_reservation(struct hstate *h,
1942                         struct vm_area_struct *vma, unsigned long addr)
1943 {
1944         return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
1945 }
1946 
1947 /*
1948  * This routine is called to restore a reservation on error paths.  In the
1949  * specific error paths, a huge page was allocated (via alloc_huge_page)
1950  * and is about to be freed.  If a reservation for the page existed,
1951  * alloc_huge_page would have consumed the reservation and set PagePrivate
1952  * in the newly allocated page.  When the page is freed via free_huge_page,
1953  * the global reservation count will be incremented if PagePrivate is set.
1954  * However, free_huge_page can not adjust the reserve map.  Adjust the
1955  * reserve map here to be consistent with global reserve count adjustments
1956  * to be made by free_huge_page.
1957  */
1958 static void restore_reserve_on_error(struct hstate *h,
1959                         struct vm_area_struct *vma, unsigned long address,
1960                         struct page *page)
1961 {
1962         if (unlikely(PagePrivate(page))) {
1963                 long rc = vma_needs_reservation(h, vma, address);
1964 
1965                 if (unlikely(rc < 0)) {
1966                         /*
1967                          * Rare out of memory condition in reserve map
1968                          * manipulation.  Clear PagePrivate so that
1969                          * global reserve count will not be incremented
1970                          * by free_huge_page.  This will make it appear
1971                          * as though the reservation for this page was
1972                          * consumed.  This may prevent the task from
1973                          * faulting in the page at a later time.  This
1974                          * is better than inconsistent global huge page
1975                          * accounting of reserve counts.
1976                          */
1977                         ClearPagePrivate(page);
1978                 } else if (rc) {
1979                         rc = vma_add_reservation(h, vma, address);
1980                         if (unlikely(rc < 0))
1981                                 /*
1982                                  * See above comment about rare out of
1983                                  * memory condition.
1984                                  */
1985                                 ClearPagePrivate(page);
1986                 } else
1987                         vma_end_reservation(h, vma, address);
1988         }
1989 }
1990 
1991 struct page *alloc_huge_page(struct vm_area_struct *vma,
1992                                     unsigned long addr, int avoid_reserve)
1993 {
1994         struct hugepage_subpool *spool = subpool_vma(vma);
1995         struct hstate *h = hstate_vma(vma);
1996         struct page *page;
1997         long map_chg, map_commit;
1998         long gbl_chg;
1999         int ret, idx;
2000         struct hugetlb_cgroup *h_cg;
2001 
2002         idx = hstate_index(h);
2003         /*
2004          * Examine the region/reserve map to determine if the process
2005          * has a reservation for the page to be allocated.  A return
2006          * code of zero indicates a reservation exists (no change).
2007          */
2008         map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2009         if (map_chg < 0)
2010                 return ERR_PTR(-ENOMEM);
2011 
2012         /*
2013          * Processes that did not create the mapping will have no
2014          * reserves as indicated by the region/reserve map. Check
2015          * that the allocation will not exceed the subpool limit.
2016          * Allocations for MAP_NORESERVE mappings also need to be
2017          * checked against any subpool limit.
2018          */
2019         if (map_chg || avoid_reserve) {
2020                 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2021                 if (gbl_chg < 0) {
2022                         vma_end_reservation(h, vma, addr);
2023                         return ERR_PTR(-ENOSPC);
2024                 }
2025 
2026                 /*
2027                  * Even though there was no reservation in the region/reserve
2028                  * map, there could be reservations associated with the
2029                  * subpool that can be used.  This would be indicated if the
2030                  * return value of hugepage_subpool_get_pages() is zero.
2031                  * However, if avoid_reserve is specified we still avoid even
2032                  * the subpool reservations.
2033                  */
2034                 if (avoid_reserve)
2035                         gbl_chg = 1;
2036         }
2037 
2038         ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2039         if (ret)
2040                 goto out_subpool_put;
2041 
2042         spin_lock(&hugetlb_lock);
2043         /*
2044          * glb_chg is passed to indicate whether or not a page must be taken
2045          * from the global free pool (global change).  gbl_chg == 0 indicates
2046          * a reservation exists for the allocation.
2047          */
2048         page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2049         if (!page) {
2050                 spin_unlock(&hugetlb_lock);
2051                 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2052                 if (!page)
2053                         goto out_uncharge_cgroup;
2054                 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2055                         SetPagePrivate(page);
2056                         h->resv_huge_pages--;
2057                 }
2058                 spin_lock(&hugetlb_lock);
2059                 list_move(&page->lru, &h->hugepage_activelist);
2060                 /* Fall through */
2061         }
2062         hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2063         spin_unlock(&hugetlb_lock);
2064 
2065         set_page_private(page, (unsigned long)spool);
2066 
2067         map_commit = vma_commit_reservation(h, vma, addr);
2068         if (unlikely(map_chg > map_commit)) {
2069                 /*
2070                  * The page was added to the reservation map between
2071                  * vma_needs_reservation and vma_commit_reservation.
2072                  * This indicates a race with hugetlb_reserve_pages.
2073                  * Adjust for the subpool count incremented above AND
2074                  * in hugetlb_reserve_pages for the same page.  Also,
2075                  * the reservation count added in hugetlb_reserve_pages
2076                  * no longer applies.
2077                  */
2078                 long rsv_adjust;
2079 
2080                 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2081                 hugetlb_acct_memory(h, -rsv_adjust);
2082         }
2083         return page;
2084 
2085 out_uncharge_cgroup:
2086         hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2087 out_subpool_put:
2088         if (map_chg || avoid_reserve)
2089                 hugepage_subpool_put_pages(spool, 1);
2090         vma_end_reservation(h, vma, addr);
2091         return ERR_PTR(-ENOSPC);
2092 }
2093 
2094 int alloc_bootmem_huge_page(struct hstate *h)
2095         __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2096 int __alloc_bootmem_huge_page(struct hstate *h)
2097 {
2098         struct huge_bootmem_page *m;
2099         int nr_nodes, node;
2100 
2101         for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2102                 void *addr;
2103 
2104                 addr = memblock_virt_alloc_try_nid_nopanic(
2105                                 huge_page_size(h), huge_page_size(h),
2106                                 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
2107                 if (addr) {
2108                         /*
2109                          * Use the beginning of the huge page to store the
2110                          * huge_bootmem_page struct (until gather_bootmem
2111                          * puts them into the mem_map).
2112                          */
2113                         m = addr;
2114                         goto found;
2115                 }
2116         }
2117         return 0;
2118 
2119 found:
2120         BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2121         /* Put them into a private list first because mem_map is not up yet */
2122         list_add(&m->list, &huge_boot_pages);
2123         m->hstate = h;
2124         return 1;
2125 }
2126 
2127 static void __init prep_compound_huge_page(struct page *page,
2128                 unsigned int order)
2129 {
2130         if (unlikely(order > (MAX_ORDER - 1)))
2131                 prep_compound_gigantic_page(page, order);
2132         else
2133                 prep_compound_page(page, order);
2134 }
2135 
2136 /* Put bootmem huge pages into the standard lists after mem_map is up */
2137 static void __init gather_bootmem_prealloc(void)
2138 {
2139         struct huge_bootmem_page *m;
2140 
2141         list_for_each_entry(m, &huge_boot_pages, list) {
2142                 struct hstate *h = m->hstate;
2143                 struct page *page;
2144 
2145 #ifdef CONFIG_HIGHMEM
2146                 page = pfn_to_page(m->phys >> PAGE_SHIFT);
2147                 memblock_free_late(__pa(m),
2148                                    sizeof(struct huge_bootmem_page));
2149 #else
2150                 page = virt_to_page(m);
2151 #endif
2152                 WARN_ON(page_count(page) != 1);
2153                 prep_compound_huge_page(page, h->order);
2154                 WARN_ON(PageReserved(page));
2155                 prep_new_huge_page(h, page, page_to_nid(page));
2156                 put_page(page); /* free it into the hugepage allocator */
2157 
2158                 /*
2159                  * If we had gigantic hugepages allocated at boot time, we need
2160                  * to restore the 'stolen' pages to totalram_pages in order to
2161                  * fix confusing memory reports from free(1) and another
2162                  * side-effects, like CommitLimit going negative.
2163                  */
2164                 if (hstate_is_gigantic(h))
2165                         adjust_managed_page_count(page, 1 << h->order);
2166                 cond_resched();
2167         }
2168 }
2169 
2170 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2171 {
2172         unsigned long i;
2173 
2174         for (i = 0; i < h->max_huge_pages; ++i) {
2175                 if (hstate_is_gigantic(h)) {
2176                         if (!alloc_bootmem_huge_page(h))
2177                                 break;
2178                 } else if (!alloc_pool_huge_page(h,
2179                                          &node_states[N_MEMORY]))
2180                         break;
2181                 cond_resched();
2182         }
2183         if (i < h->max_huge_pages) {
2184                 char buf[32];
2185 
2186                 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2187                 pr_warn("HugeTLB: allocating %lu of page size %s failed.  Only allocated %lu hugepages.\n",
2188                         h->max_huge_pages, buf, i);
2189                 h->max_huge_pages = i;
2190         }
2191 }
2192 
2193 static void __init hugetlb_init_hstates(void)
2194 {
2195         struct hstate *h;
2196 
2197         for_each_hstate(h) {
2198                 if (minimum_order > huge_page_order(h))
2199                         minimum_order = huge_page_order(h);
2200 
2201                 /* oversize hugepages were init'ed in early boot */
2202                 if (!hstate_is_gigantic(h))
2203                         hugetlb_hstate_alloc_pages(h);
2204         }
2205         VM_BUG_ON(minimum_order == UINT_MAX);
2206 }
2207 
2208 static void __init report_hugepages(void)
2209 {
2210         struct hstate *h;
2211 
2212         for_each_hstate(h) {
2213                 char buf[32];
2214 
2215                 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2216                 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2217                         buf, h->free_huge_pages);
2218         }
2219 }
2220 
2221 #ifdef CONFIG_HIGHMEM
2222 static void try_to_free_low(struct hstate *h, unsigned long count,
2223                                                 nodemask_t *nodes_allowed)
2224 {
2225         int i;
2226 
2227         if (hstate_is_gigantic(h))
2228                 return;
2229 
2230         for_each_node_mask(i, *nodes_allowed) {
2231                 struct page *page, *next;
2232                 struct list_head *freel = &h->hugepage_freelists[i];
2233                 list_for_each_entry_safe(page, next, freel, lru) {
2234                         if (count >= h->nr_huge_pages)
2235                                 return;
2236                         if (PageHighMem(page))
2237                                 continue;
2238                         list_del(&page->lru);
2239                         update_and_free_page(h, page);
2240                         h->free_huge_pages--;
2241                         h->free_huge_pages_node[page_to_nid(page)]--;
2242                 }
2243         }
2244 }
2245 #else
2246 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2247                                                 nodemask_t *nodes_allowed)
2248 {
2249 }
2250 #endif
2251 
2252 /*
2253  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
2254  * balanced by operating on them in a round-robin fashion.
2255  * Returns 1 if an adjustment was made.
2256  */
2257 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2258                                 int delta)
2259 {
2260         int nr_nodes, node;
2261 
2262         VM_BUG_ON(delta != -1 && delta != 1);
2263 
2264         if (delta < 0) {
2265                 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2266                         if (h->surplus_huge_pages_node[node])
2267                                 goto found;
2268                 }
2269         } else {
2270                 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2271                         if (h->surplus_huge_pages_node[node] <
2272                                         h->nr_huge_pages_node[node])
2273                                 goto found;
2274                 }
2275         }
2276         return 0;
2277 
2278 found:
2279         h->surplus_huge_pages += delta;
2280         h->surplus_huge_pages_node[node] += delta;
2281         return 1;
2282 }
2283 
2284 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2285 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
2286                                                 nodemask_t *nodes_allowed)
2287 {
2288         unsigned long min_count, ret;
2289 
2290         if (hstate_is_gigantic(h) && !gigantic_page_supported())
2291                 return h->max_huge_pages;
2292 
2293         /*
2294          * Increase the pool size
2295          * First take pages out of surplus state.  Then make up the
2296          * remaining difference by allocating fresh huge pages.
2297          *
2298          * We might race with alloc_surplus_huge_page() here and be unable
2299          * to convert a surplus huge page to a normal huge page. That is
2300          * not critical, though, it just means the overall size of the
2301          * pool might be one hugepage larger than it needs to be, but
2302          * within all the constraints specified by the sysctls.
2303          */
2304         spin_lock(&hugetlb_lock);
2305         while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2306                 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2307                         break;
2308         }
2309 
2310         while (count > persistent_huge_pages(h)) {
2311                 /*
2312                  * If this allocation races such that we no longer need the
2313                  * page, free_huge_page will handle it by freeing the page
2314                  * and reducing the surplus.
2315                  */
2316                 spin_unlock(&hugetlb_lock);
2317 
2318                 /* yield cpu to avoid soft lockup */
2319                 cond_resched();
2320 
2321                 ret = alloc_pool_huge_page(h, nodes_allowed);
2322                 spin_lock(&hugetlb_lock);
2323                 if (!ret)
2324                         goto out;
2325 
2326                 /* Bail for signals. Probably ctrl-c from user */
2327                 if (signal_pending(current))
2328                         goto out;
2329         }
2330 
2331         /*
2332          * Decrease the pool size
2333          * First return free pages to the buddy allocator (being careful
2334          * to keep enough around to satisfy reservations).  Then place
2335          * pages into surplus state as needed so the pool will shrink
2336          * to the desired size as pages become free.
2337          *
2338          * By placing pages into the surplus state independent of the
2339          * overcommit value, we are allowing the surplus pool size to
2340          * exceed overcommit. There are few sane options here. Since
2341          * alloc_surplus_huge_page() is checking the global counter,
2342          * though, we'll note that we're not allowed to exceed surplus
2343          * and won't grow the pool anywhere else. Not until one of the
2344          * sysctls are changed, or the surplus pages go out of use.
2345          */
2346         min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2347         min_count = max(count, min_count);
2348         try_to_free_low(h, min_count, nodes_allowed);
2349         while (min_count < persistent_huge_pages(h)) {
2350                 if (!free_pool_huge_page(h, nodes_allowed, 0))
2351                         break;
2352                 cond_resched_lock(&hugetlb_lock);
2353         }
2354         while (count < persistent_huge_pages(h)) {
2355                 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2356                         break;
2357         }
2358 out:
2359         ret = persistent_huge_pages(h);
2360         spin_unlock(&hugetlb_lock);
2361         return ret;
2362 }
2363 
2364 #define HSTATE_ATTR_RO(_name) \
2365         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2366 
2367 #define HSTATE_ATTR(_name) \
2368         static struct kobj_attribute _name##_attr = \
2369                 __ATTR(_name, 0644, _name##_show, _name##_store)
2370 
2371 static struct kobject *hugepages_kobj;
2372 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2373 
2374 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2375 
2376 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2377 {
2378         int i;
2379 
2380         for (i = 0; i < HUGE_MAX_HSTATE; i++)
2381                 if (hstate_kobjs[i] == kobj) {
2382                         if (nidp)
2383                                 *nidp = NUMA_NO_NODE;
2384                         return &hstates[i];
2385                 }
2386 
2387         return kobj_to_node_hstate(kobj, nidp);
2388 }
2389 
2390 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2391                                         struct kobj_attribute *attr, char *buf)
2392 {
2393         struct hstate *h;
2394         unsigned long nr_huge_pages;
2395         int nid;
2396 
2397         h = kobj_to_hstate(kobj, &nid);
2398         if (nid == NUMA_NO_NODE)
2399                 nr_huge_pages = h->nr_huge_pages;
2400         else
2401                 nr_huge_pages = h->nr_huge_pages_node[nid];
2402 
2403         return sprintf(buf, "%lu\n", nr_huge_pages);
2404 }
2405 
2406 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2407                                            struct hstate *h, int nid,
2408                                            unsigned long count, size_t len)
2409 {
2410         int err;
2411         NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
2412 
2413         if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
2414                 err = -EINVAL;
2415                 goto out;
2416         }
2417 
2418         if (nid == NUMA_NO_NODE) {
2419                 /*
2420                  * global hstate attribute
2421                  */
2422                 if (!(obey_mempolicy &&
2423                                 init_nodemask_of_mempolicy(nodes_allowed))) {
2424                         NODEMASK_FREE(nodes_allowed);
2425                         nodes_allowed = &node_states[N_MEMORY];
2426                 }
2427         } else if (nodes_allowed) {
2428                 /*
2429                  * per node hstate attribute: adjust count to global,
2430                  * but restrict alloc/free to the specified node.
2431                  */
2432                 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2433                 init_nodemask_of_node(nodes_allowed, nid);
2434         } else
2435                 nodes_allowed = &node_states[N_MEMORY];
2436 
2437         h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
2438 
2439         if (nodes_allowed != &node_states[N_MEMORY])
2440                 NODEMASK_FREE(nodes_allowed);
2441 
2442         return len;
2443 out:
2444         NODEMASK_FREE(nodes_allowed);
2445         return err;
2446 }
2447 
2448 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2449                                          struct kobject *kobj, const char *buf,
2450                                          size_t len)
2451 {
2452         struct hstate *h;
2453         unsigned long count;
2454         int nid;
2455         int err;
2456 
2457         err = kstrtoul(buf, 10, &count);
2458         if (err)
2459                 return err;
2460 
2461         h = kobj_to_hstate(kobj, &nid);
2462         return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2463 }
2464 
2465 static ssize_t nr_hugepages_show(struct kobject *kobj,
2466                                        struct kobj_attribute *attr, char *buf)
2467 {
2468         return nr_hugepages_show_common(kobj, attr, buf);
2469 }
2470 
2471 static ssize_t nr_hugepages_store(struct kobject *kobj,
2472                struct kobj_attribute *attr, const char *buf, size_t len)
2473 {
2474         return nr_hugepages_store_common(false, kobj, buf, len);
2475 }
2476 HSTATE_ATTR(nr_hugepages);
2477 
2478 #ifdef CONFIG_NUMA
2479 
2480 /*
2481  * hstate attribute for optionally mempolicy-based constraint on persistent
2482  * huge page alloc/free.
2483  */
2484 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2485                                        struct kobj_attribute *attr, char *buf)
2486 {
2487         return nr_hugepages_show_common(kobj, attr, buf);
2488 }
2489 
2490 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2491                struct kobj_attribute *attr, const char *buf, size_t len)
2492 {
2493         return nr_hugepages_store_common(true, kobj, buf, len);
2494 }
2495 HSTATE_ATTR(nr_hugepages_mempolicy);
2496 #endif
2497 
2498 
2499 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2500                                         struct kobj_attribute *attr, char *buf)
2501 {
2502         struct hstate *h = kobj_to_hstate(kobj, NULL);
2503         return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2504 }
2505 
2506 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2507                 struct kobj_attribute *attr, const char *buf, size_t count)
2508 {
2509         int err;
2510         unsigned long input;
2511         struct hstate *h = kobj_to_hstate(kobj, NULL);
2512 
2513         if (hstate_is_gigantic(h))
2514                 return -EINVAL;
2515 
2516         err = kstrtoul(buf, 10, &input);
2517         if (err)
2518                 return err;
2519 
2520         spin_lock(&hugetlb_lock);
2521         h->nr_overcommit_huge_pages = input;
2522         spin_unlock(&hugetlb_lock);
2523 
2524         return count;
2525 }
2526 HSTATE_ATTR(nr_overcommit_hugepages);
2527 
2528 static ssize_t free_hugepages_show(struct kobject *kobj,
2529                                         struct kobj_attribute *attr, char *buf)
2530 {
2531         struct hstate *h;
2532         unsigned long free_huge_pages;
2533         int nid;
2534 
2535         h = kobj_to_hstate(kobj, &nid);
2536         if (nid == NUMA_NO_NODE)
2537                 free_huge_pages = h->free_huge_pages;
2538         else
2539                 free_huge_pages = h->free_huge_pages_node[nid];
2540 
2541         return sprintf(buf, "%lu\n", free_huge_pages);
2542 }
2543 HSTATE_ATTR_RO(free_hugepages);
2544 
2545 static ssize_t resv_hugepages_show(struct kobject *kobj,
2546                                         struct kobj_attribute *attr, char *buf)
2547 {
2548         struct hstate *h = kobj_to_hstate(kobj, NULL);
2549         return sprintf(buf, "%lu\n", h->resv_huge_pages);
2550 }
2551 HSTATE_ATTR_RO(resv_hugepages);
2552 
2553 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2554                                         struct kobj_attribute *attr, char *buf)
2555 {
2556         struct hstate *h;
2557         unsigned long surplus_huge_pages;
2558         int nid;
2559 
2560         h = kobj_to_hstate(kobj, &nid);
2561         if (nid == NUMA_NO_NODE)
2562                 surplus_huge_pages = h->surplus_huge_pages;
2563         else
2564                 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2565 
2566         return sprintf(buf, "%lu\n", surplus_huge_pages);
2567 }
2568 HSTATE_ATTR_RO(surplus_hugepages);
2569 
2570 static struct attribute *hstate_attrs[] = {
2571         &nr_hugepages_attr.attr,
2572         &nr_overcommit_hugepages_attr.attr,
2573         &free_hugepages_attr.attr,
2574         &resv_hugepages_attr.attr,
2575         &surplus_hugepages_attr.attr,
2576 #ifdef CONFIG_NUMA
2577         &nr_hugepages_mempolicy_attr.attr,
2578 #endif
2579         NULL,
2580 };
2581 
2582 static const struct attribute_group hstate_attr_group = {
2583         .attrs = hstate_attrs,
2584 };
2585 
2586 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2587                                     struct kobject **hstate_kobjs,
2588                                     const struct attribute_group *hstate_attr_group)
2589 {
2590         int retval;
2591         int hi = hstate_index(h);
2592 
2593         hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2594         if (!hstate_kobjs[hi])
2595                 return -ENOMEM;
2596 
2597         retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2598         if (retval)
2599                 kobject_put(hstate_kobjs[hi]);
2600 
2601         return retval;
2602 }
2603 
2604 static void __init hugetlb_sysfs_init(void)
2605 {
2606         struct hstate *h;
2607         int err;
2608 
2609         hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2610         if (!hugepages_kobj)
2611                 return;
2612 
2613         for_each_hstate(h) {
2614                 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2615                                          hstate_kobjs, &hstate_attr_group);
2616                 if (err)
2617                         pr_err("Hugetlb: Unable to add hstate %s", h->name);
2618         }
2619 }
2620 
2621 #ifdef CONFIG_NUMA
2622 
2623 /*
2624  * node_hstate/s - associate per node hstate attributes, via their kobjects,
2625  * with node devices in node_devices[] using a parallel array.  The array
2626  * index of a node device or _hstate == node id.
2627  * This is here to avoid any static dependency of the node device driver, in
2628  * the base kernel, on the hugetlb module.
2629  */
2630 struct node_hstate {
2631         struct kobject          *hugepages_kobj;
2632         struct kobject          *hstate_kobjs[HUGE_MAX_HSTATE];
2633 };
2634 static struct node_hstate node_hstates[MAX_NUMNODES];
2635 
2636 /*
2637  * A subset of global hstate attributes for node devices
2638  */
2639 static struct attribute *per_node_hstate_attrs[] = {
2640         &nr_hugepages_attr.attr,
2641         &free_hugepages_attr.attr,
2642         &surplus_hugepages_attr.attr,
2643         NULL,
2644 };
2645 
2646 static const struct attribute_group per_node_hstate_attr_group = {
2647         .attrs = per_node_hstate_attrs,
2648 };
2649 
2650 /*
2651  * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2652  * Returns node id via non-NULL nidp.
2653  */
2654 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2655 {
2656         int nid;
2657 
2658         for (nid = 0; nid < nr_node_ids; nid++) {
2659                 struct node_hstate *nhs = &node_hstates[nid];
2660                 int i;
2661                 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2662                         if (nhs->hstate_kobjs[i] == kobj) {
2663                                 if (nidp)
2664                                         *nidp = nid;
2665                                 return &hstates[i];
2666                         }
2667         }
2668 
2669         BUG();
2670         return NULL;
2671 }
2672 
2673 /*
2674  * Unregister hstate attributes from a single node device.
2675  * No-op if no hstate attributes attached.
2676  */
2677 static void hugetlb_unregister_node(struct node *node)
2678 {
2679         struct hstate *h;
2680         struct node_hstate *nhs = &node_hstates[node->dev.id];
2681 
2682         if (!nhs->hugepages_kobj)
2683                 return;         /* no hstate attributes */
2684 
2685         for_each_hstate(h) {
2686                 int idx = hstate_index(h);
2687                 if (nhs->hstate_kobjs[idx]) {
2688                         kobject_put(nhs->hstate_kobjs[idx]);
2689                         nhs->hstate_kobjs[idx] = NULL;
2690                 }
2691         }
2692 
2693         kobject_put(nhs->hugepages_kobj);
2694         nhs->hugepages_kobj = NULL;
2695 }
2696 
2697 
2698 /*
2699  * Register hstate attributes for a single node device.
2700  * No-op if attributes already registered.
2701  */
2702 static void hugetlb_register_node(struct node *node)
2703 {
2704         struct hstate *h;
2705         struct node_hstate *nhs = &node_hstates[node->dev.id];
2706         int err;
2707 
2708         if (nhs->hugepages_kobj)
2709                 return;         /* already allocated */
2710 
2711         nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2712                                                         &node->dev.kobj);
2713         if (!nhs->hugepages_kobj)
2714                 return;
2715 
2716         for_each_hstate(h) {
2717                 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2718                                                 nhs->hstate_kobjs,
2719                                                 &per_node_hstate_attr_group);
2720                 if (err) {
2721                         pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2722                                 h->name, node->dev.id);
2723                         hugetlb_unregister_node(node);
2724                         break;
2725                 }
2726         }
2727 }
2728 
2729 /*
2730  * hugetlb init time:  register hstate attributes for all registered node
2731  * devices of nodes that have memory.  All on-line nodes should have
2732  * registered their associated device by this time.
2733  */
2734 static void __init hugetlb_register_all_nodes(void)
2735 {
2736         int nid;
2737 
2738         for_each_node_state(nid, N_MEMORY) {
2739                 struct node *node = node_devices[nid];
2740                 if (node->dev.id == nid)
2741                         hugetlb_register_node(node);
2742         }
2743 
2744         /*
2745          * Let the node device driver know we're here so it can
2746          * [un]register hstate attributes on node hotplug.
2747          */
2748         register_hugetlbfs_with_node(hugetlb_register_node,
2749                                      hugetlb_unregister_node);
2750 }
2751 #else   /* !CONFIG_NUMA */
2752 
2753 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2754 {
2755         BUG();
2756         if (nidp)
2757                 *nidp = -1;
2758         return NULL;
2759 }
2760 
2761 static void hugetlb_register_all_nodes(void) { }
2762 
2763 #endif
2764 
2765 static int __init hugetlb_init(void)
2766 {
2767         int i;
2768 
2769         if (!hugepages_supported())
2770                 return 0;
2771 
2772         if (!size_to_hstate(default_hstate_size)) {
2773                 if (default_hstate_size != 0) {
2774                         pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n",
2775                                default_hstate_size, HPAGE_SIZE);
2776                 }
2777 
2778                 default_hstate_size = HPAGE_SIZE;
2779                 if (!size_to_hstate(default_hstate_size))
2780                         hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2781         }
2782         default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2783         if (default_hstate_max_huge_pages) {
2784                 if (!default_hstate.max_huge_pages)
2785                         default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2786         }
2787 
2788         hugetlb_init_hstates();
2789         gather_bootmem_prealloc();
2790         report_hugepages();
2791 
2792         hugetlb_sysfs_init();
2793         hugetlb_register_all_nodes();
2794         hugetlb_cgroup_file_init();
2795 
2796 #ifdef CONFIG_SMP
2797         num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2798 #else
2799         num_fault_mutexes = 1;
2800 #endif
2801         hugetlb_fault_mutex_table =
2802                 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
2803                               GFP_KERNEL);
2804         BUG_ON(!hugetlb_fault_mutex_table);
2805 
2806         for (i = 0; i < num_fault_mutexes; i++)
2807                 mutex_init(&hugetlb_fault_mutex_table[i]);
2808         return 0;
2809 }
2810 subsys_initcall(hugetlb_init);
2811 
2812 /* Should be called on processing a hugepagesz=... option */
2813 void __init hugetlb_bad_size(void)
2814 {
2815         parsed_valid_hugepagesz = false;
2816 }
2817 
2818 void __init hugetlb_add_hstate(unsigned int order)
2819 {
2820         struct hstate *h;
2821         unsigned long i;
2822 
2823         if (size_to_hstate(PAGE_SIZE << order)) {
2824                 pr_warn("hugepagesz= specified twice, ignoring\n");
2825                 return;
2826         }
2827         BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2828         BUG_ON(order == 0);
2829         h = &hstates[hugetlb_max_hstate++];
2830         h->order = order;
2831         h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2832         h->nr_huge_pages = 0;
2833         h->free_huge_pages = 0;
2834         for (i = 0; i < MAX_NUMNODES; ++i)
2835                 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2836         INIT_LIST_HEAD(&h->hugepage_activelist);
2837         h->next_nid_to_alloc = first_memory_node;
2838         h->next_nid_to_free = first_memory_node;
2839         snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2840                                         huge_page_size(h)/1024);
2841 
2842         parsed_hstate = h;
2843 }
2844 
2845 static int __init hugetlb_nrpages_setup(char *s)
2846 {
2847         unsigned long *mhp;
2848         static unsigned long *last_mhp;
2849 
2850         if (!parsed_valid_hugepagesz) {
2851                 pr_warn("hugepages = %s preceded by "
2852                         "an unsupported hugepagesz, ignoring\n", s);
2853                 parsed_valid_hugepagesz = true;
2854                 return 1;
2855         }
2856         /*
2857          * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2858          * so this hugepages= parameter goes to the "default hstate".
2859          */
2860         else if (!hugetlb_max_hstate)
2861                 mhp = &default_hstate_max_huge_pages;
2862         else
2863                 mhp = &parsed_hstate->max_huge_pages;
2864 
2865         if (mhp == last_mhp) {
2866                 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
2867                 return 1;
2868         }
2869 
2870         if (sscanf(s, "%lu", mhp) <= 0)
2871                 *mhp = 0;
2872 
2873         /*
2874          * Global state is always initialized later in hugetlb_init.
2875          * But we need to allocate >= MAX_ORDER hstates here early to still
2876          * use the bootmem allocator.
2877          */
2878         if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2879                 hugetlb_hstate_alloc_pages(parsed_hstate);
2880 
2881         last_mhp = mhp;
2882 
2883         return 1;
2884 }
2885 __setup("hugepages=", hugetlb_nrpages_setup);
2886 
2887 static int __init hugetlb_default_setup(char *s)
2888 {
2889         default_hstate_size = memparse(s, &s);
2890         return 1;
2891 }
2892 __setup("default_hugepagesz=", hugetlb_default_setup);
2893 
2894 static unsigned int cpuset_mems_nr(unsigned int *array)
2895 {
2896         int node;
2897         unsigned int nr = 0;
2898 
2899         for_each_node_mask(node, cpuset_current_mems_allowed)
2900                 nr += array[node];
2901 
2902         return nr;
2903 }
2904 
2905 #ifdef CONFIG_SYSCTL
2906 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2907                          struct ctl_table *table, int write,
2908                          void __user *buffer, size_t *length, loff_t *ppos)
2909 {
2910         struct hstate *h = &default_hstate;
2911         unsigned long tmp = h->max_huge_pages;
2912         int ret;
2913 
2914         if (!hugepages_supported())
2915                 return -EOPNOTSUPP;
2916 
2917         table->data = &tmp;
2918         table->maxlen = sizeof(unsigned long);
2919         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2920         if (ret)
2921                 goto out;
2922 
2923         if (write)
2924                 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2925                                                   NUMA_NO_NODE, tmp, *length);
2926 out:
2927         return ret;
2928 }
2929 
2930 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2931                           void __user *buffer, size_t *length, loff_t *ppos)
2932 {
2933 
2934         return hugetlb_sysctl_handler_common(false, table, write,
2935                                                         buffer, length, ppos);
2936 }
2937 
2938 #ifdef CONFIG_NUMA
2939 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2940                           void __user *buffer, size_t *length, loff_t *ppos)
2941 {
2942         return hugetlb_sysctl_handler_common(true, table, write,
2943                                                         buffer, length, ppos);
2944 }
2945 #endif /* CONFIG_NUMA */
2946 
2947 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2948                         void __user *buffer,
2949                         size_t *length, loff_t *ppos)
2950 {
2951         struct hstate *h = &default_hstate;
2952         unsigned long tmp;
2953         int ret;
2954 
2955         if (!hugepages_supported())
2956                 return -EOPNOTSUPP;
2957 
2958         tmp = h->nr_overcommit_huge_pages;
2959 
2960         if (write && hstate_is_gigantic(h))
2961                 return -EINVAL;
2962 
2963         table->data = &tmp;
2964         table->maxlen = sizeof(unsigned long);
2965         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2966         if (ret)
2967                 goto out;
2968 
2969         if (write) {
2970                 spin_lock(&hugetlb_lock);
2971                 h->nr_overcommit_huge_pages = tmp;
2972                 spin_unlock(&hugetlb_lock);
2973         }
2974 out:
2975         return ret;
2976 }
2977 
2978 #endif /* CONFIG_SYSCTL */
2979 
2980 void hugetlb_report_meminfo(struct seq_file *m)
2981 {
2982         struct hstate *h;
2983         unsigned long total = 0;
2984 
2985         if (!hugepages_supported())
2986                 return;
2987 
2988         for_each_hstate(h) {
2989                 unsigned long count = h->nr_huge_pages;
2990 
2991                 total += (PAGE_SIZE << huge_page_order(h)) * count;
2992 
2993                 if (h == &default_hstate)
2994                         seq_printf(m,
2995                                    "HugePages_Total:   %5lu\n"
2996                                    "HugePages_Free:    %5lu\n"
2997                                    "HugePages_Rsvd:    %5lu\n"
2998                                    "HugePages_Surp:    %5lu\n"
2999                                    "Hugepagesize:   %8lu kB\n",
3000                                    count,
3001                                    h->free_huge_pages,
3002                                    h->resv_huge_pages,
3003                                    h->surplus_huge_pages,
3004                                    (PAGE_SIZE << huge_page_order(h)) / 1024);
3005         }
3006 
3007         seq_printf(m, "Hugetlb:        %8lu kB\n", total / 1024);
3008 }
3009 
3010 int hugetlb_report_node_meminfo(int nid, char *buf)
3011 {
3012         struct hstate *h = &default_hstate;
3013         if (!hugepages_supported())
3014                 return 0;
3015         return sprintf(buf,
3016                 "Node %d HugePages_Total: %5u\n"
3017                 "Node %d HugePages_Free:  %5u\n"
3018                 "Node %d HugePages_Surp:  %5u\n",
3019                 nid, h->nr_huge_pages_node[nid],
3020                 nid, h->free_huge_pages_node[nid],
3021                 nid, h->surplus_huge_pages_node[nid]);
3022 }
3023 
3024 void hugetlb_show_meminfo(void)
3025 {
3026         struct hstate *h;
3027         int nid;
3028 
3029         if (!hugepages_supported())
3030                 return;
3031 
3032         for_each_node_state(nid, N_MEMORY)
3033                 for_each_hstate(h)
3034                         pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3035                                 nid,
3036                                 h->nr_huge_pages_node[nid],
3037                                 h->free_huge_pages_node[nid],
3038                                 h->surplus_huge_pages_node[nid],
3039                                 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3040 }
3041 
3042 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3043 {
3044         seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3045                    atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3046 }
3047 
3048 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3049 unsigned long hugetlb_total_pages(void)
3050 {
3051         struct hstate *h;
3052         unsigned long nr_total_pages = 0;
3053 
3054         for_each_hstate(h)
3055                 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3056         return nr_total_pages;
3057 }
3058 
3059 static int hugetlb_acct_memory(struct hstate *h, long delta)
3060 {
3061         int ret = -ENOMEM;
3062 
3063         spin_lock(&hugetlb_lock);
3064         /*
3065          * When cpuset is configured, it breaks the strict hugetlb page
3066          * reservation as the accounting is done on a global variable. Such
3067          * reservation is completely rubbish in the presence of cpuset because
3068          * the reservation is not checked against page availability for the
3069          * current cpuset. Application can still potentially OOM'ed by kernel
3070          * with lack of free htlb page in cpuset that the task is in.
3071          * Attempt to enforce strict accounting with cpuset is almost
3072          * impossible (or too ugly) because cpuset is too fluid that
3073          * task or memory node can be dynamically moved between cpusets.
3074          *
3075          * The change of semantics for shared hugetlb mapping with cpuset is
3076          * undesirable. However, in order to preserve some of the semantics,
3077          * we fall back to check against current free page availability as
3078          * a best attempt and hopefully to minimize the impact of changing
3079          * semantics that cpuset has.
3080          */
3081         if (delta > 0) {
3082                 if (gather_surplus_pages(h, delta) < 0)
3083                         goto out;
3084 
3085                 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
3086                         return_unused_surplus_pages(h, delta);
3087                         goto out;
3088                 }
3089         }
3090 
3091         ret = 0;
3092         if (delta < 0)
3093                 return_unused_surplus_pages(h, (unsigned long) -delta);
3094 
3095 out:
3096         spin_unlock(&hugetlb_lock);
3097         return ret;
3098 }
3099 
3100 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3101 {
3102         struct resv_map *resv = vma_resv_map(vma);
3103 
3104         /*
3105          * This new VMA should share its siblings reservation map if present.
3106          * The VMA will only ever have a valid reservation map pointer where
3107          * it is being copied for another still existing VMA.  As that VMA
3108          * has a reference to the reservation map it cannot disappear until
3109          * after this open call completes.  It is therefore safe to take a
3110          * new reference here without additional locking.
3111          */
3112         if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3113                 kref_get(&resv->refs);
3114 }
3115 
3116 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3117 {
3118         struct hstate *h = hstate_vma(vma);
3119         struct resv_map *resv = vma_resv_map(vma);
3120         struct hugepage_subpool *spool = subpool_vma(vma);
3121         unsigned long reserve, start, end;
3122         long gbl_reserve;
3123 
3124         if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3125                 return;
3126 
3127         start = vma_hugecache_offset(h, vma, vma->vm_start);
3128         end = vma_hugecache_offset(h, vma, vma->vm_end);
3129 
3130         reserve = (end - start) - region_count(resv, start, end);
3131 
3132         kref_put(&resv->refs, resv_map_release);
3133 
3134         if (reserve) {
3135                 /*
3136                  * Decrement reserve counts.  The global reserve count may be
3137                  * adjusted if the subpool has a minimum size.
3138                  */
3139                 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3140                 hugetlb_acct_memory(h, -gbl_reserve);
3141         }
3142 }
3143 
3144 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3145 {
3146         if (addr & ~(huge_page_mask(hstate_vma(vma))))
3147                 return -EINVAL;
3148         return 0;
3149 }
3150 
3151 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
3152 {
3153         struct hstate *hstate = hstate_vma(vma);
3154 
3155         return 1UL << huge_page_shift(hstate);
3156 }
3157 
3158 /*
3159  * We cannot handle pagefaults against hugetlb pages at all.  They cause
3160  * handle_mm_fault() to try to instantiate regular-sized pages in the
3161  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
3162  * this far.
3163  */
3164 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
3165 {
3166         BUG();
3167         return 0;
3168 }
3169 
3170 /*
3171  * When a new function is introduced to vm_operations_struct and added
3172  * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3173  * This is because under System V memory model, mappings created via
3174  * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3175  * their original vm_ops are overwritten with shm_vm_ops.
3176  */
3177 const struct vm_operations_struct hugetlb_vm_ops = {
3178         .fault = hugetlb_vm_op_fault,
3179         .open = hugetlb_vm_op_open,
3180         .close = hugetlb_vm_op_close,
3181         .split = hugetlb_vm_op_split,
3182         .pagesize = hugetlb_vm_op_pagesize,
3183 };
3184 
3185 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3186                                 int writable)
3187 {
3188         pte_t entry;
3189 
3190         if (writable) {
3191                 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3192                                          vma->vm_page_prot)));
3193         } else {
3194                 entry = huge_pte_wrprotect(mk_huge_pte(page,
3195                                            vma->vm_page_prot));
3196         }
3197         entry = pte_mkyoung(entry);
3198         entry = pte_mkhuge(entry);
3199         entry = arch_make_huge_pte(entry, vma, page, writable);
3200 
3201         return entry;
3202 }
3203 
3204 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3205                                    unsigned long address, pte_t *ptep)
3206 {
3207         pte_t entry;
3208 
3209         entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3210         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3211                 update_mmu_cache(vma, address, ptep);
3212 }
3213 
3214 bool is_hugetlb_entry_migration(pte_t pte)
3215 {
3216         swp_entry_t swp;
3217 
3218         if (huge_pte_none(pte) || pte_present(pte))
3219                 return false;
3220         swp = pte_to_swp_entry(pte);
3221         if (non_swap_entry(swp) && is_migration_entry(swp))
3222                 return true;
3223         else
3224                 return false;
3225 }
3226 
3227 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3228 {
3229         swp_entry_t swp;
3230 
3231         if (huge_pte_none(pte) || pte_present(pte))
3232                 return 0;
3233         swp = pte_to_swp_entry(pte);
3234         if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3235                 return 1;
3236         else
3237                 return 0;
3238 }
3239 
3240 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3241                             struct vm_area_struct *vma)
3242 {
3243         pte_t *src_pte, *dst_pte, entry, dst_entry;
3244         struct page *ptepage;
3245         unsigned long addr;
3246         int cow;
3247         struct hstate *h = hstate_vma(vma);
3248         unsigned long sz = huge_page_size(h);
3249         unsigned long mmun_start;       /* For mmu_notifiers */
3250         unsigned long mmun_end;         /* For mmu_notifiers */
3251         int ret = 0;
3252 
3253         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3254 
3255         mmun_start = vma->vm_start;
3256         mmun_end = vma->vm_end;
3257         if (cow)
3258                 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
3259 
3260         for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3261                 spinlock_t *src_ptl, *dst_ptl;
3262                 src_pte = huge_pte_offset(src, addr, sz);
3263                 if (!src_pte)
3264                         continue;
3265                 dst_pte = huge_pte_alloc(dst, addr, sz);
3266                 if (!dst_pte) {
3267                         ret = -ENOMEM;
3268                         break;
3269                 }
3270 
3271                 /*
3272                  * If the pagetables are shared don't copy or take references.
3273                  * dst_pte == src_pte is the common case of src/dest sharing.
3274                  *
3275                  * However, src could have 'unshared' and dst shares with
3276                  * another vma.  If dst_pte !none, this implies sharing.
3277                  * Check here before taking page table lock, and once again
3278                  * after taking the lock below.
3279                  */
3280                 dst_entry = huge_ptep_get(dst_pte);
3281                 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
3282                         continue;
3283 
3284                 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3285                 src_ptl = huge_pte_lockptr(h, src, src_pte);
3286                 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3287                 entry = huge_ptep_get(src_pte);
3288                 dst_entry = huge_ptep_get(dst_pte);
3289                 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3290                         /*
3291                          * Skip if src entry none.  Also, skip in the
3292                          * unlikely case dst entry !none as this implies
3293                          * sharing with another vma.
3294                          */
3295                         ;
3296                 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3297                                     is_hugetlb_entry_hwpoisoned(entry))) {
3298                         swp_entry_t swp_entry = pte_to_swp_entry(entry);
3299 
3300                         if (is_write_migration_entry(swp_entry) && cow) {
3301                                 /*
3302                                  * COW mappings require pages in both
3303                                  * parent and child to be set to read.
3304                                  */
3305                                 make_migration_entry_read(&swp_entry);
3306                                 entry = swp_entry_to_pte(swp_entry);
3307                                 set_huge_swap_pte_at(src, addr, src_pte,
3308                                                      entry, sz);
3309                         }
3310                         set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3311                 } else {
3312                         if (cow) {
3313                                 /*
3314                                  * No need to notify as we are downgrading page
3315                                  * table protection not changing it to point
3316                                  * to a new page.
3317                                  *
3318                                  * See Documentation/vm/mmu_notifier.rst
3319                                  */
3320                                 huge_ptep_set_wrprotect(src, addr, src_pte);
3321                         }
3322                         entry = huge_ptep_get(src_pte);
3323                         ptepage = pte_page(entry);
3324                         get_page(ptepage);
3325                         page_dup_rmap(ptepage, true);
3326                         set_huge_pte_at(dst, addr, dst_pte, entry);
3327                         hugetlb_count_add(pages_per_huge_page(h), dst);
3328                 }
3329                 spin_unlock(src_ptl);
3330                 spin_unlock(dst_ptl);
3331         }
3332 
3333         if (cow)
3334                 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
3335 
3336         return ret;
3337 }
3338 
3339 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3340                             unsigned long start, unsigned long end,
3341                             struct page *ref_page)
3342 {
3343         struct mm_struct *mm = vma->vm_mm;
3344         unsigned long address;
3345         pte_t *ptep;
3346         pte_t pte;
3347         spinlock_t *ptl;
3348         struct page *page;
3349         struct hstate *h = hstate_vma(vma);
3350         unsigned long sz = huge_page_size(h);
3351         const unsigned long mmun_start = start; /* For mmu_notifiers */
3352         const unsigned long mmun_end   = end;   /* For mmu_notifiers */
3353 
3354         WARN_ON(!is_vm_hugetlb_page(vma));
3355         BUG_ON(start & ~huge_page_mask(h));
3356         BUG_ON(end & ~huge_page_mask(h));
3357 
3358         /*
3359          * This is a hugetlb vma, all the pte entries should point
3360          * to huge page.
3361          */
3362         tlb_remove_check_page_size_change(tlb, sz);
3363         tlb_start_vma(tlb, vma);
3364         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3365         address = start;
3366         for (; address < end; address += sz) {
3367                 ptep = huge_pte_offset(mm, address, sz);
3368                 if (!ptep)
3369                         continue;
3370 
3371                 ptl = huge_pte_lock(h, mm, ptep);
3372                 if (huge_pmd_unshare(mm, &address, ptep)) {
3373                         spin_unlock(ptl);
3374                         continue;
3375                 }
3376 
3377                 pte = huge_ptep_get(ptep);
3378                 if (huge_pte_none(pte)) {
3379                         spin_unlock(ptl);
3380                         continue;
3381                 }
3382 
3383                 /*
3384                  * Migrating hugepage or HWPoisoned hugepage is already
3385                  * unmapped and its refcount is dropped, so just clear pte here.
3386                  */
3387                 if (unlikely(!pte_present(pte))) {
3388                         huge_pte_clear(mm, address, ptep, sz);
3389                         spin_unlock(ptl);
3390                         continue;
3391                 }
3392 
3393                 page = pte_page(pte);
3394                 /*
3395                  * If a reference page is supplied, it is because a specific
3396                  * page is being unmapped, not a range. Ensure the page we
3397                  * are about to unmap is the actual page of interest.
3398                  */
3399                 if (ref_page) {
3400                         if (page != ref_page) {
3401                                 spin_unlock(ptl);
3402                                 continue;
3403                         }
3404                         /*
3405                          * Mark the VMA as having unmapped its page so that
3406                          * future faults in this VMA will fail rather than
3407                          * looking like data was lost
3408                          */
3409                         set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3410                 }
3411 
3412                 pte = huge_ptep_get_and_clear(mm, address, ptep);
3413                 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
3414                 if (huge_pte_dirty(pte))
3415                         set_page_dirty(page);
3416 
3417                 hugetlb_count_sub(pages_per_huge_page(h), mm);
3418                 page_remove_rmap(page, true);
3419 
3420                 spin_unlock(ptl);
3421                 tlb_remove_page_size(tlb, page, huge_page_size(h));
3422                 /*
3423                  * Bail out after unmapping reference page if supplied
3424                  */
3425                 if (ref_page)
3426                         break;
3427         }
3428         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3429         tlb_end_vma(tlb, vma);
3430 }
3431 
3432 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3433                           struct vm_area_struct *vma, unsigned long start,
3434                           unsigned long end, struct page *ref_page)
3435 {
3436         __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3437 
3438         /*
3439          * Clear this flag so that x86's huge_pmd_share page_table_shareable
3440          * test will fail on a vma being torn down, and not grab a page table
3441          * on its way out.  We're lucky that the flag has such an appropriate
3442          * name, and can in fact be safely cleared here. We could clear it
3443          * before the __unmap_hugepage_range above, but all that's necessary
3444          * is to clear it before releasing the i_mmap_rwsem. This works
3445          * because in the context this is called, the VMA is about to be
3446          * destroyed and the i_mmap_rwsem is held.
3447          */
3448         vma->vm_flags &= ~VM_MAYSHARE;
3449 }
3450 
3451 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3452                           unsigned long end, struct page *ref_page)
3453 {
3454         struct mm_struct *mm;
3455         struct mmu_gather tlb;
3456 
3457         mm = vma->vm_mm;
3458 
3459         tlb_gather_mmu(&tlb, mm, start, end);
3460         __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3461         tlb_finish_mmu(&tlb, start, end);
3462 }
3463 
3464 /*
3465  * This is called when the original mapper is failing to COW a MAP_PRIVATE
3466  * mappping it owns the reserve page for. The intention is to unmap the page
3467  * from other VMAs and let the children be SIGKILLed if they are faulting the
3468  * same region.
3469  */
3470 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3471                               struct page *page, unsigned long address)
3472 {
3473         struct hstate *h = hstate_vma(vma);
3474         struct vm_area_struct *iter_vma;
3475         struct address_space *mapping;
3476         pgoff_t pgoff;
3477 
3478         /*
3479          * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3480          * from page cache lookup which is in HPAGE_SIZE units.
3481          */
3482         address = address & huge_page_mask(h);
3483         pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3484                         vma->vm_pgoff;
3485         mapping = vma->vm_file->f_mapping;
3486 
3487         /*
3488          * Take the mapping lock for the duration of the table walk. As
3489          * this mapping should be shared between all the VMAs,
3490          * __unmap_hugepage_range() is called as the lock is already held
3491          */
3492         i_mmap_lock_write(mapping);
3493         vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3494                 /* Do not unmap the current VMA */
3495                 if (iter_vma == vma)
3496                         continue;
3497 
3498                 /*
3499                  * Shared VMAs have their own reserves and do not affect
3500                  * MAP_PRIVATE accounting but it is possible that a shared
3501                  * VMA is using the same page so check and skip such VMAs.
3502                  */
3503                 if (iter_vma->vm_flags & VM_MAYSHARE)
3504                         continue;
3505 
3506                 /*
3507                  * Unmap the page from other VMAs without their own reserves.
3508                  * They get marked to be SIGKILLed if they fault in these
3509                  * areas. This is because a future no-page fault on this VMA
3510                  * could insert a zeroed page instead of the data existing
3511                  * from the time of fork. This would look like data corruption
3512                  */
3513                 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3514                         unmap_hugepage_range(iter_vma, address,
3515                                              address + huge_page_size(h), page);
3516         }
3517         i_mmap_unlock_write(mapping);
3518 }
3519 
3520 /*
3521  * Hugetlb_cow() should be called with page lock of the original hugepage held.
3522  * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3523  * cannot race with other handlers or page migration.
3524  * Keep the pte_same checks anyway to make transition from the mutex easier.
3525  */
3526 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3527                        unsigned long address, pte_t *ptep,
3528                        struct page *pagecache_page, spinlock_t *ptl)
3529 {
3530         pte_t pte;
3531         struct hstate *h = hstate_vma(vma);
3532         struct page *old_page, *new_page;
3533         int ret = 0, outside_reserve = 0;
3534         unsigned long mmun_start;       /* For mmu_notifiers */
3535         unsigned long mmun_end;         /* For mmu_notifiers */
3536 
3537         pte = huge_ptep_get(ptep);
3538         old_page = pte_page(pte);
3539 
3540 retry_avoidcopy:
3541         /* If no-one else is actually using this page, avoid the copy
3542          * and just make the page writable */
3543         if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3544                 page_move_anon_rmap(old_page, vma);
3545                 set_huge_ptep_writable(vma, address, ptep);
3546                 return 0;
3547         }
3548 
3549         /*
3550          * If the process that created a MAP_PRIVATE mapping is about to
3551          * perform a COW due to a shared page count, attempt to satisfy
3552          * the allocation without using the existing reserves. The pagecache
3553          * page is used to determine if the reserve at this address was
3554          * consumed or not. If reserves were used, a partial faulted mapping
3555          * at the time of fork() could consume its reserves on COW instead
3556          * of the full address range.
3557          */
3558         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3559                         old_page != pagecache_page)
3560                 outside_reserve = 1;
3561 
3562         get_page(old_page);
3563 
3564         /*
3565          * Drop page table lock as buddy allocator may be called. It will
3566          * be acquired again before returning to the caller, as expected.
3567          */
3568         spin_unlock(ptl);
3569         new_page = alloc_huge_page(vma, address, outside_reserve);
3570 
3571         if (IS_ERR(new_page)) {
3572                 /*
3573                  * If a process owning a MAP_PRIVATE mapping fails to COW,
3574                  * it is due to references held by a child and an insufficient
3575                  * huge page pool. To guarantee the original mappers
3576                  * reliability, unmap the page from child processes. The child
3577                  * may get SIGKILLed if it later faults.
3578                  */
3579                 if (outside_reserve) {
3580                         put_page(old_page);
3581                         BUG_ON(huge_pte_none(pte));
3582                         unmap_ref_private(mm, vma, old_page, address);
3583                         BUG_ON(huge_pte_none(pte));
3584                         spin_lock(ptl);
3585                         ptep = huge_pte_offset(mm, address & huge_page_mask(h),
3586                                                huge_page_size(h));
3587                         if (likely(ptep &&
3588                                    pte_same(huge_ptep_get(ptep), pte)))
3589                                 goto retry_avoidcopy;
3590                         /*
3591                          * race occurs while re-acquiring page table
3592                          * lock, and our job is done.
3593                          */
3594                         return 0;
3595                 }
3596 
3597                 ret = (PTR_ERR(new_page) == -ENOMEM) ?
3598                         VM_FAULT_OOM : VM_FAULT_SIGBUS;
3599                 goto out_release_old;
3600         }
3601 
3602         /*
3603          * When the original hugepage is shared one, it does not have
3604          * anon_vma prepared.
3605          */
3606         if (unlikely(anon_vma_prepare(vma))) {
3607                 ret = VM_FAULT_OOM;
3608                 goto out_release_all;
3609         }
3610 
3611         copy_user_huge_page(new_page, old_page, address, vma,
3612                             pages_per_huge_page(h));
3613         __SetPageUptodate(new_page);
3614         set_page_huge_active(new_page);
3615 
3616         mmun_start = address & huge_page_mask(h);
3617         mmun_end = mmun_start + huge_page_size(h);
3618         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3619 
3620         /*
3621          * Retake the page table lock to check for racing updates
3622          * before the page tables are altered
3623          */
3624         spin_lock(ptl);
3625         ptep = huge_pte_offset(mm, address & huge_page_mask(h),
3626                                huge_page_size(h));
3627         if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3628                 ClearPagePrivate(new_page);
3629 
3630                 /* Break COW */
3631                 huge_ptep_clear_flush(vma, address, ptep);
3632                 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
3633                 set_huge_pte_at(mm, address, ptep,
3634                                 make_huge_pte(vma, new_page, 1));
3635                 page_remove_rmap(old_page, true);
3636                 hugepage_add_new_anon_rmap(new_page, vma, address);
3637                 /* Make the old page be freed below */
3638                 new_page = old_page;
3639         }
3640         spin_unlock(ptl);
3641         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3642 out_release_all:
3643         restore_reserve_on_error(h, vma, address, new_page);
3644         put_page(new_page);
3645 out_release_old:
3646         put_page(old_page);
3647 
3648         spin_lock(ptl); /* Caller expects lock to be held */
3649         return ret;
3650 }
3651 
3652 /* Return the pagecache page at a given address within a VMA */
3653 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3654                         struct vm_area_struct *vma, unsigned long address)
3655 {
3656         struct address_space *mapping;
3657         pgoff_t idx;
3658 
3659         mapping = vma->vm_file->f_mapping;
3660         idx = vma_hugecache_offset(h, vma, address);
3661 
3662         return find_lock_page(mapping, idx);
3663 }
3664 
3665 /*
3666  * Return whether there is a pagecache page to back given address within VMA.
3667  * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3668  */
3669 static bool hugetlbfs_pagecache_present(struct hstate *h,
3670                         struct vm_area_struct *vma, unsigned long address)
3671 {
3672         struct address_space *mapping;
3673         pgoff_t idx;
3674         struct page *page;
3675 
3676         mapping = vma->vm_file->f_mapping;
3677         idx = vma_hugecache_offset(h, vma, address);
3678 
3679         page = find_get_page(mapping, idx);
3680         if (page)
3681                 put_page(page);
3682         return page != NULL;
3683 }
3684 
3685 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3686                            pgoff_t idx)
3687 {
3688         struct inode *inode = mapping->host;
3689         struct hstate *h = hstate_inode(inode);
3690         int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3691 
3692         if (err)
3693                 return err;
3694         ClearPagePrivate(page);
3695 
3696         /*
3697          * set page dirty so that it will not be removed from cache/file
3698          * by non-hugetlbfs specific code paths.
3699          */
3700         set_page_dirty(page);
3701 
3702         spin_lock(&inode->i_lock);
3703         inode->i_blocks += blocks_per_huge_page(h);
3704         spin_unlock(&inode->i_lock);
3705         return 0;
3706 }
3707 
3708 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
3709                            struct address_space *mapping, pgoff_t idx,
3710                            unsigned long address, pte_t *ptep, unsigned int flags)
3711 {
3712         struct hstate *h = hstate_vma(vma);
3713         int ret = VM_FAULT_SIGBUS;
3714         int anon_rmap = 0;
3715         unsigned long size;
3716         struct page *page;
3717         pte_t new_pte;
3718         spinlock_t *ptl;
3719         unsigned long haddr = address & huge_page_mask(h);
3720 
3721         /*
3722          * Currently, we are forced to kill the process in the event the
3723          * original mapper has unmapped pages from the child due to a failed
3724          * COW. Warn that such a situation has occurred as it may not be obvious
3725          */
3726         if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3727                 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
3728                            current->pid);
3729                 return ret;
3730         }
3731 
3732         /*
3733          * Use page lock to guard against racing truncation
3734          * before we get page_table_lock.
3735          */
3736 retry:
3737         page = find_lock_page(mapping, idx);
3738         if (!page) {
3739                 size = i_size_read(mapping->host) >> huge_page_shift(h);
3740                 if (idx >= size)
3741                         goto out;
3742 
3743                 /*
3744                  * Check for page in userfault range
3745                  */
3746                 if (userfaultfd_missing(vma)) {
3747                         u32 hash;
3748                         struct vm_fault vmf = {
3749                                 .vma = vma,
3750                                 .address = haddr,
3751                                 .flags = flags,
3752                                 /*
3753                                  * Hard to debug if it ends up being
3754                                  * used by a callee that assumes
3755                                  * something about the other
3756                                  * uninitialized fields... same as in
3757                                  * memory.c
3758                                  */
3759                         };
3760 
3761                         /*
3762                          * hugetlb_fault_mutex must be dropped before
3763                          * handling userfault.  Reacquire after handling
3764                          * fault to make calling code simpler.
3765                          */
3766                         hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping,
3767                                                         idx, haddr);
3768                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3769                         ret = handle_userfault(&vmf, VM_UFFD_MISSING);
3770                         mutex_lock(&hugetlb_fault_mutex_table[hash]);
3771                         goto out;
3772                 }
3773 
3774                 page = alloc_huge_page(vma, haddr, 0);
3775                 if (IS_ERR(page)) {
3776                         ret = PTR_ERR(page);
3777                         if (ret == -ENOMEM)
3778                                 ret = VM_FAULT_OOM;
3779                         else
3780                                 ret = VM_FAULT_SIGBUS;
3781                         goto out;
3782                 }
3783                 clear_huge_page(page, address, pages_per_huge_page(h));
3784                 __SetPageUptodate(page);
3785                 set_page_huge_active(page);
3786 
3787                 if (vma->vm_flags & VM_MAYSHARE) {
3788                         int err = huge_add_to_page_cache(page, mapping, idx);
3789                         if (err) {
3790                                 put_page(page);
3791                                 if (err == -EEXIST)
3792                                         goto retry;
3793                                 goto out;
3794                         }
3795                 } else {
3796                         lock_page(page);
3797                         if (unlikely(anon_vma_prepare(vma))) {
3798                                 ret = VM_FAULT_OOM;
3799                                 goto backout_unlocked;
3800                         }
3801                         anon_rmap = 1;
3802                 }
3803         } else {
3804                 /*
3805                  * If memory error occurs between mmap() and fault, some process
3806                  * don't have hwpoisoned swap entry for errored virtual address.
3807                  * So we need to block hugepage fault by PG_hwpoison bit check.
3808                  */
3809                 if (unlikely(PageHWPoison(page))) {
3810                         ret = VM_FAULT_HWPOISON |
3811                                 VM_FAULT_SET_HINDEX(hstate_index(h));
3812                         goto backout_unlocked;
3813                 }
3814         }
3815 
3816         /*
3817          * If we are going to COW a private mapping later, we examine the
3818          * pending reservations for this page now. This will ensure that
3819          * any allocations necessary to record that reservation occur outside
3820          * the spinlock.
3821          */
3822         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3823                 if (vma_needs_reservation(h, vma, haddr) < 0) {
3824                         ret = VM_FAULT_OOM;
3825                         goto backout_unlocked;
3826                 }
3827                 /* Just decrements count, does not deallocate */
3828                 vma_end_reservation(h, vma, haddr);
3829         }
3830 
3831         ptl = huge_pte_lock(h, mm, ptep);
3832         size = i_size_read(mapping->host) >> huge_page_shift(h);
3833         if (idx >= size)
3834                 goto backout;
3835 
3836         ret = 0;
3837         if (!huge_pte_none(huge_ptep_get(ptep)))
3838                 goto backout;
3839 
3840         if (anon_rmap) {
3841                 ClearPagePrivate(page);
3842                 hugepage_add_new_anon_rmap(page, vma, haddr);
3843         } else
3844                 page_dup_rmap(page, true);
3845         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3846                                 && (vma->vm_flags & VM_SHARED)));
3847         set_huge_pte_at(mm, haddr, ptep, new_pte);
3848 
3849         hugetlb_count_add(pages_per_huge_page(h), mm);
3850         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3851                 /* Optimization, do the COW without a second fault */
3852                 ret = hugetlb_cow(mm, vma, haddr, ptep, page, ptl);
3853         }
3854 
3855         spin_unlock(ptl);
3856         unlock_page(page);
3857 out:
3858         return ret;
3859 
3860 backout:
3861         spin_unlock(ptl);
3862 backout_unlocked:
3863         unlock_page(page);
3864         restore_reserve_on_error(h, vma, haddr, page);
3865         put_page(page);
3866         goto out;
3867 }
3868 
3869 #ifdef CONFIG_SMP
3870 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3871                             struct vm_area_struct *vma,
3872                             struct address_space *mapping,
3873                             pgoff_t idx, unsigned long address)
3874 {
3875         unsigned long key[2];
3876         u32 hash;
3877 
3878         if (vma->vm_flags & VM_SHARED) {
3879                 key[0] = (unsigned long) mapping;
3880                 key[1] = idx;
3881         } else {
3882                 key[0] = (unsigned long) mm;
3883                 key[1] = address >> huge_page_shift(h);
3884         }
3885 
3886         hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3887 
3888         return hash & (num_fault_mutexes - 1);
3889 }
3890 #else
3891 /*
3892  * For uniprocesor systems we always use a single mutex, so just
3893  * return 0 and avoid the hashing overhead.
3894  */
3895 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3896                             struct vm_area_struct *vma,
3897                             struct address_space *mapping,
3898                             pgoff_t idx, unsigned long address)
3899 {
3900         return 0;
3901 }
3902 #endif
3903 
3904 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3905                         unsigned long address, unsigned int flags)
3906 {
3907         pte_t *ptep, entry;
3908         spinlock_t *ptl;
3909         int ret;
3910         u32 hash;
3911         pgoff_t idx;
3912         struct page *page = NULL;
3913         struct page *pagecache_page = NULL;
3914         struct hstate *h = hstate_vma(vma);
3915         struct address_space *mapping;
3916         int need_wait_lock = 0;
3917         unsigned long haddr = address & huge_page_mask(h);
3918 
3919         ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
3920         if (ptep) {
3921                 entry = huge_ptep_get(ptep);
3922                 if (unlikely(is_hugetlb_entry_migration(entry))) {
3923                         migration_entry_wait_huge(vma, mm, ptep);
3924                         return 0;
3925                 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
3926                         return VM_FAULT_HWPOISON_LARGE |
3927                                 VM_FAULT_SET_HINDEX(hstate_index(h));
3928         } else {
3929                 ptep = huge_pte_alloc(mm, haddr, huge_page_size(h));
3930                 if (!ptep)
3931                         return VM_FAULT_OOM;
3932         }
3933 
3934         mapping = vma->vm_file->f_mapping;
3935         idx = vma_hugecache_offset(h, vma, haddr);
3936 
3937         /*
3938          * Serialize hugepage allocation and instantiation, so that we don't
3939          * get spurious allocation failures if two CPUs race to instantiate
3940          * the same page in the page cache.
3941          */
3942         hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, haddr);
3943         mutex_lock(&hugetlb_fault_mutex_table[hash]);
3944 
3945         entry = huge_ptep_get(ptep);
3946         if (huge_pte_none(entry)) {
3947                 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
3948                 goto out_mutex;
3949         }
3950 
3951         ret = 0;
3952 
3953         /*
3954          * entry could be a migration/hwpoison entry at this point, so this
3955          * check prevents the kernel from going below assuming that we have
3956          * a active hugepage in pagecache. This goto expects the 2nd page fault,
3957          * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3958          * handle it.
3959          */
3960         if (!pte_present(entry))
3961                 goto out_mutex;
3962 
3963         /*
3964          * If we are going to COW the mapping later, we examine the pending
3965          * reservations for this page now. This will ensure that any
3966          * allocations necessary to record that reservation occur outside the
3967          * spinlock. For private mappings, we also lookup the pagecache
3968          * page now as it is used to determine if a reservation has been
3969          * consumed.
3970          */
3971         if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
3972                 if (vma_needs_reservation(h, vma, haddr) < 0) {
3973                         ret = VM_FAULT_OOM;
3974                         goto out_mutex;
3975                 }
3976                 /* Just decrements count, does not deallocate */
3977                 vma_end_reservation(h, vma, haddr);
3978 
3979                 if (!(vma->vm_flags & VM_MAYSHARE))
3980                         pagecache_page = hugetlbfs_pagecache_page(h,
3981                                                                 vma, haddr);
3982         }
3983 
3984         ptl = huge_pte_lock(h, mm, ptep);
3985 
3986         /* Check for a racing update before calling hugetlb_cow */
3987         if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
3988                 goto out_ptl;
3989 
3990         /*
3991          * hugetlb_cow() requires page locks of pte_page(entry) and
3992          * pagecache_page, so here we need take the former one
3993          * when page != pagecache_page or !pagecache_page.
3994          */
3995         page = pte_page(entry);
3996         if (page != pagecache_page)
3997                 if (!trylock_page(page)) {
3998                         need_wait_lock = 1;
3999                         goto out_ptl;
4000                 }
4001 
4002         get_page(page);
4003 
4004         if (flags & FAULT_FLAG_WRITE) {
4005                 if (!huge_pte_write(entry)) {
4006                         ret = hugetlb_cow(mm, vma, haddr, ptep,
4007                                           pagecache_page, ptl);
4008                         goto out_put_page;
4009                 }
4010                 entry = huge_pte_mkdirty(entry);
4011         }
4012         entry = pte_mkyoung(entry);
4013         if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
4014                                                 flags & FAULT_FLAG_WRITE))
4015                 update_mmu_cache(vma, haddr, ptep);
4016 out_put_page:
4017         if (page != pagecache_page)
4018                 unlock_page(page);
4019         put_page(page);
4020 out_ptl:
4021         spin_unlock(ptl);
4022 
4023         if (pagecache_page) {
4024                 unlock_page(pagecache_page);
4025                 put_page(pagecache_page);
4026         }
4027 out_mutex:
4028         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4029         /*
4030          * Generally it's safe to hold refcount during waiting page lock. But
4031          * here we just wait to defer the next page fault to avoid busy loop and
4032          * the page is not used after unlocked before returning from the current
4033          * page fault. So we are safe from accessing freed page, even if we wait
4034          * here without taking refcount.
4035          */
4036         if (need_wait_lock)
4037                 wait_on_page_locked(page);
4038         return ret;
4039 }
4040 
4041 /*
4042  * Used by userfaultfd UFFDIO_COPY.  Based on mcopy_atomic_pte with
4043  * modifications for huge pages.
4044  */
4045 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4046                             pte_t *dst_pte,
4047                             struct vm_area_struct *dst_vma,
4048                             unsigned long dst_addr,
4049                             unsigned long src_addr,
4050                             struct page **pagep)
4051 {
4052         struct address_space *mapping;
4053         pgoff_t idx;
4054         unsigned long size;
4055         int vm_shared = dst_vma->vm_flags & VM_SHARED;
4056         struct hstate *h = hstate_vma(dst_vma);
4057         pte_t _dst_pte;
4058         spinlock_t *ptl;
4059         int ret;
4060         struct page *page;
4061 
4062         if (!*pagep) {
4063                 ret = -ENOMEM;
4064                 page = alloc_huge_page(dst_vma, dst_addr, 0);
4065                 if (IS_ERR(page))
4066                         goto out;
4067 
4068                 ret = copy_huge_page_from_user(page,
4069                                                 (const void __user *) src_addr,
4070                                                 pages_per_huge_page(h), false);
4071 
4072                 /* fallback to copy_from_user outside mmap_sem */
4073                 if (unlikely(ret)) {
4074                         ret = -EFAULT;
4075                         *pagep = page;
4076                         /* don't free the page */
4077                         goto out;
4078                 }
4079         } else {
4080                 page = *pagep;
4081                 *pagep = NULL;
4082         }
4083 
4084         /*
4085          * The memory barrier inside __SetPageUptodate makes sure that
4086          * preceding stores to the page contents become visible before
4087          * the set_pte_at() write.
4088          */
4089         __SetPageUptodate(page);
4090         set_page_huge_active(page);
4091 
4092         mapping = dst_vma->vm_file->f_mapping;
4093         idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4094 
4095         /*
4096          * If shared, add to page cache
4097          */
4098         if (vm_shared) {
4099                 size = i_size_read(mapping->host) >> huge_page_shift(h);
4100                 ret = -EFAULT;
4101                 if (idx >= size)
4102                         goto out_release_nounlock;
4103 
4104                 /*
4105                  * Serialization between remove_inode_hugepages() and
4106                  * huge_add_to_page_cache() below happens through the
4107                  * hugetlb_fault_mutex_table that here must be hold by
4108                  * the caller.
4109                  */
4110                 ret = huge_add_to_page_cache(page, mapping, idx);
4111                 if (ret)
4112                         goto out_release_nounlock;
4113         }
4114 
4115         ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4116         spin_lock(ptl);
4117 
4118         /*
4119          * Recheck the i_size after holding PT lock to make sure not
4120          * to leave any page mapped (as page_mapped()) beyond the end
4121          * of the i_size (remove_inode_hugepages() is strict about
4122          * enforcing that). If we bail out here, we'll also leave a
4123          * page in the radix tree in the vm_shared case beyond the end
4124          * of the i_size, but remove_inode_hugepages() will take care
4125          * of it as soon as we drop the hugetlb_fault_mutex_table.
4126          */
4127         size = i_size_read(mapping->host) >> huge_page_shift(h);
4128         ret = -EFAULT;
4129         if (idx >= size)
4130                 goto out_release_unlock;
4131 
4132         ret = -EEXIST;
4133         if (!huge_pte_none(huge_ptep_get(dst_pte)))
4134                 goto out_release_unlock;
4135 
4136         if (vm_shared) {
4137                 page_dup_rmap(page, true);
4138         } else {
4139                 ClearPagePrivate(page);
4140                 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4141         }
4142 
4143         _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4144         if (dst_vma->vm_flags & VM_WRITE)
4145                 _dst_pte = huge_pte_mkdirty(_dst_pte);
4146         _dst_pte = pte_mkyoung(_dst_pte);
4147 
4148         set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4149 
4150         (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4151                                         dst_vma->vm_flags & VM_WRITE);
4152         hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4153 
4154         /* No need to invalidate - it was non-present before */
4155         update_mmu_cache(dst_vma, dst_addr, dst_pte);
4156 
4157         spin_unlock(ptl);
4158         if (vm_shared)
4159                 unlock_page(page);
4160         ret = 0;
4161 out:
4162         return ret;
4163 out_release_unlock:
4164         spin_unlock(ptl);
4165         if (vm_shared)
4166                 unlock_page(page);
4167 out_release_nounlock:
4168         put_page(page);
4169         goto out;
4170 }
4171 
4172 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4173                          struct page **pages, struct vm_area_struct **vmas,
4174                          unsigned long *position, unsigned long *nr_pages,
4175                          long i, unsigned int flags, int *nonblocking)
4176 {
4177         unsigned long pfn_offset;
4178         unsigned long vaddr = *position;
4179         unsigned long remainder = *nr_pages;
4180         struct hstate *h = hstate_vma(vma);
4181         int err = -EFAULT;
4182 
4183         while (vaddr < vma->vm_end && remainder) {
4184                 pte_t *pte;
4185                 spinlock_t *ptl = NULL;
4186                 int absent;
4187                 struct page *page;
4188 
4189                 /*
4190                  * If we have a pending SIGKILL, don't keep faulting pages and
4191                  * potentially allocating memory.
4192                  */
4193                 if (unlikely(fatal_signal_pending(current))) {
4194                         remainder = 0;
4195                         break;
4196                 }
4197 
4198                 /*
4199                  * Some archs (sparc64, sh*) have multiple pte_ts to
4200                  * each hugepage.  We have to make sure we get the
4201                  * first, for the page indexing below to work.
4202                  *
4203                  * Note that page table lock is not held when pte is null.
4204                  */
4205                 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4206                                       huge_page_size(h));
4207                 if (pte)
4208                         ptl = huge_pte_lock(h, mm, pte);
4209                 absent = !pte || huge_pte_none(huge_ptep_get(pte));
4210 
4211                 /*
4212                  * When coredumping, it suits get_dump_page if we just return
4213                  * an error where there's an empty slot with no huge pagecache
4214                  * to back it.  This way, we avoid allocating a hugepage, and
4215                  * the sparse dumpfile avoids allocating disk blocks, but its
4216                  * huge holes still show up with zeroes where they need to be.
4217                  */
4218                 if (absent && (flags & FOLL_DUMP) &&
4219                     !hugetlbfs_pagecache_present(h, vma, vaddr)) {
4220                         if (pte)
4221                                 spin_unlock(ptl);
4222                         remainder = 0;
4223                         break;
4224                 }
4225 
4226                 /*
4227                  * We need call hugetlb_fault for both hugepages under migration
4228                  * (in which case hugetlb_fault waits for the migration,) and
4229                  * hwpoisoned hugepages (in which case we need to prevent the
4230                  * caller from accessing to them.) In order to do this, we use
4231                  * here is_swap_pte instead of is_hugetlb_entry_migration and
4232                  * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4233                  * both cases, and because we can't follow correct pages
4234                  * directly from any kind of swap entries.
4235                  */
4236                 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4237                     ((flags & FOLL_WRITE) &&
4238                       !huge_pte_write(huge_ptep_get(pte)))) {
4239                         int ret;
4240                         unsigned int fault_flags = 0;
4241 
4242                         if (pte)
4243                                 spin_unlock(ptl);
4244                         if (flags & FOLL_WRITE)
4245                                 fault_flags |= FAULT_FLAG_WRITE;
4246                         if (nonblocking)
4247                                 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
4248                         if (flags & FOLL_NOWAIT)
4249                                 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4250                                         FAULT_FLAG_RETRY_NOWAIT;
4251                         if (flags & FOLL_TRIED) {
4252                                 VM_WARN_ON_ONCE(fault_flags &
4253                                                 FAULT_FLAG_ALLOW_RETRY);
4254                                 fault_flags |= FAULT_FLAG_TRIED;
4255                         }
4256                         ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4257                         if (ret & VM_FAULT_ERROR) {
4258                                 err = vm_fault_to_errno(ret, flags);
4259                                 remainder = 0;
4260                                 break;
4261                         }
4262                         if (ret & VM_FAULT_RETRY) {
4263                                 if (nonblocking)
4264                                         *nonblocking = 0;
4265                                 *nr_pages = 0;
4266                                 /*
4267                                  * VM_FAULT_RETRY must not return an
4268                                  * error, it will return zero
4269                                  * instead.
4270                                  *
4271                                  * No need to update "position" as the
4272                                  * caller will not check it after
4273                                  * *nr_pages is set to 0.
4274                                  */
4275                                 return i;
4276                         }
4277                         continue;
4278                 }
4279 
4280                 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4281                 page = pte_page(huge_ptep_get(pte));
4282 same_page:
4283                 if (pages) {
4284                         pages[i] = mem_map_offset(page, pfn_offset);
4285                         get_page(pages[i]);
4286                 }
4287 
4288                 if (vmas)
4289                         vmas[i] = vma;
4290 
4291                 vaddr += PAGE_SIZE;
4292                 ++pfn_offset;
4293                 --remainder;
4294                 ++i;
4295                 if (vaddr < vma->vm_end && remainder &&
4296                                 pfn_offset < pages_per_huge_page(h)) {
4297                         /*
4298                          * We use pfn_offset to avoid touching the pageframes
4299                          * of this compound page.
4300                          */
4301                         goto same_page;
4302                 }
4303                 spin_unlock(ptl);
4304         }
4305         *nr_pages = remainder;
4306         /*
4307          * setting position is actually required only if remainder is
4308          * not zero but it's faster not to add a "if (remainder)"
4309          * branch.
4310          */
4311         *position = vaddr;
4312 
4313         return i ? i : err;
4314 }
4315 
4316 #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4317 /*
4318  * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4319  * implement this.
4320  */
4321 #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4322 #endif
4323 
4324 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
4325                 unsigned long address, unsigned long end, pgprot_t newprot)
4326 {
4327         struct mm_struct *mm = vma->vm_mm;
4328         unsigned long start = address;
4329         pte_t *ptep;
4330         pte_t pte;
4331         struct hstate *h = hstate_vma(vma);
4332         unsigned long pages = 0;
4333 
4334         BUG_ON(address >= end);
4335         flush_cache_range(vma, address, end);
4336 
4337         mmu_notifier_invalidate_range_start(mm, start, end);
4338         i_mmap_lock_write(vma->vm_file->f_mapping);
4339         for (; address < end; address += huge_page_size(h)) {
4340                 spinlock_t *ptl;
4341                 ptep = huge_pte_offset(mm, address, huge_page_size(h));
4342                 if (!ptep)
4343                         continue;
4344                 ptl = huge_pte_lock(h, mm, ptep);
4345                 if (huge_pmd_unshare(mm, &address, ptep)) {
4346                         pages++;
4347                         spin_unlock(ptl);
4348                         continue;
4349                 }
4350                 pte = huge_ptep_get(ptep);
4351                 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
4352                         spin_unlock(ptl);
4353                         continue;
4354                 }
4355                 if (unlikely(is_hugetlb_entry_migration(pte))) {
4356                         swp_entry_t entry = pte_to_swp_entry(pte);
4357 
4358                         if (is_write_migration_entry(entry)) {
4359                                 pte_t newpte;
4360 
4361                                 make_migration_entry_read(&entry);
4362                                 newpte = swp_entry_to_pte(entry);
4363                                 set_huge_swap_pte_at(mm, address, ptep,
4364                                                      newpte, huge_page_size(h));
4365                                 pages++;
4366                         }
4367                         spin_unlock(ptl);
4368                         continue;
4369                 }
4370                 if (!huge_pte_none(pte)) {
4371                         pte = huge_ptep_get_and_clear(mm, address, ptep);
4372                         pte = pte_mkhuge(huge_pte_modify(pte, newprot));
4373                         pte = arch_make_huge_pte(pte, vma, NULL, 0);
4374                         set_huge_pte_at(mm, address, ptep, pte);
4375                         pages++;
4376                 }
4377                 spin_unlock(ptl);
4378         }
4379         /*
4380          * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
4381          * may have cleared our pud entry and done put_page on the page table:
4382          * once we release i_mmap_rwsem, another task can do the final put_page
4383          * and that page table be reused and filled with junk.
4384          */
4385         flush_hugetlb_tlb_range(vma, start, end);
4386         /*
4387          * No need to call mmu_notifier_invalidate_range() we are downgrading
4388          * page table protection not changing it to point to a new page.
4389          *
4390          * See Documentation/vm/mmu_notifier.rst
4391          */
4392         i_mmap_unlock_write(vma->vm_file->f_mapping);
4393         mmu_notifier_invalidate_range_end(mm, start, end);
4394 
4395         return pages << h->order;
4396 }
4397 
4398 int hugetlb_reserve_pages(struct inode *inode,
4399                                         long from, long to,
4400                                         struct vm_area_struct *vma,
4401                                         vm_flags_t vm_flags)
4402 {
4403         long ret, chg;
4404         struct hstate *h = hstate_inode(inode);
4405         struct hugepage_subpool *spool = subpool_inode(inode);
4406         struct resv_map *resv_map;
4407         long gbl_reserve;
4408 
4409         /* This should never happen */
4410         if (from > to) {
4411                 VM_WARN(1, "%s called with a negative range\n", __func__);
4412                 return -EINVAL;
4413         }
4414 
4415         /*
4416          * Only apply hugepage reservation if asked. At fault time, an
4417          * attempt will be made for VM_NORESERVE to allocate a page
4418          * without using reserves
4419          */
4420         if (vm_flags & VM_NORESERVE)
4421                 return 0;
4422 
4423         /*
4424          * Shared mappings base their reservation on the number of pages that
4425          * are already allocated on behalf of the file. Private mappings need
4426          * to reserve the full area even if read-only as mprotect() may be
4427          * called to make the mapping read-write. Assume !vma is a shm mapping
4428          */
4429         if (!vma || vma->vm_flags & VM_MAYSHARE) {
4430                 resv_map = inode_resv_map(inode);
4431 
4432                 chg = region_chg(resv_map, from, to);
4433 
4434         } else {
4435                 resv_map = resv_map_alloc();
4436                 if (!resv_map)
4437                         return -ENOMEM;
4438 
4439                 chg = to - from;
4440 
4441                 set_vma_resv_map(vma, resv_map);
4442                 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4443         }
4444 
4445         if (chg < 0) {
4446                 ret = chg;
4447                 goto out_err;
4448         }
4449 
4450         /*
4451          * There must be enough pages in the subpool for the mapping. If
4452          * the subpool has a minimum size, there may be some global
4453          * reservations already in place (gbl_reserve).
4454          */
4455         gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4456         if (gbl_reserve < 0) {
4457                 ret = -ENOSPC;
4458                 goto out_err;
4459         }
4460 
4461         /*
4462          * Check enough hugepages are available for the reservation.
4463          * Hand the pages back to the subpool if there are not
4464          */
4465         ret = hugetlb_acct_memory(h, gbl_reserve);
4466         if (ret < 0) {
4467                 /* put back original number of pages, chg */
4468                 (void)hugepage_subpool_put_pages(spool, chg);
4469                 goto out_err;
4470         }
4471 
4472         /*
4473          * Account for the reservations made. Shared mappings record regions
4474          * that have reservations as they are shared by multiple VMAs.
4475          * When the last VMA disappears, the region map says how much
4476          * the reservation was and the page cache tells how much of
4477          * the reservation was consumed. Private mappings are per-VMA and
4478          * only the consumed reservations are tracked. When the VMA
4479          * disappears, the original reservation is the VMA size and the
4480          * consumed reservations are stored in the map. Hence, nothing
4481          * else has to be done for private mappings here
4482          */
4483         if (!vma || vma->vm_flags & VM_MAYSHARE) {
4484                 long add = region_add(resv_map, from, to);
4485 
4486                 if (unlikely(chg > add)) {
4487                         /*
4488                          * pages in this range were added to the reserve
4489                          * map between region_chg and region_add.  This
4490                          * indicates a race with alloc_huge_page.  Adjust
4491                          * the subpool and reserve counts modified above
4492                          * based on the difference.
4493                          */
4494                         long rsv_adjust;
4495 
4496                         rsv_adjust = hugepage_subpool_put_pages(spool,
4497                                                                 chg - add);
4498                         hugetlb_acct_memory(h, -rsv_adjust);
4499                 }
4500         }
4501         return 0;
4502 out_err:
4503         if (!vma || vma->vm_flags & VM_MAYSHARE)
4504                 /* Don't call region_abort if region_chg failed */
4505                 if (chg >= 0)
4506                         region_abort(resv_map, from, to);
4507         if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4508                 kref_put(&resv_map->refs, resv_map_release);
4509         return ret;
4510 }
4511 
4512 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4513                                                                 long freed)
4514 {
4515         struct hstate *h = hstate_inode(inode);
4516         struct resv_map *resv_map = inode_resv_map(inode);
4517         long chg = 0;
4518         struct hugepage_subpool *spool = subpool_inode(inode);
4519         long gbl_reserve;
4520 
4521         if (resv_map) {
4522                 chg = region_del(resv_map, start, end);
4523                 /*
4524                  * region_del() can fail in the rare case where a region
4525                  * must be split and another region descriptor can not be
4526                  * allocated.  If end == LONG_MAX, it will not fail.
4527                  */
4528                 if (chg < 0)
4529                         return chg;
4530         }
4531 
4532         spin_lock(&inode->i_lock);
4533         inode->i_blocks -= (blocks_per_huge_page(h) * freed);
4534         spin_unlock(&inode->i_lock);
4535 
4536         /*
4537          * If the subpool has a minimum size, the number of global
4538          * reservations to be released may be adjusted.
4539          */
4540         gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4541         hugetlb_acct_memory(h, -gbl_reserve);
4542 
4543         return 0;
4544 }
4545 
4546 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4547 static unsigned long page_table_shareable(struct vm_area_struct *svma,
4548                                 struct vm_area_struct *vma,
4549                                 unsigned long addr, pgoff_t idx)
4550 {
4551         unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4552                                 svma->vm_start;
4553         unsigned long sbase = saddr & PUD_MASK;
4554         unsigned long s_end = sbase + PUD_SIZE;
4555 
4556         /* Allow segments to share if only one is marked locked */
4557         unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4558         unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
4559 
4560         /*
4561          * match the virtual addresses, permission and the alignment of the
4562          * page table page.
4563          */
4564         if (pmd_index(addr) != pmd_index(saddr) ||
4565             vm_flags != svm_flags ||
4566             sbase < svma->vm_start || svma->vm_end < s_end)
4567                 return 0;
4568 
4569         return saddr;
4570 }
4571 
4572 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
4573 {
4574         unsigned long base = addr & PUD_MASK;
4575         unsigned long end = base + PUD_SIZE;
4576 
4577         /*
4578          * check on proper vm_flags and page table alignment
4579          */
4580         if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
4581                 return true;
4582         return false;
4583 }
4584 
4585 /*
4586  * Determine if start,end range within vma could be mapped by shared pmd.
4587  * If yes, adjust start and end to cover range associated with possible
4588  * shared pmd mappings.
4589  */
4590 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
4591                                 unsigned long *start, unsigned long *end)
4592 {
4593         unsigned long check_addr = *start;
4594 
4595         if (!(vma->vm_flags & VM_MAYSHARE))
4596                 return;
4597 
4598         for (check_addr = *start; check_addr < *end; check_addr += PUD_SIZE) {
4599                 unsigned long a_start = check_addr & PUD_MASK;
4600                 unsigned long a_end = a_start + PUD_SIZE;
4601 
4602                 /*
4603                  * If sharing is possible, adjust start/end if necessary.
4604                  */
4605                 if (range_in_vma(vma, a_start, a_end)) {
4606                         if (a_start < *start)
4607                                 *start = a_start;
4608                         if (a_end > *end)
4609                                 *end = a_end;
4610                 }
4611         }
4612 }
4613 
4614 /*
4615  * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4616  * and returns the corresponding pte. While this is not necessary for the
4617  * !shared pmd case because we can allocate the pmd later as well, it makes the
4618  * code much cleaner. pmd allocation is essential for the shared case because
4619  * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4620  * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4621  * bad pmd for sharing.
4622  */
4623 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4624 {
4625         struct vm_area_struct *vma = find_vma(mm, addr);
4626         struct address_space *mapping = vma->vm_file->f_mapping;
4627         pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4628                         vma->vm_pgoff;
4629         struct vm_area_struct *svma;
4630         unsigned long saddr;
4631         pte_t *spte = NULL;
4632         pte_t *pte;
4633         spinlock_t *ptl;
4634 
4635         if (!vma_shareable(vma, addr))
4636                 return (pte_t *)pmd_alloc(mm, pud, addr);
4637 
4638         i_mmap_lock_write(mapping);
4639         vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4640                 if (svma == vma)
4641                         continue;
4642 
4643                 saddr = page_table_shareable(svma, vma, addr, idx);
4644                 if (saddr) {
4645                         spte = huge_pte_offset(svma->vm_mm, saddr,
4646                                                vma_mmu_pagesize(svma));
4647                         if (spte) {
4648                                 get_page(virt_to_page(spte));
4649                                 break;
4650                         }
4651                 }
4652         }
4653 
4654         if (!spte)
4655                 goto out;
4656 
4657         ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
4658         if (pud_none(*pud)) {
4659                 pud_populate(mm, pud,
4660                                 (pmd_t *)((unsigned long)spte & PAGE_MASK));
4661                 mm_inc_nr_pmds(mm);
4662         } else {
4663                 put_page(virt_to_page(spte));
4664         }
4665         spin_unlock(ptl);
4666 out:
4667         pte = (pte_t *)pmd_alloc(mm, pud, addr);
4668         i_mmap_unlock_write(mapping);
4669         return pte;
4670 }
4671 
4672 /*
4673  * unmap huge page backed by shared pte.
4674  *
4675  * Hugetlb pte page is ref counted at the time of mapping.  If pte is shared
4676  * indicated by page_count > 1, unmap is achieved by clearing pud and
4677  * decrementing the ref count. If count == 1, the pte page is not shared.
4678  *
4679  * called with page table lock held.
4680  *
4681  * returns: 1 successfully unmapped a shared pte page
4682  *          0 the underlying pte page is not shared, or it is the last user
4683  */
4684 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4685 {
4686         pgd_t *pgd = pgd_offset(mm, *addr);
4687         p4d_t *p4d = p4d_offset(pgd, *addr);
4688         pud_t *pud = pud_offset(p4d, *addr);
4689 
4690         BUG_ON(page_count(virt_to_page(ptep)) == 0);
4691         if (page_count(virt_to_page(ptep)) == 1)
4692                 return 0;
4693 
4694         pud_clear(pud);
4695         put_page(virt_to_page(ptep));
4696         mm_dec_nr_pmds(mm);
4697         *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4698         return 1;
4699 }
4700 #define want_pmd_share()        (1)
4701 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4702 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4703 {
4704         return NULL;
4705 }
4706 
4707 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4708 {
4709         return 0;
4710 }
4711 
4712 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
4713                                 unsigned long *start, unsigned long *end)
4714 {
4715 }
4716 #define want_pmd_share()        (0)
4717 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4718 
4719 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4720 pte_t *huge_pte_alloc(struct mm_struct *mm,
4721                         unsigned long addr, unsigned long sz)
4722 {
4723         pgd_t *pgd;
4724         p4d_t *p4d;
4725         pud_t *pud;
4726         pte_t *pte = NULL;
4727 
4728         pgd = pgd_offset(mm, addr);
4729         p4d = p4d_alloc(mm, pgd, addr);
4730         if (!p4d)
4731                 return NULL;
4732         pud = pud_alloc(mm, p4d, addr);
4733         if (pud) {
4734                 if (sz == PUD_SIZE) {
4735                         pte = (pte_t *)pud;
4736                 } else {
4737                         BUG_ON(sz != PMD_SIZE);
4738                         if (want_pmd_share() && pud_none(*pud))
4739                                 pte = huge_pmd_share(mm, addr, pud);
4740                         else
4741                                 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4742                 }
4743         }
4744         BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
4745 
4746         return pte;
4747 }
4748 
4749 /*
4750  * huge_pte_offset() - Walk the page table to resolve the hugepage
4751  * entry at address @addr
4752  *
4753  * Return: Pointer to page table or swap entry (PUD or PMD) for
4754  * address @addr, or NULL if a p*d_none() entry is encountered and the
4755  * size @sz doesn't match the hugepage size at this level of the page
4756  * table.
4757  */
4758 pte_t *huge_pte_offset(struct mm_struct *mm,
4759                        unsigned long addr, unsigned long sz)
4760 {
4761         pgd_t *pgd;
4762         p4d_t *p4d;
4763         pud_t *pud;
4764         pmd_t *pmd;
4765 
4766         pgd = pgd_offset(mm, addr);
4767         if (!pgd_present(*pgd))
4768                 return NULL;
4769         p4d = p4d_offset(pgd, addr);
4770         if (!p4d_present(*p4d))
4771                 return NULL;
4772 
4773         pud = pud_offset(p4d, addr);
4774         if (sz != PUD_SIZE && pud_none(*pud))
4775                 return NULL;
4776         /* hugepage or swap? */
4777         if (pud_huge(*pud) || !pud_present(*pud))
4778                 return (pte_t *)pud;
4779 
4780         pmd = pmd_offset(pud, addr);
4781         if (sz != PMD_SIZE && pmd_none(*pmd))
4782                 return NULL;
4783         /* hugepage or swap? */
4784         if (pmd_huge(*pmd) || !pmd_present(*pmd))
4785                 return (pte_t *)pmd;
4786 
4787         return NULL;
4788 }
4789 
4790 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4791 
4792 /*
4793  * These functions are overwritable if your architecture needs its own
4794  * behavior.
4795  */
4796 struct page * __weak
4797 follow_huge_addr(struct mm_struct *mm, unsigned long address,
4798                               int write)
4799 {
4800         return ERR_PTR(-EINVAL);
4801 }
4802 
4803 struct page * __weak
4804 follow_huge_pd(struct vm_area_struct *vma,
4805                unsigned long address, hugepd_t hpd, int flags, int pdshift)
4806 {
4807         WARN(1, "hugepd follow called with no support for hugepage directory format\n");
4808         return NULL;
4809 }
4810 
4811 struct page * __weak
4812 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
4813                 pmd_t *pmd, int flags)
4814 {
4815         struct page *page = NULL;
4816         spinlock_t *ptl;
4817         pte_t pte;
4818 retry:
4819         ptl = pmd_lockptr(mm, pmd);
4820         spin_lock(ptl);
4821         /*
4822          * make sure that the address range covered by this pmd is not
4823          * unmapped from other threads.
4824          */
4825         if (!pmd_huge(*pmd))
4826                 goto out;
4827         pte = huge_ptep_get((pte_t *)pmd);
4828         if (pte_present(pte)) {
4829                 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
4830                 if (flags & FOLL_GET)
4831                         get_page(page);
4832         } else {
4833                 if (is_hugetlb_entry_migration(pte)) {
4834                         spin_unlock(ptl);
4835                         __migration_entry_wait(mm, (pte_t *)pmd, ptl);
4836                         goto retry;
4837                 }
4838                 /*
4839                  * hwpoisoned entry is treated as no_page_table in
4840                  * follow_page_mask().
4841                  */
4842         }
4843 out:
4844         spin_unlock(ptl);
4845         return page;
4846 }
4847 
4848 struct page * __weak
4849 follow_huge_pud(struct mm_struct *mm, unsigned long address,
4850                 pud_t *pud, int flags)
4851 {
4852         if (flags & FOLL_GET)
4853                 return NULL;
4854 
4855         return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
4856 }
4857 
4858 struct page * __weak
4859 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
4860 {
4861         if (flags & FOLL_GET)
4862                 return NULL;
4863 
4864         return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
4865 }
4866 
4867 bool isolate_huge_page(struct page *page, struct list_head *list)
4868 {
4869         bool ret = true;
4870 
4871         VM_BUG_ON_PAGE(!PageHead(page), page);
4872         spin_lock(&hugetlb_lock);
4873         if (!page_huge_active(page) || !get_page_unless_zero(page)) {
4874                 ret = false;
4875                 goto unlock;
4876         }
4877         clear_page_huge_active(page);
4878         list_move_tail(&page->lru, list);
4879 unlock:
4880         spin_unlock(&hugetlb_lock);
4881         return ret;
4882 }
4883 
4884 void putback_active_hugepage(struct page *page)
4885 {
4886         VM_BUG_ON_PAGE(!PageHead(page), page);
4887         spin_lock(&hugetlb_lock);
4888         set_page_huge_active(page);
4889         list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
4890         spin_unlock(&hugetlb_lock);
4891         put_page(page);
4892 }
4893 
4894 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
4895 {
4896         struct hstate *h = page_hstate(oldpage);
4897 
4898         hugetlb_cgroup_migrate(oldpage, newpage);
4899         set_page_owner_migrate_reason(newpage, reason);
4900 
4901         /*
4902          * transfer temporary state of the new huge page. This is
4903          * reverse to other transitions because the newpage is going to
4904          * be final while the old one will be freed so it takes over
4905          * the temporary status.
4906          *
4907          * Also note that we have to transfer the per-node surplus state
4908          * here as well otherwise the global surplus count will not match
4909          * the per-node's.
4910          */
4911         if (PageHugeTemporary(newpage)) {
4912                 int old_nid = page_to_nid(oldpage);
4913                 int new_nid = page_to_nid(newpage);
4914 
4915                 SetPageHugeTemporary(oldpage);
4916                 ClearPageHugeTemporary(newpage);
4917 
4918                 spin_lock(&hugetlb_lock);
4919                 if (h->surplus_huge_pages_node[old_nid]) {
4920                         h->surplus_huge_pages_node[old_nid]--;
4921                         h->surplus_huge_pages_node[new_nid]++;
4922                 }
4923                 spin_unlock(&hugetlb_lock);
4924         }
4925 }
4926 

~ [ 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