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

TOMOYO Linux Cross Reference
Linux/mm/hugetlb.c

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

  1 /*
  2  * Generic hugetlb support.
  3  * (C) Nadia Yvette Chambers, April 2004
  4  */
  5 #include <linux/list.h>
  6 #include <linux/init.h>
  7 #include <linux/module.h>
  8 #include <linux/mm.h>
  9 #include <linux/seq_file.h>
 10 #include <linux/sysctl.h>
 11 #include <linux/highmem.h>
 12 #include <linux/mmu_notifier.h>
 13 #include <linux/nodemask.h>
 14 #include <linux/pagemap.h>
 15 #include <linux/mempolicy.h>
 16 #include <linux/compiler.h>
 17 #include <linux/cpuset.h>
 18 #include <linux/mutex.h>
 19 #include <linux/bootmem.h>
 20 #include <linux/sysfs.h>
 21 #include <linux/slab.h>
 22 #include <linux/rmap.h>
 23 #include <linux/swap.h>
 24 #include <linux/swapops.h>
 25 #include <linux/page-isolation.h>
 26 #include <linux/jhash.h>
 27 
 28 #include <asm/page.h>
 29 #include <asm/pgtable.h>
 30 #include <asm/tlb.h>
 31 
 32 #include <linux/io.h>
 33 #include <linux/hugetlb.h>
 34 #include <linux/hugetlb_cgroup.h>
 35 #include <linux/node.h>
 36 #include "internal.h"
 37 
 38 int hugepages_treat_as_movable;
 39 
 40 int hugetlb_max_hstate __read_mostly;
 41 unsigned int default_hstate_idx;
 42 struct hstate hstates[HUGE_MAX_HSTATE];
 43 /*
 44  * Minimum page order among possible hugepage sizes, set to a proper value
 45  * at boot time.
 46  */
 47 static unsigned int minimum_order __read_mostly = UINT_MAX;
 48 
 49 __initdata LIST_HEAD(huge_boot_pages);
 50 
 51 /* for command line parsing */
 52 static struct hstate * __initdata parsed_hstate;
 53 static unsigned long __initdata default_hstate_max_huge_pages;
 54 static unsigned long __initdata default_hstate_size;
 55 
 56 /*
 57  * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
 58  * free_huge_pages, and surplus_huge_pages.
 59  */
 60 DEFINE_SPINLOCK(hugetlb_lock);
 61 
 62 /*
 63  * Serializes faults on the same logical page.  This is used to
 64  * prevent spurious OOMs when the hugepage pool is fully utilized.
 65  */
 66 static int num_fault_mutexes;
 67 static struct mutex *htlb_fault_mutex_table ____cacheline_aligned_in_smp;
 68 
 69 /* Forward declaration */
 70 static int hugetlb_acct_memory(struct hstate *h, long delta);
 71 
 72 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
 73 {
 74         bool free = (spool->count == 0) && (spool->used_hpages == 0);
 75 
 76         spin_unlock(&spool->lock);
 77 
 78         /* If no pages are used, and no other handles to the subpool
 79          * remain, give up any reservations mased on minimum size and
 80          * free the subpool */
 81         if (free) {
 82                 if (spool->min_hpages != -1)
 83                         hugetlb_acct_memory(spool->hstate,
 84                                                 -spool->min_hpages);
 85                 kfree(spool);
 86         }
 87 }
 88 
 89 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
 90                                                 long min_hpages)
 91 {
 92         struct hugepage_subpool *spool;
 93 
 94         spool = kzalloc(sizeof(*spool), GFP_KERNEL);
 95         if (!spool)
 96                 return NULL;
 97 
 98         spin_lock_init(&spool->lock);
 99         spool->count = 1;
100         spool->max_hpages = max_hpages;
101         spool->hstate = h;
102         spool->min_hpages = min_hpages;
103 
104         if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
105                 kfree(spool);
106                 return NULL;
107         }
108         spool->rsv_hpages = min_hpages;
109 
110         return spool;
111 }
112 
113 void hugepage_put_subpool(struct hugepage_subpool *spool)
114 {
115         spin_lock(&spool->lock);
116         BUG_ON(!spool->count);
117         spool->count--;
118         unlock_or_release_subpool(spool);
119 }
120 
121 /*
122  * Subpool accounting for allocating and reserving pages.
123  * Return -ENOMEM if there are not enough resources to satisfy the
124  * the request.  Otherwise, return the number of pages by which the
125  * global pools must be adjusted (upward).  The returned value may
126  * only be different than the passed value (delta) in the case where
127  * a subpool minimum size must be manitained.
128  */
129 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
130                                       long delta)
131 {
132         long ret = delta;
133 
134         if (!spool)
135                 return ret;
136 
137         spin_lock(&spool->lock);
138 
139         if (spool->max_hpages != -1) {          /* maximum size accounting */
140                 if ((spool->used_hpages + delta) <= spool->max_hpages)
141                         spool->used_hpages += delta;
142                 else {
143                         ret = -ENOMEM;
144                         goto unlock_ret;
145                 }
146         }
147 
148         if (spool->min_hpages != -1) {          /* minimum size accounting */
149                 if (delta > spool->rsv_hpages) {
150                         /*
151                          * Asking for more reserves than those already taken on
152                          * behalf of subpool.  Return difference.
153                          */
154                         ret = delta - spool->rsv_hpages;
155                         spool->rsv_hpages = 0;
156                 } else {
157                         ret = 0;        /* reserves already accounted for */
158                         spool->rsv_hpages -= delta;
159                 }
160         }
161 
162 unlock_ret:
163         spin_unlock(&spool->lock);
164         return ret;
165 }
166 
167 /*
168  * Subpool accounting for freeing and unreserving pages.
169  * Return the number of global page reservations that must be dropped.
170  * The return value may only be different than the passed value (delta)
171  * in the case where a subpool minimum size must be maintained.
172  */
173 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
174                                        long delta)
175 {
176         long ret = delta;
177 
178         if (!spool)
179                 return delta;
180 
181         spin_lock(&spool->lock);
182 
183         if (spool->max_hpages != -1)            /* maximum size accounting */
184                 spool->used_hpages -= delta;
185 
186         if (spool->min_hpages != -1) {          /* minimum size accounting */
187                 if (spool->rsv_hpages + delta <= spool->min_hpages)
188                         ret = 0;
189                 else
190                         ret = spool->rsv_hpages + delta - spool->min_hpages;
191 
192                 spool->rsv_hpages += delta;
193                 if (spool->rsv_hpages > spool->min_hpages)
194                         spool->rsv_hpages = spool->min_hpages;
195         }
196 
197         /*
198          * If hugetlbfs_put_super couldn't free spool due to an outstanding
199          * quota reference, free it now.
200          */
201         unlock_or_release_subpool(spool);
202 
203         return ret;
204 }
205 
206 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
207 {
208         return HUGETLBFS_SB(inode->i_sb)->spool;
209 }
210 
211 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
212 {
213         return subpool_inode(file_inode(vma->vm_file));
214 }
215 
216 /*
217  * Region tracking -- allows tracking of reservations and instantiated pages
218  *                    across the pages in a mapping.
219  *
220  * The region data structures are embedded into a resv_map and protected
221  * by a resv_map's lock.  The set of regions within the resv_map represent
222  * reservations for huge pages, or huge pages that have already been
223  * instantiated within the map.  The from and to elements are huge page
224  * indicies into the associated mapping.  from indicates the starting index
225  * of the region.  to represents the first index past the end of  the region.
226  *
227  * For example, a file region structure with from == 0 and to == 4 represents
228  * four huge pages in a mapping.  It is important to note that the to element
229  * represents the first element past the end of the region. This is used in
230  * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
231  *
232  * Interval notation of the form [from, to) will be used to indicate that
233  * the endpoint from is inclusive and to is exclusive.
234  */
235 struct file_region {
236         struct list_head link;
237         long from;
238         long to;
239 };
240 
241 /*
242  * Add the huge page range represented by [f, t) to the reserve
243  * map.  Existing regions will be expanded to accommodate the
244  * specified range.  We know only existing regions need to be
245  * expanded, because region_add is only called after region_chg
246  * with the same range.  If a new file_region structure must
247  * be allocated, it is done in region_chg.
248  *
249  * Return the number of new huge pages added to the map.  This
250  * number is greater than or equal to zero.
251  */
252 static long region_add(struct resv_map *resv, long f, long t)
253 {
254         struct list_head *head = &resv->regions;
255         struct file_region *rg, *nrg, *trg;
256         long add = 0;
257 
258         spin_lock(&resv->lock);
259         /* Locate the region we are either in or before. */
260         list_for_each_entry(rg, head, link)
261                 if (f <= rg->to)
262                         break;
263 
264         /* Round our left edge to the current segment if it encloses us. */
265         if (f > rg->from)
266                 f = rg->from;
267 
268         /* Check for and consume any regions we now overlap with. */
269         nrg = rg;
270         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
271                 if (&rg->link == head)
272                         break;
273                 if (rg->from > t)
274                         break;
275 
276                 /* If this area reaches higher then extend our area to
277                  * include it completely.  If this is not the first area
278                  * which we intend to reuse, free it. */
279                 if (rg->to > t)
280                         t = rg->to;
281                 if (rg != nrg) {
282                         /* Decrement return value by the deleted range.
283                          * Another range will span this area so that by
284                          * end of routine add will be >= zero
285                          */
286                         add -= (rg->to - rg->from);
287                         list_del(&rg->link);
288                         kfree(rg);
289                 }
290         }
291 
292         add += (nrg->from - f);         /* Added to beginning of region */
293         nrg->from = f;
294         add += t - nrg->to;             /* Added to end of region */
295         nrg->to = t;
296 
297         spin_unlock(&resv->lock);
298         VM_BUG_ON(add < 0);
299         return add;
300 }
301 
302 /*
303  * Examine the existing reserve map and determine how many
304  * huge pages in the specified range [f, t) are NOT currently
305  * represented.  This routine is called before a subsequent
306  * call to region_add that will actually modify the reserve
307  * map to add the specified range [f, t).  region_chg does
308  * not change the number of huge pages represented by the
309  * map.  However, if the existing regions in the map can not
310  * be expanded to represent the new range, a new file_region
311  * structure is added to the map as a placeholder.  This is
312  * so that the subsequent region_add call will have all the
313  * regions it needs and will not fail.
314  *
315  * Returns the number of huge pages that need to be added
316  * to the existing reservation map for the range [f, t).
317  * This number is greater or equal to zero.  -ENOMEM is
318  * returned if a new file_region structure is needed and can
319  * not be allocated.
320  */
321 static long region_chg(struct resv_map *resv, long f, long t)
322 {
323         struct list_head *head = &resv->regions;
324         struct file_region *rg, *nrg = NULL;
325         long chg = 0;
326 
327 retry:
328         spin_lock(&resv->lock);
329         /* Locate the region we are before or in. */
330         list_for_each_entry(rg, head, link)
331                 if (f <= rg->to)
332                         break;
333 
334         /* If we are below the current region then a new region is required.
335          * Subtle, allocate a new region at the position but make it zero
336          * size such that we can guarantee to record the reservation. */
337         if (&rg->link == head || t < rg->from) {
338                 if (!nrg) {
339                         spin_unlock(&resv->lock);
340                         nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
341                         if (!nrg)
342                                 return -ENOMEM;
343 
344                         nrg->from = f;
345                         nrg->to   = f;
346                         INIT_LIST_HEAD(&nrg->link);
347                         goto retry;
348                 }
349 
350                 list_add(&nrg->link, rg->link.prev);
351                 chg = t - f;
352                 goto out_nrg;
353         }
354 
355         /* Round our left edge to the current segment if it encloses us. */
356         if (f > rg->from)
357                 f = rg->from;
358         chg = t - f;
359 
360         /* Check for and consume any regions we now overlap with. */
361         list_for_each_entry(rg, rg->link.prev, link) {
362                 if (&rg->link == head)
363                         break;
364                 if (rg->from > t)
365                         goto out;
366 
367                 /* We overlap with this area, if it extends further than
368                  * us then we must extend ourselves.  Account for its
369                  * existing reservation. */
370                 if (rg->to > t) {
371                         chg += rg->to - t;
372                         t = rg->to;
373                 }
374                 chg -= rg->to - rg->from;
375         }
376 
377 out:
378         spin_unlock(&resv->lock);
379         /*  We already know we raced and no longer need the new region */
380         kfree(nrg);
381         return chg;
382 out_nrg:
383         spin_unlock(&resv->lock);
384         return chg;
385 }
386 
387 /*
388  * Truncate the reserve map at index 'end'.  Modify/truncate any
389  * region which contains end.  Delete any regions past end.
390  * Return the number of huge pages removed from the map.
391  */
392 static long region_truncate(struct resv_map *resv, long end)
393 {
394         struct list_head *head = &resv->regions;
395         struct file_region *rg, *trg;
396         long chg = 0;
397 
398         spin_lock(&resv->lock);
399         /* Locate the region we are either in or before. */
400         list_for_each_entry(rg, head, link)
401                 if (end <= rg->to)
402                         break;
403         if (&rg->link == head)
404                 goto out;
405 
406         /* If we are in the middle of a region then adjust it. */
407         if (end > rg->from) {
408                 chg = rg->to - end;
409                 rg->to = end;
410                 rg = list_entry(rg->link.next, typeof(*rg), link);
411         }
412 
413         /* Drop any remaining regions. */
414         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
415                 if (&rg->link == head)
416                         break;
417                 chg += rg->to - rg->from;
418                 list_del(&rg->link);
419                 kfree(rg);
420         }
421 
422 out:
423         spin_unlock(&resv->lock);
424         return chg;
425 }
426 
427 /*
428  * Count and return the number of huge pages in the reserve map
429  * that intersect with the range [f, t).
430  */
431 static long region_count(struct resv_map *resv, long f, long t)
432 {
433         struct list_head *head = &resv->regions;
434         struct file_region *rg;
435         long chg = 0;
436 
437         spin_lock(&resv->lock);
438         /* Locate each segment we overlap with, and count that overlap. */
439         list_for_each_entry(rg, head, link) {
440                 long seg_from;
441                 long seg_to;
442 
443                 if (rg->to <= f)
444                         continue;
445                 if (rg->from >= t)
446                         break;
447 
448                 seg_from = max(rg->from, f);
449                 seg_to = min(rg->to, t);
450 
451                 chg += seg_to - seg_from;
452         }
453         spin_unlock(&resv->lock);
454 
455         return chg;
456 }
457 
458 /*
459  * Convert the address within this vma to the page offset within
460  * the mapping, in pagecache page units; huge pages here.
461  */
462 static pgoff_t vma_hugecache_offset(struct hstate *h,
463                         struct vm_area_struct *vma, unsigned long address)
464 {
465         return ((address - vma->vm_start) >> huge_page_shift(h)) +
466                         (vma->vm_pgoff >> huge_page_order(h));
467 }
468 
469 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
470                                      unsigned long address)
471 {
472         return vma_hugecache_offset(hstate_vma(vma), vma, address);
473 }
474 
475 /*
476  * Return the size of the pages allocated when backing a VMA. In the majority
477  * cases this will be same size as used by the page table entries.
478  */
479 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
480 {
481         struct hstate *hstate;
482 
483         if (!is_vm_hugetlb_page(vma))
484                 return PAGE_SIZE;
485 
486         hstate = hstate_vma(vma);
487 
488         return 1UL << huge_page_shift(hstate);
489 }
490 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
491 
492 /*
493  * Return the page size being used by the MMU to back a VMA. In the majority
494  * of cases, the page size used by the kernel matches the MMU size. On
495  * architectures where it differs, an architecture-specific version of this
496  * function is required.
497  */
498 #ifndef vma_mmu_pagesize
499 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
500 {
501         return vma_kernel_pagesize(vma);
502 }
503 #endif
504 
505 /*
506  * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
507  * bits of the reservation map pointer, which are always clear due to
508  * alignment.
509  */
510 #define HPAGE_RESV_OWNER    (1UL << 0)
511 #define HPAGE_RESV_UNMAPPED (1UL << 1)
512 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
513 
514 /*
515  * These helpers are used to track how many pages are reserved for
516  * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
517  * is guaranteed to have their future faults succeed.
518  *
519  * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
520  * the reserve counters are updated with the hugetlb_lock held. It is safe
521  * to reset the VMA at fork() time as it is not in use yet and there is no
522  * chance of the global counters getting corrupted as a result of the values.
523  *
524  * The private mapping reservation is represented in a subtly different
525  * manner to a shared mapping.  A shared mapping has a region map associated
526  * with the underlying file, this region map represents the backing file
527  * pages which have ever had a reservation assigned which this persists even
528  * after the page is instantiated.  A private mapping has a region map
529  * associated with the original mmap which is attached to all VMAs which
530  * reference it, this region map represents those offsets which have consumed
531  * reservation ie. where pages have been instantiated.
532  */
533 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
534 {
535         return (unsigned long)vma->vm_private_data;
536 }
537 
538 static void set_vma_private_data(struct vm_area_struct *vma,
539                                                         unsigned long value)
540 {
541         vma->vm_private_data = (void *)value;
542 }
543 
544 struct resv_map *resv_map_alloc(void)
545 {
546         struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
547         if (!resv_map)
548                 return NULL;
549 
550         kref_init(&resv_map->refs);
551         spin_lock_init(&resv_map->lock);
552         INIT_LIST_HEAD(&resv_map->regions);
553 
554         return resv_map;
555 }
556 
557 void resv_map_release(struct kref *ref)
558 {
559         struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
560 
561         /* Clear out any active regions before we release the map. */
562         region_truncate(resv_map, 0);
563         kfree(resv_map);
564 }
565 
566 static inline struct resv_map *inode_resv_map(struct inode *inode)
567 {
568         return inode->i_mapping->private_data;
569 }
570 
571 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
572 {
573         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
574         if (vma->vm_flags & VM_MAYSHARE) {
575                 struct address_space *mapping = vma->vm_file->f_mapping;
576                 struct inode *inode = mapping->host;
577 
578                 return inode_resv_map(inode);
579 
580         } else {
581                 return (struct resv_map *)(get_vma_private_data(vma) &
582                                                         ~HPAGE_RESV_MASK);
583         }
584 }
585 
586 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
587 {
588         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
589         VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
590 
591         set_vma_private_data(vma, (get_vma_private_data(vma) &
592                                 HPAGE_RESV_MASK) | (unsigned long)map);
593 }
594 
595 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
596 {
597         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
598         VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
599 
600         set_vma_private_data(vma, get_vma_private_data(vma) | flags);
601 }
602 
603 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
604 {
605         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
606 
607         return (get_vma_private_data(vma) & flag) != 0;
608 }
609 
610 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
611 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
612 {
613         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
614         if (!(vma->vm_flags & VM_MAYSHARE))
615                 vma->vm_private_data = (void *)0;
616 }
617 
618 /* Returns true if the VMA has associated reserve pages */
619 static int vma_has_reserves(struct vm_area_struct *vma, long chg)
620 {
621         if (vma->vm_flags & VM_NORESERVE) {
622                 /*
623                  * This address is already reserved by other process(chg == 0),
624                  * so, we should decrement reserved count. Without decrementing,
625                  * reserve count remains after releasing inode, because this
626                  * allocated page will go into page cache and is regarded as
627                  * coming from reserved pool in releasing step.  Currently, we
628                  * don't have any other solution to deal with this situation
629                  * properly, so add work-around here.
630                  */
631                 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
632                         return 1;
633                 else
634                         return 0;
635         }
636 
637         /* Shared mappings always use reserves */
638         if (vma->vm_flags & VM_MAYSHARE)
639                 return 1;
640 
641         /*
642          * Only the process that called mmap() has reserves for
643          * private mappings.
644          */
645         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
646                 return 1;
647 
648         return 0;
649 }
650 
651 static void enqueue_huge_page(struct hstate *h, struct page *page)
652 {
653         int nid = page_to_nid(page);
654         list_move(&page->lru, &h->hugepage_freelists[nid]);
655         h->free_huge_pages++;
656         h->free_huge_pages_node[nid]++;
657 }
658 
659 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
660 {
661         struct page *page;
662 
663         list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
664                 if (!is_migrate_isolate_page(page))
665                         break;
666         /*
667          * if 'non-isolated free hugepage' not found on the list,
668          * the allocation fails.
669          */
670         if (&h->hugepage_freelists[nid] == &page->lru)
671                 return NULL;
672         list_move(&page->lru, &h->hugepage_activelist);
673         set_page_refcounted(page);
674         h->free_huge_pages--;
675         h->free_huge_pages_node[nid]--;
676         return page;
677 }
678 
679 /* Movability of hugepages depends on migration support. */
680 static inline gfp_t htlb_alloc_mask(struct hstate *h)
681 {
682         if (hugepages_treat_as_movable || hugepage_migration_supported(h))
683                 return GFP_HIGHUSER_MOVABLE;
684         else
685                 return GFP_HIGHUSER;
686 }
687 
688 static struct page *dequeue_huge_page_vma(struct hstate *h,
689                                 struct vm_area_struct *vma,
690                                 unsigned long address, int avoid_reserve,
691                                 long chg)
692 {
693         struct page *page = NULL;
694         struct mempolicy *mpol;
695         nodemask_t *nodemask;
696         struct zonelist *zonelist;
697         struct zone *zone;
698         struct zoneref *z;
699         unsigned int cpuset_mems_cookie;
700 
701         /*
702          * A child process with MAP_PRIVATE mappings created by their parent
703          * have no page reserves. This check ensures that reservations are
704          * not "stolen". The child may still get SIGKILLed
705          */
706         if (!vma_has_reserves(vma, chg) &&
707                         h->free_huge_pages - h->resv_huge_pages == 0)
708                 goto err;
709 
710         /* If reserves cannot be used, ensure enough pages are in the pool */
711         if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
712                 goto err;
713 
714 retry_cpuset:
715         cpuset_mems_cookie = read_mems_allowed_begin();
716         zonelist = huge_zonelist(vma, address,
717                                         htlb_alloc_mask(h), &mpol, &nodemask);
718 
719         for_each_zone_zonelist_nodemask(zone, z, zonelist,
720                                                 MAX_NR_ZONES - 1, nodemask) {
721                 if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) {
722                         page = dequeue_huge_page_node(h, zone_to_nid(zone));
723                         if (page) {
724                                 if (avoid_reserve)
725                                         break;
726                                 if (!vma_has_reserves(vma, chg))
727                                         break;
728 
729                                 SetPagePrivate(page);
730                                 h->resv_huge_pages--;
731                                 break;
732                         }
733                 }
734         }
735 
736         mpol_cond_put(mpol);
737         if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
738                 goto retry_cpuset;
739         return page;
740 
741 err:
742         return NULL;
743 }
744 
745 /*
746  * common helper functions for hstate_next_node_to_{alloc|free}.
747  * We may have allocated or freed a huge page based on a different
748  * nodes_allowed previously, so h->next_node_to_{alloc|free} might
749  * be outside of *nodes_allowed.  Ensure that we use an allowed
750  * node for alloc or free.
751  */
752 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
753 {
754         nid = next_node(nid, *nodes_allowed);
755         if (nid == MAX_NUMNODES)
756                 nid = first_node(*nodes_allowed);
757         VM_BUG_ON(nid >= MAX_NUMNODES);
758 
759         return nid;
760 }
761 
762 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
763 {
764         if (!node_isset(nid, *nodes_allowed))
765                 nid = next_node_allowed(nid, nodes_allowed);
766         return nid;
767 }
768 
769 /*
770  * returns the previously saved node ["this node"] from which to
771  * allocate a persistent huge page for the pool and advance the
772  * next node from which to allocate, handling wrap at end of node
773  * mask.
774  */
775 static int hstate_next_node_to_alloc(struct hstate *h,
776                                         nodemask_t *nodes_allowed)
777 {
778         int nid;
779 
780         VM_BUG_ON(!nodes_allowed);
781 
782         nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
783         h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
784 
785         return nid;
786 }
787 
788 /*
789  * helper for free_pool_huge_page() - return the previously saved
790  * node ["this node"] from which to free a huge page.  Advance the
791  * next node id whether or not we find a free huge page to free so
792  * that the next attempt to free addresses the next node.
793  */
794 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
795 {
796         int nid;
797 
798         VM_BUG_ON(!nodes_allowed);
799 
800         nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
801         h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
802 
803         return nid;
804 }
805 
806 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask)           \
807         for (nr_nodes = nodes_weight(*mask);                            \
808                 nr_nodes > 0 &&                                         \
809                 ((node = hstate_next_node_to_alloc(hs, mask)) || 1);    \
810                 nr_nodes--)
811 
812 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask)            \
813         for (nr_nodes = nodes_weight(*mask);                            \
814                 nr_nodes > 0 &&                                         \
815                 ((node = hstate_next_node_to_free(hs, mask)) || 1);     \
816                 nr_nodes--)
817 
818 #if defined(CONFIG_CMA) && defined(CONFIG_X86_64)
819 static void destroy_compound_gigantic_page(struct page *page,
820                                         unsigned long order)
821 {
822         int i;
823         int nr_pages = 1 << order;
824         struct page *p = page + 1;
825 
826         for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
827                 __ClearPageTail(p);
828                 set_page_refcounted(p);
829                 p->first_page = NULL;
830         }
831 
832         set_compound_order(page, 0);
833         __ClearPageHead(page);
834 }
835 
836 static void free_gigantic_page(struct page *page, unsigned order)
837 {
838         free_contig_range(page_to_pfn(page), 1 << order);
839 }
840 
841 static int __alloc_gigantic_page(unsigned long start_pfn,
842                                 unsigned long nr_pages)
843 {
844         unsigned long end_pfn = start_pfn + nr_pages;
845         return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
846 }
847 
848 static bool pfn_range_valid_gigantic(unsigned long start_pfn,
849                                 unsigned long nr_pages)
850 {
851         unsigned long i, end_pfn = start_pfn + nr_pages;
852         struct page *page;
853 
854         for (i = start_pfn; i < end_pfn; i++) {
855                 if (!pfn_valid(i))
856                         return false;
857 
858                 page = pfn_to_page(i);
859 
860                 if (PageReserved(page))
861                         return false;
862 
863                 if (page_count(page) > 0)
864                         return false;
865 
866                 if (PageHuge(page))
867                         return false;
868         }
869 
870         return true;
871 }
872 
873 static bool zone_spans_last_pfn(const struct zone *zone,
874                         unsigned long start_pfn, unsigned long nr_pages)
875 {
876         unsigned long last_pfn = start_pfn + nr_pages - 1;
877         return zone_spans_pfn(zone, last_pfn);
878 }
879 
880 static struct page *alloc_gigantic_page(int nid, unsigned order)
881 {
882         unsigned long nr_pages = 1 << order;
883         unsigned long ret, pfn, flags;
884         struct zone *z;
885 
886         z = NODE_DATA(nid)->node_zones;
887         for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) {
888                 spin_lock_irqsave(&z->lock, flags);
889 
890                 pfn = ALIGN(z->zone_start_pfn, nr_pages);
891                 while (zone_spans_last_pfn(z, pfn, nr_pages)) {
892                         if (pfn_range_valid_gigantic(pfn, nr_pages)) {
893                                 /*
894                                  * We release the zone lock here because
895                                  * alloc_contig_range() will also lock the zone
896                                  * at some point. If there's an allocation
897                                  * spinning on this lock, it may win the race
898                                  * and cause alloc_contig_range() to fail...
899                                  */
900                                 spin_unlock_irqrestore(&z->lock, flags);
901                                 ret = __alloc_gigantic_page(pfn, nr_pages);
902                                 if (!ret)
903                                         return pfn_to_page(pfn);
904                                 spin_lock_irqsave(&z->lock, flags);
905                         }
906                         pfn += nr_pages;
907                 }
908 
909                 spin_unlock_irqrestore(&z->lock, flags);
910         }
911 
912         return NULL;
913 }
914 
915 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
916 static void prep_compound_gigantic_page(struct page *page, unsigned long order);
917 
918 static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid)
919 {
920         struct page *page;
921 
922         page = alloc_gigantic_page(nid, huge_page_order(h));
923         if (page) {
924                 prep_compound_gigantic_page(page, huge_page_order(h));
925                 prep_new_huge_page(h, page, nid);
926         }
927 
928         return page;
929 }
930 
931 static int alloc_fresh_gigantic_page(struct hstate *h,
932                                 nodemask_t *nodes_allowed)
933 {
934         struct page *page = NULL;
935         int nr_nodes, node;
936 
937         for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
938                 page = alloc_fresh_gigantic_page_node(h, node);
939                 if (page)
940                         return 1;
941         }
942 
943         return 0;
944 }
945 
946 static inline bool gigantic_page_supported(void) { return true; }
947 #else
948 static inline bool gigantic_page_supported(void) { return false; }
949 static inline void free_gigantic_page(struct page *page, unsigned order) { }
950 static inline void destroy_compound_gigantic_page(struct page *page,
951                                                 unsigned long order) { }
952 static inline int alloc_fresh_gigantic_page(struct hstate *h,
953                                         nodemask_t *nodes_allowed) { return 0; }
954 #endif
955 
956 static void update_and_free_page(struct hstate *h, struct page *page)
957 {
958         int i;
959 
960         if (hstate_is_gigantic(h) && !gigantic_page_supported())
961                 return;
962 
963         h->nr_huge_pages--;
964         h->nr_huge_pages_node[page_to_nid(page)]--;
965         for (i = 0; i < pages_per_huge_page(h); i++) {
966                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
967                                 1 << PG_referenced | 1 << PG_dirty |
968                                 1 << PG_active | 1 << PG_private |
969                                 1 << PG_writeback);
970         }
971         VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
972         set_compound_page_dtor(page, NULL);
973         set_page_refcounted(page);
974         if (hstate_is_gigantic(h)) {
975                 destroy_compound_gigantic_page(page, huge_page_order(h));
976                 free_gigantic_page(page, huge_page_order(h));
977         } else {
978                 __free_pages(page, huge_page_order(h));
979         }
980 }
981 
982 struct hstate *size_to_hstate(unsigned long size)
983 {
984         struct hstate *h;
985 
986         for_each_hstate(h) {
987                 if (huge_page_size(h) == size)
988                         return h;
989         }
990         return NULL;
991 }
992 
993 /*
994  * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
995  * to hstate->hugepage_activelist.)
996  *
997  * This function can be called for tail pages, but never returns true for them.
998  */
999 bool page_huge_active(struct page *page)
1000 {
1001         VM_BUG_ON_PAGE(!PageHuge(page), page);
1002         return PageHead(page) && PagePrivate(&page[1]);
1003 }
1004 
1005 /* never called for tail page */
1006 static void set_page_huge_active(struct page *page)
1007 {
1008         VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1009         SetPagePrivate(&page[1]);
1010 }
1011 
1012 static void clear_page_huge_active(struct page *page)
1013 {
1014         VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1015         ClearPagePrivate(&page[1]);
1016 }
1017 
1018 void free_huge_page(struct page *page)
1019 {
1020         /*
1021          * Can't pass hstate in here because it is called from the
1022          * compound page destructor.
1023          */
1024         struct hstate *h = page_hstate(page);
1025         int nid = page_to_nid(page);
1026         struct hugepage_subpool *spool =
1027                 (struct hugepage_subpool *)page_private(page);
1028         bool restore_reserve;
1029 
1030         set_page_private(page, 0);
1031         page->mapping = NULL;
1032         BUG_ON(page_count(page));
1033         BUG_ON(page_mapcount(page));
1034         restore_reserve = PagePrivate(page);
1035         ClearPagePrivate(page);
1036 
1037         /*
1038          * A return code of zero implies that the subpool will be under its
1039          * minimum size if the reservation is not restored after page is free.
1040          * Therefore, force restore_reserve operation.
1041          */
1042         if (hugepage_subpool_put_pages(spool, 1) == 0)
1043                 restore_reserve = true;
1044 
1045         spin_lock(&hugetlb_lock);
1046         clear_page_huge_active(page);
1047         hugetlb_cgroup_uncharge_page(hstate_index(h),
1048                                      pages_per_huge_page(h), page);
1049         if (restore_reserve)
1050                 h->resv_huge_pages++;
1051 
1052         if (h->surplus_huge_pages_node[nid]) {
1053                 /* remove the page from active list */
1054                 list_del(&page->lru);
1055                 update_and_free_page(h, page);
1056                 h->surplus_huge_pages--;
1057                 h->surplus_huge_pages_node[nid]--;
1058         } else {
1059                 arch_clear_hugepage_flags(page);
1060                 enqueue_huge_page(h, page);
1061         }
1062         spin_unlock(&hugetlb_lock);
1063 }
1064 
1065 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1066 {
1067         INIT_LIST_HEAD(&page->lru);
1068         set_compound_page_dtor(page, free_huge_page);
1069         spin_lock(&hugetlb_lock);
1070         set_hugetlb_cgroup(page, NULL);
1071         h->nr_huge_pages++;
1072         h->nr_huge_pages_node[nid]++;
1073         spin_unlock(&hugetlb_lock);
1074         put_page(page); /* free it into the hugepage allocator */
1075 }
1076 
1077 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
1078 {
1079         int i;
1080         int nr_pages = 1 << order;
1081         struct page *p = page + 1;
1082 
1083         /* we rely on prep_new_huge_page to set the destructor */
1084         set_compound_order(page, order);
1085         __SetPageHead(page);
1086         __ClearPageReserved(page);
1087         for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1088                 /*
1089                  * For gigantic hugepages allocated through bootmem at
1090                  * boot, it's safer to be consistent with the not-gigantic
1091                  * hugepages and clear the PG_reserved bit from all tail pages
1092                  * too.  Otherwse drivers using get_user_pages() to access tail
1093                  * pages may get the reference counting wrong if they see
1094                  * PG_reserved set on a tail page (despite the head page not
1095                  * having PG_reserved set).  Enforcing this consistency between
1096                  * head and tail pages allows drivers to optimize away a check
1097                  * on the head page when they need know if put_page() is needed
1098                  * after get_user_pages().
1099                  */
1100                 __ClearPageReserved(p);
1101                 set_page_count(p, 0);
1102                 p->first_page = page;
1103                 /* Make sure p->first_page is always valid for PageTail() */
1104                 smp_wmb();
1105                 __SetPageTail(p);
1106         }
1107 }
1108 
1109 /*
1110  * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1111  * transparent huge pages.  See the PageTransHuge() documentation for more
1112  * details.
1113  */
1114 int PageHuge(struct page *page)
1115 {
1116         if (!PageCompound(page))
1117                 return 0;
1118 
1119         page = compound_head(page);
1120         return get_compound_page_dtor(page) == free_huge_page;
1121 }
1122 EXPORT_SYMBOL_GPL(PageHuge);
1123 
1124 /*
1125  * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1126  * normal or transparent huge pages.
1127  */
1128 int PageHeadHuge(struct page *page_head)
1129 {
1130         if (!PageHead(page_head))
1131                 return 0;
1132 
1133         return get_compound_page_dtor(page_head) == free_huge_page;
1134 }
1135 
1136 pgoff_t __basepage_index(struct page *page)
1137 {
1138         struct page *page_head = compound_head(page);
1139         pgoff_t index = page_index(page_head);
1140         unsigned long compound_idx;
1141 
1142         if (!PageHuge(page_head))
1143                 return page_index(page);
1144 
1145         if (compound_order(page_head) >= MAX_ORDER)
1146                 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1147         else
1148                 compound_idx = page - page_head;
1149 
1150         return (index << compound_order(page_head)) + compound_idx;
1151 }
1152 
1153 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
1154 {
1155         struct page *page;
1156 
1157         page = alloc_pages_exact_node(nid,
1158                 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
1159                                                 __GFP_REPEAT|__GFP_NOWARN,
1160                 huge_page_order(h));
1161         if (page) {
1162                 prep_new_huge_page(h, page, nid);
1163         }
1164 
1165         return page;
1166 }
1167 
1168 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1169 {
1170         struct page *page;
1171         int nr_nodes, node;
1172         int ret = 0;
1173 
1174         for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1175                 page = alloc_fresh_huge_page_node(h, node);
1176                 if (page) {
1177                         ret = 1;
1178                         break;
1179                 }
1180         }
1181 
1182         if (ret)
1183                 count_vm_event(HTLB_BUDDY_PGALLOC);
1184         else
1185                 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1186 
1187         return ret;
1188 }
1189 
1190 /*
1191  * Free huge page from pool from next node to free.
1192  * Attempt to keep persistent huge pages more or less
1193  * balanced over allowed nodes.
1194  * Called with hugetlb_lock locked.
1195  */
1196 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1197                                                          bool acct_surplus)
1198 {
1199         int nr_nodes, node;
1200         int ret = 0;
1201 
1202         for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1203                 /*
1204                  * If we're returning unused surplus pages, only examine
1205                  * nodes with surplus pages.
1206                  */
1207                 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1208                     !list_empty(&h->hugepage_freelists[node])) {
1209                         struct page *page =
1210                                 list_entry(h->hugepage_freelists[node].next,
1211                                           struct page, lru);
1212                         list_del(&page->lru);
1213                         h->free_huge_pages--;
1214                         h->free_huge_pages_node[node]--;
1215                         if (acct_surplus) {
1216                                 h->surplus_huge_pages--;
1217                                 h->surplus_huge_pages_node[node]--;
1218                         }
1219                         update_and_free_page(h, page);
1220                         ret = 1;
1221                         break;
1222                 }
1223         }
1224 
1225         return ret;
1226 }
1227 
1228 /*
1229  * Dissolve a given free hugepage into free buddy pages. This function does
1230  * nothing for in-use (including surplus) hugepages.
1231  */
1232 static void dissolve_free_huge_page(struct page *page)
1233 {
1234         spin_lock(&hugetlb_lock);
1235         if (PageHuge(page) && !page_count(page)) {
1236                 struct hstate *h = page_hstate(page);
1237                 int nid = page_to_nid(page);
1238                 list_del(&page->lru);
1239                 h->free_huge_pages--;
1240                 h->free_huge_pages_node[nid]--;
1241                 update_and_free_page(h, page);
1242         }
1243         spin_unlock(&hugetlb_lock);
1244 }
1245 
1246 /*
1247  * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1248  * make specified memory blocks removable from the system.
1249  * Note that start_pfn should aligned with (minimum) hugepage size.
1250  */
1251 void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1252 {
1253         unsigned long pfn;
1254 
1255         if (!hugepages_supported())
1256                 return;
1257 
1258         VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << minimum_order));
1259         for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order)
1260                 dissolve_free_huge_page(pfn_to_page(pfn));
1261 }
1262 
1263 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
1264 {
1265         struct page *page;
1266         unsigned int r_nid;
1267 
1268         if (hstate_is_gigantic(h))
1269                 return NULL;
1270 
1271         /*
1272          * Assume we will successfully allocate the surplus page to
1273          * prevent racing processes from causing the surplus to exceed
1274          * overcommit
1275          *
1276          * This however introduces a different race, where a process B
1277          * tries to grow the static hugepage pool while alloc_pages() is
1278          * called by process A. B will only examine the per-node
1279          * counters in determining if surplus huge pages can be
1280          * converted to normal huge pages in adjust_pool_surplus(). A
1281          * won't be able to increment the per-node counter, until the
1282          * lock is dropped by B, but B doesn't drop hugetlb_lock until
1283          * no more huge pages can be converted from surplus to normal
1284          * state (and doesn't try to convert again). Thus, we have a
1285          * case where a surplus huge page exists, the pool is grown, and
1286          * the surplus huge page still exists after, even though it
1287          * should just have been converted to a normal huge page. This
1288          * does not leak memory, though, as the hugepage will be freed
1289          * once it is out of use. It also does not allow the counters to
1290          * go out of whack in adjust_pool_surplus() as we don't modify
1291          * the node values until we've gotten the hugepage and only the
1292          * per-node value is checked there.
1293          */
1294         spin_lock(&hugetlb_lock);
1295         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1296                 spin_unlock(&hugetlb_lock);
1297                 return NULL;
1298         } else {
1299                 h->nr_huge_pages++;
1300                 h->surplus_huge_pages++;
1301         }
1302         spin_unlock(&hugetlb_lock);
1303 
1304         if (nid == NUMA_NO_NODE)
1305                 page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP|
1306                                    __GFP_REPEAT|__GFP_NOWARN,
1307                                    huge_page_order(h));
1308         else
1309                 page = alloc_pages_exact_node(nid,
1310                         htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
1311                         __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
1312 
1313         spin_lock(&hugetlb_lock);
1314         if (page) {
1315                 INIT_LIST_HEAD(&page->lru);
1316                 r_nid = page_to_nid(page);
1317                 set_compound_page_dtor(page, free_huge_page);
1318                 set_hugetlb_cgroup(page, NULL);
1319                 /*
1320                  * We incremented the global counters already
1321                  */
1322                 h->nr_huge_pages_node[r_nid]++;
1323                 h->surplus_huge_pages_node[r_nid]++;
1324                 __count_vm_event(HTLB_BUDDY_PGALLOC);
1325         } else {
1326                 h->nr_huge_pages--;
1327                 h->surplus_huge_pages--;
1328                 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1329         }
1330         spin_unlock(&hugetlb_lock);
1331 
1332         return page;
1333 }
1334 
1335 /*
1336  * This allocation function is useful in the context where vma is irrelevant.
1337  * E.g. soft-offlining uses this function because it only cares physical
1338  * address of error page.
1339  */
1340 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1341 {
1342         struct page *page = NULL;
1343 
1344         spin_lock(&hugetlb_lock);
1345         if (h->free_huge_pages - h->resv_huge_pages > 0)
1346                 page = dequeue_huge_page_node(h, nid);
1347         spin_unlock(&hugetlb_lock);
1348 
1349         if (!page)
1350                 page = alloc_buddy_huge_page(h, nid);
1351 
1352         return page;
1353 }
1354 
1355 /*
1356  * Increase the hugetlb pool such that it can accommodate a reservation
1357  * of size 'delta'.
1358  */
1359 static int gather_surplus_pages(struct hstate *h, int delta)
1360 {
1361         struct list_head surplus_list;
1362         struct page *page, *tmp;
1363         int ret, i;
1364         int needed, allocated;
1365         bool alloc_ok = true;
1366 
1367         needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1368         if (needed <= 0) {
1369                 h->resv_huge_pages += delta;
1370                 return 0;
1371         }
1372 
1373         allocated = 0;
1374         INIT_LIST_HEAD(&surplus_list);
1375 
1376         ret = -ENOMEM;
1377 retry:
1378         spin_unlock(&hugetlb_lock);
1379         for (i = 0; i < needed; i++) {
1380                 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1381                 if (!page) {
1382                         alloc_ok = false;
1383                         break;
1384                 }
1385                 list_add(&page->lru, &surplus_list);
1386         }
1387         allocated += i;
1388 
1389         /*
1390          * After retaking hugetlb_lock, we need to recalculate 'needed'
1391          * because either resv_huge_pages or free_huge_pages may have changed.
1392          */
1393         spin_lock(&hugetlb_lock);
1394         needed = (h->resv_huge_pages + delta) -
1395                         (h->free_huge_pages + allocated);
1396         if (needed > 0) {
1397                 if (alloc_ok)
1398                         goto retry;
1399                 /*
1400                  * We were not able to allocate enough pages to
1401                  * satisfy the entire reservation so we free what
1402                  * we've allocated so far.
1403                  */
1404                 goto free;
1405         }
1406         /*
1407          * The surplus_list now contains _at_least_ the number of extra pages
1408          * needed to accommodate the reservation.  Add the appropriate number
1409          * of pages to the hugetlb pool and free the extras back to the buddy
1410          * allocator.  Commit the entire reservation here to prevent another
1411          * process from stealing the pages as they are added to the pool but
1412          * before they are reserved.
1413          */
1414         needed += allocated;
1415         h->resv_huge_pages += delta;
1416         ret = 0;
1417 
1418         /* Free the needed pages to the hugetlb pool */
1419         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1420                 if ((--needed) < 0)
1421                         break;
1422                 /*
1423                  * This page is now managed by the hugetlb allocator and has
1424                  * no users -- drop the buddy allocator's reference.
1425                  */
1426                 put_page_testzero(page);
1427                 VM_BUG_ON_PAGE(page_count(page), page);
1428                 enqueue_huge_page(h, page);
1429         }
1430 free:
1431         spin_unlock(&hugetlb_lock);
1432 
1433         /* Free unnecessary surplus pages to the buddy allocator */
1434         list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1435                 put_page(page);
1436         spin_lock(&hugetlb_lock);
1437 
1438         return ret;
1439 }
1440 
1441 /*
1442  * When releasing a hugetlb pool reservation, any surplus pages that were
1443  * allocated to satisfy the reservation must be explicitly freed if they were
1444  * never used.
1445  * Called with hugetlb_lock held.
1446  */
1447 static void return_unused_surplus_pages(struct hstate *h,
1448                                         unsigned long unused_resv_pages)
1449 {
1450         unsigned long nr_pages;
1451 
1452         /* Uncommit the reservation */
1453         h->resv_huge_pages -= unused_resv_pages;
1454 
1455         /* Cannot return gigantic pages currently */
1456         if (hstate_is_gigantic(h))
1457                 return;
1458 
1459         nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1460 
1461         /*
1462          * We want to release as many surplus pages as possible, spread
1463          * evenly across all nodes with memory. Iterate across these nodes
1464          * until we can no longer free unreserved surplus pages. This occurs
1465          * when the nodes with surplus pages have no free pages.
1466          * free_pool_huge_page() will balance the the freed pages across the
1467          * on-line nodes with memory and will handle the hstate accounting.
1468          */
1469         while (nr_pages--) {
1470                 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1471                         break;
1472                 cond_resched_lock(&hugetlb_lock);
1473         }
1474 }
1475 
1476 /*
1477  * vma_needs_reservation and vma_commit_reservation are used by the huge
1478  * page allocation routines to manage reservations.
1479  *
1480  * vma_needs_reservation is called to determine if the huge page at addr
1481  * within the vma has an associated reservation.  If a reservation is
1482  * needed, the value 1 is returned.  The caller is then responsible for
1483  * managing the global reservation and subpool usage counts.  After
1484  * the huge page has been allocated, vma_commit_reservation is called
1485  * to add the page to the reservation map.
1486  *
1487  * In the normal case, vma_commit_reservation returns the same value
1488  * as the preceding vma_needs_reservation call.  The only time this
1489  * is not the case is if a reserve map was changed between calls.  It
1490  * is the responsibility of the caller to notice the difference and
1491  * take appropriate action.
1492  */
1493 static long __vma_reservation_common(struct hstate *h,
1494                                 struct vm_area_struct *vma, unsigned long addr,
1495                                 bool commit)
1496 {
1497         struct resv_map *resv;
1498         pgoff_t idx;
1499         long ret;
1500 
1501         resv = vma_resv_map(vma);
1502         if (!resv)
1503                 return 1;
1504 
1505         idx = vma_hugecache_offset(h, vma, addr);
1506         if (commit)
1507                 ret = region_add(resv, idx, idx + 1);
1508         else
1509                 ret = region_chg(resv, idx, idx + 1);
1510 
1511         if (vma->vm_flags & VM_MAYSHARE)
1512                 return ret;
1513         else
1514                 return ret < 0 ? ret : 0;
1515 }
1516 
1517 static long vma_needs_reservation(struct hstate *h,
1518                         struct vm_area_struct *vma, unsigned long addr)
1519 {
1520         return __vma_reservation_common(h, vma, addr, false);
1521 }
1522 
1523 static long vma_commit_reservation(struct hstate *h,
1524                         struct vm_area_struct *vma, unsigned long addr)
1525 {
1526         return __vma_reservation_common(h, vma, addr, true);
1527 }
1528 
1529 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1530                                     unsigned long addr, int avoid_reserve)
1531 {
1532         struct hugepage_subpool *spool = subpool_vma(vma);
1533         struct hstate *h = hstate_vma(vma);
1534         struct page *page;
1535         long chg, commit;
1536         int ret, idx;
1537         struct hugetlb_cgroup *h_cg;
1538 
1539         idx = hstate_index(h);
1540         /*
1541          * Processes that did not create the mapping will have no
1542          * reserves and will not have accounted against subpool
1543          * limit. Check that the subpool limit can be made before
1544          * satisfying the allocation MAP_NORESERVE mappings may also
1545          * need pages and subpool limit allocated allocated if no reserve
1546          * mapping overlaps.
1547          */
1548         chg = vma_needs_reservation(h, vma, addr);
1549         if (chg < 0)
1550                 return ERR_PTR(-ENOMEM);
1551         if (chg || avoid_reserve)
1552                 if (hugepage_subpool_get_pages(spool, 1) < 0)
1553                         return ERR_PTR(-ENOSPC);
1554 
1555         ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1556         if (ret)
1557                 goto out_subpool_put;
1558 
1559         spin_lock(&hugetlb_lock);
1560         page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg);
1561         if (!page) {
1562                 spin_unlock(&hugetlb_lock);
1563                 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1564                 if (!page)
1565                         goto out_uncharge_cgroup;
1566 
1567                 spin_lock(&hugetlb_lock);
1568                 list_move(&page->lru, &h->hugepage_activelist);
1569                 /* Fall through */
1570         }
1571         hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1572         spin_unlock(&hugetlb_lock);
1573 
1574         set_page_private(page, (unsigned long)spool);
1575 
1576         commit = vma_commit_reservation(h, vma, addr);
1577         if (unlikely(chg > commit)) {
1578                 /*
1579                  * The page was added to the reservation map between
1580                  * vma_needs_reservation and vma_commit_reservation.
1581                  * This indicates a race with hugetlb_reserve_pages.
1582                  * Adjust for the subpool count incremented above AND
1583                  * in hugetlb_reserve_pages for the same page.  Also,
1584                  * the reservation count added in hugetlb_reserve_pages
1585                  * no longer applies.
1586                  */
1587                 long rsv_adjust;
1588 
1589                 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
1590                 hugetlb_acct_memory(h, -rsv_adjust);
1591         }
1592         return page;
1593 
1594 out_uncharge_cgroup:
1595         hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
1596 out_subpool_put:
1597         if (chg || avoid_reserve)
1598                 hugepage_subpool_put_pages(spool, 1);
1599         return ERR_PTR(-ENOSPC);
1600 }
1601 
1602 /*
1603  * alloc_huge_page()'s wrapper which simply returns the page if allocation
1604  * succeeds, otherwise NULL. This function is called from new_vma_page(),
1605  * where no ERR_VALUE is expected to be returned.
1606  */
1607 struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
1608                                 unsigned long addr, int avoid_reserve)
1609 {
1610         struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
1611         if (IS_ERR(page))
1612                 page = NULL;
1613         return page;
1614 }
1615 
1616 int __weak alloc_bootmem_huge_page(struct hstate *h)
1617 {
1618         struct huge_bootmem_page *m;
1619         int nr_nodes, node;
1620 
1621         for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
1622                 void *addr;
1623 
1624                 addr = memblock_virt_alloc_try_nid_nopanic(
1625                                 huge_page_size(h), huge_page_size(h),
1626                                 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
1627                 if (addr) {
1628                         /*
1629                          * Use the beginning of the huge page to store the
1630                          * huge_bootmem_page struct (until gather_bootmem
1631                          * puts them into the mem_map).
1632                          */
1633                         m = addr;
1634                         goto found;
1635                 }
1636         }
1637         return 0;
1638 
1639 found:
1640         BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
1641         /* Put them into a private list first because mem_map is not up yet */
1642         list_add(&m->list, &huge_boot_pages);
1643         m->hstate = h;
1644         return 1;
1645 }
1646 
1647 static void __init prep_compound_huge_page(struct page *page, int order)
1648 {
1649         if (unlikely(order > (MAX_ORDER - 1)))
1650                 prep_compound_gigantic_page(page, order);
1651         else
1652                 prep_compound_page(page, order);
1653 }
1654 
1655 /* Put bootmem huge pages into the standard lists after mem_map is up */
1656 static void __init gather_bootmem_prealloc(void)
1657 {
1658         struct huge_bootmem_page *m;
1659 
1660         list_for_each_entry(m, &huge_boot_pages, list) {
1661                 struct hstate *h = m->hstate;
1662                 struct page *page;
1663 
1664 #ifdef CONFIG_HIGHMEM
1665                 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1666                 memblock_free_late(__pa(m),
1667                                    sizeof(struct huge_bootmem_page));
1668 #else
1669                 page = virt_to_page(m);
1670 #endif
1671                 WARN_ON(page_count(page) != 1);
1672                 prep_compound_huge_page(page, h->order);
1673                 WARN_ON(PageReserved(page));
1674                 prep_new_huge_page(h, page, page_to_nid(page));
1675                 /*
1676                  * If we had gigantic hugepages allocated at boot time, we need
1677                  * to restore the 'stolen' pages to totalram_pages in order to
1678                  * fix confusing memory reports from free(1) and another
1679                  * side-effects, like CommitLimit going negative.
1680                  */
1681                 if (hstate_is_gigantic(h))
1682                         adjust_managed_page_count(page, 1 << h->order);
1683         }
1684 }
1685 
1686 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1687 {
1688         unsigned long i;
1689 
1690         for (i = 0; i < h->max_huge_pages; ++i) {
1691                 if (hstate_is_gigantic(h)) {
1692                         if (!alloc_bootmem_huge_page(h))
1693                                 break;
1694                 } else if (!alloc_fresh_huge_page(h,
1695                                          &node_states[N_MEMORY]))
1696                         break;
1697         }
1698         h->max_huge_pages = i;
1699 }
1700 
1701 static void __init hugetlb_init_hstates(void)
1702 {
1703         struct hstate *h;
1704 
1705         for_each_hstate(h) {
1706                 if (minimum_order > huge_page_order(h))
1707                         minimum_order = huge_page_order(h);
1708 
1709                 /* oversize hugepages were init'ed in early boot */
1710                 if (!hstate_is_gigantic(h))
1711                         hugetlb_hstate_alloc_pages(h);
1712         }
1713         VM_BUG_ON(minimum_order == UINT_MAX);
1714 }
1715 
1716 static char * __init memfmt(char *buf, unsigned long n)
1717 {
1718         if (n >= (1UL << 30))
1719                 sprintf(buf, "%lu GB", n >> 30);
1720         else if (n >= (1UL << 20))
1721                 sprintf(buf, "%lu MB", n >> 20);
1722         else
1723                 sprintf(buf, "%lu KB", n >> 10);
1724         return buf;
1725 }
1726 
1727 static void __init report_hugepages(void)
1728 {
1729         struct hstate *h;
1730 
1731         for_each_hstate(h) {
1732                 char buf[32];
1733                 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1734                         memfmt(buf, huge_page_size(h)),
1735                         h->free_huge_pages);
1736         }
1737 }
1738 
1739 #ifdef CONFIG_HIGHMEM
1740 static void try_to_free_low(struct hstate *h, unsigned long count,
1741                                                 nodemask_t *nodes_allowed)
1742 {
1743         int i;
1744 
1745         if (hstate_is_gigantic(h))
1746                 return;
1747 
1748         for_each_node_mask(i, *nodes_allowed) {
1749                 struct page *page, *next;
1750                 struct list_head *freel = &h->hugepage_freelists[i];
1751                 list_for_each_entry_safe(page, next, freel, lru) {
1752                         if (count >= h->nr_huge_pages)
1753                                 return;
1754                         if (PageHighMem(page))
1755                                 continue;
1756                         list_del(&page->lru);
1757                         update_and_free_page(h, page);
1758                         h->free_huge_pages--;
1759                         h->free_huge_pages_node[page_to_nid(page)]--;
1760                 }
1761         }
1762 }
1763 #else
1764 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1765                                                 nodemask_t *nodes_allowed)
1766 {
1767 }
1768 #endif
1769 
1770 /*
1771  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
1772  * balanced by operating on them in a round-robin fashion.
1773  * Returns 1 if an adjustment was made.
1774  */
1775 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1776                                 int delta)
1777 {
1778         int nr_nodes, node;
1779 
1780         VM_BUG_ON(delta != -1 && delta != 1);
1781 
1782         if (delta < 0) {
1783                 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1784                         if (h->surplus_huge_pages_node[node])
1785                                 goto found;
1786                 }
1787         } else {
1788                 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1789                         if (h->surplus_huge_pages_node[node] <
1790                                         h->nr_huge_pages_node[node])
1791                                 goto found;
1792                 }
1793         }
1794         return 0;
1795 
1796 found:
1797         h->surplus_huge_pages += delta;
1798         h->surplus_huge_pages_node[node] += delta;
1799         return 1;
1800 }
1801 
1802 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1803 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1804                                                 nodemask_t *nodes_allowed)
1805 {
1806         unsigned long min_count, ret;
1807 
1808         if (hstate_is_gigantic(h) && !gigantic_page_supported())
1809                 return h->max_huge_pages;
1810 
1811         /*
1812          * Increase the pool size
1813          * First take pages out of surplus state.  Then make up the
1814          * remaining difference by allocating fresh huge pages.
1815          *
1816          * We might race with alloc_buddy_huge_page() here and be unable
1817          * to convert a surplus huge page to a normal huge page. That is
1818          * not critical, though, it just means the overall size of the
1819          * pool might be one hugepage larger than it needs to be, but
1820          * within all the constraints specified by the sysctls.
1821          */
1822         spin_lock(&hugetlb_lock);
1823         while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1824                 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1825                         break;
1826         }
1827 
1828         while (count > persistent_huge_pages(h)) {
1829                 /*
1830                  * If this allocation races such that we no longer need the
1831                  * page, free_huge_page will handle it by freeing the page
1832                  * and reducing the surplus.
1833                  */
1834                 spin_unlock(&hugetlb_lock);
1835                 if (hstate_is_gigantic(h))
1836                         ret = alloc_fresh_gigantic_page(h, nodes_allowed);
1837                 else
1838                         ret = alloc_fresh_huge_page(h, nodes_allowed);
1839                 spin_lock(&hugetlb_lock);
1840                 if (!ret)
1841                         goto out;
1842 
1843                 /* Bail for signals. Probably ctrl-c from user */
1844                 if (signal_pending(current))
1845                         goto out;
1846         }
1847 
1848         /*
1849          * Decrease the pool size
1850          * First return free pages to the buddy allocator (being careful
1851          * to keep enough around to satisfy reservations).  Then place
1852          * pages into surplus state as needed so the pool will shrink
1853          * to the desired size as pages become free.
1854          *
1855          * By placing pages into the surplus state independent of the
1856          * overcommit value, we are allowing the surplus pool size to
1857          * exceed overcommit. There are few sane options here. Since
1858          * alloc_buddy_huge_page() is checking the global counter,
1859          * though, we'll note that we're not allowed to exceed surplus
1860          * and won't grow the pool anywhere else. Not until one of the
1861          * sysctls are changed, or the surplus pages go out of use.
1862          */
1863         min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1864         min_count = max(count, min_count);
1865         try_to_free_low(h, min_count, nodes_allowed);
1866         while (min_count < persistent_huge_pages(h)) {
1867                 if (!free_pool_huge_page(h, nodes_allowed, 0))
1868                         break;
1869                 cond_resched_lock(&hugetlb_lock);
1870         }
1871         while (count < persistent_huge_pages(h)) {
1872                 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1873                         break;
1874         }
1875 out:
1876         ret = persistent_huge_pages(h);
1877         spin_unlock(&hugetlb_lock);
1878         return ret;
1879 }
1880 
1881 #define HSTATE_ATTR_RO(_name) \
1882         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1883 
1884 #define HSTATE_ATTR(_name) \
1885         static struct kobj_attribute _name##_attr = \
1886                 __ATTR(_name, 0644, _name##_show, _name##_store)
1887 
1888 static struct kobject *hugepages_kobj;
1889 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1890 
1891 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1892 
1893 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1894 {
1895         int i;
1896 
1897         for (i = 0; i < HUGE_MAX_HSTATE; i++)
1898                 if (hstate_kobjs[i] == kobj) {
1899                         if (nidp)
1900                                 *nidp = NUMA_NO_NODE;
1901                         return &hstates[i];
1902                 }
1903 
1904         return kobj_to_node_hstate(kobj, nidp);
1905 }
1906 
1907 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1908                                         struct kobj_attribute *attr, char *buf)
1909 {
1910         struct hstate *h;
1911         unsigned long nr_huge_pages;
1912         int nid;
1913 
1914         h = kobj_to_hstate(kobj, &nid);
1915         if (nid == NUMA_NO_NODE)
1916                 nr_huge_pages = h->nr_huge_pages;
1917         else
1918                 nr_huge_pages = h->nr_huge_pages_node[nid];
1919 
1920         return sprintf(buf, "%lu\n", nr_huge_pages);
1921 }
1922 
1923 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
1924                                            struct hstate *h, int nid,
1925                                            unsigned long count, size_t len)
1926 {
1927         int err;
1928         NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1929 
1930         if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
1931                 err = -EINVAL;
1932                 goto out;
1933         }
1934 
1935         if (nid == NUMA_NO_NODE) {
1936                 /*
1937                  * global hstate attribute
1938                  */
1939                 if (!(obey_mempolicy &&
1940                                 init_nodemask_of_mempolicy(nodes_allowed))) {
1941                         NODEMASK_FREE(nodes_allowed);
1942                         nodes_allowed = &node_states[N_MEMORY];
1943                 }
1944         } else if (nodes_allowed) {
1945                 /*
1946                  * per node hstate attribute: adjust count to global,
1947                  * but restrict alloc/free to the specified node.
1948                  */
1949                 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1950                 init_nodemask_of_node(nodes_allowed, nid);
1951         } else
1952                 nodes_allowed = &node_states[N_MEMORY];
1953 
1954         h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1955 
1956         if (nodes_allowed != &node_states[N_MEMORY])
1957                 NODEMASK_FREE(nodes_allowed);
1958 
1959         return len;
1960 out:
1961         NODEMASK_FREE(nodes_allowed);
1962         return err;
1963 }
1964 
1965 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1966                                          struct kobject *kobj, const char *buf,
1967                                          size_t len)
1968 {
1969         struct hstate *h;
1970         unsigned long count;
1971         int nid;
1972         int err;
1973 
1974         err = kstrtoul(buf, 10, &count);
1975         if (err)
1976                 return err;
1977 
1978         h = kobj_to_hstate(kobj, &nid);
1979         return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
1980 }
1981 
1982 static ssize_t nr_hugepages_show(struct kobject *kobj,
1983                                        struct kobj_attribute *attr, char *buf)
1984 {
1985         return nr_hugepages_show_common(kobj, attr, buf);
1986 }
1987 
1988 static ssize_t nr_hugepages_store(struct kobject *kobj,
1989                struct kobj_attribute *attr, const char *buf, size_t len)
1990 {
1991         return nr_hugepages_store_common(false, kobj, buf, len);
1992 }
1993 HSTATE_ATTR(nr_hugepages);
1994 
1995 #ifdef CONFIG_NUMA
1996 
1997 /*
1998  * hstate attribute for optionally mempolicy-based constraint on persistent
1999  * huge page alloc/free.
2000  */
2001 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2002                                        struct kobj_attribute *attr, char *buf)
2003 {
2004         return nr_hugepages_show_common(kobj, attr, buf);
2005 }
2006 
2007 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2008                struct kobj_attribute *attr, const char *buf, size_t len)
2009 {
2010         return nr_hugepages_store_common(true, kobj, buf, len);
2011 }
2012 HSTATE_ATTR(nr_hugepages_mempolicy);
2013 #endif
2014 
2015 
2016 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2017                                         struct kobj_attribute *attr, char *buf)
2018 {
2019         struct hstate *h = kobj_to_hstate(kobj, NULL);
2020         return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2021 }
2022 
2023 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2024                 struct kobj_attribute *attr, const char *buf, size_t count)
2025 {
2026         int err;
2027         unsigned long input;
2028         struct hstate *h = kobj_to_hstate(kobj, NULL);
2029 
2030         if (hstate_is_gigantic(h))
2031                 return -EINVAL;
2032 
2033         err = kstrtoul(buf, 10, &input);
2034         if (err)
2035                 return err;
2036 
2037         spin_lock(&hugetlb_lock);
2038         h->nr_overcommit_huge_pages = input;
2039         spin_unlock(&hugetlb_lock);
2040 
2041         return count;
2042 }
2043 HSTATE_ATTR(nr_overcommit_hugepages);
2044 
2045 static ssize_t free_hugepages_show(struct kobject *kobj,
2046                                         struct kobj_attribute *attr, char *buf)
2047 {
2048         struct hstate *h;
2049         unsigned long free_huge_pages;
2050         int nid;
2051 
2052         h = kobj_to_hstate(kobj, &nid);
2053         if (nid == NUMA_NO_NODE)
2054                 free_huge_pages = h->free_huge_pages;
2055         else
2056                 free_huge_pages = h->free_huge_pages_node[nid];
2057 
2058         return sprintf(buf, "%lu\n", free_huge_pages);
2059 }
2060 HSTATE_ATTR_RO(free_hugepages);
2061 
2062 static ssize_t resv_hugepages_show(struct kobject *kobj,
2063                                         struct kobj_attribute *attr, char *buf)
2064 {
2065         struct hstate *h = kobj_to_hstate(kobj, NULL);
2066         return sprintf(buf, "%lu\n", h->resv_huge_pages);
2067 }
2068 HSTATE_ATTR_RO(resv_hugepages);
2069 
2070 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2071                                         struct kobj_attribute *attr, char *buf)
2072 {
2073         struct hstate *h;
2074         unsigned long surplus_huge_pages;
2075         int nid;
2076 
2077         h = kobj_to_hstate(kobj, &nid);
2078         if (nid == NUMA_NO_NODE)
2079                 surplus_huge_pages = h->surplus_huge_pages;
2080         else
2081                 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2082 
2083         return sprintf(buf, "%lu\n", surplus_huge_pages);
2084 }
2085 HSTATE_ATTR_RO(surplus_hugepages);
2086 
2087 static struct attribute *hstate_attrs[] = {
2088         &nr_hugepages_attr.attr,
2089         &nr_overcommit_hugepages_attr.attr,
2090         &free_hugepages_attr.attr,
2091         &resv_hugepages_attr.attr,
2092         &surplus_hugepages_attr.attr,
2093 #ifdef CONFIG_NUMA
2094         &nr_hugepages_mempolicy_attr.attr,
2095 #endif
2096         NULL,
2097 };
2098 
2099 static struct attribute_group hstate_attr_group = {
2100         .attrs = hstate_attrs,
2101 };
2102 
2103 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2104                                     struct kobject **hstate_kobjs,
2105                                     struct attribute_group *hstate_attr_group)
2106 {
2107         int retval;
2108         int hi = hstate_index(h);
2109 
2110         hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2111         if (!hstate_kobjs[hi])
2112                 return -ENOMEM;
2113 
2114         retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2115         if (retval)
2116                 kobject_put(hstate_kobjs[hi]);
2117 
2118         return retval;
2119 }
2120 
2121 static void __init hugetlb_sysfs_init(void)
2122 {
2123         struct hstate *h;
2124         int err;
2125 
2126         hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2127         if (!hugepages_kobj)
2128                 return;
2129 
2130         for_each_hstate(h) {
2131                 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2132                                          hstate_kobjs, &hstate_attr_group);
2133                 if (err)
2134                         pr_err("Hugetlb: Unable to add hstate %s", h->name);
2135         }
2136 }
2137 
2138 #ifdef CONFIG_NUMA
2139 
2140 /*
2141  * node_hstate/s - associate per node hstate attributes, via their kobjects,
2142  * with node devices in node_devices[] using a parallel array.  The array
2143  * index of a node device or _hstate == node id.
2144  * This is here to avoid any static dependency of the node device driver, in
2145  * the base kernel, on the hugetlb module.
2146  */
2147 struct node_hstate {
2148         struct kobject          *hugepages_kobj;
2149         struct kobject          *hstate_kobjs[HUGE_MAX_HSTATE];
2150 };
2151 struct node_hstate node_hstates[MAX_NUMNODES];
2152 
2153 /*
2154  * A subset of global hstate attributes for node devices
2155  */
2156 static struct attribute *per_node_hstate_attrs[] = {
2157         &nr_hugepages_attr.attr,
2158         &free_hugepages_attr.attr,
2159         &surplus_hugepages_attr.attr,
2160         NULL,
2161 };
2162 
2163 static struct attribute_group per_node_hstate_attr_group = {
2164         .attrs = per_node_hstate_attrs,
2165 };
2166 
2167 /*
2168  * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2169  * Returns node id via non-NULL nidp.
2170  */
2171 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2172 {
2173         int nid;
2174 
2175         for (nid = 0; nid < nr_node_ids; nid++) {
2176                 struct node_hstate *nhs = &node_hstates[nid];
2177                 int i;
2178                 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2179                         if (nhs->hstate_kobjs[i] == kobj) {
2180                                 if (nidp)
2181                                         *nidp = nid;
2182                                 return &hstates[i];
2183                         }
2184         }
2185 
2186         BUG();
2187         return NULL;
2188 }
2189 
2190 /*
2191  * Unregister hstate attributes from a single node device.
2192  * No-op if no hstate attributes attached.
2193  */
2194 static void hugetlb_unregister_node(struct node *node)
2195 {
2196         struct hstate *h;
2197         struct node_hstate *nhs = &node_hstates[node->dev.id];
2198 
2199         if (!nhs->hugepages_kobj)
2200                 return;         /* no hstate attributes */
2201 
2202         for_each_hstate(h) {
2203                 int idx = hstate_index(h);
2204                 if (nhs->hstate_kobjs[idx]) {
2205                         kobject_put(nhs->hstate_kobjs[idx]);
2206                         nhs->hstate_kobjs[idx] = NULL;
2207                 }
2208         }
2209 
2210         kobject_put(nhs->hugepages_kobj);
2211         nhs->hugepages_kobj = NULL;
2212 }
2213 
2214 /*
2215  * hugetlb module exit:  unregister hstate attributes from node devices
2216  * that have them.
2217  */
2218 static void hugetlb_unregister_all_nodes(void)
2219 {
2220         int nid;
2221 
2222         /*
2223          * disable node device registrations.
2224          */
2225         register_hugetlbfs_with_node(NULL, NULL);
2226 
2227         /*
2228          * remove hstate attributes from any nodes that have them.
2229          */
2230         for (nid = 0; nid < nr_node_ids; nid++)
2231                 hugetlb_unregister_node(node_devices[nid]);
2232 }
2233 
2234 /*
2235  * Register hstate attributes for a single node device.
2236  * No-op if attributes already registered.
2237  */
2238 static void hugetlb_register_node(struct node *node)
2239 {
2240         struct hstate *h;
2241         struct node_hstate *nhs = &node_hstates[node->dev.id];
2242         int err;
2243 
2244         if (nhs->hugepages_kobj)
2245                 return;         /* already allocated */
2246 
2247         nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2248                                                         &node->dev.kobj);
2249         if (!nhs->hugepages_kobj)
2250                 return;
2251 
2252         for_each_hstate(h) {
2253                 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2254                                                 nhs->hstate_kobjs,
2255                                                 &per_node_hstate_attr_group);
2256                 if (err) {
2257                         pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2258                                 h->name, node->dev.id);
2259                         hugetlb_unregister_node(node);
2260                         break;
2261                 }
2262         }
2263 }
2264 
2265 /*
2266  * hugetlb init time:  register hstate attributes for all registered node
2267  * devices of nodes that have memory.  All on-line nodes should have
2268  * registered their associated device by this time.
2269  */
2270 static void __init hugetlb_register_all_nodes(void)
2271 {
2272         int nid;
2273 
2274         for_each_node_state(nid, N_MEMORY) {
2275                 struct node *node = node_devices[nid];
2276                 if (node->dev.id == nid)
2277                         hugetlb_register_node(node);
2278         }
2279 
2280         /*
2281          * Let the node device driver know we're here so it can
2282          * [un]register hstate attributes on node hotplug.
2283          */
2284         register_hugetlbfs_with_node(hugetlb_register_node,
2285                                      hugetlb_unregister_node);
2286 }
2287 #else   /* !CONFIG_NUMA */
2288 
2289 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2290 {
2291         BUG();
2292         if (nidp)
2293                 *nidp = -1;
2294         return NULL;
2295 }
2296 
2297 static void hugetlb_unregister_all_nodes(void) { }
2298 
2299 static void hugetlb_register_all_nodes(void) { }
2300 
2301 #endif
2302 
2303 static void __exit hugetlb_exit(void)
2304 {
2305         struct hstate *h;
2306 
2307         hugetlb_unregister_all_nodes();
2308 
2309         for_each_hstate(h) {
2310                 kobject_put(hstate_kobjs[hstate_index(h)]);
2311         }
2312 
2313         kobject_put(hugepages_kobj);
2314         kfree(htlb_fault_mutex_table);
2315 }
2316 module_exit(hugetlb_exit);
2317 
2318 static int __init hugetlb_init(void)
2319 {
2320         int i;
2321 
2322         if (!hugepages_supported())
2323                 return 0;
2324 
2325         if (!size_to_hstate(default_hstate_size)) {
2326                 default_hstate_size = HPAGE_SIZE;
2327                 if (!size_to_hstate(default_hstate_size))
2328                         hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2329         }
2330         default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2331         if (default_hstate_max_huge_pages)
2332                 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2333 
2334         hugetlb_init_hstates();
2335         gather_bootmem_prealloc();
2336         report_hugepages();
2337 
2338         hugetlb_sysfs_init();
2339         hugetlb_register_all_nodes();
2340         hugetlb_cgroup_file_init();
2341 
2342 #ifdef CONFIG_SMP
2343         num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2344 #else
2345         num_fault_mutexes = 1;
2346 #endif
2347         htlb_fault_mutex_table =
2348                 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
2349         BUG_ON(!htlb_fault_mutex_table);
2350 
2351         for (i = 0; i < num_fault_mutexes; i++)
2352                 mutex_init(&htlb_fault_mutex_table[i]);
2353         return 0;
2354 }
2355 module_init(hugetlb_init);
2356 
2357 /* Should be called on processing a hugepagesz=... option */
2358 void __init hugetlb_add_hstate(unsigned order)
2359 {
2360         struct hstate *h;
2361         unsigned long i;
2362 
2363         if (size_to_hstate(PAGE_SIZE << order)) {
2364                 pr_warning("hugepagesz= specified twice, ignoring\n");
2365                 return;
2366         }
2367         BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2368         BUG_ON(order == 0);
2369         h = &hstates[hugetlb_max_hstate++];
2370         h->order = order;
2371         h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2372         h->nr_huge_pages = 0;
2373         h->free_huge_pages = 0;
2374         for (i = 0; i < MAX_NUMNODES; ++i)
2375                 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2376         INIT_LIST_HEAD(&h->hugepage_activelist);
2377         h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
2378         h->next_nid_to_free = first_node(node_states[N_MEMORY]);
2379         snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2380                                         huge_page_size(h)/1024);
2381 
2382         parsed_hstate = h;
2383 }
2384 
2385 static int __init hugetlb_nrpages_setup(char *s)
2386 {
2387         unsigned long *mhp;
2388         static unsigned long *last_mhp;
2389 
2390         /*
2391          * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2392          * so this hugepages= parameter goes to the "default hstate".
2393          */
2394         if (!hugetlb_max_hstate)
2395                 mhp = &default_hstate_max_huge_pages;
2396         else
2397                 mhp = &parsed_hstate->max_huge_pages;
2398 
2399         if (mhp == last_mhp) {
2400                 pr_warning("hugepages= specified twice without "
2401                            "interleaving hugepagesz=, ignoring\n");
2402                 return 1;
2403         }
2404 
2405         if (sscanf(s, "%lu", mhp) <= 0)
2406                 *mhp = 0;
2407 
2408         /*
2409          * Global state is always initialized later in hugetlb_init.
2410          * But we need to allocate >= MAX_ORDER hstates here early to still
2411          * use the bootmem allocator.
2412          */
2413         if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2414                 hugetlb_hstate_alloc_pages(parsed_hstate);
2415 
2416         last_mhp = mhp;
2417 
2418         return 1;
2419 }
2420 __setup("hugepages=", hugetlb_nrpages_setup);
2421 
2422 static int __init hugetlb_default_setup(char *s)
2423 {
2424         default_hstate_size = memparse(s, &s);
2425         return 1;
2426 }
2427 __setup("default_hugepagesz=", hugetlb_default_setup);
2428 
2429 static unsigned int cpuset_mems_nr(unsigned int *array)
2430 {
2431         int node;
2432         unsigned int nr = 0;
2433 
2434         for_each_node_mask(node, cpuset_current_mems_allowed)
2435                 nr += array[node];
2436 
2437         return nr;
2438 }
2439 
2440 #ifdef CONFIG_SYSCTL
2441 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2442                          struct ctl_table *table, int write,
2443                          void __user *buffer, size_t *length, loff_t *ppos)
2444 {
2445         struct hstate *h = &default_hstate;
2446         unsigned long tmp = h->max_huge_pages;
2447         int ret;
2448 
2449         if (!hugepages_supported())
2450                 return -ENOTSUPP;
2451 
2452         table->data = &tmp;
2453         table->maxlen = sizeof(unsigned long);
2454         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2455         if (ret)
2456                 goto out;
2457 
2458         if (write)
2459                 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2460                                                   NUMA_NO_NODE, tmp, *length);
2461 out:
2462         return ret;
2463 }
2464 
2465 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2466                           void __user *buffer, size_t *length, loff_t *ppos)
2467 {
2468 
2469         return hugetlb_sysctl_handler_common(false, table, write,
2470                                                         buffer, length, ppos);
2471 }
2472 
2473 #ifdef CONFIG_NUMA
2474 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2475                           void __user *buffer, size_t *length, loff_t *ppos)
2476 {
2477         return hugetlb_sysctl_handler_common(true, table, write,
2478                                                         buffer, length, ppos);
2479 }
2480 #endif /* CONFIG_NUMA */
2481 
2482 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2483                         void __user *buffer,
2484                         size_t *length, loff_t *ppos)
2485 {
2486         struct hstate *h = &default_hstate;
2487         unsigned long tmp;
2488         int ret;
2489 
2490         if (!hugepages_supported())
2491                 return -ENOTSUPP;
2492 
2493         tmp = h->nr_overcommit_huge_pages;
2494 
2495         if (write && hstate_is_gigantic(h))
2496                 return -EINVAL;
2497 
2498         table->data = &tmp;
2499         table->maxlen = sizeof(unsigned long);
2500         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2501         if (ret)
2502                 goto out;
2503 
2504         if (write) {
2505                 spin_lock(&hugetlb_lock);
2506                 h->nr_overcommit_huge_pages = tmp;
2507                 spin_unlock(&hugetlb_lock);
2508         }
2509 out:
2510         return ret;
2511 }
2512 
2513 #endif /* CONFIG_SYSCTL */
2514 
2515 void hugetlb_report_meminfo(struct seq_file *m)
2516 {
2517         struct hstate *h = &default_hstate;
2518         if (!hugepages_supported())
2519                 return;
2520         seq_printf(m,
2521                         "HugePages_Total:   %5lu\n"
2522                         "HugePages_Free:    %5lu\n"
2523                         "HugePages_Rsvd:    %5lu\n"
2524                         "HugePages_Surp:    %5lu\n"
2525                         "Hugepagesize:   %8lu kB\n",
2526                         h->nr_huge_pages,
2527                         h->free_huge_pages,
2528                         h->resv_huge_pages,
2529                         h->surplus_huge_pages,
2530                         1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2531 }
2532 
2533 int hugetlb_report_node_meminfo(int nid, char *buf)
2534 {
2535         struct hstate *h = &default_hstate;
2536         if (!hugepages_supported())
2537                 return 0;
2538         return sprintf(buf,
2539                 "Node %d HugePages_Total: %5u\n"
2540                 "Node %d HugePages_Free:  %5u\n"
2541                 "Node %d HugePages_Surp:  %5u\n",
2542                 nid, h->nr_huge_pages_node[nid],
2543                 nid, h->free_huge_pages_node[nid],
2544                 nid, h->surplus_huge_pages_node[nid]);
2545 }
2546 
2547 void hugetlb_show_meminfo(void)
2548 {
2549         struct hstate *h;
2550         int nid;
2551 
2552         if (!hugepages_supported())
2553                 return;
2554 
2555         for_each_node_state(nid, N_MEMORY)
2556                 for_each_hstate(h)
2557                         pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2558                                 nid,
2559                                 h->nr_huge_pages_node[nid],
2560                                 h->free_huge_pages_node[nid],
2561                                 h->surplus_huge_pages_node[nid],
2562                                 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2563 }
2564 
2565 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2566 unsigned long hugetlb_total_pages(void)
2567 {
2568         struct hstate *h;
2569         unsigned long nr_total_pages = 0;
2570 
2571         for_each_hstate(h)
2572                 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2573         return nr_total_pages;
2574 }
2575 
2576 static int hugetlb_acct_memory(struct hstate *h, long delta)
2577 {
2578         int ret = -ENOMEM;
2579 
2580         spin_lock(&hugetlb_lock);
2581         /*
2582          * When cpuset is configured, it breaks the strict hugetlb page
2583          * reservation as the accounting is done on a global variable. Such
2584          * reservation is completely rubbish in the presence of cpuset because
2585          * the reservation is not checked against page availability for the
2586          * current cpuset. Application can still potentially OOM'ed by kernel
2587          * with lack of free htlb page in cpuset that the task is in.
2588          * Attempt to enforce strict accounting with cpuset is almost
2589          * impossible (or too ugly) because cpuset is too fluid that
2590          * task or memory node can be dynamically moved between cpusets.
2591          *
2592          * The change of semantics for shared hugetlb mapping with cpuset is
2593          * undesirable. However, in order to preserve some of the semantics,
2594          * we fall back to check against current free page availability as
2595          * a best attempt and hopefully to minimize the impact of changing
2596          * semantics that cpuset has.
2597          */
2598         if (delta > 0) {
2599                 if (gather_surplus_pages(h, delta) < 0)
2600                         goto out;
2601 
2602                 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2603                         return_unused_surplus_pages(h, delta);
2604                         goto out;
2605                 }
2606         }
2607 
2608         ret = 0;
2609         if (delta < 0)
2610                 return_unused_surplus_pages(h, (unsigned long) -delta);
2611 
2612 out:
2613         spin_unlock(&hugetlb_lock);
2614         return ret;
2615 }
2616 
2617 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2618 {
2619         struct resv_map *resv = vma_resv_map(vma);
2620 
2621         /*
2622          * This new VMA should share its siblings reservation map if present.
2623          * The VMA will only ever have a valid reservation map pointer where
2624          * it is being copied for another still existing VMA.  As that VMA
2625          * has a reference to the reservation map it cannot disappear until
2626          * after this open call completes.  It is therefore safe to take a
2627          * new reference here without additional locking.
2628          */
2629         if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2630                 kref_get(&resv->refs);
2631 }
2632 
2633 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2634 {
2635         struct hstate *h = hstate_vma(vma);
2636         struct resv_map *resv = vma_resv_map(vma);
2637         struct hugepage_subpool *spool = subpool_vma(vma);
2638         unsigned long reserve, start, end;
2639         long gbl_reserve;
2640 
2641         if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2642                 return;
2643 
2644         start = vma_hugecache_offset(h, vma, vma->vm_start);
2645         end = vma_hugecache_offset(h, vma, vma->vm_end);
2646 
2647         reserve = (end - start) - region_count(resv, start, end);
2648 
2649         kref_put(&resv->refs, resv_map_release);
2650 
2651         if (reserve) {
2652                 /*
2653                  * Decrement reserve counts.  The global reserve count may be
2654                  * adjusted if the subpool has a minimum size.
2655                  */
2656                 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
2657                 hugetlb_acct_memory(h, -gbl_reserve);
2658         }
2659 }
2660 
2661 /*
2662  * We cannot handle pagefaults against hugetlb pages at all.  They cause
2663  * handle_mm_fault() to try to instantiate regular-sized pages in the
2664  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
2665  * this far.
2666  */
2667 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2668 {
2669         BUG();
2670         return 0;
2671 }
2672 
2673 const struct vm_operations_struct hugetlb_vm_ops = {
2674         .fault = hugetlb_vm_op_fault,
2675         .open = hugetlb_vm_op_open,
2676         .close = hugetlb_vm_op_close,
2677 };
2678 
2679 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2680                                 int writable)
2681 {
2682         pte_t entry;
2683 
2684         if (writable) {
2685                 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2686                                          vma->vm_page_prot)));
2687         } else {
2688                 entry = huge_pte_wrprotect(mk_huge_pte(page,
2689                                            vma->vm_page_prot));
2690         }
2691         entry = pte_mkyoung(entry);
2692         entry = pte_mkhuge(entry);
2693         entry = arch_make_huge_pte(entry, vma, page, writable);
2694 
2695         return entry;
2696 }
2697 
2698 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2699                                    unsigned long address, pte_t *ptep)
2700 {
2701         pte_t entry;
2702 
2703         entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
2704         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2705                 update_mmu_cache(vma, address, ptep);
2706 }
2707 
2708 static int is_hugetlb_entry_migration(pte_t pte)
2709 {
2710         swp_entry_t swp;
2711 
2712         if (huge_pte_none(pte) || pte_present(pte))
2713                 return 0;
2714         swp = pte_to_swp_entry(pte);
2715         if (non_swap_entry(swp) && is_migration_entry(swp))
2716                 return 1;
2717         else
2718                 return 0;
2719 }
2720 
2721 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2722 {
2723         swp_entry_t swp;
2724 
2725         if (huge_pte_none(pte) || pte_present(pte))
2726                 return 0;
2727         swp = pte_to_swp_entry(pte);
2728         if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2729                 return 1;
2730         else
2731                 return 0;
2732 }
2733 
2734 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2735                             struct vm_area_struct *vma)
2736 {
2737         pte_t *src_pte, *dst_pte, entry;
2738         struct page *ptepage;
2739         unsigned long addr;
2740         int cow;
2741         struct hstate *h = hstate_vma(vma);
2742         unsigned long sz = huge_page_size(h);
2743         unsigned long mmun_start;       /* For mmu_notifiers */
2744         unsigned long mmun_end;         /* For mmu_notifiers */
2745         int ret = 0;
2746 
2747         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2748 
2749         mmun_start = vma->vm_start;
2750         mmun_end = vma->vm_end;
2751         if (cow)
2752                 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
2753 
2754         for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2755                 spinlock_t *src_ptl, *dst_ptl;
2756                 src_pte = huge_pte_offset(src, addr);
2757                 if (!src_pte)
2758                         continue;
2759                 dst_pte = huge_pte_alloc(dst, addr, sz);
2760                 if (!dst_pte) {
2761                         ret = -ENOMEM;
2762                         break;
2763                 }
2764 
2765                 /* If the pagetables are shared don't copy or take references */
2766                 if (dst_pte == src_pte)
2767                         continue;
2768 
2769                 dst_ptl = huge_pte_lock(h, dst, dst_pte);
2770                 src_ptl = huge_pte_lockptr(h, src, src_pte);
2771                 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
2772                 entry = huge_ptep_get(src_pte);
2773                 if (huge_pte_none(entry)) { /* skip none entry */
2774                         ;
2775                 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
2776                                     is_hugetlb_entry_hwpoisoned(entry))) {
2777                         swp_entry_t swp_entry = pte_to_swp_entry(entry);
2778 
2779                         if (is_write_migration_entry(swp_entry) && cow) {
2780                                 /*
2781                                  * COW mappings require pages in both
2782                                  * parent and child to be set to read.
2783                                  */
2784                                 make_migration_entry_read(&swp_entry);
2785                                 entry = swp_entry_to_pte(swp_entry);
2786                                 set_huge_pte_at(src, addr, src_pte, entry);
2787                         }
2788                         set_huge_pte_at(dst, addr, dst_pte, entry);
2789                 } else {
2790                         if (cow) {
2791                                 huge_ptep_set_wrprotect(src, addr, src_pte);
2792                                 mmu_notifier_invalidate_range(src, mmun_start,
2793                                                                    mmun_end);
2794                         }
2795                         entry = huge_ptep_get(src_pte);
2796                         ptepage = pte_page(entry);
2797                         get_page(ptepage);
2798                         page_dup_rmap(ptepage);
2799                         set_huge_pte_at(dst, addr, dst_pte, entry);
2800                 }
2801                 spin_unlock(src_ptl);
2802                 spin_unlock(dst_ptl);
2803         }
2804 
2805         if (cow)
2806                 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
2807 
2808         return ret;
2809 }
2810 
2811 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2812                             unsigned long start, unsigned long end,
2813                             struct page *ref_page)
2814 {
2815         int force_flush = 0;
2816         struct mm_struct *mm = vma->vm_mm;
2817         unsigned long address;
2818         pte_t *ptep;
2819         pte_t pte;
2820         spinlock_t *ptl;
2821         struct page *page;
2822         struct hstate *h = hstate_vma(vma);
2823         unsigned long sz = huge_page_size(h);
2824         const unsigned long mmun_start = start; /* For mmu_notifiers */
2825         const unsigned long mmun_end   = end;   /* For mmu_notifiers */
2826 
2827         WARN_ON(!is_vm_hugetlb_page(vma));
2828         BUG_ON(start & ~huge_page_mask(h));
2829         BUG_ON(end & ~huge_page_mask(h));
2830 
2831         tlb_start_vma(tlb, vma);
2832         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2833         address = start;
2834 again:
2835         for (; address < end; address += sz) {
2836                 ptep = huge_pte_offset(mm, address);
2837                 if (!ptep)
2838                         continue;
2839 
2840                 ptl = huge_pte_lock(h, mm, ptep);
2841                 if (huge_pmd_unshare(mm, &address, ptep))
2842                         goto unlock;
2843 
2844                 pte = huge_ptep_get(ptep);
2845                 if (huge_pte_none(pte))
2846                         goto unlock;
2847 
2848                 /*
2849                  * Migrating hugepage or HWPoisoned hugepage is already
2850                  * unmapped and its refcount is dropped, so just clear pte here.
2851                  */
2852                 if (unlikely(!pte_present(pte))) {
2853                         huge_pte_clear(mm, address, ptep);
2854                         goto unlock;
2855                 }
2856 
2857                 page = pte_page(pte);
2858                 /*
2859                  * If a reference page is supplied, it is because a specific
2860                  * page is being unmapped, not a range. Ensure the page we
2861                  * are about to unmap is the actual page of interest.
2862                  */
2863                 if (ref_page) {
2864                         if (page != ref_page)
2865                                 goto unlock;
2866 
2867                         /*
2868                          * Mark the VMA as having unmapped its page so that
2869                          * future faults in this VMA will fail rather than
2870                          * looking like data was lost
2871                          */
2872                         set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2873                 }
2874 
2875                 pte = huge_ptep_get_and_clear(mm, address, ptep);
2876                 tlb_remove_tlb_entry(tlb, ptep, address);
2877                 if (huge_pte_dirty(pte))
2878                         set_page_dirty(page);
2879 
2880                 page_remove_rmap(page);
2881                 force_flush = !__tlb_remove_page(tlb, page);
2882                 if (force_flush) {
2883                         address += sz;
2884                         spin_unlock(ptl);
2885                         break;
2886                 }
2887                 /* Bail out after unmapping reference page if supplied */
2888                 if (ref_page) {
2889                         spin_unlock(ptl);
2890                         break;
2891                 }
2892 unlock:
2893                 spin_unlock(ptl);
2894         }
2895         /*
2896          * mmu_gather ran out of room to batch pages, we break out of
2897          * the PTE lock to avoid doing the potential expensive TLB invalidate
2898          * and page-free while holding it.
2899          */
2900         if (force_flush) {
2901                 force_flush = 0;
2902                 tlb_flush_mmu(tlb);
2903                 if (address < end && !ref_page)
2904                         goto again;
2905         }
2906         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2907         tlb_end_vma(tlb, vma);
2908 }
2909 
2910 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
2911                           struct vm_area_struct *vma, unsigned long start,
2912                           unsigned long end, struct page *ref_page)
2913 {
2914         __unmap_hugepage_range(tlb, vma, start, end, ref_page);
2915 
2916         /*
2917          * Clear this flag so that x86's huge_pmd_share page_table_shareable
2918          * test will fail on a vma being torn down, and not grab a page table
2919          * on its way out.  We're lucky that the flag has such an appropriate
2920          * name, and can in fact be safely cleared here. We could clear it
2921          * before the __unmap_hugepage_range above, but all that's necessary
2922          * is to clear it before releasing the i_mmap_rwsem. This works
2923          * because in the context this is called, the VMA is about to be
2924          * destroyed and the i_mmap_rwsem is held.
2925          */
2926         vma->vm_flags &= ~VM_MAYSHARE;
2927 }
2928 
2929 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2930                           unsigned long end, struct page *ref_page)
2931 {
2932         struct mm_struct *mm;
2933         struct mmu_gather tlb;
2934 
2935         mm = vma->vm_mm;
2936 
2937         tlb_gather_mmu(&tlb, mm, start, end);
2938         __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2939         tlb_finish_mmu(&tlb, start, end);
2940 }
2941 
2942 /*
2943  * This is called when the original mapper is failing to COW a MAP_PRIVATE
2944  * mappping it owns the reserve page for. The intention is to unmap the page
2945  * from other VMAs and let the children be SIGKILLed if they are faulting the
2946  * same region.
2947  */
2948 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2949                               struct page *page, unsigned long address)
2950 {
2951         struct hstate *h = hstate_vma(vma);
2952         struct vm_area_struct *iter_vma;
2953         struct address_space *mapping;
2954         pgoff_t pgoff;
2955 
2956         /*
2957          * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2958          * from page cache lookup which is in HPAGE_SIZE units.
2959          */
2960         address = address & huge_page_mask(h);
2961         pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2962                         vma->vm_pgoff;
2963         mapping = file_inode(vma->vm_file)->i_mapping;
2964 
2965         /*
2966          * Take the mapping lock for the duration of the table walk. As
2967          * this mapping should be shared between all the VMAs,
2968          * __unmap_hugepage_range() is called as the lock is already held
2969          */
2970         i_mmap_lock_write(mapping);
2971         vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
2972                 /* Do not unmap the current VMA */
2973                 if (iter_vma == vma)
2974                         continue;
2975 
2976                 /*
2977                  * Shared VMAs have their own reserves and do not affect
2978                  * MAP_PRIVATE accounting but it is possible that a shared
2979                  * VMA is using the same page so check and skip such VMAs.
2980                  */
2981                 if (iter_vma->vm_flags & VM_MAYSHARE)
2982                         continue;
2983 
2984                 /*
2985                  * Unmap the page from other VMAs without their own reserves.
2986                  * They get marked to be SIGKILLed if they fault in these
2987                  * areas. This is because a future no-page fault on this VMA
2988                  * could insert a zeroed page instead of the data existing
2989                  * from the time of fork. This would look like data corruption
2990                  */
2991                 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2992                         unmap_hugepage_range(iter_vma, address,
2993                                              address + huge_page_size(h), page);
2994         }
2995         i_mmap_unlock_write(mapping);
2996 }
2997 
2998 /*
2999  * Hugetlb_cow() should be called with page lock of the original hugepage held.
3000  * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3001  * cannot race with other handlers or page migration.
3002  * Keep the pte_same checks anyway to make transition from the mutex easier.
3003  */
3004 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3005                         unsigned long address, pte_t *ptep, pte_t pte,
3006                         struct page *pagecache_page, spinlock_t *ptl)
3007 {
3008         struct hstate *h = hstate_vma(vma);
3009         struct page *old_page, *new_page;
3010         int ret = 0, outside_reserve = 0;
3011         unsigned long mmun_start;       /* For mmu_notifiers */
3012         unsigned long mmun_end;         /* For mmu_notifiers */
3013 
3014         old_page = pte_page(pte);
3015 
3016 retry_avoidcopy:
3017         /* If no-one else is actually using this page, avoid the copy
3018          * and just make the page writable */
3019         if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3020                 page_move_anon_rmap(old_page, vma, address);
3021                 set_huge_ptep_writable(vma, address, ptep);
3022                 return 0;
3023         }
3024 
3025         /*
3026          * If the process that created a MAP_PRIVATE mapping is about to
3027          * perform a COW due to a shared page count, attempt to satisfy
3028          * the allocation without using the existing reserves. The pagecache
3029          * page is used to determine if the reserve at this address was
3030          * consumed or not. If reserves were used, a partial faulted mapping
3031          * at the time of fork() could consume its reserves on COW instead
3032          * of the full address range.
3033          */
3034         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3035                         old_page != pagecache_page)
3036                 outside_reserve = 1;
3037 
3038         page_cache_get(old_page);
3039 
3040         /*
3041          * Drop page table lock as buddy allocator may be called. It will
3042          * be acquired again before returning to the caller, as expected.
3043          */
3044         spin_unlock(ptl);
3045         new_page = alloc_huge_page(vma, address, outside_reserve);
3046 
3047         if (IS_ERR(new_page)) {
3048                 /*
3049                  * If a process owning a MAP_PRIVATE mapping fails to COW,
3050                  * it is due to references held by a child and an insufficient
3051                  * huge page pool. To guarantee the original mappers
3052                  * reliability, unmap the page from child processes. The child
3053                  * may get SIGKILLed if it later faults.
3054                  */
3055                 if (outside_reserve) {
3056                         page_cache_release(old_page);
3057                         BUG_ON(huge_pte_none(pte));
3058                         unmap_ref_private(mm, vma, old_page, address);
3059                         BUG_ON(huge_pte_none(pte));
3060                         spin_lock(ptl);
3061                         ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3062                         if (likely(ptep &&
3063                                    pte_same(huge_ptep_get(ptep), pte)))
3064                                 goto retry_avoidcopy;
3065                         /*
3066                          * race occurs while re-acquiring page table
3067                          * lock, and our job is done.
3068                          */
3069                         return 0;
3070                 }
3071 
3072                 ret = (PTR_ERR(new_page) == -ENOMEM) ?
3073                         VM_FAULT_OOM : VM_FAULT_SIGBUS;
3074                 goto out_release_old;
3075         }
3076 
3077         /*
3078          * When the original hugepage is shared one, it does not have
3079          * anon_vma prepared.
3080          */
3081         if (unlikely(anon_vma_prepare(vma))) {
3082                 ret = VM_FAULT_OOM;
3083                 goto out_release_all;
3084         }
3085 
3086         copy_user_huge_page(new_page, old_page, address, vma,
3087                             pages_per_huge_page(h));
3088         __SetPageUptodate(new_page);
3089         set_page_huge_active(new_page);
3090 
3091         mmun_start = address & huge_page_mask(h);
3092         mmun_end = mmun_start + huge_page_size(h);
3093         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3094 
3095         /*
3096          * Retake the page table lock to check for racing updates
3097          * before the page tables are altered
3098          */
3099         spin_lock(ptl);
3100         ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3101         if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3102                 ClearPagePrivate(new_page);
3103 
3104                 /* Break COW */
3105                 huge_ptep_clear_flush(vma, address, ptep);
3106                 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
3107                 set_huge_pte_at(mm, address, ptep,
3108                                 make_huge_pte(vma, new_page, 1));
3109                 page_remove_rmap(old_page);
3110                 hugepage_add_new_anon_rmap(new_page, vma, address);
3111                 /* Make the old page be freed below */
3112                 new_page = old_page;
3113         }
3114         spin_unlock(ptl);
3115         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3116 out_release_all:
3117         page_cache_release(new_page);
3118 out_release_old:
3119         page_cache_release(old_page);
3120 
3121         spin_lock(ptl); /* Caller expects lock to be held */
3122         return ret;
3123 }
3124 
3125 /* Return the pagecache page at a given address within a VMA */
3126 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3127                         struct vm_area_struct *vma, unsigned long address)
3128 {
3129         struct address_space *mapping;
3130         pgoff_t idx;
3131 
3132         mapping = vma->vm_file->f_mapping;
3133         idx = vma_hugecache_offset(h, vma, address);
3134 
3135         return find_lock_page(mapping, idx);
3136 }
3137 
3138 /*
3139  * Return whether there is a pagecache page to back given address within VMA.
3140  * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3141  */
3142 static bool hugetlbfs_pagecache_present(struct hstate *h,
3143                         struct vm_area_struct *vma, unsigned long address)
3144 {
3145         struct address_space *mapping;
3146         pgoff_t idx;
3147         struct page *page;
3148 
3149         mapping = vma->vm_file->f_mapping;
3150         idx = vma_hugecache_offset(h, vma, address);
3151 
3152         page = find_get_page(mapping, idx);
3153         if (page)
3154                 put_page(page);
3155         return page != NULL;
3156 }
3157 
3158 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
3159                            struct address_space *mapping, pgoff_t idx,
3160                            unsigned long address, pte_t *ptep, unsigned int flags)
3161 {
3162         struct hstate *h = hstate_vma(vma);
3163         int ret = VM_FAULT_SIGBUS;
3164         int anon_rmap = 0;
3165         unsigned long size;
3166         struct page *page;
3167         pte_t new_pte;
3168         spinlock_t *ptl;
3169 
3170         /*
3171          * Currently, we are forced to kill the process in the event the
3172          * original mapper has unmapped pages from the child due to a failed
3173          * COW. Warn that such a situation has occurred as it may not be obvious
3174          */
3175         if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3176                 pr_warning("PID %d killed due to inadequate hugepage pool\n",
3177                            current->pid);
3178                 return ret;
3179         }
3180 
3181         /*
3182          * Use page lock to guard against racing truncation
3183          * before we get page_table_lock.
3184          */
3185 retry:
3186         page = find_lock_page(mapping, idx);
3187         if (!page) {
3188                 size = i_size_read(mapping->host) >> huge_page_shift(h);
3189                 if (idx >= size)
3190                         goto out;
3191                 page = alloc_huge_page(vma, address, 0);
3192                 if (IS_ERR(page)) {
3193                         ret = PTR_ERR(page);
3194                         if (ret == -ENOMEM)
3195                                 ret = VM_FAULT_OOM;
3196                         else
3197                                 ret = VM_FAULT_SIGBUS;
3198                         goto out;
3199                 }
3200                 clear_huge_page(page, address, pages_per_huge_page(h));
3201                 __SetPageUptodate(page);
3202                 set_page_huge_active(page);
3203 
3204                 if (vma->vm_flags & VM_MAYSHARE) {
3205                         int err;
3206                         struct inode *inode = mapping->host;
3207 
3208                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3209                         if (err) {
3210                                 put_page(page);
3211                                 if (err == -EEXIST)
3212                                         goto retry;
3213                                 goto out;
3214                         }
3215                         ClearPagePrivate(page);
3216 
3217                         spin_lock(&inode->i_lock);
3218                         inode->i_blocks += blocks_per_huge_page(h);
3219                         spin_unlock(&inode->i_lock);
3220                 } else {
3221                         lock_page(page);
3222                         if (unlikely(anon_vma_prepare(vma))) {
3223                                 ret = VM_FAULT_OOM;
3224                                 goto backout_unlocked;
3225                         }
3226                         anon_rmap = 1;
3227                 }
3228         } else {
3229                 /*
3230                  * If memory error occurs between mmap() and fault, some process
3231                  * don't have hwpoisoned swap entry for errored virtual address.
3232                  * So we need to block hugepage fault by PG_hwpoison bit check.
3233                  */
3234                 if (unlikely(PageHWPoison(page))) {
3235                         ret = VM_FAULT_HWPOISON |
3236                                 VM_FAULT_SET_HINDEX(hstate_index(h));
3237                         goto backout_unlocked;
3238                 }
3239         }
3240 
3241         /*
3242          * If we are going to COW a private mapping later, we examine the
3243          * pending reservations for this page now. This will ensure that
3244          * any allocations necessary to record that reservation occur outside
3245          * the spinlock.
3246          */
3247         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
3248                 if (vma_needs_reservation(h, vma, address) < 0) {
3249                         ret = VM_FAULT_OOM;
3250                         goto backout_unlocked;
3251                 }
3252 
3253         ptl = huge_pte_lockptr(h, mm, ptep);
3254         spin_lock(ptl);
3255         size = i_size_read(mapping->host) >> huge_page_shift(h);
3256         if (idx >= size)
3257                 goto backout;
3258 
3259         ret = 0;
3260         if (!huge_pte_none(huge_ptep_get(ptep)))
3261                 goto backout;
3262 
3263         if (anon_rmap) {
3264                 ClearPagePrivate(page);
3265                 hugepage_add_new_anon_rmap(page, vma, address);
3266         } else
3267                 page_dup_rmap(page);
3268         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3269                                 && (vma->vm_flags & VM_SHARED)));
3270         set_huge_pte_at(mm, address, ptep, new_pte);
3271 
3272         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3273                 /* Optimization, do the COW without a second fault */
3274                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl);
3275         }
3276 
3277         spin_unlock(ptl);
3278         unlock_page(page);
3279 out:
3280         return ret;
3281 
3282 backout:
3283         spin_unlock(ptl);
3284 backout_unlocked:
3285         unlock_page(page);
3286         put_page(page);
3287         goto out;
3288 }
3289 
3290 #ifdef CONFIG_SMP
3291 static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3292                             struct vm_area_struct *vma,
3293                             struct address_space *mapping,
3294                             pgoff_t idx, unsigned long address)
3295 {
3296         unsigned long key[2];
3297         u32 hash;
3298 
3299         if (vma->vm_flags & VM_SHARED) {
3300                 key[0] = (unsigned long) mapping;
3301                 key[1] = idx;
3302         } else {
3303                 key[0] = (unsigned long) mm;
3304                 key[1] = address >> huge_page_shift(h);
3305         }
3306 
3307         hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3308 
3309         return hash & (num_fault_mutexes - 1);
3310 }
3311 #else
3312 /*
3313  * For uniprocesor systems we always use a single mutex, so just
3314  * return 0 and avoid the hashing overhead.
3315  */
3316 static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3317                             struct vm_area_struct *vma,
3318                             struct address_space *mapping,
3319                             pgoff_t idx, unsigned long address)
3320 {
3321         return 0;
3322 }
3323 #endif
3324 
3325 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3326                         unsigned long address, unsigned int flags)
3327 {
3328         pte_t *ptep, entry;
3329         spinlock_t *ptl;
3330         int ret;
3331         u32 hash;
3332         pgoff_t idx;
3333         struct page *page = NULL;
3334         struct page *pagecache_page = NULL;
3335         struct hstate *h = hstate_vma(vma);
3336         struct address_space *mapping;
3337         int need_wait_lock = 0;
3338 
3339         address &= huge_page_mask(h);
3340 
3341         ptep = huge_pte_offset(mm, address);
3342         if (ptep) {
3343                 entry = huge_ptep_get(ptep);
3344                 if (unlikely(is_hugetlb_entry_migration(entry))) {
3345                         migration_entry_wait_huge(vma, mm, ptep);
3346                         return 0;
3347                 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
3348                         return VM_FAULT_HWPOISON_LARGE |
3349                                 VM_FAULT_SET_HINDEX(hstate_index(h));
3350         }
3351 
3352         ptep = huge_pte_alloc(mm, address, huge_page_size(h));
3353         if (!ptep)
3354                 return VM_FAULT_OOM;
3355 
3356         mapping = vma->vm_file->f_mapping;
3357         idx = vma_hugecache_offset(h, vma, address);
3358 
3359         /*
3360          * Serialize hugepage allocation and instantiation, so that we don't
3361          * get spurious allocation failures if two CPUs race to instantiate
3362          * the same page in the page cache.
3363          */
3364         hash = fault_mutex_hash(h, mm, vma, mapping, idx, address);
3365         mutex_lock(&htlb_fault_mutex_table[hash]);
3366 
3367         entry = huge_ptep_get(ptep);
3368         if (huge_pte_none(entry)) {
3369                 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
3370                 goto out_mutex;
3371         }
3372 
3373         ret = 0;
3374 
3375         /*
3376          * entry could be a migration/hwpoison entry at this point, so this
3377          * check prevents the kernel from going below assuming that we have
3378          * a active hugepage in pagecache. This goto expects the 2nd page fault,
3379          * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3380          * handle it.
3381          */
3382         if (!pte_present(entry))
3383                 goto out_mutex;
3384 
3385         /*
3386          * If we are going to COW the mapping later, we examine the pending
3387          * reservations for this page now. This will ensure that any
3388          * allocations necessary to record that reservation occur outside the
3389          * spinlock. For private mappings, we also lookup the pagecache
3390          * page now as it is used to determine if a reservation has been
3391          * consumed.
3392          */
3393         if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
3394                 if (vma_needs_reservation(h, vma, address) < 0) {
3395                         ret = VM_FAULT_OOM;
3396                         goto out_mutex;
3397                 }
3398 
3399                 if (!(vma->vm_flags & VM_MAYSHARE))
3400                         pagecache_page = hugetlbfs_pagecache_page(h,
3401                                                                 vma, address);
3402         }
3403 
3404         ptl = huge_pte_lock(h, mm, ptep);
3405 
3406         /* Check for a racing update before calling hugetlb_cow */
3407         if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
3408                 goto out_ptl;
3409 
3410         /*
3411          * hugetlb_cow() requires page locks of pte_page(entry) and
3412          * pagecache_page, so here we need take the former one
3413          * when page != pagecache_page or !pagecache_page.
3414          */
3415         page = pte_page(entry);
3416         if (page != pagecache_page)
3417                 if (!trylock_page(page)) {
3418                         need_wait_lock = 1;
3419                         goto out_ptl;
3420                 }
3421 
3422         get_page(page);
3423 
3424         if (flags & FAULT_FLAG_WRITE) {
3425                 if (!huge_pte_write(entry)) {
3426                         ret = hugetlb_cow(mm, vma, address, ptep, entry,
3427                                         pagecache_page, ptl);
3428                         goto out_put_page;
3429                 }
3430                 entry = huge_pte_mkdirty(entry);
3431         }
3432         entry = pte_mkyoung(entry);
3433         if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3434                                                 flags & FAULT_FLAG_WRITE))
3435                 update_mmu_cache(vma, address, ptep);
3436 out_put_page:
3437         if (page != pagecache_page)
3438                 unlock_page(page);
3439         put_page(page);
3440 out_ptl:
3441         spin_unlock(ptl);
3442 
3443         if (pagecache_page) {
3444                 unlock_page(pagecache_page);
3445                 put_page(pagecache_page);
3446         }
3447 out_mutex:
3448         mutex_unlock(&htlb_fault_mutex_table[hash]);
3449         /*
3450          * Generally it's safe to hold refcount during waiting page lock. But
3451          * here we just wait to defer the next page fault to avoid busy loop and
3452          * the page is not used after unlocked before returning from the current
3453          * page fault. So we are safe from accessing freed page, even if we wait
3454          * here without taking refcount.
3455          */
3456         if (need_wait_lock)
3457                 wait_on_page_locked(page);
3458         return ret;
3459 }
3460 
3461 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3462                          struct page **pages, struct vm_area_struct **vmas,
3463                          unsigned long *position, unsigned long *nr_pages,
3464                          long i, unsigned int flags)
3465 {
3466         unsigned long pfn_offset;
3467         unsigned long vaddr = *position;
3468         unsigned long remainder = *nr_pages;
3469         struct hstate *h = hstate_vma(vma);
3470 
3471         while (vaddr < vma->vm_end && remainder) {
3472                 pte_t *pte;
3473                 spinlock_t *ptl = NULL;
3474                 int absent;
3475                 struct page *page;
3476 
3477                 /*
3478                  * If we have a pending SIGKILL, don't keep faulting pages and
3479                  * potentially allocating memory.
3480                  */
3481                 if (unlikely(fatal_signal_pending(current))) {
3482                         remainder = 0;
3483                         break;
3484                 }
3485 
3486                 /*
3487                  * Some archs (sparc64, sh*) have multiple pte_ts to
3488                  * each hugepage.  We have to make sure we get the
3489                  * first, for the page indexing below to work.
3490                  *
3491                  * Note that page table lock is not held when pte is null.
3492                  */
3493                 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
3494                 if (pte)
3495                         ptl = huge_pte_lock(h, mm, pte);
3496                 absent = !pte || huge_pte_none(huge_ptep_get(pte));
3497 
3498                 /*
3499                  * When coredumping, it suits get_dump_page if we just return
3500                  * an error where there's an empty slot with no huge pagecache
3501                  * to back it.  This way, we avoid allocating a hugepage, and
3502                  * the sparse dumpfile avoids allocating disk blocks, but its
3503                  * huge holes still show up with zeroes where they need to be.
3504                  */
3505                 if (absent && (flags & FOLL_DUMP) &&
3506                     !hugetlbfs_pagecache_present(h, vma, vaddr)) {
3507                         if (pte)
3508                                 spin_unlock(ptl);
3509                         remainder = 0;
3510                         break;
3511                 }
3512 
3513                 /*
3514                  * We need call hugetlb_fault for both hugepages under migration
3515                  * (in which case hugetlb_fault waits for the migration,) and
3516                  * hwpoisoned hugepages (in which case we need to prevent the
3517                  * caller from accessing to them.) In order to do this, we use
3518                  * here is_swap_pte instead of is_hugetlb_entry_migration and
3519                  * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3520                  * both cases, and because we can't follow correct pages
3521                  * directly from any kind of swap entries.
3522                  */
3523                 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
3524                     ((flags & FOLL_WRITE) &&
3525                       !huge_pte_write(huge_ptep_get(pte)))) {
3526                         int ret;
3527 
3528                         if (pte)
3529                                 spin_unlock(ptl);
3530                         ret = hugetlb_fault(mm, vma, vaddr,
3531                                 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
3532                         if (!(ret & VM_FAULT_ERROR))
3533                                 continue;
3534 
3535                         remainder = 0;
3536                         break;
3537                 }
3538 
3539                 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
3540                 page = pte_page(huge_ptep_get(pte));
3541 same_page:
3542                 if (pages) {
3543                         pages[i] = mem_map_offset(page, pfn_offset);
3544                         get_page_foll(pages[i]);
3545                 }
3546 
3547                 if (vmas)
3548                         vmas[i] = vma;
3549 
3550                 vaddr += PAGE_SIZE;
3551                 ++pfn_offset;
3552                 --remainder;
3553                 ++i;
3554                 if (vaddr < vma->vm_end && remainder &&
3555                                 pfn_offset < pages_per_huge_page(h)) {
3556                         /*
3557                          * We use pfn_offset to avoid touching the pageframes
3558                          * of this compound page.
3559                          */
3560                         goto same_page;
3561                 }
3562                 spin_unlock(ptl);
3563         }
3564         *nr_pages = remainder;
3565         *position = vaddr;
3566 
3567         return i ? i : -EFAULT;
3568 }
3569 
3570 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
3571                 unsigned long address, unsigned long end, pgprot_t newprot)
3572 {
3573         struct mm_struct *mm = vma->vm_mm;
3574         unsigned long start = address;
3575         pte_t *ptep;
3576         pte_t pte;
3577         struct hstate *h = hstate_vma(vma);
3578         unsigned long pages = 0;
3579 
3580         BUG_ON(address >= end);
3581         flush_cache_range(vma, address, end);
3582 
3583         mmu_notifier_invalidate_range_start(mm, start, end);
3584         i_mmap_lock_write(vma->vm_file->f_mapping);
3585         for (; address < end; address += huge_page_size(h)) {
3586                 spinlock_t *ptl;
3587                 ptep = huge_pte_offset(mm, address);
3588                 if (!ptep)
3589                         continue;
3590                 ptl = huge_pte_lock(h, mm, ptep);
3591                 if (huge_pmd_unshare(mm, &address, ptep)) {
3592                         pages++;
3593                         spin_unlock(ptl);
3594                         continue;
3595                 }
3596                 pte = huge_ptep_get(ptep);
3597                 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
3598                         spin_unlock(ptl);
3599                         continue;
3600                 }
3601                 if (unlikely(is_hugetlb_entry_migration(pte))) {
3602                         swp_entry_t entry = pte_to_swp_entry(pte);
3603 
3604                         if (is_write_migration_entry(entry)) {
3605                                 pte_t newpte;
3606 
3607                                 make_migration_entry_read(&entry);
3608                                 newpte = swp_entry_to_pte(entry);
3609                                 set_huge_pte_at(mm, address, ptep, newpte);
3610                                 pages++;
3611                         }
3612                         spin_unlock(ptl);
3613                         continue;
3614                 }
3615                 if (!huge_pte_none(pte)) {
3616                         pte = huge_ptep_get_and_clear(mm, address, ptep);
3617                         pte = pte_mkhuge(huge_pte_modify(pte, newprot));
3618                         pte = arch_make_huge_pte(pte, vma, NULL, 0);
3619                         set_huge_pte_at(mm, address, ptep, pte);
3620                         pages++;
3621                 }
3622                 spin_unlock(ptl);
3623         }
3624         /*
3625          * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
3626          * may have cleared our pud entry and done put_page on the page table:
3627          * once we release i_mmap_rwsem, another task can do the final put_page
3628          * and that page table be reused and filled with junk.
3629          */
3630         flush_tlb_range(vma, start, end);
3631         mmu_notifier_invalidate_range(mm, start, end);
3632         i_mmap_unlock_write(vma->vm_file->f_mapping);
3633         mmu_notifier_invalidate_range_end(mm, start, end);
3634 
3635         return pages << h->order;
3636 }
3637 
3638 int hugetlb_reserve_pages(struct inode *inode,
3639                                         long from, long to,
3640                                         struct vm_area_struct *vma,
3641                                         vm_flags_t vm_flags)
3642 {
3643         long ret, chg;
3644         struct hstate *h = hstate_inode(inode);
3645         struct hugepage_subpool *spool = subpool_inode(inode);
3646         struct resv_map *resv_map;
3647         long gbl_reserve;
3648 
3649         /*
3650          * Only apply hugepage reservation if asked. At fault time, an
3651          * attempt will be made for VM_NORESERVE to allocate a page
3652          * without using reserves
3653          */
3654         if (vm_flags & VM_NORESERVE)
3655                 return 0;
3656 
3657         /*
3658          * Shared mappings base their reservation on the number of pages that
3659          * are already allocated on behalf of the file. Private mappings need
3660          * to reserve the full area even if read-only as mprotect() may be
3661          * called to make the mapping read-write. Assume !vma is a shm mapping
3662          */
3663         if (!vma || vma->vm_flags & VM_MAYSHARE) {
3664                 resv_map = inode_resv_map(inode);
3665 
3666                 chg = region_chg(resv_map, from, to);
3667 
3668         } else {
3669                 resv_map = resv_map_alloc();
3670                 if (!resv_map)
3671                         return -ENOMEM;
3672 
3673                 chg = to - from;
3674 
3675                 set_vma_resv_map(vma, resv_map);
3676                 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3677         }
3678 
3679         if (chg < 0) {
3680                 ret = chg;
3681                 goto out_err;
3682         }
3683 
3684         /*
3685          * There must be enough pages in the subpool for the mapping. If
3686          * the subpool has a minimum size, there may be some global
3687          * reservations already in place (gbl_reserve).
3688          */
3689         gbl_reserve = hugepage_subpool_get_pages(spool, chg);
3690         if (gbl_reserve < 0) {
3691                 ret = -ENOSPC;
3692                 goto out_err;
3693         }
3694 
3695         /*
3696          * Check enough hugepages are available for the reservation.
3697          * Hand the pages back to the subpool if there are not
3698          */
3699         ret = hugetlb_acct_memory(h, gbl_reserve);
3700         if (ret < 0) {
3701                 /* put back original number of pages, chg */
3702                 (void)hugepage_subpool_put_pages(spool, chg);
3703                 goto out_err;
3704         }
3705 
3706         /*
3707          * Account for the reservations made. Shared mappings record regions
3708          * that have reservations as they are shared by multiple VMAs.
3709          * When the last VMA disappears, the region map says how much
3710          * the reservation was and the page cache tells how much of
3711          * the reservation was consumed. Private mappings are per-VMA and
3712          * only the consumed reservations are tracked. When the VMA
3713          * disappears, the original reservation is the VMA size and the
3714          * consumed reservations are stored in the map. Hence, nothing
3715          * else has to be done for private mappings here
3716          */
3717         if (!vma || vma->vm_flags & VM_MAYSHARE) {
3718                 long add = region_add(resv_map, from, to);
3719 
3720                 if (unlikely(chg > add)) {
3721                         /*
3722                          * pages in this range were added to the reserve
3723                          * map between region_chg and region_add.  This
3724                          * indicates a race with alloc_huge_page.  Adjust
3725                          * the subpool and reserve counts modified above
3726                          * based on the difference.
3727                          */
3728                         long rsv_adjust;
3729 
3730                         rsv_adjust = hugepage_subpool_put_pages(spool,
3731                                                                 chg - add);
3732                         hugetlb_acct_memory(h, -rsv_adjust);
3733                 }
3734         }
3735         return 0;
3736 out_err:
3737         if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3738                 kref_put(&resv_map->refs, resv_map_release);
3739         return ret;
3740 }
3741 
3742 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3743 {
3744         struct hstate *h = hstate_inode(inode);
3745         struct resv_map *resv_map = inode_resv_map(inode);
3746         long chg = 0;
3747         struct hugepage_subpool *spool = subpool_inode(inode);
3748         long gbl_reserve;
3749 
3750         if (resv_map)
3751                 chg = region_truncate(resv_map, offset);
3752         spin_lock(&inode->i_lock);
3753         inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3754         spin_unlock(&inode->i_lock);
3755 
3756         /*
3757          * If the subpool has a minimum size, the number of global
3758          * reservations to be released may be adjusted.
3759          */
3760         gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
3761         hugetlb_acct_memory(h, -gbl_reserve);
3762 }
3763 
3764 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3765 static unsigned long page_table_shareable(struct vm_area_struct *svma,
3766                                 struct vm_area_struct *vma,
3767                                 unsigned long addr, pgoff_t idx)
3768 {
3769         unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
3770                                 svma->vm_start;
3771         unsigned long sbase = saddr & PUD_MASK;
3772         unsigned long s_end = sbase + PUD_SIZE;
3773 
3774         /* Allow segments to share if only one is marked locked */
3775         unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED;
3776         unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED;
3777 
3778         /*
3779          * match the virtual addresses, permission and the alignment of the
3780          * page table page.
3781          */
3782         if (pmd_index(addr) != pmd_index(saddr) ||
3783             vm_flags != svm_flags ||
3784             sbase < svma->vm_start || svma->vm_end < s_end)
3785                 return 0;
3786 
3787         return saddr;
3788 }
3789 
3790 static int vma_shareable(struct vm_area_struct *vma, unsigned long addr)
3791 {
3792         unsigned long base = addr & PUD_MASK;
3793         unsigned long end = base + PUD_SIZE;
3794 
3795         /*
3796          * check on proper vm_flags and page table alignment
3797          */
3798         if (vma->vm_flags & VM_MAYSHARE &&
3799             vma->vm_start <= base && end <= vma->vm_end)
3800                 return 1;
3801         return 0;
3802 }
3803 
3804 /*
3805  * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3806  * and returns the corresponding pte. While this is not necessary for the
3807  * !shared pmd case because we can allocate the pmd later as well, it makes the
3808  * code much cleaner. pmd allocation is essential for the shared case because
3809  * pud has to be populated inside the same i_mmap_rwsem section - otherwise
3810  * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3811  * bad pmd for sharing.
3812  */
3813 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3814 {
3815         struct vm_area_struct *vma = find_vma(mm, addr);
3816         struct address_space *mapping = vma->vm_file->f_mapping;
3817         pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
3818                         vma->vm_pgoff;
3819         struct vm_area_struct *svma;
3820         unsigned long saddr;
3821         pte_t *spte = NULL;
3822         pte_t *pte;
3823         spinlock_t *ptl;
3824 
3825         if (!vma_shareable(vma, addr))
3826                 return (pte_t *)pmd_alloc(mm, pud, addr);
3827 
3828         i_mmap_lock_write(mapping);
3829         vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
3830                 if (svma == vma)
3831                         continue;
3832 
3833                 saddr = page_table_shareable(svma, vma, addr, idx);
3834                 if (saddr) {
3835                         spte = huge_pte_offset(svma->vm_mm, saddr);
3836                         if (spte) {
3837                                 mm_inc_nr_pmds(mm);
3838                                 get_page(virt_to_page(spte));
3839                                 break;
3840                         }
3841                 }
3842         }
3843 
3844         if (!spte)
3845                 goto out;
3846 
3847         ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte);
3848         spin_lock(ptl);
3849         if (pud_none(*pud)) {
3850                 pud_populate(mm, pud,
3851                                 (pmd_t *)((unsigned long)spte & PAGE_MASK));
3852         } else {
3853                 put_page(virt_to_page(spte));
3854                 mm_inc_nr_pmds(mm);
3855         }
3856         spin_unlock(ptl);
3857 out:
3858         pte = (pte_t *)pmd_alloc(mm, pud, addr);
3859         i_mmap_unlock_write(mapping);
3860         return pte;
3861 }
3862 
3863 /*
3864  * unmap huge page backed by shared pte.
3865  *
3866  * Hugetlb pte page is ref counted at the time of mapping.  If pte is shared
3867  * indicated by page_count > 1, unmap is achieved by clearing pud and
3868  * decrementing the ref count. If count == 1, the pte page is not shared.
3869  *
3870  * called with page table lock held.
3871  *
3872  * returns: 1 successfully unmapped a shared pte page
3873  *          0 the underlying pte page is not shared, or it is the last user
3874  */
3875 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
3876 {
3877         pgd_t *pgd = pgd_offset(mm, *addr);
3878         pud_t *pud = pud_offset(pgd, *addr);
3879 
3880         BUG_ON(page_count(virt_to_page(ptep)) == 0);
3881         if (page_count(virt_to_page(ptep)) == 1)
3882                 return 0;
3883 
3884         pud_clear(pud);
3885         put_page(virt_to_page(ptep));
3886         mm_dec_nr_pmds(mm);
3887         *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
3888         return 1;
3889 }
3890 #define want_pmd_share()        (1)
3891 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3892 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3893 {
3894         return NULL;
3895 }
3896 
3897 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
3898 {
3899         return 0;
3900 }
3901 #define want_pmd_share()        (0)
3902 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3903 
3904 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3905 pte_t *huge_pte_alloc(struct mm_struct *mm,
3906                         unsigned long addr, unsigned long sz)
3907 {
3908         pgd_t *pgd;
3909         pud_t *pud;
3910         pte_t *pte = NULL;
3911 
3912         pgd = pgd_offset(mm, addr);
3913         pud = pud_alloc(mm, pgd, addr);
3914         if (pud) {
3915                 if (sz == PUD_SIZE) {
3916                         pte = (pte_t *)pud;
3917                 } else {
3918                         BUG_ON(sz != PMD_SIZE);
3919                         if (want_pmd_share() && pud_none(*pud))
3920                                 pte = huge_pmd_share(mm, addr, pud);
3921                         else
3922                                 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3923                 }
3924         }
3925         BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
3926 
3927         return pte;
3928 }
3929 
3930 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
3931 {
3932         pgd_t *pgd;
3933         pud_t *pud;
3934         pmd_t *pmd = NULL;
3935 
3936         pgd = pgd_offset(mm, addr);
3937         if (pgd_present(*pgd)) {
3938                 pud = pud_offset(pgd, addr);
3939                 if (pud_present(*pud)) {
3940                         if (pud_huge(*pud))
3941                                 return (pte_t *)pud;
3942                         pmd = pmd_offset(pud, addr);
3943                 }
3944         }
3945         return (pte_t *) pmd;
3946 }
3947 
3948 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3949 
3950 /*
3951  * These functions are overwritable if your architecture needs its own
3952  * behavior.
3953  */
3954 struct page * __weak
3955 follow_huge_addr(struct mm_struct *mm, unsigned long address,
3956                               int write)
3957 {
3958         return ERR_PTR(-EINVAL);
3959 }
3960 
3961 struct page * __weak
3962 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
3963                 pmd_t *pmd, int flags)
3964 {
3965         struct page *page = NULL;
3966         spinlock_t *ptl;
3967 retry:
3968         ptl = pmd_lockptr(mm, pmd);
3969         spin_lock(ptl);
3970         /*
3971          * make sure that the address range covered by this pmd is not
3972          * unmapped from other threads.
3973          */
3974         if (!pmd_huge(*pmd))
3975                 goto out;
3976         if (pmd_present(*pmd)) {
3977                 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
3978                 if (flags & FOLL_GET)
3979                         get_page(page);
3980         } else {
3981                 if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) {
3982                         spin_unlock(ptl);
3983                         __migration_entry_wait(mm, (pte_t *)pmd, ptl);
3984                         goto retry;
3985                 }
3986                 /*
3987                  * hwpoisoned entry is treated as no_page_table in
3988                  * follow_page_mask().
3989                  */
3990         }
3991 out:
3992         spin_unlock(ptl);
3993         return page;
3994 }
3995 
3996 struct page * __weak
3997 follow_huge_pud(struct mm_struct *mm, unsigned long address,
3998                 pud_t *pud, int flags)
3999 {
4000         if (flags & FOLL_GET)
4001                 return NULL;
4002 
4003         return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
4004 }
4005 
4006 #ifdef CONFIG_MEMORY_FAILURE
4007 
4008 /*
4009  * This function is called from memory failure code.
4010  * Assume the caller holds page lock of the head page.
4011  */
4012 int dequeue_hwpoisoned_huge_page(struct page *hpage)
4013 {
4014         struct hstate *h = page_hstate(hpage);
4015         int nid = page_to_nid(hpage);
4016         int ret = -EBUSY;
4017 
4018         spin_lock(&hugetlb_lock);
4019         /*
4020          * Just checking !page_huge_active is not enough, because that could be
4021          * an isolated/hwpoisoned hugepage (which have >0 refcount).
4022          */
4023         if (!page_huge_active(hpage) && !page_count(hpage)) {
4024                 /*
4025                  * Hwpoisoned hugepage isn't linked to activelist or freelist,
4026                  * but dangling hpage->lru can trigger list-debug warnings
4027                  * (this happens when we call unpoison_memory() on it),
4028                  * so let it point to itself with list_del_init().
4029                  */
4030                 list_del_init(&hpage->lru);
4031                 set_page_refcounted(hpage);
4032                 h->free_huge_pages--;
4033                 h->free_huge_pages_node[nid]--;
4034                 ret = 0;
4035         }
4036         spin_unlock(&hugetlb_lock);
4037         return ret;
4038 }
4039 #endif
4040 
4041 bool isolate_huge_page(struct page *page, struct list_head *list)
4042 {
4043         bool ret = true;
4044 
4045         VM_BUG_ON_PAGE(!PageHead(page), page);
4046         spin_lock(&hugetlb_lock);
4047         if (!page_huge_active(page) || !get_page_unless_zero(page)) {
4048                 ret = false;
4049                 goto unlock;
4050         }
4051         clear_page_huge_active(page);
4052         list_move_tail(&page->lru, list);
4053 unlock:
4054         spin_unlock(&hugetlb_lock);
4055         return ret;
4056 }
4057 
4058 void putback_active_hugepage(struct page *page)
4059 {
4060         VM_BUG_ON_PAGE(!PageHead(page), page);
4061         spin_lock(&hugetlb_lock);
4062         set_page_huge_active(page);
4063         list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
4064         spin_unlock(&hugetlb_lock);
4065         put_page(page);
4066 }
4067 

~ [ 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