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

TOMOYO Linux Cross Reference
Linux/mm/hugetlb.c

Version: ~ [ linux-5.2-rc4 ] ~ [ linux-5.1.9 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.50 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.125 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.181 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.181 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.140 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.68 ] ~ [ linux-3.15.10 ] ~ [ linux-3.14.79 ] ~ [ linux-3.13.11 ] ~ [ linux-3.12.74 ] ~ [ linux-3.11.10 ] ~ [ linux-3.10.108 ] ~ [ linux-3.9.11 ] ~ [ linux-3.8.13 ] ~ [ linux-3.7.10 ] ~ [ linux-3.6.11 ] ~ [ linux-3.5.7 ] ~ [ linux-3.4.113 ] ~ [ linux-3.3.8 ] ~ [ linux-3.2.102 ] ~ [ linux-3.1.10 ] ~ [ linux-3.0.101 ] ~ [ linux-2.6.39.4 ] ~ [ linux-2.6.38.8 ] ~ [ linux-2.6.37.6 ] ~ [ linux-2.6.36.4 ] ~ [ linux-2.6.35.14 ] ~ [ linux-2.6.34.15 ] ~ [ linux-2.6.33.20 ] ~ [ 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/cpuset.h>
 17 #include <linux/mutex.h>
 18 #include <linux/bootmem.h>
 19 #include <linux/sysfs.h>
 20 #include <linux/slab.h>
 21 #include <linux/rmap.h>
 22 #include <linux/swap.h>
 23 #include <linux/swapops.h>
 24 #include <linux/page-isolation.h>
 25 
 26 #include <asm/page.h>
 27 #include <asm/pgtable.h>
 28 #include <asm/tlb.h>
 29 
 30 #include <linux/io.h>
 31 #include <linux/hugetlb.h>
 32 #include <linux/hugetlb_cgroup.h>
 33 #include <linux/node.h>
 34 #include "internal.h"
 35 
 36 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
 37 unsigned long hugepages_treat_as_movable;
 38 
 39 int hugetlb_max_hstate __read_mostly;
 40 unsigned int default_hstate_idx;
 41 struct hstate hstates[HUGE_MAX_HSTATE];
 42 
 43 __initdata LIST_HEAD(huge_boot_pages);
 44 
 45 /* for command line parsing */
 46 static struct hstate * __initdata parsed_hstate;
 47 static unsigned long __initdata default_hstate_max_huge_pages;
 48 static unsigned long __initdata default_hstate_size;
 49 
 50 /*
 51  * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
 52  * free_huge_pages, and surplus_huge_pages.
 53  */
 54 DEFINE_SPINLOCK(hugetlb_lock);
 55 
 56 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
 57 {
 58         bool free = (spool->count == 0) && (spool->used_hpages == 0);
 59 
 60         spin_unlock(&spool->lock);
 61 
 62         /* If no pages are used, and no other handles to the subpool
 63          * remain, free the subpool the subpool remain */
 64         if (free)
 65                 kfree(spool);
 66 }
 67 
 68 struct hugepage_subpool *hugepage_new_subpool(long nr_blocks)
 69 {
 70         struct hugepage_subpool *spool;
 71 
 72         spool = kmalloc(sizeof(*spool), GFP_KERNEL);
 73         if (!spool)
 74                 return NULL;
 75 
 76         spin_lock_init(&spool->lock);
 77         spool->count = 1;
 78         spool->max_hpages = nr_blocks;
 79         spool->used_hpages = 0;
 80 
 81         return spool;
 82 }
 83 
 84 void hugepage_put_subpool(struct hugepage_subpool *spool)
 85 {
 86         spin_lock(&spool->lock);
 87         BUG_ON(!spool->count);
 88         spool->count--;
 89         unlock_or_release_subpool(spool);
 90 }
 91 
 92 static int hugepage_subpool_get_pages(struct hugepage_subpool *spool,
 93                                       long delta)
 94 {
 95         int ret = 0;
 96 
 97         if (!spool)
 98                 return 0;
 99 
100         spin_lock(&spool->lock);
101         if ((spool->used_hpages + delta) <= spool->max_hpages) {
102                 spool->used_hpages += delta;
103         } else {
104                 ret = -ENOMEM;
105         }
106         spin_unlock(&spool->lock);
107 
108         return ret;
109 }
110 
111 static void hugepage_subpool_put_pages(struct hugepage_subpool *spool,
112                                        long delta)
113 {
114         if (!spool)
115                 return;
116 
117         spin_lock(&spool->lock);
118         spool->used_hpages -= delta;
119         /* If hugetlbfs_put_super couldn't free spool due to
120         * an outstanding quota reference, free it now. */
121         unlock_or_release_subpool(spool);
122 }
123 
124 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
125 {
126         return HUGETLBFS_SB(inode->i_sb)->spool;
127 }
128 
129 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
130 {
131         return subpool_inode(file_inode(vma->vm_file));
132 }
133 
134 /*
135  * Region tracking -- allows tracking of reservations and instantiated pages
136  *                    across the pages in a mapping.
137  *
138  * The region data structures are protected by a combination of the mmap_sem
139  * and the hugetlb_instantiation_mutex.  To access or modify a region the caller
140  * must either hold the mmap_sem for write, or the mmap_sem for read and
141  * the hugetlb_instantiation_mutex:
142  *
143  *      down_write(&mm->mmap_sem);
144  * or
145  *      down_read(&mm->mmap_sem);
146  *      mutex_lock(&hugetlb_instantiation_mutex);
147  */
148 struct file_region {
149         struct list_head link;
150         long from;
151         long to;
152 };
153 
154 static long region_add(struct list_head *head, long f, long t)
155 {
156         struct file_region *rg, *nrg, *trg;
157 
158         /* Locate the region we are either in or before. */
159         list_for_each_entry(rg, head, link)
160                 if (f <= rg->to)
161                         break;
162 
163         /* Round our left edge to the current segment if it encloses us. */
164         if (f > rg->from)
165                 f = rg->from;
166 
167         /* Check for and consume any regions we now overlap with. */
168         nrg = rg;
169         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
170                 if (&rg->link == head)
171                         break;
172                 if (rg->from > t)
173                         break;
174 
175                 /* If this area reaches higher then extend our area to
176                  * include it completely.  If this is not the first area
177                  * which we intend to reuse, free it. */
178                 if (rg->to > t)
179                         t = rg->to;
180                 if (rg != nrg) {
181                         list_del(&rg->link);
182                         kfree(rg);
183                 }
184         }
185         nrg->from = f;
186         nrg->to = t;
187         return 0;
188 }
189 
190 static long region_chg(struct list_head *head, long f, long t)
191 {
192         struct file_region *rg, *nrg;
193         long chg = 0;
194 
195         /* Locate the region we are before or in. */
196         list_for_each_entry(rg, head, link)
197                 if (f <= rg->to)
198                         break;
199 
200         /* If we are below the current region then a new region is required.
201          * Subtle, allocate a new region at the position but make it zero
202          * size such that we can guarantee to record the reservation. */
203         if (&rg->link == head || t < rg->from) {
204                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
205                 if (!nrg)
206                         return -ENOMEM;
207                 nrg->from = f;
208                 nrg->to   = f;
209                 INIT_LIST_HEAD(&nrg->link);
210                 list_add(&nrg->link, rg->link.prev);
211 
212                 return t - f;
213         }
214 
215         /* Round our left edge to the current segment if it encloses us. */
216         if (f > rg->from)
217                 f = rg->from;
218         chg = t - f;
219 
220         /* Check for and consume any regions we now overlap with. */
221         list_for_each_entry(rg, rg->link.prev, link) {
222                 if (&rg->link == head)
223                         break;
224                 if (rg->from > t)
225                         return chg;
226 
227                 /* We overlap with this area, if it extends further than
228                  * us then we must extend ourselves.  Account for its
229                  * existing reservation. */
230                 if (rg->to > t) {
231                         chg += rg->to - t;
232                         t = rg->to;
233                 }
234                 chg -= rg->to - rg->from;
235         }
236         return chg;
237 }
238 
239 static long region_truncate(struct list_head *head, long end)
240 {
241         struct file_region *rg, *trg;
242         long chg = 0;
243 
244         /* Locate the region we are either in or before. */
245         list_for_each_entry(rg, head, link)
246                 if (end <= rg->to)
247                         break;
248         if (&rg->link == head)
249                 return 0;
250 
251         /* If we are in the middle of a region then adjust it. */
252         if (end > rg->from) {
253                 chg = rg->to - end;
254                 rg->to = end;
255                 rg = list_entry(rg->link.next, typeof(*rg), link);
256         }
257 
258         /* Drop any remaining regions. */
259         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
260                 if (&rg->link == head)
261                         break;
262                 chg += rg->to - rg->from;
263                 list_del(&rg->link);
264                 kfree(rg);
265         }
266         return chg;
267 }
268 
269 static long region_count(struct list_head *head, long f, long t)
270 {
271         struct file_region *rg;
272         long chg = 0;
273 
274         /* Locate each segment we overlap with, and count that overlap. */
275         list_for_each_entry(rg, head, link) {
276                 long seg_from;
277                 long seg_to;
278 
279                 if (rg->to <= f)
280                         continue;
281                 if (rg->from >= t)
282                         break;
283 
284                 seg_from = max(rg->from, f);
285                 seg_to = min(rg->to, t);
286 
287                 chg += seg_to - seg_from;
288         }
289 
290         return chg;
291 }
292 
293 /*
294  * Convert the address within this vma to the page offset within
295  * the mapping, in pagecache page units; huge pages here.
296  */
297 static pgoff_t vma_hugecache_offset(struct hstate *h,
298                         struct vm_area_struct *vma, unsigned long address)
299 {
300         return ((address - vma->vm_start) >> huge_page_shift(h)) +
301                         (vma->vm_pgoff >> huge_page_order(h));
302 }
303 
304 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
305                                      unsigned long address)
306 {
307         return vma_hugecache_offset(hstate_vma(vma), vma, address);
308 }
309 
310 /*
311  * Return the size of the pages allocated when backing a VMA. In the majority
312  * cases this will be same size as used by the page table entries.
313  */
314 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
315 {
316         struct hstate *hstate;
317 
318         if (!is_vm_hugetlb_page(vma))
319                 return PAGE_SIZE;
320 
321         hstate = hstate_vma(vma);
322 
323         return 1UL << huge_page_shift(hstate);
324 }
325 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
326 
327 /*
328  * Return the page size being used by the MMU to back a VMA. In the majority
329  * of cases, the page size used by the kernel matches the MMU size. On
330  * architectures where it differs, an architecture-specific version of this
331  * function is required.
332  */
333 #ifndef vma_mmu_pagesize
334 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
335 {
336         return vma_kernel_pagesize(vma);
337 }
338 #endif
339 
340 /*
341  * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
342  * bits of the reservation map pointer, which are always clear due to
343  * alignment.
344  */
345 #define HPAGE_RESV_OWNER    (1UL << 0)
346 #define HPAGE_RESV_UNMAPPED (1UL << 1)
347 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
348 
349 /*
350  * These helpers are used to track how many pages are reserved for
351  * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
352  * is guaranteed to have their future faults succeed.
353  *
354  * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
355  * the reserve counters are updated with the hugetlb_lock held. It is safe
356  * to reset the VMA at fork() time as it is not in use yet and there is no
357  * chance of the global counters getting corrupted as a result of the values.
358  *
359  * The private mapping reservation is represented in a subtly different
360  * manner to a shared mapping.  A shared mapping has a region map associated
361  * with the underlying file, this region map represents the backing file
362  * pages which have ever had a reservation assigned which this persists even
363  * after the page is instantiated.  A private mapping has a region map
364  * associated with the original mmap which is attached to all VMAs which
365  * reference it, this region map represents those offsets which have consumed
366  * reservation ie. where pages have been instantiated.
367  */
368 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
369 {
370         return (unsigned long)vma->vm_private_data;
371 }
372 
373 static void set_vma_private_data(struct vm_area_struct *vma,
374                                                         unsigned long value)
375 {
376         vma->vm_private_data = (void *)value;
377 }
378 
379 struct resv_map {
380         struct kref refs;
381         struct list_head regions;
382 };
383 
384 static struct resv_map *resv_map_alloc(void)
385 {
386         struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
387         if (!resv_map)
388                 return NULL;
389 
390         kref_init(&resv_map->refs);
391         INIT_LIST_HEAD(&resv_map->regions);
392 
393         return resv_map;
394 }
395 
396 static void resv_map_release(struct kref *ref)
397 {
398         struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
399 
400         /* Clear out any active regions before we release the map. */
401         region_truncate(&resv_map->regions, 0);
402         kfree(resv_map);
403 }
404 
405 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
406 {
407         VM_BUG_ON(!is_vm_hugetlb_page(vma));
408         if (!(vma->vm_flags & VM_MAYSHARE))
409                 return (struct resv_map *)(get_vma_private_data(vma) &
410                                                         ~HPAGE_RESV_MASK);
411         return NULL;
412 }
413 
414 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
415 {
416         VM_BUG_ON(!is_vm_hugetlb_page(vma));
417         VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
418 
419         set_vma_private_data(vma, (get_vma_private_data(vma) &
420                                 HPAGE_RESV_MASK) | (unsigned long)map);
421 }
422 
423 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
424 {
425         VM_BUG_ON(!is_vm_hugetlb_page(vma));
426         VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
427 
428         set_vma_private_data(vma, get_vma_private_data(vma) | flags);
429 }
430 
431 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
432 {
433         VM_BUG_ON(!is_vm_hugetlb_page(vma));
434 
435         return (get_vma_private_data(vma) & flag) != 0;
436 }
437 
438 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
439 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
440 {
441         VM_BUG_ON(!is_vm_hugetlb_page(vma));
442         if (!(vma->vm_flags & VM_MAYSHARE))
443                 vma->vm_private_data = (void *)0;
444 }
445 
446 /* Returns true if the VMA has associated reserve pages */
447 static int vma_has_reserves(struct vm_area_struct *vma, long chg)
448 {
449         if (vma->vm_flags & VM_NORESERVE) {
450                 /*
451                  * This address is already reserved by other process(chg == 0),
452                  * so, we should decrement reserved count. Without decrementing,
453                  * reserve count remains after releasing inode, because this
454                  * allocated page will go into page cache and is regarded as
455                  * coming from reserved pool in releasing step.  Currently, we
456                  * don't have any other solution to deal with this situation
457                  * properly, so add work-around here.
458                  */
459                 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
460                         return 1;
461                 else
462                         return 0;
463         }
464 
465         /* Shared mappings always use reserves */
466         if (vma->vm_flags & VM_MAYSHARE)
467                 return 1;
468 
469         /*
470          * Only the process that called mmap() has reserves for
471          * private mappings.
472          */
473         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
474                 return 1;
475 
476         return 0;
477 }
478 
479 static void enqueue_huge_page(struct hstate *h, struct page *page)
480 {
481         int nid = page_to_nid(page);
482         list_move(&page->lru, &h->hugepage_freelists[nid]);
483         h->free_huge_pages++;
484         h->free_huge_pages_node[nid]++;
485 }
486 
487 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
488 {
489         struct page *page;
490 
491         list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
492                 if (!is_migrate_isolate_page(page))
493                         break;
494         /*
495          * if 'non-isolated free hugepage' not found on the list,
496          * the allocation fails.
497          */
498         if (&h->hugepage_freelists[nid] == &page->lru)
499                 return NULL;
500         list_move(&page->lru, &h->hugepage_activelist);
501         set_page_refcounted(page);
502         h->free_huge_pages--;
503         h->free_huge_pages_node[nid]--;
504         return page;
505 }
506 
507 /* Movability of hugepages depends on migration support. */
508 static inline gfp_t htlb_alloc_mask(struct hstate *h)
509 {
510         if (hugepages_treat_as_movable || hugepage_migration_support(h))
511                 return GFP_HIGHUSER_MOVABLE;
512         else
513                 return GFP_HIGHUSER;
514 }
515 
516 static struct page *dequeue_huge_page_vma(struct hstate *h,
517                                 struct vm_area_struct *vma,
518                                 unsigned long address, int avoid_reserve,
519                                 long chg)
520 {
521         struct page *page = NULL;
522         struct mempolicy *mpol;
523         nodemask_t *nodemask;
524         struct zonelist *zonelist;
525         struct zone *zone;
526         struct zoneref *z;
527         unsigned int cpuset_mems_cookie;
528 
529         /*
530          * A child process with MAP_PRIVATE mappings created by their parent
531          * have no page reserves. This check ensures that reservations are
532          * not "stolen". The child may still get SIGKILLed
533          */
534         if (!vma_has_reserves(vma, chg) &&
535                         h->free_huge_pages - h->resv_huge_pages == 0)
536                 goto err;
537 
538         /* If reserves cannot be used, ensure enough pages are in the pool */
539         if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
540                 goto err;
541 
542 retry_cpuset:
543         cpuset_mems_cookie = read_mems_allowed_begin();
544         zonelist = huge_zonelist(vma, address,
545                                         htlb_alloc_mask(h), &mpol, &nodemask);
546 
547         for_each_zone_zonelist_nodemask(zone, z, zonelist,
548                                                 MAX_NR_ZONES - 1, nodemask) {
549                 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask(h))) {
550                         page = dequeue_huge_page_node(h, zone_to_nid(zone));
551                         if (page) {
552                                 if (avoid_reserve)
553                                         break;
554                                 if (!vma_has_reserves(vma, chg))
555                                         break;
556 
557                                 SetPagePrivate(page);
558                                 h->resv_huge_pages--;
559                                 break;
560                         }
561                 }
562         }
563 
564         mpol_cond_put(mpol);
565         if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
566                 goto retry_cpuset;
567         return page;
568 
569 err:
570         return NULL;
571 }
572 
573 static void update_and_free_page(struct hstate *h, struct page *page)
574 {
575         int i;
576 
577         VM_BUG_ON(h->order >= MAX_ORDER);
578 
579         h->nr_huge_pages--;
580         h->nr_huge_pages_node[page_to_nid(page)]--;
581         for (i = 0; i < pages_per_huge_page(h); i++) {
582                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
583                                 1 << PG_referenced | 1 << PG_dirty |
584                                 1 << PG_active | 1 << PG_reserved |
585                                 1 << PG_private | 1 << PG_writeback);
586         }
587         VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
588         set_compound_page_dtor(page, NULL);
589         set_page_refcounted(page);
590         arch_release_hugepage(page);
591         __free_pages(page, huge_page_order(h));
592 }
593 
594 struct hstate *size_to_hstate(unsigned long size)
595 {
596         struct hstate *h;
597 
598         for_each_hstate(h) {
599                 if (huge_page_size(h) == size)
600                         return h;
601         }
602         return NULL;
603 }
604 
605 static void free_huge_page(struct page *page)
606 {
607         /*
608          * Can't pass hstate in here because it is called from the
609          * compound page destructor.
610          */
611         struct hstate *h = page_hstate(page);
612         int nid = page_to_nid(page);
613         struct hugepage_subpool *spool =
614                 (struct hugepage_subpool *)page_private(page);
615         bool restore_reserve;
616 
617         set_page_private(page, 0);
618         page->mapping = NULL;
619         BUG_ON(page_count(page));
620         BUG_ON(page_mapcount(page));
621         restore_reserve = PagePrivate(page);
622         ClearPagePrivate(page);
623 
624         spin_lock(&hugetlb_lock);
625         hugetlb_cgroup_uncharge_page(hstate_index(h),
626                                      pages_per_huge_page(h), page);
627         if (restore_reserve)
628                 h->resv_huge_pages++;
629 
630         if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
631                 /* remove the page from active list */
632                 list_del(&page->lru);
633                 update_and_free_page(h, page);
634                 h->surplus_huge_pages--;
635                 h->surplus_huge_pages_node[nid]--;
636         } else {
637                 arch_clear_hugepage_flags(page);
638                 enqueue_huge_page(h, page);
639         }
640         spin_unlock(&hugetlb_lock);
641         hugepage_subpool_put_pages(spool, 1);
642 }
643 
644 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
645 {
646         INIT_LIST_HEAD(&page->lru);
647         set_compound_page_dtor(page, free_huge_page);
648         spin_lock(&hugetlb_lock);
649         set_hugetlb_cgroup(page, NULL);
650         h->nr_huge_pages++;
651         h->nr_huge_pages_node[nid]++;
652         spin_unlock(&hugetlb_lock);
653         put_page(page); /* free it into the hugepage allocator */
654 }
655 
656 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
657 {
658         int i;
659         int nr_pages = 1 << order;
660         struct page *p = page + 1;
661 
662         /* we rely on prep_new_huge_page to set the destructor */
663         set_compound_order(page, order);
664         __SetPageHead(page);
665         __ClearPageReserved(page);
666         for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
667                 __SetPageTail(p);
668                 /*
669                  * For gigantic hugepages allocated through bootmem at
670                  * boot, it's safer to be consistent with the not-gigantic
671                  * hugepages and clear the PG_reserved bit from all tail pages
672                  * too.  Otherwse drivers using get_user_pages() to access tail
673                  * pages may get the reference counting wrong if they see
674                  * PG_reserved set on a tail page (despite the head page not
675                  * having PG_reserved set).  Enforcing this consistency between
676                  * head and tail pages allows drivers to optimize away a check
677                  * on the head page when they need know if put_page() is needed
678                  * after get_user_pages().
679                  */
680                 __ClearPageReserved(p);
681                 set_page_count(p, 0);
682                 p->first_page = page;
683         }
684 }
685 
686 /*
687  * PageHuge() only returns true for hugetlbfs pages, but not for normal or
688  * transparent huge pages.  See the PageTransHuge() documentation for more
689  * details.
690  */
691 int PageHuge(struct page *page)
692 {
693         if (!PageCompound(page))
694                 return 0;
695 
696         page = compound_head(page);
697         return get_compound_page_dtor(page) == free_huge_page;
698 }
699 EXPORT_SYMBOL_GPL(PageHuge);
700 
701 /*
702  * PageHeadHuge() only returns true for hugetlbfs head page, but not for
703  * normal or transparent huge pages.
704  */
705 int PageHeadHuge(struct page *page_head)
706 {
707         if (!PageHead(page_head))
708                 return 0;
709 
710         return get_compound_page_dtor(page_head) == free_huge_page;
711 }
712 
713 pgoff_t __basepage_index(struct page *page)
714 {
715         struct page *page_head = compound_head(page);
716         pgoff_t index = page_index(page_head);
717         unsigned long compound_idx;
718 
719         if (!PageHuge(page_head))
720                 return page_index(page);
721 
722         if (compound_order(page_head) >= MAX_ORDER)
723                 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
724         else
725                 compound_idx = page - page_head;
726 
727         return (index << compound_order(page_head)) + compound_idx;
728 }
729 
730 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
731 {
732         struct page *page;
733 
734         if (h->order >= MAX_ORDER)
735                 return NULL;
736 
737         page = alloc_pages_exact_node(nid,
738                 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
739                                                 __GFP_REPEAT|__GFP_NOWARN,
740                 huge_page_order(h));
741         if (page) {
742                 if (arch_prepare_hugepage(page)) {
743                         __free_pages(page, huge_page_order(h));
744                         return NULL;
745                 }
746                 prep_new_huge_page(h, page, nid);
747         }
748 
749         return page;
750 }
751 
752 /*
753  * common helper functions for hstate_next_node_to_{alloc|free}.
754  * We may have allocated or freed a huge page based on a different
755  * nodes_allowed previously, so h->next_node_to_{alloc|free} might
756  * be outside of *nodes_allowed.  Ensure that we use an allowed
757  * node for alloc or free.
758  */
759 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
760 {
761         nid = next_node(nid, *nodes_allowed);
762         if (nid == MAX_NUMNODES)
763                 nid = first_node(*nodes_allowed);
764         VM_BUG_ON(nid >= MAX_NUMNODES);
765 
766         return nid;
767 }
768 
769 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
770 {
771         if (!node_isset(nid, *nodes_allowed))
772                 nid = next_node_allowed(nid, nodes_allowed);
773         return nid;
774 }
775 
776 /*
777  * returns the previously saved node ["this node"] from which to
778  * allocate a persistent huge page for the pool and advance the
779  * next node from which to allocate, handling wrap at end of node
780  * mask.
781  */
782 static int hstate_next_node_to_alloc(struct hstate *h,
783                                         nodemask_t *nodes_allowed)
784 {
785         int nid;
786 
787         VM_BUG_ON(!nodes_allowed);
788 
789         nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
790         h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
791 
792         return nid;
793 }
794 
795 /*
796  * helper for free_pool_huge_page() - return the previously saved
797  * node ["this node"] from which to free a huge page.  Advance the
798  * next node id whether or not we find a free huge page to free so
799  * that the next attempt to free addresses the next node.
800  */
801 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
802 {
803         int nid;
804 
805         VM_BUG_ON(!nodes_allowed);
806 
807         nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
808         h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
809 
810         return nid;
811 }
812 
813 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask)           \
814         for (nr_nodes = nodes_weight(*mask);                            \
815                 nr_nodes > 0 &&                                         \
816                 ((node = hstate_next_node_to_alloc(hs, mask)) || 1);    \
817                 nr_nodes--)
818 
819 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask)            \
820         for (nr_nodes = nodes_weight(*mask);                            \
821                 nr_nodes > 0 &&                                         \
822                 ((node = hstate_next_node_to_free(hs, mask)) || 1);     \
823                 nr_nodes--)
824 
825 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
826 {
827         struct page *page;
828         int nr_nodes, node;
829         int ret = 0;
830 
831         for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
832                 page = alloc_fresh_huge_page_node(h, node);
833                 if (page) {
834                         ret = 1;
835                         break;
836                 }
837         }
838 
839         if (ret)
840                 count_vm_event(HTLB_BUDDY_PGALLOC);
841         else
842                 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
843 
844         return ret;
845 }
846 
847 /*
848  * Free huge page from pool from next node to free.
849  * Attempt to keep persistent huge pages more or less
850  * balanced over allowed nodes.
851  * Called with hugetlb_lock locked.
852  */
853 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
854                                                          bool acct_surplus)
855 {
856         int nr_nodes, node;
857         int ret = 0;
858 
859         for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
860                 /*
861                  * If we're returning unused surplus pages, only examine
862                  * nodes with surplus pages.
863                  */
864                 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
865                     !list_empty(&h->hugepage_freelists[node])) {
866                         struct page *page =
867                                 list_entry(h->hugepage_freelists[node].next,
868                                           struct page, lru);
869                         list_del(&page->lru);
870                         h->free_huge_pages--;
871                         h->free_huge_pages_node[node]--;
872                         if (acct_surplus) {
873                                 h->surplus_huge_pages--;
874                                 h->surplus_huge_pages_node[node]--;
875                         }
876                         update_and_free_page(h, page);
877                         ret = 1;
878                         break;
879                 }
880         }
881 
882         return ret;
883 }
884 
885 /*
886  * Dissolve a given free hugepage into free buddy pages. This function does
887  * nothing for in-use (including surplus) hugepages.
888  */
889 static void dissolve_free_huge_page(struct page *page)
890 {
891         spin_lock(&hugetlb_lock);
892         if (PageHuge(page) && !page_count(page)) {
893                 struct hstate *h = page_hstate(page);
894                 int nid = page_to_nid(page);
895                 list_del(&page->lru);
896                 h->free_huge_pages--;
897                 h->free_huge_pages_node[nid]--;
898                 update_and_free_page(h, page);
899         }
900         spin_unlock(&hugetlb_lock);
901 }
902 
903 /*
904  * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
905  * make specified memory blocks removable from the system.
906  * Note that start_pfn should aligned with (minimum) hugepage size.
907  */
908 void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
909 {
910         unsigned int order = 8 * sizeof(void *);
911         unsigned long pfn;
912         struct hstate *h;
913 
914         /* Set scan step to minimum hugepage size */
915         for_each_hstate(h)
916                 if (order > huge_page_order(h))
917                         order = huge_page_order(h);
918         VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << order));
919         for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order)
920                 dissolve_free_huge_page(pfn_to_page(pfn));
921 }
922 
923 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
924 {
925         struct page *page;
926         unsigned int r_nid;
927 
928         if (h->order >= MAX_ORDER)
929                 return NULL;
930 
931         /*
932          * Assume we will successfully allocate the surplus page to
933          * prevent racing processes from causing the surplus to exceed
934          * overcommit
935          *
936          * This however introduces a different race, where a process B
937          * tries to grow the static hugepage pool while alloc_pages() is
938          * called by process A. B will only examine the per-node
939          * counters in determining if surplus huge pages can be
940          * converted to normal huge pages in adjust_pool_surplus(). A
941          * won't be able to increment the per-node counter, until the
942          * lock is dropped by B, but B doesn't drop hugetlb_lock until
943          * no more huge pages can be converted from surplus to normal
944          * state (and doesn't try to convert again). Thus, we have a
945          * case where a surplus huge page exists, the pool is grown, and
946          * the surplus huge page still exists after, even though it
947          * should just have been converted to a normal huge page. This
948          * does not leak memory, though, as the hugepage will be freed
949          * once it is out of use. It also does not allow the counters to
950          * go out of whack in adjust_pool_surplus() as we don't modify
951          * the node values until we've gotten the hugepage and only the
952          * per-node value is checked there.
953          */
954         spin_lock(&hugetlb_lock);
955         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
956                 spin_unlock(&hugetlb_lock);
957                 return NULL;
958         } else {
959                 h->nr_huge_pages++;
960                 h->surplus_huge_pages++;
961         }
962         spin_unlock(&hugetlb_lock);
963 
964         if (nid == NUMA_NO_NODE)
965                 page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP|
966                                    __GFP_REPEAT|__GFP_NOWARN,
967                                    huge_page_order(h));
968         else
969                 page = alloc_pages_exact_node(nid,
970                         htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
971                         __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
972 
973         if (page && arch_prepare_hugepage(page)) {
974                 __free_pages(page, huge_page_order(h));
975                 page = NULL;
976         }
977 
978         spin_lock(&hugetlb_lock);
979         if (page) {
980                 INIT_LIST_HEAD(&page->lru);
981                 r_nid = page_to_nid(page);
982                 set_compound_page_dtor(page, free_huge_page);
983                 set_hugetlb_cgroup(page, NULL);
984                 /*
985                  * We incremented the global counters already
986                  */
987                 h->nr_huge_pages_node[r_nid]++;
988                 h->surplus_huge_pages_node[r_nid]++;
989                 __count_vm_event(HTLB_BUDDY_PGALLOC);
990         } else {
991                 h->nr_huge_pages--;
992                 h->surplus_huge_pages--;
993                 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
994         }
995         spin_unlock(&hugetlb_lock);
996 
997         return page;
998 }
999 
1000 /*
1001  * This allocation function is useful in the context where vma is irrelevant.
1002  * E.g. soft-offlining uses this function because it only cares physical
1003  * address of error page.
1004  */
1005 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1006 {
1007         struct page *page = NULL;
1008 
1009         spin_lock(&hugetlb_lock);
1010         if (h->free_huge_pages - h->resv_huge_pages > 0)
1011                 page = dequeue_huge_page_node(h, nid);
1012         spin_unlock(&hugetlb_lock);
1013 
1014         if (!page)
1015                 page = alloc_buddy_huge_page(h, nid);
1016 
1017         return page;
1018 }
1019 
1020 /*
1021  * Increase the hugetlb pool such that it can accommodate a reservation
1022  * of size 'delta'.
1023  */
1024 static int gather_surplus_pages(struct hstate *h, int delta)
1025 {
1026         struct list_head surplus_list;
1027         struct page *page, *tmp;
1028         int ret, i;
1029         int needed, allocated;
1030         bool alloc_ok = true;
1031 
1032         needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1033         if (needed <= 0) {
1034                 h->resv_huge_pages += delta;
1035                 return 0;
1036         }
1037 
1038         allocated = 0;
1039         INIT_LIST_HEAD(&surplus_list);
1040 
1041         ret = -ENOMEM;
1042 retry:
1043         spin_unlock(&hugetlb_lock);
1044         for (i = 0; i < needed; i++) {
1045                 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1046                 if (!page) {
1047                         alloc_ok = false;
1048                         break;
1049                 }
1050                 list_add(&page->lru, &surplus_list);
1051         }
1052         allocated += i;
1053 
1054         /*
1055          * After retaking hugetlb_lock, we need to recalculate 'needed'
1056          * because either resv_huge_pages or free_huge_pages may have changed.
1057          */
1058         spin_lock(&hugetlb_lock);
1059         needed = (h->resv_huge_pages + delta) -
1060                         (h->free_huge_pages + allocated);
1061         if (needed > 0) {
1062                 if (alloc_ok)
1063                         goto retry;
1064                 /*
1065                  * We were not able to allocate enough pages to
1066                  * satisfy the entire reservation so we free what
1067                  * we've allocated so far.
1068                  */
1069                 goto free;
1070         }
1071         /*
1072          * The surplus_list now contains _at_least_ the number of extra pages
1073          * needed to accommodate the reservation.  Add the appropriate number
1074          * of pages to the hugetlb pool and free the extras back to the buddy
1075          * allocator.  Commit the entire reservation here to prevent another
1076          * process from stealing the pages as they are added to the pool but
1077          * before they are reserved.
1078          */
1079         needed += allocated;
1080         h->resv_huge_pages += delta;
1081         ret = 0;
1082 
1083         /* Free the needed pages to the hugetlb pool */
1084         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1085                 if ((--needed) < 0)
1086                         break;
1087                 /*
1088                  * This page is now managed by the hugetlb allocator and has
1089                  * no users -- drop the buddy allocator's reference.
1090                  */
1091                 put_page_testzero(page);
1092                 VM_BUG_ON_PAGE(page_count(page), page);
1093                 enqueue_huge_page(h, page);
1094         }
1095 free:
1096         spin_unlock(&hugetlb_lock);
1097 
1098         /* Free unnecessary surplus pages to the buddy allocator */
1099         list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1100                 put_page(page);
1101         spin_lock(&hugetlb_lock);
1102 
1103         return ret;
1104 }
1105 
1106 /*
1107  * When releasing a hugetlb pool reservation, any surplus pages that were
1108  * allocated to satisfy the reservation must be explicitly freed if they were
1109  * never used.
1110  * Called with hugetlb_lock held.
1111  */
1112 static void return_unused_surplus_pages(struct hstate *h,
1113                                         unsigned long unused_resv_pages)
1114 {
1115         unsigned long nr_pages;
1116 
1117         /* Uncommit the reservation */
1118         h->resv_huge_pages -= unused_resv_pages;
1119 
1120         /* Cannot return gigantic pages currently */
1121         if (h->order >= MAX_ORDER)
1122                 return;
1123 
1124         nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1125 
1126         /*
1127          * We want to release as many surplus pages as possible, spread
1128          * evenly across all nodes with memory. Iterate across these nodes
1129          * until we can no longer free unreserved surplus pages. This occurs
1130          * when the nodes with surplus pages have no free pages.
1131          * free_pool_huge_page() will balance the the freed pages across the
1132          * on-line nodes with memory and will handle the hstate accounting.
1133          */
1134         while (nr_pages--) {
1135                 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1136                         break;
1137                 cond_resched_lock(&hugetlb_lock);
1138         }
1139 }
1140 
1141 /*
1142  * Determine if the huge page at addr within the vma has an associated
1143  * reservation.  Where it does not we will need to logically increase
1144  * reservation and actually increase subpool usage before an allocation
1145  * can occur.  Where any new reservation would be required the
1146  * reservation change is prepared, but not committed.  Once the page
1147  * has been allocated from the subpool and instantiated the change should
1148  * be committed via vma_commit_reservation.  No action is required on
1149  * failure.
1150  */
1151 static long vma_needs_reservation(struct hstate *h,
1152                         struct vm_area_struct *vma, unsigned long addr)
1153 {
1154         struct address_space *mapping = vma->vm_file->f_mapping;
1155         struct inode *inode = mapping->host;
1156 
1157         if (vma->vm_flags & VM_MAYSHARE) {
1158                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1159                 return region_chg(&inode->i_mapping->private_list,
1160                                                         idx, idx + 1);
1161 
1162         } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1163                 return 1;
1164 
1165         } else  {
1166                 long err;
1167                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1168                 struct resv_map *resv = vma_resv_map(vma);
1169 
1170                 err = region_chg(&resv->regions, idx, idx + 1);
1171                 if (err < 0)
1172                         return err;
1173                 return 0;
1174         }
1175 }
1176 static void vma_commit_reservation(struct hstate *h,
1177                         struct vm_area_struct *vma, unsigned long addr)
1178 {
1179         struct address_space *mapping = vma->vm_file->f_mapping;
1180         struct inode *inode = mapping->host;
1181 
1182         if (vma->vm_flags & VM_MAYSHARE) {
1183                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1184                 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1185 
1186         } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1187                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1188                 struct resv_map *resv = vma_resv_map(vma);
1189 
1190                 /* Mark this page used in the map. */
1191                 region_add(&resv->regions, idx, idx + 1);
1192         }
1193 }
1194 
1195 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1196                                     unsigned long addr, int avoid_reserve)
1197 {
1198         struct hugepage_subpool *spool = subpool_vma(vma);
1199         struct hstate *h = hstate_vma(vma);
1200         struct page *page;
1201         long chg;
1202         int ret, idx;
1203         struct hugetlb_cgroup *h_cg;
1204 
1205         idx = hstate_index(h);
1206         /*
1207          * Processes that did not create the mapping will have no
1208          * reserves and will not have accounted against subpool
1209          * limit. Check that the subpool limit can be made before
1210          * satisfying the allocation MAP_NORESERVE mappings may also
1211          * need pages and subpool limit allocated allocated if no reserve
1212          * mapping overlaps.
1213          */
1214         chg = vma_needs_reservation(h, vma, addr);
1215         if (chg < 0)
1216                 return ERR_PTR(-ENOMEM);
1217         if (chg || avoid_reserve)
1218                 if (hugepage_subpool_get_pages(spool, 1))
1219                         return ERR_PTR(-ENOSPC);
1220 
1221         ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1222         if (ret) {
1223                 if (chg || avoid_reserve)
1224                         hugepage_subpool_put_pages(spool, 1);
1225                 return ERR_PTR(-ENOSPC);
1226         }
1227         spin_lock(&hugetlb_lock);
1228         page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg);
1229         if (!page) {
1230                 spin_unlock(&hugetlb_lock);
1231                 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1232                 if (!page) {
1233                         hugetlb_cgroup_uncharge_cgroup(idx,
1234                                                        pages_per_huge_page(h),
1235                                                        h_cg);
1236                         if (chg || avoid_reserve)
1237                                 hugepage_subpool_put_pages(spool, 1);
1238                         return ERR_PTR(-ENOSPC);
1239                 }
1240                 spin_lock(&hugetlb_lock);
1241                 list_move(&page->lru, &h->hugepage_activelist);
1242                 /* Fall through */
1243         }
1244         hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1245         spin_unlock(&hugetlb_lock);
1246 
1247         set_page_private(page, (unsigned long)spool);
1248 
1249         vma_commit_reservation(h, vma, addr);
1250         return page;
1251 }
1252 
1253 /*
1254  * alloc_huge_page()'s wrapper which simply returns the page if allocation
1255  * succeeds, otherwise NULL. This function is called from new_vma_page(),
1256  * where no ERR_VALUE is expected to be returned.
1257  */
1258 struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
1259                                 unsigned long addr, int avoid_reserve)
1260 {
1261         struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
1262         if (IS_ERR(page))
1263                 page = NULL;
1264         return page;
1265 }
1266 
1267 int __weak alloc_bootmem_huge_page(struct hstate *h)
1268 {
1269         struct huge_bootmem_page *m;
1270         int nr_nodes, node;
1271 
1272         for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
1273                 void *addr;
1274 
1275                 addr = memblock_virt_alloc_try_nid_nopanic(
1276                                 huge_page_size(h), huge_page_size(h),
1277                                 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
1278                 if (addr) {
1279                         /*
1280                          * Use the beginning of the huge page to store the
1281                          * huge_bootmem_page struct (until gather_bootmem
1282                          * puts them into the mem_map).
1283                          */
1284                         m = addr;
1285                         goto found;
1286                 }
1287         }
1288         return 0;
1289 
1290 found:
1291         BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1292         /* Put them into a private list first because mem_map is not up yet */
1293         list_add(&m->list, &huge_boot_pages);
1294         m->hstate = h;
1295         return 1;
1296 }
1297 
1298 static void prep_compound_huge_page(struct page *page, int order)
1299 {
1300         if (unlikely(order > (MAX_ORDER - 1)))
1301                 prep_compound_gigantic_page(page, order);
1302         else
1303                 prep_compound_page(page, order);
1304 }
1305 
1306 /* Put bootmem huge pages into the standard lists after mem_map is up */
1307 static void __init gather_bootmem_prealloc(void)
1308 {
1309         struct huge_bootmem_page *m;
1310 
1311         list_for_each_entry(m, &huge_boot_pages, list) {
1312                 struct hstate *h = m->hstate;
1313                 struct page *page;
1314 
1315 #ifdef CONFIG_HIGHMEM
1316                 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1317                 memblock_free_late(__pa(m),
1318                                    sizeof(struct huge_bootmem_page));
1319 #else
1320                 page = virt_to_page(m);
1321 #endif
1322                 WARN_ON(page_count(page) != 1);
1323                 prep_compound_huge_page(page, h->order);
1324                 WARN_ON(PageReserved(page));
1325                 prep_new_huge_page(h, page, page_to_nid(page));
1326                 /*
1327                  * If we had gigantic hugepages allocated at boot time, we need
1328                  * to restore the 'stolen' pages to totalram_pages in order to
1329                  * fix confusing memory reports from free(1) and another
1330                  * side-effects, like CommitLimit going negative.
1331                  */
1332                 if (h->order > (MAX_ORDER - 1))
1333                         adjust_managed_page_count(page, 1 << h->order);
1334         }
1335 }
1336 
1337 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1338 {
1339         unsigned long i;
1340 
1341         for (i = 0; i < h->max_huge_pages; ++i) {
1342                 if (h->order >= MAX_ORDER) {
1343                         if (!alloc_bootmem_huge_page(h))
1344                                 break;
1345                 } else if (!alloc_fresh_huge_page(h,
1346                                          &node_states[N_MEMORY]))
1347                         break;
1348         }
1349         h->max_huge_pages = i;
1350 }
1351 
1352 static void __init hugetlb_init_hstates(void)
1353 {
1354         struct hstate *h;
1355 
1356         for_each_hstate(h) {
1357                 /* oversize hugepages were init'ed in early boot */
1358                 if (h->order < MAX_ORDER)
1359                         hugetlb_hstate_alloc_pages(h);
1360         }
1361 }
1362 
1363 static char * __init memfmt(char *buf, unsigned long n)
1364 {
1365         if (n >= (1UL << 30))
1366                 sprintf(buf, "%lu GB", n >> 30);
1367         else if (n >= (1UL << 20))
1368                 sprintf(buf, "%lu MB", n >> 20);
1369         else
1370                 sprintf(buf, "%lu KB", n >> 10);
1371         return buf;
1372 }
1373 
1374 static void __init report_hugepages(void)
1375 {
1376         struct hstate *h;
1377 
1378         for_each_hstate(h) {
1379                 char buf[32];
1380                 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1381                         memfmt(buf, huge_page_size(h)),
1382                         h->free_huge_pages);
1383         }
1384 }
1385 
1386 #ifdef CONFIG_HIGHMEM
1387 static void try_to_free_low(struct hstate *h, unsigned long count,
1388                                                 nodemask_t *nodes_allowed)
1389 {
1390         int i;
1391 
1392         if (h->order >= MAX_ORDER)
1393                 return;
1394 
1395         for_each_node_mask(i, *nodes_allowed) {
1396                 struct page *page, *next;
1397                 struct list_head *freel = &h->hugepage_freelists[i];
1398                 list_for_each_entry_safe(page, next, freel, lru) {
1399                         if (count >= h->nr_huge_pages)
1400                                 return;
1401                         if (PageHighMem(page))
1402                                 continue;
1403                         list_del(&page->lru);
1404                         update_and_free_page(h, page);
1405                         h->free_huge_pages--;
1406                         h->free_huge_pages_node[page_to_nid(page)]--;
1407                 }
1408         }
1409 }
1410 #else
1411 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1412                                                 nodemask_t *nodes_allowed)
1413 {
1414 }
1415 #endif
1416 
1417 /*
1418  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
1419  * balanced by operating on them in a round-robin fashion.
1420  * Returns 1 if an adjustment was made.
1421  */
1422 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1423                                 int delta)
1424 {
1425         int nr_nodes, node;
1426 
1427         VM_BUG_ON(delta != -1 && delta != 1);
1428 
1429         if (delta < 0) {
1430                 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1431                         if (h->surplus_huge_pages_node[node])
1432                                 goto found;
1433                 }
1434         } else {
1435                 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1436                         if (h->surplus_huge_pages_node[node] <
1437                                         h->nr_huge_pages_node[node])
1438                                 goto found;
1439                 }
1440         }
1441         return 0;
1442 
1443 found:
1444         h->surplus_huge_pages += delta;
1445         h->surplus_huge_pages_node[node] += delta;
1446         return 1;
1447 }
1448 
1449 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1450 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1451                                                 nodemask_t *nodes_allowed)
1452 {
1453         unsigned long min_count, ret;
1454 
1455         if (h->order >= MAX_ORDER)
1456                 return h->max_huge_pages;
1457 
1458         /*
1459          * Increase the pool size
1460          * First take pages out of surplus state.  Then make up the
1461          * remaining difference by allocating fresh huge pages.
1462          *
1463          * We might race with alloc_buddy_huge_page() here and be unable
1464          * to convert a surplus huge page to a normal huge page. That is
1465          * not critical, though, it just means the overall size of the
1466          * pool might be one hugepage larger than it needs to be, but
1467          * within all the constraints specified by the sysctls.
1468          */
1469         spin_lock(&hugetlb_lock);
1470         while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1471                 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1472                         break;
1473         }
1474 
1475         while (count > persistent_huge_pages(h)) {
1476                 /*
1477                  * If this allocation races such that we no longer need the
1478                  * page, free_huge_page will handle it by freeing the page
1479                  * and reducing the surplus.
1480                  */
1481                 spin_unlock(&hugetlb_lock);
1482                 ret = alloc_fresh_huge_page(h, nodes_allowed);
1483                 spin_lock(&hugetlb_lock);
1484                 if (!ret)
1485                         goto out;
1486 
1487                 /* Bail for signals. Probably ctrl-c from user */
1488                 if (signal_pending(current))
1489                         goto out;
1490         }
1491 
1492         /*
1493          * Decrease the pool size
1494          * First return free pages to the buddy allocator (being careful
1495          * to keep enough around to satisfy reservations).  Then place
1496          * pages into surplus state as needed so the pool will shrink
1497          * to the desired size as pages become free.
1498          *
1499          * By placing pages into the surplus state independent of the
1500          * overcommit value, we are allowing the surplus pool size to
1501          * exceed overcommit. There are few sane options here. Since
1502          * alloc_buddy_huge_page() is checking the global counter,
1503          * though, we'll note that we're not allowed to exceed surplus
1504          * and won't grow the pool anywhere else. Not until one of the
1505          * sysctls are changed, or the surplus pages go out of use.
1506          */
1507         min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1508         min_count = max(count, min_count);
1509         try_to_free_low(h, min_count, nodes_allowed);
1510         while (min_count < persistent_huge_pages(h)) {
1511                 if (!free_pool_huge_page(h, nodes_allowed, 0))
1512                         break;
1513                 cond_resched_lock(&hugetlb_lock);
1514         }
1515         while (count < persistent_huge_pages(h)) {
1516                 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1517                         break;
1518         }
1519 out:
1520         ret = persistent_huge_pages(h);
1521         spin_unlock(&hugetlb_lock);
1522         return ret;
1523 }
1524 
1525 #define HSTATE_ATTR_RO(_name) \
1526         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1527 
1528 #define HSTATE_ATTR(_name) \
1529         static struct kobj_attribute _name##_attr = \
1530                 __ATTR(_name, 0644, _name##_show, _name##_store)
1531 
1532 static struct kobject *hugepages_kobj;
1533 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1534 
1535 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1536 
1537 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1538 {
1539         int i;
1540 
1541         for (i = 0; i < HUGE_MAX_HSTATE; i++)
1542                 if (hstate_kobjs[i] == kobj) {
1543                         if (nidp)
1544                                 *nidp = NUMA_NO_NODE;
1545                         return &hstates[i];
1546                 }
1547 
1548         return kobj_to_node_hstate(kobj, nidp);
1549 }
1550 
1551 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1552                                         struct kobj_attribute *attr, char *buf)
1553 {
1554         struct hstate *h;
1555         unsigned long nr_huge_pages;
1556         int nid;
1557 
1558         h = kobj_to_hstate(kobj, &nid);
1559         if (nid == NUMA_NO_NODE)
1560                 nr_huge_pages = h->nr_huge_pages;
1561         else
1562                 nr_huge_pages = h->nr_huge_pages_node[nid];
1563 
1564         return sprintf(buf, "%lu\n", nr_huge_pages);
1565 }
1566 
1567 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1568                         struct kobject *kobj, struct kobj_attribute *attr,
1569                         const char *buf, size_t len)
1570 {
1571         int err;
1572         int nid;
1573         unsigned long count;
1574         struct hstate *h;
1575         NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1576 
1577         err = kstrtoul(buf, 10, &count);
1578         if (err)
1579                 goto out;
1580 
1581         h = kobj_to_hstate(kobj, &nid);
1582         if (h->order >= MAX_ORDER) {
1583                 err = -EINVAL;
1584                 goto out;
1585         }
1586 
1587         if (nid == NUMA_NO_NODE) {
1588                 /*
1589                  * global hstate attribute
1590                  */
1591                 if (!(obey_mempolicy &&
1592                                 init_nodemask_of_mempolicy(nodes_allowed))) {
1593                         NODEMASK_FREE(nodes_allowed);
1594                         nodes_allowed = &node_states[N_MEMORY];
1595                 }
1596         } else if (nodes_allowed) {
1597                 /*
1598                  * per node hstate attribute: adjust count to global,
1599                  * but restrict alloc/free to the specified node.
1600                  */
1601                 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1602                 init_nodemask_of_node(nodes_allowed, nid);
1603         } else
1604                 nodes_allowed = &node_states[N_MEMORY];
1605 
1606         h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1607 
1608         if (nodes_allowed != &node_states[N_MEMORY])
1609                 NODEMASK_FREE(nodes_allowed);
1610 
1611         return len;
1612 out:
1613         NODEMASK_FREE(nodes_allowed);
1614         return err;
1615 }
1616 
1617 static ssize_t nr_hugepages_show(struct kobject *kobj,
1618                                        struct kobj_attribute *attr, char *buf)
1619 {
1620         return nr_hugepages_show_common(kobj, attr, buf);
1621 }
1622 
1623 static ssize_t nr_hugepages_store(struct kobject *kobj,
1624                struct kobj_attribute *attr, const char *buf, size_t len)
1625 {
1626         return nr_hugepages_store_common(false, kobj, attr, buf, len);
1627 }
1628 HSTATE_ATTR(nr_hugepages);
1629 
1630 #ifdef CONFIG_NUMA
1631 
1632 /*
1633  * hstate attribute for optionally mempolicy-based constraint on persistent
1634  * huge page alloc/free.
1635  */
1636 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1637                                        struct kobj_attribute *attr, char *buf)
1638 {
1639         return nr_hugepages_show_common(kobj, attr, buf);
1640 }
1641 
1642 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1643                struct kobj_attribute *attr, const char *buf, size_t len)
1644 {
1645         return nr_hugepages_store_common(true, kobj, attr, buf, len);
1646 }
1647 HSTATE_ATTR(nr_hugepages_mempolicy);
1648 #endif
1649 
1650 
1651 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1652                                         struct kobj_attribute *attr, char *buf)
1653 {
1654         struct hstate *h = kobj_to_hstate(kobj, NULL);
1655         return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1656 }
1657 
1658 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1659                 struct kobj_attribute *attr, const char *buf, size_t count)
1660 {
1661         int err;
1662         unsigned long input;
1663         struct hstate *h = kobj_to_hstate(kobj, NULL);
1664 
1665         if (h->order >= MAX_ORDER)
1666                 return -EINVAL;
1667 
1668         err = kstrtoul(buf, 10, &input);
1669         if (err)
1670                 return err;
1671 
1672         spin_lock(&hugetlb_lock);
1673         h->nr_overcommit_huge_pages = input;
1674         spin_unlock(&hugetlb_lock);
1675 
1676         return count;
1677 }
1678 HSTATE_ATTR(nr_overcommit_hugepages);
1679 
1680 static ssize_t free_hugepages_show(struct kobject *kobj,
1681                                         struct kobj_attribute *attr, char *buf)
1682 {
1683         struct hstate *h;
1684         unsigned long free_huge_pages;
1685         int nid;
1686 
1687         h = kobj_to_hstate(kobj, &nid);
1688         if (nid == NUMA_NO_NODE)
1689                 free_huge_pages = h->free_huge_pages;
1690         else
1691                 free_huge_pages = h->free_huge_pages_node[nid];
1692 
1693         return sprintf(buf, "%lu\n", free_huge_pages);
1694 }
1695 HSTATE_ATTR_RO(free_hugepages);
1696 
1697 static ssize_t resv_hugepages_show(struct kobject *kobj,
1698                                         struct kobj_attribute *attr, char *buf)
1699 {
1700         struct hstate *h = kobj_to_hstate(kobj, NULL);
1701         return sprintf(buf, "%lu\n", h->resv_huge_pages);
1702 }
1703 HSTATE_ATTR_RO(resv_hugepages);
1704 
1705 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1706                                         struct kobj_attribute *attr, char *buf)
1707 {
1708         struct hstate *h;
1709         unsigned long surplus_huge_pages;
1710         int nid;
1711 
1712         h = kobj_to_hstate(kobj, &nid);
1713         if (nid == NUMA_NO_NODE)
1714                 surplus_huge_pages = h->surplus_huge_pages;
1715         else
1716                 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1717 
1718         return sprintf(buf, "%lu\n", surplus_huge_pages);
1719 }
1720 HSTATE_ATTR_RO(surplus_hugepages);
1721 
1722 static struct attribute *hstate_attrs[] = {
1723         &nr_hugepages_attr.attr,
1724         &nr_overcommit_hugepages_attr.attr,
1725         &free_hugepages_attr.attr,
1726         &resv_hugepages_attr.attr,
1727         &surplus_hugepages_attr.attr,
1728 #ifdef CONFIG_NUMA
1729         &nr_hugepages_mempolicy_attr.attr,
1730 #endif
1731         NULL,
1732 };
1733 
1734 static struct attribute_group hstate_attr_group = {
1735         .attrs = hstate_attrs,
1736 };
1737 
1738 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1739                                     struct kobject **hstate_kobjs,
1740                                     struct attribute_group *hstate_attr_group)
1741 {
1742         int retval;
1743         int hi = hstate_index(h);
1744 
1745         hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1746         if (!hstate_kobjs[hi])
1747                 return -ENOMEM;
1748 
1749         retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1750         if (retval)
1751                 kobject_put(hstate_kobjs[hi]);
1752 
1753         return retval;
1754 }
1755 
1756 static void __init hugetlb_sysfs_init(void)
1757 {
1758         struct hstate *h;
1759         int err;
1760 
1761         hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1762         if (!hugepages_kobj)
1763                 return;
1764 
1765         for_each_hstate(h) {
1766                 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1767                                          hstate_kobjs, &hstate_attr_group);
1768                 if (err)
1769                         pr_err("Hugetlb: Unable to add hstate %s", h->name);
1770         }
1771 }
1772 
1773 #ifdef CONFIG_NUMA
1774 
1775 /*
1776  * node_hstate/s - associate per node hstate attributes, via their kobjects,
1777  * with node devices in node_devices[] using a parallel array.  The array
1778  * index of a node device or _hstate == node id.
1779  * This is here to avoid any static dependency of the node device driver, in
1780  * the base kernel, on the hugetlb module.
1781  */
1782 struct node_hstate {
1783         struct kobject          *hugepages_kobj;
1784         struct kobject          *hstate_kobjs[HUGE_MAX_HSTATE];
1785 };
1786 struct node_hstate node_hstates[MAX_NUMNODES];
1787 
1788 /*
1789  * A subset of global hstate attributes for node devices
1790  */
1791 static struct attribute *per_node_hstate_attrs[] = {
1792         &nr_hugepages_attr.attr,
1793         &free_hugepages_attr.attr,
1794         &surplus_hugepages_attr.attr,
1795         NULL,
1796 };
1797 
1798 static struct attribute_group per_node_hstate_attr_group = {
1799         .attrs = per_node_hstate_attrs,
1800 };
1801 
1802 /*
1803  * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1804  * Returns node id via non-NULL nidp.
1805  */
1806 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1807 {
1808         int nid;
1809 
1810         for (nid = 0; nid < nr_node_ids; nid++) {
1811                 struct node_hstate *nhs = &node_hstates[nid];
1812                 int i;
1813                 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1814                         if (nhs->hstate_kobjs[i] == kobj) {
1815                                 if (nidp)
1816                                         *nidp = nid;
1817                                 return &hstates[i];
1818                         }
1819         }
1820 
1821         BUG();
1822         return NULL;
1823 }
1824 
1825 /*
1826  * Unregister hstate attributes from a single node device.
1827  * No-op if no hstate attributes attached.
1828  */
1829 static void hugetlb_unregister_node(struct node *node)
1830 {
1831         struct hstate *h;
1832         struct node_hstate *nhs = &node_hstates[node->dev.id];
1833 
1834         if (!nhs->hugepages_kobj)
1835                 return;         /* no hstate attributes */
1836 
1837         for_each_hstate(h) {
1838                 int idx = hstate_index(h);
1839                 if (nhs->hstate_kobjs[idx]) {
1840                         kobject_put(nhs->hstate_kobjs[idx]);
1841                         nhs->hstate_kobjs[idx] = NULL;
1842                 }
1843         }
1844 
1845         kobject_put(nhs->hugepages_kobj);
1846         nhs->hugepages_kobj = NULL;
1847 }
1848 
1849 /*
1850  * hugetlb module exit:  unregister hstate attributes from node devices
1851  * that have them.
1852  */
1853 static void hugetlb_unregister_all_nodes(void)
1854 {
1855         int nid;
1856 
1857         /*
1858          * disable node device registrations.
1859          */
1860         register_hugetlbfs_with_node(NULL, NULL);
1861 
1862         /*
1863          * remove hstate attributes from any nodes that have them.
1864          */
1865         for (nid = 0; nid < nr_node_ids; nid++)
1866                 hugetlb_unregister_node(node_devices[nid]);
1867 }
1868 
1869 /*
1870  * Register hstate attributes for a single node device.
1871  * No-op if attributes already registered.
1872  */
1873 static void hugetlb_register_node(struct node *node)
1874 {
1875         struct hstate *h;
1876         struct node_hstate *nhs = &node_hstates[node->dev.id];
1877         int err;
1878 
1879         if (nhs->hugepages_kobj)
1880                 return;         /* already allocated */
1881 
1882         nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1883                                                         &node->dev.kobj);
1884         if (!nhs->hugepages_kobj)
1885                 return;
1886 
1887         for_each_hstate(h) {
1888                 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1889                                                 nhs->hstate_kobjs,
1890                                                 &per_node_hstate_attr_group);
1891                 if (err) {
1892                         pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
1893                                 h->name, node->dev.id);
1894                         hugetlb_unregister_node(node);
1895                         break;
1896                 }
1897         }
1898 }
1899 
1900 /*
1901  * hugetlb init time:  register hstate attributes for all registered node
1902  * devices of nodes that have memory.  All on-line nodes should have
1903  * registered their associated device by this time.
1904  */
1905 static void hugetlb_register_all_nodes(void)
1906 {
1907         int nid;
1908 
1909         for_each_node_state(nid, N_MEMORY) {
1910                 struct node *node = node_devices[nid];
1911                 if (node->dev.id == nid)
1912                         hugetlb_register_node(node);
1913         }
1914 
1915         /*
1916          * Let the node device driver know we're here so it can
1917          * [un]register hstate attributes on node hotplug.
1918          */
1919         register_hugetlbfs_with_node(hugetlb_register_node,
1920                                      hugetlb_unregister_node);
1921 }
1922 #else   /* !CONFIG_NUMA */
1923 
1924 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1925 {
1926         BUG();
1927         if (nidp)
1928                 *nidp = -1;
1929         return NULL;
1930 }
1931 
1932 static void hugetlb_unregister_all_nodes(void) { }
1933 
1934 static void hugetlb_register_all_nodes(void) { }
1935 
1936 #endif
1937 
1938 static void __exit hugetlb_exit(void)
1939 {
1940         struct hstate *h;
1941 
1942         hugetlb_unregister_all_nodes();
1943 
1944         for_each_hstate(h) {
1945                 kobject_put(hstate_kobjs[hstate_index(h)]);
1946         }
1947 
1948         kobject_put(hugepages_kobj);
1949 }
1950 module_exit(hugetlb_exit);
1951 
1952 static int __init hugetlb_init(void)
1953 {
1954         /* Some platform decide whether they support huge pages at boot
1955          * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1956          * there is no such support
1957          */
1958         if (HPAGE_SHIFT == 0)
1959                 return 0;
1960 
1961         if (!size_to_hstate(default_hstate_size)) {
1962                 default_hstate_size = HPAGE_SIZE;
1963                 if (!size_to_hstate(default_hstate_size))
1964                         hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1965         }
1966         default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
1967         if (default_hstate_max_huge_pages)
1968                 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1969 
1970         hugetlb_init_hstates();
1971         gather_bootmem_prealloc();
1972         report_hugepages();
1973 
1974         hugetlb_sysfs_init();
1975         hugetlb_register_all_nodes();
1976         hugetlb_cgroup_file_init();
1977 
1978         return 0;
1979 }
1980 module_init(hugetlb_init);
1981 
1982 /* Should be called on processing a hugepagesz=... option */
1983 void __init hugetlb_add_hstate(unsigned order)
1984 {
1985         struct hstate *h;
1986         unsigned long i;
1987 
1988         if (size_to_hstate(PAGE_SIZE << order)) {
1989                 pr_warning("hugepagesz= specified twice, ignoring\n");
1990                 return;
1991         }
1992         BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
1993         BUG_ON(order == 0);
1994         h = &hstates[hugetlb_max_hstate++];
1995         h->order = order;
1996         h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1997         h->nr_huge_pages = 0;
1998         h->free_huge_pages = 0;
1999         for (i = 0; i < MAX_NUMNODES; ++i)
2000                 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2001         INIT_LIST_HEAD(&h->hugepage_activelist);
2002         h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
2003         h->next_nid_to_free = first_node(node_states[N_MEMORY]);
2004         snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2005                                         huge_page_size(h)/1024);
2006 
2007         parsed_hstate = h;
2008 }
2009 
2010 static int __init hugetlb_nrpages_setup(char *s)
2011 {
2012         unsigned long *mhp;
2013         static unsigned long *last_mhp;
2014 
2015         /*
2016          * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2017          * so this hugepages= parameter goes to the "default hstate".
2018          */
2019         if (!hugetlb_max_hstate)
2020                 mhp = &default_hstate_max_huge_pages;
2021         else
2022                 mhp = &parsed_hstate->max_huge_pages;
2023 
2024         if (mhp == last_mhp) {
2025                 pr_warning("hugepages= specified twice without "
2026                            "interleaving hugepagesz=, ignoring\n");
2027                 return 1;
2028         }
2029 
2030         if (sscanf(s, "%lu", mhp) <= 0)
2031                 *mhp = 0;
2032 
2033         /*
2034          * Global state is always initialized later in hugetlb_init.
2035          * But we need to allocate >= MAX_ORDER hstates here early to still
2036          * use the bootmem allocator.
2037          */
2038         if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2039                 hugetlb_hstate_alloc_pages(parsed_hstate);
2040 
2041         last_mhp = mhp;
2042 
2043         return 1;
2044 }
2045 __setup("hugepages=", hugetlb_nrpages_setup);
2046 
2047 static int __init hugetlb_default_setup(char *s)
2048 {
2049         default_hstate_size = memparse(s, &s);
2050         return 1;
2051 }
2052 __setup("default_hugepagesz=", hugetlb_default_setup);
2053 
2054 static unsigned int cpuset_mems_nr(unsigned int *array)
2055 {
2056         int node;
2057         unsigned int nr = 0;
2058 
2059         for_each_node_mask(node, cpuset_current_mems_allowed)
2060                 nr += array[node];
2061 
2062         return nr;
2063 }
2064 
2065 #ifdef CONFIG_SYSCTL
2066 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2067                          struct ctl_table *table, int write,
2068                          void __user *buffer, size_t *length, loff_t *ppos)
2069 {
2070         struct hstate *h = &default_hstate;
2071         unsigned long tmp;
2072         int ret;
2073 
2074         if (!hugepages_supported())
2075                 return -ENOTSUPP;
2076 
2077         tmp = h->max_huge_pages;
2078 
2079         if (write && h->order >= MAX_ORDER)
2080                 return -EINVAL;
2081 
2082         table->data = &tmp;
2083         table->maxlen = sizeof(unsigned long);
2084         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2085         if (ret)
2086                 goto out;
2087 
2088         if (write) {
2089                 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
2090                                                 GFP_KERNEL | __GFP_NORETRY);
2091                 if (!(obey_mempolicy &&
2092                                init_nodemask_of_mempolicy(nodes_allowed))) {
2093                         NODEMASK_FREE(nodes_allowed);
2094                         nodes_allowed = &node_states[N_MEMORY];
2095                 }
2096                 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2097 
2098                 if (nodes_allowed != &node_states[N_MEMORY])
2099                         NODEMASK_FREE(nodes_allowed);
2100         }
2101 out:
2102         return ret;
2103 }
2104 
2105 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2106                           void __user *buffer, size_t *length, loff_t *ppos)
2107 {
2108 
2109         return hugetlb_sysctl_handler_common(false, table, write,
2110                                                         buffer, length, ppos);
2111 }
2112 
2113 #ifdef CONFIG_NUMA
2114 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2115                           void __user *buffer, size_t *length, loff_t *ppos)
2116 {
2117         return hugetlb_sysctl_handler_common(true, table, write,
2118                                                         buffer, length, ppos);
2119 }
2120 #endif /* CONFIG_NUMA */
2121 
2122 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2123                         void __user *buffer,
2124                         size_t *length, loff_t *ppos)
2125 {
2126         struct hstate *h = &default_hstate;
2127         unsigned long tmp;
2128         int ret;
2129 
2130         if (!hugepages_supported())
2131                 return -ENOTSUPP;
2132 
2133         tmp = h->nr_overcommit_huge_pages;
2134 
2135         if (write && h->order >= MAX_ORDER)
2136                 return -EINVAL;
2137 
2138         table->data = &tmp;
2139         table->maxlen = sizeof(unsigned long);
2140         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2141         if (ret)
2142                 goto out;
2143 
2144         if (write) {
2145                 spin_lock(&hugetlb_lock);
2146                 h->nr_overcommit_huge_pages = tmp;
2147                 spin_unlock(&hugetlb_lock);
2148         }
2149 out:
2150         return ret;
2151 }
2152 
2153 #endif /* CONFIG_SYSCTL */
2154 
2155 void hugetlb_report_meminfo(struct seq_file *m)
2156 {
2157         struct hstate *h = &default_hstate;
2158         if (!hugepages_supported())
2159                 return;
2160         seq_printf(m,
2161                         "HugePages_Total:   %5lu\n"
2162                         "HugePages_Free:    %5lu\n"
2163                         "HugePages_Rsvd:    %5lu\n"
2164                         "HugePages_Surp:    %5lu\n"
2165                         "Hugepagesize:   %8lu kB\n",
2166                         h->nr_huge_pages,
2167                         h->free_huge_pages,
2168                         h->resv_huge_pages,
2169                         h->surplus_huge_pages,
2170                         1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2171 }
2172 
2173 int hugetlb_report_node_meminfo(int nid, char *buf)
2174 {
2175         struct hstate *h = &default_hstate;
2176         if (!hugepages_supported())
2177                 return 0;
2178         return sprintf(buf,
2179                 "Node %d HugePages_Total: %5u\n"
2180                 "Node %d HugePages_Free:  %5u\n"
2181                 "Node %d HugePages_Surp:  %5u\n",
2182                 nid, h->nr_huge_pages_node[nid],
2183                 nid, h->free_huge_pages_node[nid],
2184                 nid, h->surplus_huge_pages_node[nid]);
2185 }
2186 
2187 void hugetlb_show_meminfo(void)
2188 {
2189         struct hstate *h;
2190         int nid;
2191 
2192         if (!hugepages_supported())
2193                 return;
2194 
2195         for_each_node_state(nid, N_MEMORY)
2196                 for_each_hstate(h)
2197                         pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2198                                 nid,
2199                                 h->nr_huge_pages_node[nid],
2200                                 h->free_huge_pages_node[nid],
2201                                 h->surplus_huge_pages_node[nid],
2202                                 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2203 }
2204 
2205 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2206 unsigned long hugetlb_total_pages(void)
2207 {
2208         struct hstate *h;
2209         unsigned long nr_total_pages = 0;
2210 
2211         for_each_hstate(h)
2212                 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2213         return nr_total_pages;
2214 }
2215 
2216 static int hugetlb_acct_memory(struct hstate *h, long delta)
2217 {
2218         int ret = -ENOMEM;
2219 
2220         spin_lock(&hugetlb_lock);
2221         /*
2222          * When cpuset is configured, it breaks the strict hugetlb page
2223          * reservation as the accounting is done on a global variable. Such
2224          * reservation is completely rubbish in the presence of cpuset because
2225          * the reservation is not checked against page availability for the
2226          * current cpuset. Application can still potentially OOM'ed by kernel
2227          * with lack of free htlb page in cpuset that the task is in.
2228          * Attempt to enforce strict accounting with cpuset is almost
2229          * impossible (or too ugly) because cpuset is too fluid that
2230          * task or memory node can be dynamically moved between cpusets.
2231          *
2232          * The change of semantics for shared hugetlb mapping with cpuset is
2233          * undesirable. However, in order to preserve some of the semantics,
2234          * we fall back to check against current free page availability as
2235          * a best attempt and hopefully to minimize the impact of changing
2236          * semantics that cpuset has.
2237          */
2238         if (delta > 0) {
2239                 if (gather_surplus_pages(h, delta) < 0)
2240                         goto out;
2241 
2242                 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2243                         return_unused_surplus_pages(h, delta);
2244                         goto out;
2245                 }
2246         }
2247 
2248         ret = 0;
2249         if (delta < 0)
2250                 return_unused_surplus_pages(h, (unsigned long) -delta);
2251 
2252 out:
2253         spin_unlock(&hugetlb_lock);
2254         return ret;
2255 }
2256 
2257 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2258 {
2259         struct resv_map *resv = vma_resv_map(vma);
2260 
2261         /*
2262          * This new VMA should share its siblings reservation map if present.
2263          * The VMA will only ever have a valid reservation map pointer where
2264          * it is being copied for another still existing VMA.  As that VMA
2265          * has a reference to the reservation map it cannot disappear until
2266          * after this open call completes.  It is therefore safe to take a
2267          * new reference here without additional locking.
2268          */
2269         if (resv)
2270                 kref_get(&resv->refs);
2271 }
2272 
2273 static void resv_map_put(struct vm_area_struct *vma)
2274 {
2275         struct resv_map *resv = vma_resv_map(vma);
2276 
2277         if (!resv)
2278                 return;
2279         kref_put(&resv->refs, resv_map_release);
2280 }
2281 
2282 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2283 {
2284         struct hstate *h = hstate_vma(vma);
2285         struct resv_map *resv = vma_resv_map(vma);
2286         struct hugepage_subpool *spool = subpool_vma(vma);
2287         unsigned long reserve;
2288         unsigned long start;
2289         unsigned long end;
2290 
2291         if (resv) {
2292                 start = vma_hugecache_offset(h, vma, vma->vm_start);
2293                 end = vma_hugecache_offset(h, vma, vma->vm_end);
2294 
2295                 reserve = (end - start) -
2296                         region_count(&resv->regions, start, end);
2297 
2298                 resv_map_put(vma);
2299 
2300                 if (reserve) {
2301                         hugetlb_acct_memory(h, -reserve);
2302                         hugepage_subpool_put_pages(spool, reserve);
2303                 }
2304         }
2305 }
2306 
2307 /*
2308  * We cannot handle pagefaults against hugetlb pages at all.  They cause
2309  * handle_mm_fault() to try to instantiate regular-sized pages in the
2310  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
2311  * this far.
2312  */
2313 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2314 {
2315         BUG();
2316         return 0;
2317 }
2318 
2319 const struct vm_operations_struct hugetlb_vm_ops = {
2320         .fault = hugetlb_vm_op_fault,
2321         .open = hugetlb_vm_op_open,
2322         .close = hugetlb_vm_op_close,
2323 };
2324 
2325 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2326                                 int writable)
2327 {
2328         pte_t entry;
2329 
2330         if (writable) {
2331                 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2332                                          vma->vm_page_prot)));
2333         } else {
2334                 entry = huge_pte_wrprotect(mk_huge_pte(page,
2335                                            vma->vm_page_prot));
2336         }
2337         entry = pte_mkyoung(entry);
2338         entry = pte_mkhuge(entry);
2339         entry = arch_make_huge_pte(entry, vma, page, writable);
2340 
2341         return entry;
2342 }
2343 
2344 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2345                                    unsigned long address, pte_t *ptep)
2346 {
2347         pte_t entry;
2348 
2349         entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
2350         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2351                 update_mmu_cache(vma, address, ptep);
2352 }
2353 
2354 static int is_hugetlb_entry_migration(pte_t pte)
2355 {
2356         swp_entry_t swp;
2357 
2358         if (huge_pte_none(pte) || pte_present(pte))
2359                 return 0;
2360         swp = pte_to_swp_entry(pte);
2361         if (non_swap_entry(swp) && is_migration_entry(swp))
2362                 return 1;
2363         else
2364                 return 0;
2365 }
2366 
2367 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2368 {
2369         swp_entry_t swp;
2370 
2371         if (huge_pte_none(pte) || pte_present(pte))
2372                 return 0;
2373         swp = pte_to_swp_entry(pte);
2374         if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2375                 return 1;
2376         else
2377                 return 0;
2378 }
2379 
2380 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2381                             struct vm_area_struct *vma)
2382 {
2383         pte_t *src_pte, *dst_pte, entry;
2384         struct page *ptepage;
2385         unsigned long addr;
2386         int cow;
2387         struct hstate *h = hstate_vma(vma);
2388         unsigned long sz = huge_page_size(h);
2389         unsigned long mmun_start;       /* For mmu_notifiers */
2390         unsigned long mmun_end;         /* For mmu_notifiers */
2391         int ret = 0;
2392 
2393         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2394 
2395         mmun_start = vma->vm_start;
2396         mmun_end = vma->vm_end;
2397         if (cow)
2398                 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
2399 
2400         for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2401                 spinlock_t *src_ptl, *dst_ptl;
2402                 src_pte = huge_pte_offset(src, addr);
2403                 if (!src_pte)
2404                         continue;
2405                 dst_pte = huge_pte_alloc(dst, addr, sz);
2406                 if (!dst_pte) {
2407                         ret = -ENOMEM;
2408                         break;
2409                 }
2410 
2411                 /* If the pagetables are shared don't copy or take references */
2412                 if (dst_pte == src_pte)
2413                         continue;
2414 
2415                 dst_ptl = huge_pte_lock(h, dst, dst_pte);
2416                 src_ptl = huge_pte_lockptr(h, src, src_pte);
2417                 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
2418                 entry = huge_ptep_get(src_pte);
2419                 if (huge_pte_none(entry)) { /* skip none entry */
2420                         ;
2421                 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
2422                                     is_hugetlb_entry_hwpoisoned(entry))) {
2423                         swp_entry_t swp_entry = pte_to_swp_entry(entry);
2424 
2425                         if (is_write_migration_entry(swp_entry) && cow) {
2426                                 /*
2427                                  * COW mappings require pages in both
2428                                  * parent and child to be set to read.
2429                                  */
2430                                 make_migration_entry_read(&swp_entry);
2431                                 entry = swp_entry_to_pte(swp_entry);
2432                                 set_huge_pte_at(src, addr, src_pte, entry);
2433                         }
2434                         set_huge_pte_at(dst, addr, dst_pte, entry);
2435                 } else {
2436                         if (cow)
2437                                 huge_ptep_set_wrprotect(src, addr, src_pte);
2438                         entry = huge_ptep_get(src_pte);
2439                         ptepage = pte_page(entry);
2440                         get_page(ptepage);
2441                         page_dup_rmap(ptepage);
2442                         set_huge_pte_at(dst, addr, dst_pte, entry);
2443                 }
2444                 spin_unlock(src_ptl);
2445                 spin_unlock(dst_ptl);
2446         }
2447 
2448         if (cow)
2449                 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
2450 
2451         return ret;
2452 }
2453 
2454 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2455                             unsigned long start, unsigned long end,
2456                             struct page *ref_page)
2457 {
2458         int force_flush = 0;
2459         struct mm_struct *mm = vma->vm_mm;
2460         unsigned long address;
2461         pte_t *ptep;
2462         pte_t pte;
2463         spinlock_t *ptl;
2464         struct page *page;
2465         struct hstate *h = hstate_vma(vma);
2466         unsigned long sz = huge_page_size(h);
2467         const unsigned long mmun_start = start; /* For mmu_notifiers */
2468         const unsigned long mmun_end   = end;   /* For mmu_notifiers */
2469 
2470         WARN_ON(!is_vm_hugetlb_page(vma));
2471         BUG_ON(start & ~huge_page_mask(h));
2472         BUG_ON(end & ~huge_page_mask(h));
2473 
2474         tlb_start_vma(tlb, vma);
2475         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2476 again:
2477         for (address = start; address < end; address += sz) {
2478                 ptep = huge_pte_offset(mm, address);
2479                 if (!ptep)
2480                         continue;
2481 
2482                 ptl = huge_pte_lock(h, mm, ptep);
2483                 if (huge_pmd_unshare(mm, &address, ptep))
2484                         goto unlock;
2485 
2486                 pte = huge_ptep_get(ptep);
2487                 if (huge_pte_none(pte))
2488                         goto unlock;
2489 
2490                 /*
2491                  * Migrating hugepage or HWPoisoned hugepage is already
2492                  * unmapped and its refcount is dropped, so just clear pte here.
2493                  */
2494                 if (unlikely(!pte_present(pte))) {
2495                         huge_pte_clear(mm, address, ptep);
2496                         goto unlock;
2497                 }
2498 
2499                 page = pte_page(pte);
2500                 /*
2501                  * If a reference page is supplied, it is because a specific
2502                  * page is being unmapped, not a range. Ensure the page we
2503                  * are about to unmap is the actual page of interest.
2504                  */
2505                 if (ref_page) {
2506                         if (page != ref_page)
2507                                 goto unlock;
2508 
2509                         /*
2510                          * Mark the VMA as having unmapped its page so that
2511                          * future faults in this VMA will fail rather than
2512                          * looking like data was lost
2513                          */
2514                         set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2515                 }
2516 
2517                 pte = huge_ptep_get_and_clear(mm, address, ptep);
2518                 tlb_remove_tlb_entry(tlb, ptep, address);
2519                 if (huge_pte_dirty(pte))
2520                         set_page_dirty(page);
2521 
2522                 page_remove_rmap(page);
2523                 force_flush = !__tlb_remove_page(tlb, page);
2524                 if (force_flush) {
2525                         spin_unlock(ptl);
2526                         break;
2527                 }
2528                 /* Bail out after unmapping reference page if supplied */
2529                 if (ref_page) {
2530                         spin_unlock(ptl);
2531                         break;
2532                 }
2533 unlock:
2534                 spin_unlock(ptl);
2535         }
2536         /*
2537          * mmu_gather ran out of room to batch pages, we break out of
2538          * the PTE lock to avoid doing the potential expensive TLB invalidate
2539          * and page-free while holding it.
2540          */
2541         if (force_flush) {
2542                 force_flush = 0;
2543                 tlb_flush_mmu(tlb);
2544                 if (address < end && !ref_page)
2545                         goto again;
2546         }
2547         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2548         tlb_end_vma(tlb, vma);
2549 }
2550 
2551 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
2552                           struct vm_area_struct *vma, unsigned long start,
2553                           unsigned long end, struct page *ref_page)
2554 {
2555         __unmap_hugepage_range(tlb, vma, start, end, ref_page);
2556 
2557         /*
2558          * Clear this flag so that x86's huge_pmd_share page_table_shareable
2559          * test will fail on a vma being torn down, and not grab a page table
2560          * on its way out.  We're lucky that the flag has such an appropriate
2561          * name, and can in fact be safely cleared here. We could clear it
2562          * before the __unmap_hugepage_range above, but all that's necessary
2563          * is to clear it before releasing the i_mmap_mutex. This works
2564          * because in the context this is called, the VMA is about to be
2565          * destroyed and the i_mmap_mutex is held.
2566          */
2567         vma->vm_flags &= ~VM_MAYSHARE;
2568 }
2569 
2570 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2571                           unsigned long end, struct page *ref_page)
2572 {
2573         struct mm_struct *mm;
2574         struct mmu_gather tlb;
2575 
2576         mm = vma->vm_mm;
2577 
2578         tlb_gather_mmu(&tlb, mm, start, end);
2579         __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2580         tlb_finish_mmu(&tlb, start, end);
2581 }
2582 
2583 /*
2584  * This is called when the original mapper is failing to COW a MAP_PRIVATE
2585  * mappping it owns the reserve page for. The intention is to unmap the page
2586  * from other VMAs and let the children be SIGKILLed if they are faulting the
2587  * same region.
2588  */
2589 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2590                                 struct page *page, unsigned long address)
2591 {
2592         struct hstate *h = hstate_vma(vma);
2593         struct vm_area_struct *iter_vma;
2594         struct address_space *mapping;
2595         pgoff_t pgoff;
2596 
2597         /*
2598          * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2599          * from page cache lookup which is in HPAGE_SIZE units.
2600          */
2601         address = address & huge_page_mask(h);
2602         pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2603                         vma->vm_pgoff;
2604         mapping = file_inode(vma->vm_file)->i_mapping;
2605 
2606         /*
2607          * Take the mapping lock for the duration of the table walk. As
2608          * this mapping should be shared between all the VMAs,
2609          * __unmap_hugepage_range() is called as the lock is already held
2610          */
2611         mutex_lock(&mapping->i_mmap_mutex);
2612         vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
2613                 /* Do not unmap the current VMA */
2614                 if (iter_vma == vma)
2615                         continue;
2616 
2617                 /*
2618                  * Shared VMAs have their own reserves and do not affect
2619                  * MAP_PRIVATE accounting but it is possible that a shared
2620                  * VMA is using the same page so check and skip such VMAs.
2621                  */
2622                 if (iter_vma->vm_flags & VM_MAYSHARE)
2623                         continue;
2624 
2625                 /*
2626                  * Unmap the page from other VMAs without their own reserves.
2627                  * They get marked to be SIGKILLed if they fault in these
2628                  * areas. This is because a future no-page fault on this VMA
2629                  * could insert a zeroed page instead of the data existing
2630                  * from the time of fork. This would look like data corruption
2631                  */
2632                 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2633                         unmap_hugepage_range(iter_vma, address,
2634                                              address + huge_page_size(h), page);
2635         }
2636         mutex_unlock(&mapping->i_mmap_mutex);
2637 
2638         return 1;
2639 }
2640 
2641 /*
2642  * Hugetlb_cow() should be called with page lock of the original hugepage held.
2643  * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2644  * cannot race with other handlers or page migration.
2645  * Keep the pte_same checks anyway to make transition from the mutex easier.
2646  */
2647 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2648                         unsigned long address, pte_t *ptep, pte_t pte,
2649                         struct page *pagecache_page, spinlock_t *ptl)
2650 {
2651         struct hstate *h = hstate_vma(vma);
2652         struct page *old_page, *new_page;
2653         int outside_reserve = 0;
2654         unsigned long mmun_start;       /* For mmu_notifiers */
2655         unsigned long mmun_end;         /* For mmu_notifiers */
2656 
2657         old_page = pte_page(pte);
2658 
2659 retry_avoidcopy:
2660         /* If no-one else is actually using this page, avoid the copy
2661          * and just make the page writable */
2662         if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
2663                 page_move_anon_rmap(old_page, vma, address);
2664                 set_huge_ptep_writable(vma, address, ptep);
2665                 return 0;
2666         }
2667 
2668         /*
2669          * If the process that created a MAP_PRIVATE mapping is about to
2670          * perform a COW due to a shared page count, attempt to satisfy
2671          * the allocation without using the existing reserves. The pagecache
2672          * page is used to determine if the reserve at this address was
2673          * consumed or not. If reserves were used, a partial faulted mapping
2674          * at the time of fork() could consume its reserves on COW instead
2675          * of the full address range.
2676          */
2677         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2678                         old_page != pagecache_page)
2679                 outside_reserve = 1;
2680 
2681         page_cache_get(old_page);
2682 
2683         /* Drop page table lock as buddy allocator may be called */
2684         spin_unlock(ptl);
2685         new_page = alloc_huge_page(vma, address, outside_reserve);
2686 
2687         if (IS_ERR(new_page)) {
2688                 long err = PTR_ERR(new_page);
2689                 page_cache_release(old_page);
2690 
2691                 /*
2692                  * If a process owning a MAP_PRIVATE mapping fails to COW,
2693                  * it is due to references held by a child and an insufficient
2694                  * huge page pool. To guarantee the original mappers
2695                  * reliability, unmap the page from child processes. The child
2696                  * may get SIGKILLed if it later faults.
2697                  */
2698                 if (outside_reserve) {
2699                         BUG_ON(huge_pte_none(pte));
2700                         if (unmap_ref_private(mm, vma, old_page, address)) {
2701                                 BUG_ON(huge_pte_none(pte));
2702                                 spin_lock(ptl);
2703                                 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2704                                 if (likely(pte_same(huge_ptep_get(ptep), pte)))
2705                                         goto retry_avoidcopy;
2706                                 /*
2707                                  * race occurs while re-acquiring page table
2708                                  * lock, and our job is done.
2709                                  */
2710                                 return 0;
2711                         }
2712                         WARN_ON_ONCE(1);
2713                 }
2714 
2715                 /* Caller expects lock to be held */
2716                 spin_lock(ptl);
2717                 if (err == -ENOMEM)
2718                         return VM_FAULT_OOM;
2719                 else
2720                         return VM_FAULT_SIGBUS;
2721         }
2722 
2723         /*
2724          * When the original hugepage is shared one, it does not have
2725          * anon_vma prepared.
2726          */
2727         if (unlikely(anon_vma_prepare(vma))) {
2728                 page_cache_release(new_page);
2729                 page_cache_release(old_page);
2730                 /* Caller expects lock to be held */
2731                 spin_lock(ptl);
2732                 return VM_FAULT_OOM;
2733         }
2734 
2735         copy_user_huge_page(new_page, old_page, address, vma,
2736                             pages_per_huge_page(h));
2737         __SetPageUptodate(new_page);
2738 
2739         mmun_start = address & huge_page_mask(h);
2740         mmun_end = mmun_start + huge_page_size(h);
2741         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2742         /*
2743          * Retake the page table lock to check for racing updates
2744          * before the page tables are altered
2745          */
2746         spin_lock(ptl);
2747         ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2748         if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2749                 ClearPagePrivate(new_page);
2750 
2751                 /* Break COW */
2752                 huge_ptep_clear_flush(vma, address, ptep);
2753                 set_huge_pte_at(mm, address, ptep,
2754                                 make_huge_pte(vma, new_page, 1));
2755                 page_remove_rmap(old_page);
2756                 hugepage_add_new_anon_rmap(new_page, vma, address);
2757                 /* Make the old page be freed below */
2758                 new_page = old_page;
2759         }
2760         spin_unlock(ptl);
2761         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2762         page_cache_release(new_page);
2763         page_cache_release(old_page);
2764 
2765         /* Caller expects lock to be held */
2766         spin_lock(ptl);
2767         return 0;
2768 }
2769 
2770 /* Return the pagecache page at a given address within a VMA */
2771 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2772                         struct vm_area_struct *vma, unsigned long address)
2773 {
2774         struct address_space *mapping;
2775         pgoff_t idx;
2776 
2777         mapping = vma->vm_file->f_mapping;
2778         idx = vma_hugecache_offset(h, vma, address);
2779 
2780         return find_lock_page(mapping, idx);
2781 }
2782 
2783 /*
2784  * Return whether there is a pagecache page to back given address within VMA.
2785  * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2786  */
2787 static bool hugetlbfs_pagecache_present(struct hstate *h,
2788                         struct vm_area_struct *vma, unsigned long address)
2789 {
2790         struct address_space *mapping;
2791         pgoff_t idx;
2792         struct page *page;
2793 
2794         mapping = vma->vm_file->f_mapping;
2795         idx = vma_hugecache_offset(h, vma, address);
2796 
2797         page = find_get_page(mapping, idx);
2798         if (page)
2799                 put_page(page);
2800         return page != NULL;
2801 }
2802 
2803 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2804                         unsigned long address, pte_t *ptep, unsigned int flags)
2805 {
2806         struct hstate *h = hstate_vma(vma);
2807         int ret = VM_FAULT_SIGBUS;
2808         int anon_rmap = 0;
2809         pgoff_t idx;
2810         unsigned long size;
2811         struct page *page;
2812         struct address_space *mapping;
2813         pte_t new_pte;
2814         spinlock_t *ptl;
2815 
2816         /*
2817          * Currently, we are forced to kill the process in the event the
2818          * original mapper has unmapped pages from the child due to a failed
2819          * COW. Warn that such a situation has occurred as it may not be obvious
2820          */
2821         if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2822                 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2823                            current->pid);
2824                 return ret;
2825         }
2826 
2827         mapping = vma->vm_file->f_mapping;
2828         idx = vma_hugecache_offset(h, vma, address);
2829 
2830         /*
2831          * Use page lock to guard against racing truncation
2832          * before we get page_table_lock.
2833          */
2834 retry:
2835         page = find_lock_page(mapping, idx);
2836         if (!page) {
2837                 size = i_size_read(mapping->host) >> huge_page_shift(h);
2838                 if (idx >= size)
2839                         goto out;
2840                 page = alloc_huge_page(vma, address, 0);
2841                 if (IS_ERR(page)) {
2842                         ret = PTR_ERR(page);
2843                         if (ret == -ENOMEM)
2844                                 ret = VM_FAULT_OOM;
2845                         else
2846                                 ret = VM_FAULT_SIGBUS;
2847                         goto out;
2848                 }
2849                 clear_huge_page(page, address, pages_per_huge_page(h));
2850                 __SetPageUptodate(page);
2851 
2852                 if (vma->vm_flags & VM_MAYSHARE) {
2853                         int err;
2854                         struct inode *inode = mapping->host;
2855 
2856                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2857                         if (err) {
2858                                 put_page(page);
2859                                 if (err == -EEXIST)
2860                                         goto retry;
2861                                 goto out;
2862                         }
2863                         ClearPagePrivate(page);
2864 
2865                         spin_lock(&inode->i_lock);
2866                         inode->i_blocks += blocks_per_huge_page(h);
2867                         spin_unlock(&inode->i_lock);
2868                 } else {
2869                         lock_page(page);
2870                         if (unlikely(anon_vma_prepare(vma))) {
2871                                 ret = VM_FAULT_OOM;
2872                                 goto backout_unlocked;
2873                         }
2874                         anon_rmap = 1;
2875                 }
2876         } else {
2877                 /*
2878                  * If memory error occurs between mmap() and fault, some process
2879                  * don't have hwpoisoned swap entry for errored virtual address.
2880                  * So we need to block hugepage fault by PG_hwpoison bit check.
2881                  */
2882                 if (unlikely(PageHWPoison(page))) {
2883                         ret = VM_FAULT_HWPOISON |
2884                                 VM_FAULT_SET_HINDEX(hstate_index(h));
2885                         goto backout_unlocked;
2886                 }
2887         }
2888 
2889         /*
2890          * If we are going to COW a private mapping later, we examine the
2891          * pending reservations for this page now. This will ensure that
2892          * any allocations necessary to record that reservation occur outside
2893          * the spinlock.
2894          */
2895         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2896                 if (vma_needs_reservation(h, vma, address) < 0) {
2897                         ret = VM_FAULT_OOM;
2898                         goto backout_unlocked;
2899                 }
2900 
2901         ptl = huge_pte_lockptr(h, mm, ptep);
2902         spin_lock(ptl);
2903         size = i_size_read(mapping->host) >> huge_page_shift(h);
2904         if (idx >= size)
2905                 goto backout;
2906 
2907         ret = 0;
2908         if (!huge_pte_none(huge_ptep_get(ptep)))
2909                 goto backout;
2910 
2911         if (anon_rmap) {
2912                 ClearPagePrivate(page);
2913                 hugepage_add_new_anon_rmap(page, vma, address);
2914         }
2915         else
2916                 page_dup_rmap(page);
2917         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2918                                 && (vma->vm_flags & VM_SHARED)));
2919         set_huge_pte_at(mm, address, ptep, new_pte);
2920 
2921         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2922                 /* Optimization, do the COW without a second fault */
2923                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl);
2924         }
2925 
2926         spin_unlock(ptl);
2927         unlock_page(page);
2928 out:
2929         return ret;
2930 
2931 backout:
2932         spin_unlock(ptl);
2933 backout_unlocked:
2934         unlock_page(page);
2935         put_page(page);
2936         goto out;
2937 }
2938 
2939 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2940                         unsigned long address, unsigned int flags)
2941 {
2942         pte_t *ptep;
2943         pte_t entry;
2944         spinlock_t *ptl;
2945         int ret;
2946         struct page *page = NULL;
2947         struct page *pagecache_page = NULL;
2948         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2949         struct hstate *h = hstate_vma(vma);
2950 
2951         address &= huge_page_mask(h);
2952 
2953         ptep = huge_pte_offset(mm, address);
2954         if (ptep) {
2955                 entry = huge_ptep_get(ptep);
2956                 if (unlikely(is_hugetlb_entry_migration(entry))) {
2957                         migration_entry_wait_huge(vma, mm, ptep);
2958                         return 0;
2959                 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2960                         return VM_FAULT_HWPOISON_LARGE |
2961                                 VM_FAULT_SET_HINDEX(hstate_index(h));
2962         }
2963 
2964         ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2965         if (!ptep)
2966                 return VM_FAULT_OOM;
2967 
2968         /*
2969          * Serialize hugepage allocation and instantiation, so that we don't
2970          * get spurious allocation failures if two CPUs race to instantiate
2971          * the same page in the page cache.
2972          */
2973         mutex_lock(&hugetlb_instantiation_mutex);
2974         entry = huge_ptep_get(ptep);
2975         if (huge_pte_none(entry)) {
2976                 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2977                 goto out_mutex;
2978         }
2979 
2980         ret = 0;
2981 
2982         /*
2983          * If we are going to COW the mapping later, we examine the pending
2984          * reservations for this page now. This will ensure that any
2985          * allocations necessary to record that reservation occur outside the
2986          * spinlock. For private mappings, we also lookup the pagecache
2987          * page now as it is used to determine if a reservation has been
2988          * consumed.
2989          */
2990         if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
2991                 if (vma_needs_reservation(h, vma, address) < 0) {
2992                         ret = VM_FAULT_OOM;
2993                         goto out_mutex;
2994                 }
2995 
2996                 if (!(vma->vm_flags & VM_MAYSHARE))
2997                         pagecache_page = hugetlbfs_pagecache_page(h,
2998                                                                 vma, address);
2999         }
3000 
3001         /*
3002          * hugetlb_cow() requires page locks of pte_page(entry) and
3003          * pagecache_page, so here we need take the former one
3004          * when page != pagecache_page or !pagecache_page.
3005          * Note that locking order is always pagecache_page -> page,
3006          * so no worry about deadlock.
3007          */
3008         page = pte_page(entry);
3009         get_page(page);
3010         if (page != pagecache_page)
3011                 lock_page(page);
3012 
3013         ptl = huge_pte_lockptr(h, mm, ptep);
3014         spin_lock(ptl);
3015         /* Check for a racing update before calling hugetlb_cow */
3016         if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
3017                 goto out_ptl;
3018 
3019 
3020         if (flags & FAULT_FLAG_WRITE) {
3021                 if (!huge_pte_write(entry)) {
3022                         ret = hugetlb_cow(mm, vma, address, ptep, entry,
3023                                         pagecache_page, ptl);
3024                         goto out_ptl;
3025                 }
3026                 entry = huge_pte_mkdirty(entry);
3027         }
3028         entry = pte_mkyoung(entry);
3029         if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3030                                                 flags & FAULT_FLAG_WRITE))
3031                 update_mmu_cache(vma, address, ptep);
3032 
3033 out_ptl:
3034         spin_unlock(ptl);
3035 
3036         if (pagecache_page) {
3037                 unlock_page(pagecache_page);
3038                 put_page(pagecache_page);
3039         }
3040         if (page != pagecache_page)
3041                 unlock_page(page);
3042         put_page(page);
3043 
3044 out_mutex:
3045         mutex_unlock(&hugetlb_instantiation_mutex);
3046 
3047         return ret;
3048 }
3049 
3050 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3051                          struct page **pages, struct vm_area_struct **vmas,
3052                          unsigned long *position, unsigned long *nr_pages,
3053                          long i, unsigned int flags)
3054 {
3055         unsigned long pfn_offset;
3056         unsigned long vaddr = *position;
3057         unsigned long remainder = *nr_pages;
3058         struct hstate *h = hstate_vma(vma);
3059 
3060         while (vaddr < vma->vm_end && remainder) {
3061                 pte_t *pte;
3062                 spinlock_t *ptl = NULL;
3063                 int absent;
3064                 struct page *page;
3065 
3066                 /*
3067                  * Some archs (sparc64, sh*) have multiple pte_ts to
3068                  * each hugepage.  We have to make sure we get the
3069                  * first, for the page indexing below to work.
3070                  *
3071                  * Note that page table lock is not held when pte is null.
3072                  */
3073                 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
3074                 if (pte)
3075                         ptl = huge_pte_lock(h, mm, pte);
3076                 absent = !pte || huge_pte_none(huge_ptep_get(pte));
3077 
3078                 /*
3079                  * When coredumping, it suits get_dump_page if we just return
3080                  * an error where there's an empty slot with no huge pagecache
3081                  * to back it.  This way, we avoid allocating a hugepage, and
3082                  * the sparse dumpfile avoids allocating disk blocks, but its
3083                  * huge holes still show up with zeroes where they need to be.
3084                  */
3085                 if (absent && (flags & FOLL_DUMP) &&
3086                     !hugetlbfs_pagecache_present(h, vma, vaddr)) {
3087                         if (pte)
3088                                 spin_unlock(ptl);
3089                         remainder = 0;
3090                         break;
3091                 }
3092 
3093                 /*
3094                  * We need call hugetlb_fault for both hugepages under migration
3095                  * (in which case hugetlb_fault waits for the migration,) and
3096                  * hwpoisoned hugepages (in which case we need to prevent the
3097                  * caller from accessing to them.) In order to do this, we use
3098                  * here is_swap_pte instead of is_hugetlb_entry_migration and
3099                  * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3100                  * both cases, and because we can't follow correct pages
3101                  * directly from any kind of swap entries.
3102                  */
3103                 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
3104                     ((flags & FOLL_WRITE) &&
3105                       !huge_pte_write(huge_ptep_get(pte)))) {
3106                         int ret;
3107 
3108                         if (pte)
3109                                 spin_unlock(ptl);
3110                         ret = hugetlb_fault(mm, vma, vaddr,
3111                                 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
3112                         if (!(ret & VM_FAULT_ERROR))
3113                                 continue;
3114 
3115                         remainder = 0;
3116                         break;
3117                 }
3118 
3119                 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
3120                 page = pte_page(huge_ptep_get(pte));
3121 same_page:
3122                 if (pages) {
3123                         pages[i] = mem_map_offset(page, pfn_offset);
3124                         get_page_foll(pages[i]);
3125                 }
3126 
3127                 if (vmas)
3128                         vmas[i] = vma;
3129 
3130                 vaddr += PAGE_SIZE;
3131                 ++pfn_offset;
3132                 --remainder;
3133                 ++i;
3134                 if (vaddr < vma->vm_end && remainder &&
3135                                 pfn_offset < pages_per_huge_page(h)) {
3136                         /*
3137                          * We use pfn_offset to avoid touching the pageframes
3138                          * of this compound page.
3139                          */
3140                         goto same_page;
3141                 }
3142                 spin_unlock(ptl);
3143         }
3144         *nr_pages = remainder;
3145         *position = vaddr;
3146 
3147         return i ? i : -EFAULT;
3148 }
3149 
3150 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
3151                 unsigned long address, unsigned long end, pgprot_t newprot)
3152 {
3153         struct mm_struct *mm = vma->vm_mm;
3154         unsigned long start = address;
3155         pte_t *ptep;
3156         pte_t pte;
3157         struct hstate *h = hstate_vma(vma);
3158         unsigned long pages = 0;
3159 
3160         BUG_ON(address >= end);
3161         flush_cache_range(vma, address, end);
3162 
3163         mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
3164         for (; address < end; address += huge_page_size(h)) {
3165                 spinlock_t *ptl;
3166                 ptep = huge_pte_offset(mm, address);
3167                 if (!ptep)
3168                         continue;
3169                 ptl = huge_pte_lock(h, mm, ptep);
3170                 if (huge_pmd_unshare(mm, &address, ptep)) {
3171                         pages++;
3172                         spin_unlock(ptl);
3173                         continue;
3174                 }
3175                 pte = huge_ptep_get(ptep);
3176                 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
3177                         spin_unlock(ptl);
3178                         continue;
3179                 }
3180                 if (unlikely(is_hugetlb_entry_migration(pte))) {
3181                         swp_entry_t entry = pte_to_swp_entry(pte);
3182 
3183                         if (is_write_migration_entry(entry)) {
3184                                 pte_t newpte;
3185 
3186                                 make_migration_entry_read(&entry);
3187                                 newpte = swp_entry_to_pte(entry);
3188                                 set_huge_pte_at(mm, address, ptep, newpte);
3189                                 pages++;
3190                         }
3191                         spin_unlock(ptl);
3192                         continue;
3193                 }
3194                 if (!huge_pte_none(pte)) {
3195                         pte = huge_ptep_get_and_clear(mm, address, ptep);
3196                         pte = pte_mkhuge(huge_pte_modify(pte, newprot));
3197                         pte = arch_make_huge_pte(pte, vma, NULL, 0);
3198                         set_huge_pte_at(mm, address, ptep, pte);
3199                         pages++;
3200                 }
3201                 spin_unlock(ptl);
3202         }
3203         /*
3204          * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3205          * may have cleared our pud entry and done put_page on the page table:
3206          * once we release i_mmap_mutex, another task can do the final put_page
3207          * and that page table be reused and filled with junk.
3208          */
3209         flush_tlb_range(vma, start, end);
3210         mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
3211 
3212         return pages << h->order;
3213 }
3214 
3215 int hugetlb_reserve_pages(struct inode *inode,
3216                                         long from, long to,
3217                                         struct vm_area_struct *vma,
3218                                         vm_flags_t vm_flags)
3219 {
3220         long ret, chg;
3221         struct hstate *h = hstate_inode(inode);
3222         struct hugepage_subpool *spool = subpool_inode(inode);
3223 
3224         /*
3225          * Only apply hugepage reservation if asked. At fault time, an
3226          * attempt will be made for VM_NORESERVE to allocate a page
3227          * without using reserves
3228          */
3229         if (vm_flags & VM_NORESERVE)
3230                 return 0;
3231 
3232         /*
3233          * Shared mappings base their reservation on the number of pages that
3234          * are already allocated on behalf of the file. Private mappings need
3235          * to reserve the full area even if read-only as mprotect() may be
3236          * called to make the mapping read-write. Assume !vma is a shm mapping
3237          */
3238         if (!vma || vma->vm_flags & VM_MAYSHARE)
3239                 chg = region_chg(&inode->i_mapping->private_list, from, to);
3240         else {
3241                 struct resv_map *resv_map = resv_map_alloc();
3242                 if (!resv_map)
3243                         return -ENOMEM;
3244 
3245                 chg = to - from;
3246 
3247                 set_vma_resv_map(vma, resv_map);
3248                 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3249         }
3250 
3251         if (chg < 0) {
3252                 ret = chg;
3253                 goto out_err;
3254         }
3255 
3256         /* There must be enough pages in the subpool for the mapping */
3257         if (hugepage_subpool_get_pages(spool, chg)) {
3258                 ret = -ENOSPC;
3259                 goto out_err;
3260         }
3261 
3262         /*
3263          * Check enough hugepages are available for the reservation.
3264          * Hand the pages back to the subpool if there are not
3265          */
3266         ret = hugetlb_acct_memory(h, chg);
3267         if (ret < 0) {
3268                 hugepage_subpool_put_pages(spool, chg);
3269                 goto out_err;
3270         }
3271 
3272         /*
3273          * Account for the reservations made. Shared mappings record regions
3274          * that have reservations as they are shared by multiple VMAs.
3275          * When the last VMA disappears, the region map says how much
3276          * the reservation was and the page cache tells how much of
3277          * the reservation was consumed. Private mappings are per-VMA and
3278          * only the consumed reservations are tracked. When the VMA
3279          * disappears, the original reservation is the VMA size and the
3280          * consumed reservations are stored in the map. Hence, nothing
3281          * else has to be done for private mappings here
3282          */
3283         if (!vma || vma->vm_flags & VM_MAYSHARE)
3284                 region_add(&inode->i_mapping->private_list, from, to);
3285         return 0;
3286 out_err:
3287         if (vma)
3288                 resv_map_put(vma);
3289         return ret;
3290 }
3291 
3292 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3293 {
3294         struct hstate *h = hstate_inode(inode);
3295         long chg = region_truncate(&inode->i_mapping->private_list, offset);
3296         struct hugepage_subpool *spool = subpool_inode(inode);
3297 
3298         spin_lock(&inode->i_lock);
3299         inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3300         spin_unlock(&inode->i_lock);
3301 
3302         hugepage_subpool_put_pages(spool, (chg - freed));
3303         hugetlb_acct_memory(h, -(chg - freed));
3304 }
3305 
3306 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3307 static unsigned long page_table_shareable(struct vm_area_struct *svma,
3308                                 struct vm_area_struct *vma,
3309                                 unsigned long addr, pgoff_t idx)
3310 {
3311         unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
3312                                 svma->vm_start;
3313         unsigned long sbase = saddr & PUD_MASK;
3314         unsigned long s_end = sbase + PUD_SIZE;
3315 
3316         /* Allow segments to share if only one is marked locked */
3317         unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED;
3318         unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED;
3319 
3320         /*
3321          * match the virtual addresses, permission and the alignment of the
3322          * page table page.
3323          */
3324         if (pmd_index(addr) != pmd_index(saddr) ||
3325             vm_flags != svm_flags ||
3326             sbase < svma->vm_start || svma->vm_end < s_end)
3327                 return 0;
3328 
3329         return saddr;
3330 }
3331 
3332 static int vma_shareable(struct vm_area_struct *vma, unsigned long addr)
3333 {
3334         unsigned long base = addr & PUD_MASK;
3335         unsigned long end = base + PUD_SIZE;
3336 
3337         /*
3338          * check on proper vm_flags and page table alignment
3339          */
3340         if (vma->vm_flags & VM_MAYSHARE &&
3341             vma->vm_start <= base && end <= vma->vm_end)
3342                 return 1;
3343         return 0;
3344 }
3345 
3346 /*
3347  * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3348  * and returns the corresponding pte. While this is not necessary for the
3349  * !shared pmd case because we can allocate the pmd later as well, it makes the
3350  * code much cleaner. pmd allocation is essential for the shared case because
3351  * pud has to be populated inside the same i_mmap_mutex section - otherwise
3352  * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3353  * bad pmd for sharing.
3354  */
3355 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3356 {
3357         struct vm_area_struct *vma = find_vma(mm, addr);
3358         struct address_space *mapping = vma->vm_file->f_mapping;
3359         pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
3360                         vma->vm_pgoff;
3361         struct vm_area_struct *svma;
3362         unsigned long saddr;
3363         pte_t *spte = NULL;
3364         pte_t *pte;
3365         spinlock_t *ptl;
3366 
3367         if (!vma_shareable(vma, addr))
3368                 return (pte_t *)pmd_alloc(mm, pud, addr);
3369 
3370         mutex_lock(&mapping->i_mmap_mutex);
3371         vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
3372                 if (svma == vma)
3373                         continue;
3374 
3375                 saddr = page_table_shareable(svma, vma, addr, idx);
3376                 if (saddr) {
3377                         spte = huge_pte_offset(svma->vm_mm, saddr);
3378                         if (spte) {
3379                                 get_page(virt_to_page(spte));
3380                                 break;
3381                         }
3382                 }
3383         }
3384 
3385         if (!spte)
3386                 goto out;
3387 
3388         ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte);
3389         spin_lock(ptl);
3390         if (pud_none(*pud))
3391                 pud_populate(mm, pud,
3392                                 (pmd_t *)((unsigned long)spte & PAGE_MASK));
3393         else
3394                 put_page(virt_to_page(spte));
3395         spin_unlock(ptl);
3396 out:
3397         pte = (pte_t *)pmd_alloc(mm, pud, addr);
3398         mutex_unlock(&mapping->i_mmap_mutex);
3399         return pte;
3400 }
3401 
3402 /*
3403  * unmap huge page backed by shared pte.
3404  *
3405  * Hugetlb pte page is ref counted at the time of mapping.  If pte is shared
3406  * indicated by page_count > 1, unmap is achieved by clearing pud and
3407  * decrementing the ref count. If count == 1, the pte page is not shared.
3408  *
3409  * called with page table lock held.
3410  *
3411  * returns: 1 successfully unmapped a shared pte page
3412  *          0 the underlying pte page is not shared, or it is the last user
3413  */
3414 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
3415 {
3416         pgd_t *pgd = pgd_offset(mm, *addr);
3417         pud_t *pud = pud_offset(pgd, *addr);
3418 
3419         BUG_ON(page_count(virt_to_page(ptep)) == 0);
3420         if (page_count(virt_to_page(ptep)) == 1)
3421                 return 0;
3422 
3423         pud_clear(pud);
3424         put_page(virt_to_page(ptep));
3425         *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
3426         return 1;
3427 }
3428 #define want_pmd_share()        (1)
3429 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3430 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3431 {
3432         return NULL;
3433 }
3434 #define want_pmd_share()        (0)
3435 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3436 
3437 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3438 pte_t *huge_pte_alloc(struct mm_struct *mm,
3439                         unsigned long addr, unsigned long sz)
3440 {
3441         pgd_t *pgd;
3442         pud_t *pud;
3443         pte_t *pte = NULL;
3444 
3445         pgd = pgd_offset(mm, addr);
3446         pud = pud_alloc(mm, pgd, addr);
3447         if (pud) {
3448                 if (sz == PUD_SIZE) {
3449                         pte = (pte_t *)pud;
3450                 } else {
3451                         BUG_ON(sz != PMD_SIZE);
3452                         if (want_pmd_share() && pud_none(*pud))
3453                                 pte = huge_pmd_share(mm, addr, pud);
3454                         else
3455                                 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3456                 }
3457         }
3458         BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
3459 
3460         return pte;
3461 }
3462 
3463 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
3464 {
3465         pgd_t *pgd;
3466         pud_t *pud;
3467         pmd_t *pmd = NULL;
3468 
3469         pgd = pgd_offset(mm, addr);
3470         if (pgd_present(*pgd)) {
3471                 pud = pud_offset(pgd, addr);
3472                 if (pud_present(*pud)) {
3473                         if (pud_huge(*pud))
3474                                 return (pte_t *)pud;
3475                         pmd = pmd_offset(pud, addr);
3476                 }
3477         }
3478         return (pte_t *) pmd;
3479 }
3480 
3481 struct page *
3482 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
3483                 pmd_t *pmd, int write)
3484 {
3485         struct page *page;
3486 
3487         if (!pmd_present(*pmd))
3488                 return NULL;
3489         page = pte_page(*(pte_t *)pmd);
3490         if (page)
3491                 page += ((address & ~PMD_MASK) >> PAGE_SHIFT);
3492         return page;
3493 }
3494 
3495 struct page *
3496 follow_huge_pud(struct mm_struct *mm, unsigned long address,
3497                 pud_t *pud, int write)
3498 {
3499         struct page *page;
3500 
3501         page = pte_page(*(pte_t *)pud);
3502         if (page)
3503                 page += ((address & ~PUD_MASK) >> PAGE_SHIFT);
3504         return page;
3505 }
3506 
3507 #else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3508 
3509 /* Can be overriden by architectures */
3510 __attribute__((weak)) struct page *
3511 follow_huge_pud(struct mm_struct *mm, unsigned long address,
3512                pud_t *pud, int write)
3513 {
3514         BUG();
3515         return NULL;
3516 }
3517 
3518 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3519 
3520 #ifdef CONFIG_MEMORY_FAILURE
3521 
3522 /* Should be called in hugetlb_lock */
3523 static int is_hugepage_on_freelist(struct page *hpage)
3524 {
3525         struct page *page;
3526         struct page *tmp;
3527         struct hstate *h = page_hstate(hpage);
3528         int nid = page_to_nid(hpage);
3529 
3530         list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3531                 if (page == hpage)
3532                         return 1;
3533         return 0;
3534 }
3535 
3536 /*
3537  * This function is called from memory failure code.
3538  * Assume the caller holds page lock of the head page.
3539  */
3540 int dequeue_hwpoisoned_huge_page(struct page *hpage)
3541 {
3542         struct hstate *h = page_hstate(hpage);
3543         int nid = page_to_nid(hpage);
3544         int ret = -EBUSY;
3545 
3546         spin_lock(&hugetlb_lock);
3547         if (is_hugepage_on_freelist(hpage)) {
3548                 /*
3549                  * Hwpoisoned hugepage isn't linked to activelist or freelist,
3550                  * but dangling hpage->lru can trigger list-debug warnings
3551                  * (this happens when we call unpoison_memory() on it),
3552                  * so let it point to itself with list_del_init().
3553                  */
3554                 list_del_init(&hpage->lru);
3555                 set_page_refcounted(hpage);
3556                 h->free_huge_pages--;
3557                 h->free_huge_pages_node[nid]--;
3558                 ret = 0;
3559         }
3560         spin_unlock(&hugetlb_lock);
3561         return ret;
3562 }
3563 #endif
3564 
3565 bool isolate_huge_page(struct page *page, struct list_head *list)
3566 {
3567         VM_BUG_ON_PAGE(!PageHead(page), page);
3568         if (!get_page_unless_zero(page))
3569                 return false;
3570         spin_lock(&hugetlb_lock);
3571         list_move_tail(&page->lru, list);
3572         spin_unlock(&hugetlb_lock);
3573         return true;
3574 }
3575 
3576 void putback_active_hugepage(struct page *page)
3577 {
3578         VM_BUG_ON_PAGE(!PageHead(page), page);
3579         spin_lock(&hugetlb_lock);
3580         list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
3581         spin_unlock(&hugetlb_lock);
3582         put_page(page);
3583 }
3584 
3585 bool is_hugepage_active(struct page *page)
3586 {
3587         VM_BUG_ON_PAGE(!PageHuge(page), page);
3588         /*
3589          * This function can be called for a tail page because the caller,
3590          * scan_movable_pages, scans through a given pfn-range which typically
3591          * covers one memory block. In systems using gigantic hugepage (1GB
3592          * for x86_64,) a hugepage is larger than a memory block, and we don't
3593          * support migrating such large hugepages for now, so return false
3594          * when called for tail pages.
3595          */
3596         if (PageTail(page))
3597                 return false;
3598         /*
3599          * Refcount of a hwpoisoned hugepages is 1, but they are not active,
3600          * so we should return false for them.
3601          */
3602         if (unlikely(PageHWPoison(page)))
3603                 return false;
3604         return page_count(page) > 0;
3605 }
3606 

~ [ 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