Back to home page

LXR

 
 

    


0001 /*
0002  * Generic hugetlb support.
0003  * (C) Nadia Yvette Chambers, April 2004
0004  */
0005 #include <linux/list.h>
0006 #include <linux/init.h>
0007 #include <linux/mm.h>
0008 #include <linux/seq_file.h>
0009 #include <linux/sysctl.h>
0010 #include <linux/highmem.h>
0011 #include <linux/mmu_notifier.h>
0012 #include <linux/nodemask.h>
0013 #include <linux/pagemap.h>
0014 #include <linux/mempolicy.h>
0015 #include <linux/compiler.h>
0016 #include <linux/cpuset.h>
0017 #include <linux/mutex.h>
0018 #include <linux/bootmem.h>
0019 #include <linux/sysfs.h>
0020 #include <linux/slab.h>
0021 #include <linux/rmap.h>
0022 #include <linux/swap.h>
0023 #include <linux/swapops.h>
0024 #include <linux/page-isolation.h>
0025 #include <linux/jhash.h>
0026 
0027 #include <asm/page.h>
0028 #include <asm/pgtable.h>
0029 #include <asm/tlb.h>
0030 
0031 #include <linux/io.h>
0032 #include <linux/hugetlb.h>
0033 #include <linux/hugetlb_cgroup.h>
0034 #include <linux/node.h>
0035 #include "internal.h"
0036 
0037 int hugepages_treat_as_movable;
0038 
0039 int hugetlb_max_hstate __read_mostly;
0040 unsigned int default_hstate_idx;
0041 struct hstate hstates[HUGE_MAX_HSTATE];
0042 /*
0043  * Minimum page order among possible hugepage sizes, set to a proper value
0044  * at boot time.
0045  */
0046 static unsigned int minimum_order __read_mostly = UINT_MAX;
0047 
0048 __initdata LIST_HEAD(huge_boot_pages);
0049 
0050 /* for command line parsing */
0051 static struct hstate * __initdata parsed_hstate;
0052 static unsigned long __initdata default_hstate_max_huge_pages;
0053 static unsigned long __initdata default_hstate_size;
0054 static bool __initdata parsed_valid_hugepagesz = true;
0055 
0056 /*
0057  * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
0058  * free_huge_pages, and surplus_huge_pages.
0059  */
0060 DEFINE_SPINLOCK(hugetlb_lock);
0061 
0062 /*
0063  * Serializes faults on the same logical page.  This is used to
0064  * prevent spurious OOMs when the hugepage pool is fully utilized.
0065  */
0066 static int num_fault_mutexes;
0067 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
0068 
0069 /* Forward declaration */
0070 static int hugetlb_acct_memory(struct hstate *h, long delta);
0071 
0072 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
0073 {
0074     bool free = (spool->count == 0) && (spool->used_hpages == 0);
0075 
0076     spin_unlock(&spool->lock);
0077 
0078     /* If no pages are used, and no other handles to the subpool
0079      * remain, give up any reservations mased on minimum size and
0080      * free the subpool */
0081     if (free) {
0082         if (spool->min_hpages != -1)
0083             hugetlb_acct_memory(spool->hstate,
0084                         -spool->min_hpages);
0085         kfree(spool);
0086     }
0087 }
0088 
0089 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
0090                         long min_hpages)
0091 {
0092     struct hugepage_subpool *spool;
0093 
0094     spool = kzalloc(sizeof(*spool), GFP_KERNEL);
0095     if (!spool)
0096         return NULL;
0097 
0098     spin_lock_init(&spool->lock);
0099     spool->count = 1;
0100     spool->max_hpages = max_hpages;
0101     spool->hstate = h;
0102     spool->min_hpages = min_hpages;
0103 
0104     if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
0105         kfree(spool);
0106         return NULL;
0107     }
0108     spool->rsv_hpages = min_hpages;
0109 
0110     return spool;
0111 }
0112 
0113 void hugepage_put_subpool(struct hugepage_subpool *spool)
0114 {
0115     spin_lock(&spool->lock);
0116     BUG_ON(!spool->count);
0117     spool->count--;
0118     unlock_or_release_subpool(spool);
0119 }
0120 
0121 /*
0122  * Subpool accounting for allocating and reserving pages.
0123  * Return -ENOMEM if there are not enough resources to satisfy the
0124  * the request.  Otherwise, return the number of pages by which the
0125  * global pools must be adjusted (upward).  The returned value may
0126  * only be different than the passed value (delta) in the case where
0127  * a subpool minimum size must be manitained.
0128  */
0129 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
0130                       long delta)
0131 {
0132     long ret = delta;
0133 
0134     if (!spool)
0135         return ret;
0136 
0137     spin_lock(&spool->lock);
0138 
0139     if (spool->max_hpages != -1) {      /* maximum size accounting */
0140         if ((spool->used_hpages + delta) <= spool->max_hpages)
0141             spool->used_hpages += delta;
0142         else {
0143             ret = -ENOMEM;
0144             goto unlock_ret;
0145         }
0146     }
0147 
0148     /* minimum size accounting */
0149     if (spool->min_hpages != -1 && spool->rsv_hpages) {
0150         if (delta > spool->rsv_hpages) {
0151             /*
0152              * Asking for more reserves than those already taken on
0153              * behalf of subpool.  Return difference.
0154              */
0155             ret = delta - spool->rsv_hpages;
0156             spool->rsv_hpages = 0;
0157         } else {
0158             ret = 0;    /* reserves already accounted for */
0159             spool->rsv_hpages -= delta;
0160         }
0161     }
0162 
0163 unlock_ret:
0164     spin_unlock(&spool->lock);
0165     return ret;
0166 }
0167 
0168 /*
0169  * Subpool accounting for freeing and unreserving pages.
0170  * Return the number of global page reservations that must be dropped.
0171  * The return value may only be different than the passed value (delta)
0172  * in the case where a subpool minimum size must be maintained.
0173  */
0174 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
0175                        long delta)
0176 {
0177     long ret = delta;
0178 
0179     if (!spool)
0180         return delta;
0181 
0182     spin_lock(&spool->lock);
0183 
0184     if (spool->max_hpages != -1)        /* maximum size accounting */
0185         spool->used_hpages -= delta;
0186 
0187      /* minimum size accounting */
0188     if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
0189         if (spool->rsv_hpages + delta <= spool->min_hpages)
0190             ret = 0;
0191         else
0192             ret = spool->rsv_hpages + delta - spool->min_hpages;
0193 
0194         spool->rsv_hpages += delta;
0195         if (spool->rsv_hpages > spool->min_hpages)
0196             spool->rsv_hpages = spool->min_hpages;
0197     }
0198 
0199     /*
0200      * If hugetlbfs_put_super couldn't free spool due to an outstanding
0201      * quota reference, free it now.
0202      */
0203     unlock_or_release_subpool(spool);
0204 
0205     return ret;
0206 }
0207 
0208 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
0209 {
0210     return HUGETLBFS_SB(inode->i_sb)->spool;
0211 }
0212 
0213 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
0214 {
0215     return subpool_inode(file_inode(vma->vm_file));
0216 }
0217 
0218 /*
0219  * Region tracking -- allows tracking of reservations and instantiated pages
0220  *                    across the pages in a mapping.
0221  *
0222  * The region data structures are embedded into a resv_map and protected
0223  * by a resv_map's lock.  The set of regions within the resv_map represent
0224  * reservations for huge pages, or huge pages that have already been
0225  * instantiated within the map.  The from and to elements are huge page
0226  * indicies into the associated mapping.  from indicates the starting index
0227  * of the region.  to represents the first index past the end of  the region.
0228  *
0229  * For example, a file region structure with from == 0 and to == 4 represents
0230  * four huge pages in a mapping.  It is important to note that the to element
0231  * represents the first element past the end of the region. This is used in
0232  * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
0233  *
0234  * Interval notation of the form [from, to) will be used to indicate that
0235  * the endpoint from is inclusive and to is exclusive.
0236  */
0237 struct file_region {
0238     struct list_head link;
0239     long from;
0240     long to;
0241 };
0242 
0243 /*
0244  * Add the huge page range represented by [f, t) to the reserve
0245  * map.  In the normal case, existing regions will be expanded
0246  * to accommodate the specified range.  Sufficient regions should
0247  * exist for expansion due to the previous call to region_chg
0248  * with the same range.  However, it is possible that region_del
0249  * could have been called after region_chg and modifed the map
0250  * in such a way that no region exists to be expanded.  In this
0251  * case, pull a region descriptor from the cache associated with
0252  * the map and use that for the new range.
0253  *
0254  * Return the number of new huge pages added to the map.  This
0255  * number is greater than or equal to zero.
0256  */
0257 static long region_add(struct resv_map *resv, long f, long t)
0258 {
0259     struct list_head *head = &resv->regions;
0260     struct file_region *rg, *nrg, *trg;
0261     long add = 0;
0262 
0263     spin_lock(&resv->lock);
0264     /* Locate the region we are either in or before. */
0265     list_for_each_entry(rg, head, link)
0266         if (f <= rg->to)
0267             break;
0268 
0269     /*
0270      * If no region exists which can be expanded to include the
0271      * specified range, the list must have been modified by an
0272      * interleving call to region_del().  Pull a region descriptor
0273      * from the cache and use it for this range.
0274      */
0275     if (&rg->link == head || t < rg->from) {
0276         VM_BUG_ON(resv->region_cache_count <= 0);
0277 
0278         resv->region_cache_count--;
0279         nrg = list_first_entry(&resv->region_cache, struct file_region,
0280                     link);
0281         list_del(&nrg->link);
0282 
0283         nrg->from = f;
0284         nrg->to = t;
0285         list_add(&nrg->link, rg->link.prev);
0286 
0287         add += t - f;
0288         goto out_locked;
0289     }
0290 
0291     /* Round our left edge to the current segment if it encloses us. */
0292     if (f > rg->from)
0293         f = rg->from;
0294 
0295     /* Check for and consume any regions we now overlap with. */
0296     nrg = rg;
0297     list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
0298         if (&rg->link == head)
0299             break;
0300         if (rg->from > t)
0301             break;
0302 
0303         /* If this area reaches higher then extend our area to
0304          * include it completely.  If this is not the first area
0305          * which we intend to reuse, free it. */
0306         if (rg->to > t)
0307             t = rg->to;
0308         if (rg != nrg) {
0309             /* Decrement return value by the deleted range.
0310              * Another range will span this area so that by
0311              * end of routine add will be >= zero
0312              */
0313             add -= (rg->to - rg->from);
0314             list_del(&rg->link);
0315             kfree(rg);
0316         }
0317     }
0318 
0319     add += (nrg->from - f);     /* Added to beginning of region */
0320     nrg->from = f;
0321     add += t - nrg->to;     /* Added to end of region */
0322     nrg->to = t;
0323 
0324 out_locked:
0325     resv->adds_in_progress--;
0326     spin_unlock(&resv->lock);
0327     VM_BUG_ON(add < 0);
0328     return add;
0329 }
0330 
0331 /*
0332  * Examine the existing reserve map and determine how many
0333  * huge pages in the specified range [f, t) are NOT currently
0334  * represented.  This routine is called before a subsequent
0335  * call to region_add that will actually modify the reserve
0336  * map to add the specified range [f, t).  region_chg does
0337  * not change the number of huge pages represented by the
0338  * map.  However, if the existing regions in the map can not
0339  * be expanded to represent the new range, a new file_region
0340  * structure is added to the map as a placeholder.  This is
0341  * so that the subsequent region_add call will have all the
0342  * regions it needs and will not fail.
0343  *
0344  * Upon entry, region_chg will also examine the cache of region descriptors
0345  * associated with the map.  If there are not enough descriptors cached, one
0346  * will be allocated for the in progress add operation.
0347  *
0348  * Returns the number of huge pages that need to be added to the existing
0349  * reservation map for the range [f, t).  This number is greater or equal to
0350  * zero.  -ENOMEM is returned if a new file_region structure or cache entry
0351  * is needed and can not be allocated.
0352  */
0353 static long region_chg(struct resv_map *resv, long f, long t)
0354 {
0355     struct list_head *head = &resv->regions;
0356     struct file_region *rg, *nrg = NULL;
0357     long chg = 0;
0358 
0359 retry:
0360     spin_lock(&resv->lock);
0361 retry_locked:
0362     resv->adds_in_progress++;
0363 
0364     /*
0365      * Check for sufficient descriptors in the cache to accommodate
0366      * the number of in progress add operations.
0367      */
0368     if (resv->adds_in_progress > resv->region_cache_count) {
0369         struct file_region *trg;
0370 
0371         VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1);
0372         /* Must drop lock to allocate a new descriptor. */
0373         resv->adds_in_progress--;
0374         spin_unlock(&resv->lock);
0375 
0376         trg = kmalloc(sizeof(*trg), GFP_KERNEL);
0377         if (!trg) {
0378             kfree(nrg);
0379             return -ENOMEM;
0380         }
0381 
0382         spin_lock(&resv->lock);
0383         list_add(&trg->link, &resv->region_cache);
0384         resv->region_cache_count++;
0385         goto retry_locked;
0386     }
0387 
0388     /* Locate the region we are before or in. */
0389     list_for_each_entry(rg, head, link)
0390         if (f <= rg->to)
0391             break;
0392 
0393     /* If we are below the current region then a new region is required.
0394      * Subtle, allocate a new region at the position but make it zero
0395      * size such that we can guarantee to record the reservation. */
0396     if (&rg->link == head || t < rg->from) {
0397         if (!nrg) {
0398             resv->adds_in_progress--;
0399             spin_unlock(&resv->lock);
0400             nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
0401             if (!nrg)
0402                 return -ENOMEM;
0403 
0404             nrg->from = f;
0405             nrg->to   = f;
0406             INIT_LIST_HEAD(&nrg->link);
0407             goto retry;
0408         }
0409 
0410         list_add(&nrg->link, rg->link.prev);
0411         chg = t - f;
0412         goto out_nrg;
0413     }
0414 
0415     /* Round our left edge to the current segment if it encloses us. */
0416     if (f > rg->from)
0417         f = rg->from;
0418     chg = t - f;
0419 
0420     /* Check for and consume any regions we now overlap with. */
0421     list_for_each_entry(rg, rg->link.prev, link) {
0422         if (&rg->link == head)
0423             break;
0424         if (rg->from > t)
0425             goto out;
0426 
0427         /* We overlap with this area, if it extends further than
0428          * us then we must extend ourselves.  Account for its
0429          * existing reservation. */
0430         if (rg->to > t) {
0431             chg += rg->to - t;
0432             t = rg->to;
0433         }
0434         chg -= rg->to - rg->from;
0435     }
0436 
0437 out:
0438     spin_unlock(&resv->lock);
0439     /*  We already know we raced and no longer need the new region */
0440     kfree(nrg);
0441     return chg;
0442 out_nrg:
0443     spin_unlock(&resv->lock);
0444     return chg;
0445 }
0446 
0447 /*
0448  * Abort the in progress add operation.  The adds_in_progress field
0449  * of the resv_map keeps track of the operations in progress between
0450  * calls to region_chg and region_add.  Operations are sometimes
0451  * aborted after the call to region_chg.  In such cases, region_abort
0452  * is called to decrement the adds_in_progress counter.
0453  *
0454  * NOTE: The range arguments [f, t) are not needed or used in this
0455  * routine.  They are kept to make reading the calling code easier as
0456  * arguments will match the associated region_chg call.
0457  */
0458 static void region_abort(struct resv_map *resv, long f, long t)
0459 {
0460     spin_lock(&resv->lock);
0461     VM_BUG_ON(!resv->region_cache_count);
0462     resv->adds_in_progress--;
0463     spin_unlock(&resv->lock);
0464 }
0465 
0466 /*
0467  * Delete the specified range [f, t) from the reserve map.  If the
0468  * t parameter is LONG_MAX, this indicates that ALL regions after f
0469  * should be deleted.  Locate the regions which intersect [f, t)
0470  * and either trim, delete or split the existing regions.
0471  *
0472  * Returns the number of huge pages deleted from the reserve map.
0473  * In the normal case, the return value is zero or more.  In the
0474  * case where a region must be split, a new region descriptor must
0475  * be allocated.  If the allocation fails, -ENOMEM will be returned.
0476  * NOTE: If the parameter t == LONG_MAX, then we will never split
0477  * a region and possibly return -ENOMEM.  Callers specifying
0478  * t == LONG_MAX do not need to check for -ENOMEM error.
0479  */
0480 static long region_del(struct resv_map *resv, long f, long t)
0481 {
0482     struct list_head *head = &resv->regions;
0483     struct file_region *rg, *trg;
0484     struct file_region *nrg = NULL;
0485     long del = 0;
0486 
0487 retry:
0488     spin_lock(&resv->lock);
0489     list_for_each_entry_safe(rg, trg, head, link) {
0490         /*
0491          * Skip regions before the range to be deleted.  file_region
0492          * ranges are normally of the form [from, to).  However, there
0493          * may be a "placeholder" entry in the map which is of the form
0494          * (from, to) with from == to.  Check for placeholder entries
0495          * at the beginning of the range to be deleted.
0496          */
0497         if (rg->to <= f && (rg->to != rg->from || rg->to != f))
0498             continue;
0499 
0500         if (rg->from >= t)
0501             break;
0502 
0503         if (f > rg->from && t < rg->to) { /* Must split region */
0504             /*
0505              * Check for an entry in the cache before dropping
0506              * lock and attempting allocation.
0507              */
0508             if (!nrg &&
0509                 resv->region_cache_count > resv->adds_in_progress) {
0510                 nrg = list_first_entry(&resv->region_cache,
0511                             struct file_region,
0512                             link);
0513                 list_del(&nrg->link);
0514                 resv->region_cache_count--;
0515             }
0516 
0517             if (!nrg) {
0518                 spin_unlock(&resv->lock);
0519                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
0520                 if (!nrg)
0521                     return -ENOMEM;
0522                 goto retry;
0523             }
0524 
0525             del += t - f;
0526 
0527             /* New entry for end of split region */
0528             nrg->from = t;
0529             nrg->to = rg->to;
0530             INIT_LIST_HEAD(&nrg->link);
0531 
0532             /* Original entry is trimmed */
0533             rg->to = f;
0534 
0535             list_add(&nrg->link, &rg->link);
0536             nrg = NULL;
0537             break;
0538         }
0539 
0540         if (f <= rg->from && t >= rg->to) { /* Remove entire region */
0541             del += rg->to - rg->from;
0542             list_del(&rg->link);
0543             kfree(rg);
0544             continue;
0545         }
0546 
0547         if (f <= rg->from) {    /* Trim beginning of region */
0548             del += t - rg->from;
0549             rg->from = t;
0550         } else {        /* Trim end of region */
0551             del += rg->to - f;
0552             rg->to = f;
0553         }
0554     }
0555 
0556     spin_unlock(&resv->lock);
0557     kfree(nrg);
0558     return del;
0559 }
0560 
0561 /*
0562  * A rare out of memory error was encountered which prevented removal of
0563  * the reserve map region for a page.  The huge page itself was free'ed
0564  * and removed from the page cache.  This routine will adjust the subpool
0565  * usage count, and the global reserve count if needed.  By incrementing
0566  * these counts, the reserve map entry which could not be deleted will
0567  * appear as a "reserved" entry instead of simply dangling with incorrect
0568  * counts.
0569  */
0570 void hugetlb_fix_reserve_counts(struct inode *inode)
0571 {
0572     struct hugepage_subpool *spool = subpool_inode(inode);
0573     long rsv_adjust;
0574 
0575     rsv_adjust = hugepage_subpool_get_pages(spool, 1);
0576     if (rsv_adjust) {
0577         struct hstate *h = hstate_inode(inode);
0578 
0579         hugetlb_acct_memory(h, 1);
0580     }
0581 }
0582 
0583 /*
0584  * Count and return the number of huge pages in the reserve map
0585  * that intersect with the range [f, t).
0586  */
0587 static long region_count(struct resv_map *resv, long f, long t)
0588 {
0589     struct list_head *head = &resv->regions;
0590     struct file_region *rg;
0591     long chg = 0;
0592 
0593     spin_lock(&resv->lock);
0594     /* Locate each segment we overlap with, and count that overlap. */
0595     list_for_each_entry(rg, head, link) {
0596         long seg_from;
0597         long seg_to;
0598 
0599         if (rg->to <= f)
0600             continue;
0601         if (rg->from >= t)
0602             break;
0603 
0604         seg_from = max(rg->from, f);
0605         seg_to = min(rg->to, t);
0606 
0607         chg += seg_to - seg_from;
0608     }
0609     spin_unlock(&resv->lock);
0610 
0611     return chg;
0612 }
0613 
0614 /*
0615  * Convert the address within this vma to the page offset within
0616  * the mapping, in pagecache page units; huge pages here.
0617  */
0618 static pgoff_t vma_hugecache_offset(struct hstate *h,
0619             struct vm_area_struct *vma, unsigned long address)
0620 {
0621     return ((address - vma->vm_start) >> huge_page_shift(h)) +
0622             (vma->vm_pgoff >> huge_page_order(h));
0623 }
0624 
0625 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
0626                      unsigned long address)
0627 {
0628     return vma_hugecache_offset(hstate_vma(vma), vma, address);
0629 }
0630 EXPORT_SYMBOL_GPL(linear_hugepage_index);
0631 
0632 /*
0633  * Return the size of the pages allocated when backing a VMA. In the majority
0634  * cases this will be same size as used by the page table entries.
0635  */
0636 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
0637 {
0638     struct hstate *hstate;
0639 
0640     if (!is_vm_hugetlb_page(vma))
0641         return PAGE_SIZE;
0642 
0643     hstate = hstate_vma(vma);
0644 
0645     return 1UL << huge_page_shift(hstate);
0646 }
0647 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
0648 
0649 /*
0650  * Return the page size being used by the MMU to back a VMA. In the majority
0651  * of cases, the page size used by the kernel matches the MMU size. On
0652  * architectures where it differs, an architecture-specific version of this
0653  * function is required.
0654  */
0655 #ifndef vma_mmu_pagesize
0656 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
0657 {
0658     return vma_kernel_pagesize(vma);
0659 }
0660 #endif
0661 
0662 /*
0663  * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
0664  * bits of the reservation map pointer, which are always clear due to
0665  * alignment.
0666  */
0667 #define HPAGE_RESV_OWNER    (1UL << 0)
0668 #define HPAGE_RESV_UNMAPPED (1UL << 1)
0669 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
0670 
0671 /*
0672  * These helpers are used to track how many pages are reserved for
0673  * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
0674  * is guaranteed to have their future faults succeed.
0675  *
0676  * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
0677  * the reserve counters are updated with the hugetlb_lock held. It is safe
0678  * to reset the VMA at fork() time as it is not in use yet and there is no
0679  * chance of the global counters getting corrupted as a result of the values.
0680  *
0681  * The private mapping reservation is represented in a subtly different
0682  * manner to a shared mapping.  A shared mapping has a region map associated
0683  * with the underlying file, this region map represents the backing file
0684  * pages which have ever had a reservation assigned which this persists even
0685  * after the page is instantiated.  A private mapping has a region map
0686  * associated with the original mmap which is attached to all VMAs which
0687  * reference it, this region map represents those offsets which have consumed
0688  * reservation ie. where pages have been instantiated.
0689  */
0690 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
0691 {
0692     return (unsigned long)vma->vm_private_data;
0693 }
0694 
0695 static void set_vma_private_data(struct vm_area_struct *vma,
0696                             unsigned long value)
0697 {
0698     vma->vm_private_data = (void *)value;
0699 }
0700 
0701 struct resv_map *resv_map_alloc(void)
0702 {
0703     struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
0704     struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
0705 
0706     if (!resv_map || !rg) {
0707         kfree(resv_map);
0708         kfree(rg);
0709         return NULL;
0710     }
0711 
0712     kref_init(&resv_map->refs);
0713     spin_lock_init(&resv_map->lock);
0714     INIT_LIST_HEAD(&resv_map->regions);
0715 
0716     resv_map->adds_in_progress = 0;
0717 
0718     INIT_LIST_HEAD(&resv_map->region_cache);
0719     list_add(&rg->link, &resv_map->region_cache);
0720     resv_map->region_cache_count = 1;
0721 
0722     return resv_map;
0723 }
0724 
0725 void resv_map_release(struct kref *ref)
0726 {
0727     struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
0728     struct list_head *head = &resv_map->region_cache;
0729     struct file_region *rg, *trg;
0730 
0731     /* Clear out any active regions before we release the map. */
0732     region_del(resv_map, 0, LONG_MAX);
0733 
0734     /* ... and any entries left in the cache */
0735     list_for_each_entry_safe(rg, trg, head, link) {
0736         list_del(&rg->link);
0737         kfree(rg);
0738     }
0739 
0740     VM_BUG_ON(resv_map->adds_in_progress);
0741 
0742     kfree(resv_map);
0743 }
0744 
0745 static inline struct resv_map *inode_resv_map(struct inode *inode)
0746 {
0747     return inode->i_mapping->private_data;
0748 }
0749 
0750 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
0751 {
0752     VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
0753     if (vma->vm_flags & VM_MAYSHARE) {
0754         struct address_space *mapping = vma->vm_file->f_mapping;
0755         struct inode *inode = mapping->host;
0756 
0757         return inode_resv_map(inode);
0758 
0759     } else {
0760         return (struct resv_map *)(get_vma_private_data(vma) &
0761                             ~HPAGE_RESV_MASK);
0762     }
0763 }
0764 
0765 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
0766 {
0767     VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
0768     VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
0769 
0770     set_vma_private_data(vma, (get_vma_private_data(vma) &
0771                 HPAGE_RESV_MASK) | (unsigned long)map);
0772 }
0773 
0774 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
0775 {
0776     VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
0777     VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
0778 
0779     set_vma_private_data(vma, get_vma_private_data(vma) | flags);
0780 }
0781 
0782 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
0783 {
0784     VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
0785 
0786     return (get_vma_private_data(vma) & flag) != 0;
0787 }
0788 
0789 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
0790 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
0791 {
0792     VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
0793     if (!(vma->vm_flags & VM_MAYSHARE))
0794         vma->vm_private_data = (void *)0;
0795 }
0796 
0797 /* Returns true if the VMA has associated reserve pages */
0798 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
0799 {
0800     if (vma->vm_flags & VM_NORESERVE) {
0801         /*
0802          * This address is already reserved by other process(chg == 0),
0803          * so, we should decrement reserved count. Without decrementing,
0804          * reserve count remains after releasing inode, because this
0805          * allocated page will go into page cache and is regarded as
0806          * coming from reserved pool in releasing step.  Currently, we
0807          * don't have any other solution to deal with this situation
0808          * properly, so add work-around here.
0809          */
0810         if (vma->vm_flags & VM_MAYSHARE && chg == 0)
0811             return true;
0812         else
0813             return false;
0814     }
0815 
0816     /* Shared mappings always use reserves */
0817     if (vma->vm_flags & VM_MAYSHARE) {
0818         /*
0819          * We know VM_NORESERVE is not set.  Therefore, there SHOULD
0820          * be a region map for all pages.  The only situation where
0821          * there is no region map is if a hole was punched via
0822          * fallocate.  In this case, there really are no reverves to
0823          * use.  This situation is indicated if chg != 0.
0824          */
0825         if (chg)
0826             return false;
0827         else
0828             return true;
0829     }
0830 
0831     /*
0832      * Only the process that called mmap() has reserves for
0833      * private mappings.
0834      */
0835     if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
0836         /*
0837          * Like the shared case above, a hole punch or truncate
0838          * could have been performed on the private mapping.
0839          * Examine the value of chg to determine if reserves
0840          * actually exist or were previously consumed.
0841          * Very Subtle - The value of chg comes from a previous
0842          * call to vma_needs_reserves().  The reserve map for
0843          * private mappings has different (opposite) semantics
0844          * than that of shared mappings.  vma_needs_reserves()
0845          * has already taken this difference in semantics into
0846          * account.  Therefore, the meaning of chg is the same
0847          * as in the shared case above.  Code could easily be
0848          * combined, but keeping it separate draws attention to
0849          * subtle differences.
0850          */
0851         if (chg)
0852             return false;
0853         else
0854             return true;
0855     }
0856 
0857     return false;
0858 }
0859 
0860 static void enqueue_huge_page(struct hstate *h, struct page *page)
0861 {
0862     int nid = page_to_nid(page);
0863     list_move(&page->lru, &h->hugepage_freelists[nid]);
0864     h->free_huge_pages++;
0865     h->free_huge_pages_node[nid]++;
0866 }
0867 
0868 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
0869 {
0870     struct page *page;
0871 
0872     list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
0873         if (!is_migrate_isolate_page(page))
0874             break;
0875     /*
0876      * if 'non-isolated free hugepage' not found on the list,
0877      * the allocation fails.
0878      */
0879     if (&h->hugepage_freelists[nid] == &page->lru)
0880         return NULL;
0881     list_move(&page->lru, &h->hugepage_activelist);
0882     set_page_refcounted(page);
0883     h->free_huge_pages--;
0884     h->free_huge_pages_node[nid]--;
0885     return page;
0886 }
0887 
0888 /* Movability of hugepages depends on migration support. */
0889 static inline gfp_t htlb_alloc_mask(struct hstate *h)
0890 {
0891     if (hugepages_treat_as_movable || hugepage_migration_supported(h))
0892         return GFP_HIGHUSER_MOVABLE;
0893     else
0894         return GFP_HIGHUSER;
0895 }
0896 
0897 static struct page *dequeue_huge_page_vma(struct hstate *h,
0898                 struct vm_area_struct *vma,
0899                 unsigned long address, int avoid_reserve,
0900                 long chg)
0901 {
0902     struct page *page = NULL;
0903     struct mempolicy *mpol;
0904     nodemask_t *nodemask;
0905     struct zonelist *zonelist;
0906     struct zone *zone;
0907     struct zoneref *z;
0908     unsigned int cpuset_mems_cookie;
0909 
0910     /*
0911      * A child process with MAP_PRIVATE mappings created by their parent
0912      * have no page reserves. This check ensures that reservations are
0913      * not "stolen". The child may still get SIGKILLed
0914      */
0915     if (!vma_has_reserves(vma, chg) &&
0916             h->free_huge_pages - h->resv_huge_pages == 0)
0917         goto err;
0918 
0919     /* If reserves cannot be used, ensure enough pages are in the pool */
0920     if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
0921         goto err;
0922 
0923 retry_cpuset:
0924     cpuset_mems_cookie = read_mems_allowed_begin();
0925     zonelist = huge_zonelist(vma, address,
0926                     htlb_alloc_mask(h), &mpol, &nodemask);
0927 
0928     for_each_zone_zonelist_nodemask(zone, z, zonelist,
0929                         MAX_NR_ZONES - 1, nodemask) {
0930         if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) {
0931             page = dequeue_huge_page_node(h, zone_to_nid(zone));
0932             if (page) {
0933                 if (avoid_reserve)
0934                     break;
0935                 if (!vma_has_reserves(vma, chg))
0936                     break;
0937 
0938                 SetPagePrivate(page);
0939                 h->resv_huge_pages--;
0940                 break;
0941             }
0942         }
0943     }
0944 
0945     mpol_cond_put(mpol);
0946     if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
0947         goto retry_cpuset;
0948     return page;
0949 
0950 err:
0951     return NULL;
0952 }
0953 
0954 /*
0955  * common helper functions for hstate_next_node_to_{alloc|free}.
0956  * We may have allocated or freed a huge page based on a different
0957  * nodes_allowed previously, so h->next_node_to_{alloc|free} might
0958  * be outside of *nodes_allowed.  Ensure that we use an allowed
0959  * node for alloc or free.
0960  */
0961 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
0962 {
0963     nid = next_node_in(nid, *nodes_allowed);
0964     VM_BUG_ON(nid >= MAX_NUMNODES);
0965 
0966     return nid;
0967 }
0968 
0969 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
0970 {
0971     if (!node_isset(nid, *nodes_allowed))
0972         nid = next_node_allowed(nid, nodes_allowed);
0973     return nid;
0974 }
0975 
0976 /*
0977  * returns the previously saved node ["this node"] from which to
0978  * allocate a persistent huge page for the pool and advance the
0979  * next node from which to allocate, handling wrap at end of node
0980  * mask.
0981  */
0982 static int hstate_next_node_to_alloc(struct hstate *h,
0983                     nodemask_t *nodes_allowed)
0984 {
0985     int nid;
0986 
0987     VM_BUG_ON(!nodes_allowed);
0988 
0989     nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
0990     h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
0991 
0992     return nid;
0993 }
0994 
0995 /*
0996  * helper for free_pool_huge_page() - return the previously saved
0997  * node ["this node"] from which to free a huge page.  Advance the
0998  * next node id whether or not we find a free huge page to free so
0999  * that the next attempt to free addresses the next node.
1000  */
1001 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1002 {
1003     int nid;
1004 
1005     VM_BUG_ON(!nodes_allowed);
1006 
1007     nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1008     h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1009 
1010     return nid;
1011 }
1012 
1013 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask)       \
1014     for (nr_nodes = nodes_weight(*mask);                \
1015         nr_nodes > 0 &&                     \
1016         ((node = hstate_next_node_to_alloc(hs, mask)) || 1);    \
1017         nr_nodes--)
1018 
1019 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask)        \
1020     for (nr_nodes = nodes_weight(*mask);                \
1021         nr_nodes > 0 &&                     \
1022         ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1023         nr_nodes--)
1024 
1025 #if defined(CONFIG_ARCH_HAS_GIGANTIC_PAGE) && \
1026     ((defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || \
1027     defined(CONFIG_CMA))
1028 static void destroy_compound_gigantic_page(struct page *page,
1029                     unsigned int order)
1030 {
1031     int i;
1032     int nr_pages = 1 << order;
1033     struct page *p = page + 1;
1034 
1035     atomic_set(compound_mapcount_ptr(page), 0);
1036     for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1037         clear_compound_head(p);
1038         set_page_refcounted(p);
1039     }
1040 
1041     set_compound_order(page, 0);
1042     __ClearPageHead(page);
1043 }
1044 
1045 static void free_gigantic_page(struct page *page, unsigned int order)
1046 {
1047     free_contig_range(page_to_pfn(page), 1 << order);
1048 }
1049 
1050 static int __alloc_gigantic_page(unsigned long start_pfn,
1051                 unsigned long nr_pages)
1052 {
1053     unsigned long end_pfn = start_pfn + nr_pages;
1054     return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
1055 }
1056 
1057 static bool pfn_range_valid_gigantic(struct zone *z,
1058             unsigned long start_pfn, unsigned long nr_pages)
1059 {
1060     unsigned long i, end_pfn = start_pfn + nr_pages;
1061     struct page *page;
1062 
1063     for (i = start_pfn; i < end_pfn; i++) {
1064         if (!pfn_valid(i))
1065             return false;
1066 
1067         page = pfn_to_page(i);
1068 
1069         if (page_zone(page) != z)
1070             return false;
1071 
1072         if (PageReserved(page))
1073             return false;
1074 
1075         if (page_count(page) > 0)
1076             return false;
1077 
1078         if (PageHuge(page))
1079             return false;
1080     }
1081 
1082     return true;
1083 }
1084 
1085 static bool zone_spans_last_pfn(const struct zone *zone,
1086             unsigned long start_pfn, unsigned long nr_pages)
1087 {
1088     unsigned long last_pfn = start_pfn + nr_pages - 1;
1089     return zone_spans_pfn(zone, last_pfn);
1090 }
1091 
1092 static struct page *alloc_gigantic_page(int nid, unsigned int order)
1093 {
1094     unsigned long nr_pages = 1 << order;
1095     unsigned long ret, pfn, flags;
1096     struct zone *z;
1097 
1098     z = NODE_DATA(nid)->node_zones;
1099     for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) {
1100         spin_lock_irqsave(&z->lock, flags);
1101 
1102         pfn = ALIGN(z->zone_start_pfn, nr_pages);
1103         while (zone_spans_last_pfn(z, pfn, nr_pages)) {
1104             if (pfn_range_valid_gigantic(z, pfn, nr_pages)) {
1105                 /*
1106                  * We release the zone lock here because
1107                  * alloc_contig_range() will also lock the zone
1108                  * at some point. If there's an allocation
1109                  * spinning on this lock, it may win the race
1110                  * and cause alloc_contig_range() to fail...
1111                  */
1112                 spin_unlock_irqrestore(&z->lock, flags);
1113                 ret = __alloc_gigantic_page(pfn, nr_pages);
1114                 if (!ret)
1115                     return pfn_to_page(pfn);
1116                 spin_lock_irqsave(&z->lock, flags);
1117             }
1118             pfn += nr_pages;
1119         }
1120 
1121         spin_unlock_irqrestore(&z->lock, flags);
1122     }
1123 
1124     return NULL;
1125 }
1126 
1127 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1128 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1129 
1130 static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid)
1131 {
1132     struct page *page;
1133 
1134     page = alloc_gigantic_page(nid, huge_page_order(h));
1135     if (page) {
1136         prep_compound_gigantic_page(page, huge_page_order(h));
1137         prep_new_huge_page(h, page, nid);
1138     }
1139 
1140     return page;
1141 }
1142 
1143 static int alloc_fresh_gigantic_page(struct hstate *h,
1144                 nodemask_t *nodes_allowed)
1145 {
1146     struct page *page = NULL;
1147     int nr_nodes, node;
1148 
1149     for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1150         page = alloc_fresh_gigantic_page_node(h, node);
1151         if (page)
1152             return 1;
1153     }
1154 
1155     return 0;
1156 }
1157 
1158 static inline bool gigantic_page_supported(void) { return true; }
1159 #else
1160 static inline bool gigantic_page_supported(void) { return false; }
1161 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1162 static inline void destroy_compound_gigantic_page(struct page *page,
1163                         unsigned int order) { }
1164 static inline int alloc_fresh_gigantic_page(struct hstate *h,
1165                     nodemask_t *nodes_allowed) { return 0; }
1166 #endif
1167 
1168 static void update_and_free_page(struct hstate *h, struct page *page)
1169 {
1170     int i;
1171 
1172     if (hstate_is_gigantic(h) && !gigantic_page_supported())
1173         return;
1174 
1175     h->nr_huge_pages--;
1176     h->nr_huge_pages_node[page_to_nid(page)]--;
1177     for (i = 0; i < pages_per_huge_page(h); i++) {
1178         page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1179                 1 << PG_referenced | 1 << PG_dirty |
1180                 1 << PG_active | 1 << PG_private |
1181                 1 << PG_writeback);
1182     }
1183     VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1184     set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1185     set_page_refcounted(page);
1186     if (hstate_is_gigantic(h)) {
1187         destroy_compound_gigantic_page(page, huge_page_order(h));
1188         free_gigantic_page(page, huge_page_order(h));
1189     } else {
1190         __free_pages(page, huge_page_order(h));
1191     }
1192 }
1193 
1194 struct hstate *size_to_hstate(unsigned long size)
1195 {
1196     struct hstate *h;
1197 
1198     for_each_hstate(h) {
1199         if (huge_page_size(h) == size)
1200             return h;
1201     }
1202     return NULL;
1203 }
1204 
1205 /*
1206  * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1207  * to hstate->hugepage_activelist.)
1208  *
1209  * This function can be called for tail pages, but never returns true for them.
1210  */
1211 bool page_huge_active(struct page *page)
1212 {
1213     VM_BUG_ON_PAGE(!PageHuge(page), page);
1214     return PageHead(page) && PagePrivate(&page[1]);
1215 }
1216 
1217 /* never called for tail page */
1218 static void set_page_huge_active(struct page *page)
1219 {
1220     VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1221     SetPagePrivate(&page[1]);
1222 }
1223 
1224 static void clear_page_huge_active(struct page *page)
1225 {
1226     VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1227     ClearPagePrivate(&page[1]);
1228 }
1229 
1230 void free_huge_page(struct page *page)
1231 {
1232     /*
1233      * Can't pass hstate in here because it is called from the
1234      * compound page destructor.
1235      */
1236     struct hstate *h = page_hstate(page);
1237     int nid = page_to_nid(page);
1238     struct hugepage_subpool *spool =
1239         (struct hugepage_subpool *)page_private(page);
1240     bool restore_reserve;
1241 
1242     set_page_private(page, 0);
1243     page->mapping = NULL;
1244     VM_BUG_ON_PAGE(page_count(page), page);
1245     VM_BUG_ON_PAGE(page_mapcount(page), page);
1246     restore_reserve = PagePrivate(page);
1247     ClearPagePrivate(page);
1248 
1249     /*
1250      * A return code of zero implies that the subpool will be under its
1251      * minimum size if the reservation is not restored after page is free.
1252      * Therefore, force restore_reserve operation.
1253      */
1254     if (hugepage_subpool_put_pages(spool, 1) == 0)
1255         restore_reserve = true;
1256 
1257     spin_lock(&hugetlb_lock);
1258     clear_page_huge_active(page);
1259     hugetlb_cgroup_uncharge_page(hstate_index(h),
1260                      pages_per_huge_page(h), page);
1261     if (restore_reserve)
1262         h->resv_huge_pages++;
1263 
1264     if (h->surplus_huge_pages_node[nid]) {
1265         /* remove the page from active list */
1266         list_del(&page->lru);
1267         update_and_free_page(h, page);
1268         h->surplus_huge_pages--;
1269         h->surplus_huge_pages_node[nid]--;
1270     } else {
1271         arch_clear_hugepage_flags(page);
1272         enqueue_huge_page(h, page);
1273     }
1274     spin_unlock(&hugetlb_lock);
1275 }
1276 
1277 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1278 {
1279     INIT_LIST_HEAD(&page->lru);
1280     set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1281     spin_lock(&hugetlb_lock);
1282     set_hugetlb_cgroup(page, NULL);
1283     h->nr_huge_pages++;
1284     h->nr_huge_pages_node[nid]++;
1285     spin_unlock(&hugetlb_lock);
1286     put_page(page); /* free it into the hugepage allocator */
1287 }
1288 
1289 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1290 {
1291     int i;
1292     int nr_pages = 1 << order;
1293     struct page *p = page + 1;
1294 
1295     /* we rely on prep_new_huge_page to set the destructor */
1296     set_compound_order(page, order);
1297     __ClearPageReserved(page);
1298     __SetPageHead(page);
1299     for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1300         /*
1301          * For gigantic hugepages allocated through bootmem at
1302          * boot, it's safer to be consistent with the not-gigantic
1303          * hugepages and clear the PG_reserved bit from all tail pages
1304          * too.  Otherwse drivers using get_user_pages() to access tail
1305          * pages may get the reference counting wrong if they see
1306          * PG_reserved set on a tail page (despite the head page not
1307          * having PG_reserved set).  Enforcing this consistency between
1308          * head and tail pages allows drivers to optimize away a check
1309          * on the head page when they need know if put_page() is needed
1310          * after get_user_pages().
1311          */
1312         __ClearPageReserved(p);
1313         set_page_count(p, 0);
1314         set_compound_head(p, page);
1315     }
1316     atomic_set(compound_mapcount_ptr(page), -1);
1317 }
1318 
1319 /*
1320  * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1321  * transparent huge pages.  See the PageTransHuge() documentation for more
1322  * details.
1323  */
1324 int PageHuge(struct page *page)
1325 {
1326     if (!PageCompound(page))
1327         return 0;
1328 
1329     page = compound_head(page);
1330     return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1331 }
1332 EXPORT_SYMBOL_GPL(PageHuge);
1333 
1334 /*
1335  * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1336  * normal or transparent huge pages.
1337  */
1338 int PageHeadHuge(struct page *page_head)
1339 {
1340     if (!PageHead(page_head))
1341         return 0;
1342 
1343     return get_compound_page_dtor(page_head) == free_huge_page;
1344 }
1345 
1346 pgoff_t __basepage_index(struct page *page)
1347 {
1348     struct page *page_head = compound_head(page);
1349     pgoff_t index = page_index(page_head);
1350     unsigned long compound_idx;
1351 
1352     if (!PageHuge(page_head))
1353         return page_index(page);
1354 
1355     if (compound_order(page_head) >= MAX_ORDER)
1356         compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1357     else
1358         compound_idx = page - page_head;
1359 
1360     return (index << compound_order(page_head)) + compound_idx;
1361 }
1362 
1363 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
1364 {
1365     struct page *page;
1366 
1367     page = __alloc_pages_node(nid,
1368         htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
1369                         __GFP_REPEAT|__GFP_NOWARN,
1370         huge_page_order(h));
1371     if (page) {
1372         prep_new_huge_page(h, page, nid);
1373     }
1374 
1375     return page;
1376 }
1377 
1378 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1379 {
1380     struct page *page;
1381     int nr_nodes, node;
1382     int ret = 0;
1383 
1384     for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1385         page = alloc_fresh_huge_page_node(h, node);
1386         if (page) {
1387             ret = 1;
1388             break;
1389         }
1390     }
1391 
1392     if (ret)
1393         count_vm_event(HTLB_BUDDY_PGALLOC);
1394     else
1395         count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1396 
1397     return ret;
1398 }
1399 
1400 /*
1401  * Free huge page from pool from next node to free.
1402  * Attempt to keep persistent huge pages more or less
1403  * balanced over allowed nodes.
1404  * Called with hugetlb_lock locked.
1405  */
1406 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1407                              bool acct_surplus)
1408 {
1409     int nr_nodes, node;
1410     int ret = 0;
1411 
1412     for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1413         /*
1414          * If we're returning unused surplus pages, only examine
1415          * nodes with surplus pages.
1416          */
1417         if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1418             !list_empty(&h->hugepage_freelists[node])) {
1419             struct page *page =
1420                 list_entry(h->hugepage_freelists[node].next,
1421                       struct page, lru);
1422             list_del(&page->lru);
1423             h->free_huge_pages--;
1424             h->free_huge_pages_node[node]--;
1425             if (acct_surplus) {
1426                 h->surplus_huge_pages--;
1427                 h->surplus_huge_pages_node[node]--;
1428             }
1429             update_and_free_page(h, page);
1430             ret = 1;
1431             break;
1432         }
1433     }
1434 
1435     return ret;
1436 }
1437 
1438 /*
1439  * Dissolve a given free hugepage into free buddy pages. This function does
1440  * nothing for in-use (including surplus) hugepages. Returns -EBUSY if the
1441  * number of free hugepages would be reduced below the number of reserved
1442  * hugepages.
1443  */
1444 static int dissolve_free_huge_page(struct page *page)
1445 {
1446     int rc = 0;
1447 
1448     spin_lock(&hugetlb_lock);
1449     if (PageHuge(page) && !page_count(page)) {
1450         struct page *head = compound_head(page);
1451         struct hstate *h = page_hstate(head);
1452         int nid = page_to_nid(head);
1453         if (h->free_huge_pages - h->resv_huge_pages == 0) {
1454             rc = -EBUSY;
1455             goto out;
1456         }
1457         list_del(&head->lru);
1458         h->free_huge_pages--;
1459         h->free_huge_pages_node[nid]--;
1460         h->max_huge_pages--;
1461         update_and_free_page(h, head);
1462     }
1463 out:
1464     spin_unlock(&hugetlb_lock);
1465     return rc;
1466 }
1467 
1468 /*
1469  * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1470  * make specified memory blocks removable from the system.
1471  * Note that this will dissolve a free gigantic hugepage completely, if any
1472  * part of it lies within the given range.
1473  * Also note that if dissolve_free_huge_page() returns with an error, all
1474  * free hugepages that were dissolved before that error are lost.
1475  */
1476 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1477 {
1478     unsigned long pfn;
1479     struct page *page;
1480     int rc = 0;
1481 
1482     if (!hugepages_supported())
1483         return rc;
1484 
1485     for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1486         page = pfn_to_page(pfn);
1487         if (PageHuge(page) && !page_count(page)) {
1488             rc = dissolve_free_huge_page(page);
1489             if (rc)
1490                 break;
1491         }
1492     }
1493 
1494     return rc;
1495 }
1496 
1497 /*
1498  * There are 3 ways this can get called:
1499  * 1. With vma+addr: we use the VMA's memory policy
1500  * 2. With !vma, but nid=NUMA_NO_NODE:  We try to allocate a huge
1501  *    page from any node, and let the buddy allocator itself figure
1502  *    it out.
1503  * 3. With !vma, but nid!=NUMA_NO_NODE.  We allocate a huge page
1504  *    strictly from 'nid'
1505  */
1506 static struct page *__hugetlb_alloc_buddy_huge_page(struct hstate *h,
1507         struct vm_area_struct *vma, unsigned long addr, int nid)
1508 {
1509     int order = huge_page_order(h);
1510     gfp_t gfp = htlb_alloc_mask(h)|__GFP_COMP|__GFP_REPEAT|__GFP_NOWARN;
1511     unsigned int cpuset_mems_cookie;
1512 
1513     /*
1514      * We need a VMA to get a memory policy.  If we do not
1515      * have one, we use the 'nid' argument.
1516      *
1517      * The mempolicy stuff below has some non-inlined bits
1518      * and calls ->vm_ops.  That makes it hard to optimize at
1519      * compile-time, even when NUMA is off and it does
1520      * nothing.  This helps the compiler optimize it out.
1521      */
1522     if (!IS_ENABLED(CONFIG_NUMA) || !vma) {
1523         /*
1524          * If a specific node is requested, make sure to
1525          * get memory from there, but only when a node
1526          * is explicitly specified.
1527          */
1528         if (nid != NUMA_NO_NODE)
1529             gfp |= __GFP_THISNODE;
1530         /*
1531          * Make sure to call something that can handle
1532          * nid=NUMA_NO_NODE
1533          */
1534         return alloc_pages_node(nid, gfp, order);
1535     }
1536 
1537     /*
1538      * OK, so we have a VMA.  Fetch the mempolicy and try to
1539      * allocate a huge page with it.  We will only reach this
1540      * when CONFIG_NUMA=y.
1541      */
1542     do {
1543         struct page *page;
1544         struct mempolicy *mpol;
1545         struct zonelist *zl;
1546         nodemask_t *nodemask;
1547 
1548         cpuset_mems_cookie = read_mems_allowed_begin();
1549         zl = huge_zonelist(vma, addr, gfp, &mpol, &nodemask);
1550         mpol_cond_put(mpol);
1551         page = __alloc_pages_nodemask(gfp, order, zl, nodemask);
1552         if (page)
1553             return page;
1554     } while (read_mems_allowed_retry(cpuset_mems_cookie));
1555 
1556     return NULL;
1557 }
1558 
1559 /*
1560  * There are two ways to allocate a huge page:
1561  * 1. When you have a VMA and an address (like a fault)
1562  * 2. When you have no VMA (like when setting /proc/.../nr_hugepages)
1563  *
1564  * 'vma' and 'addr' are only for (1).  'nid' is always NUMA_NO_NODE in
1565  * this case which signifies that the allocation should be done with
1566  * respect for the VMA's memory policy.
1567  *
1568  * For (2), we ignore 'vma' and 'addr' and use 'nid' exclusively. This
1569  * implies that memory policies will not be taken in to account.
1570  */
1571 static struct page *__alloc_buddy_huge_page(struct hstate *h,
1572         struct vm_area_struct *vma, unsigned long addr, int nid)
1573 {
1574     struct page *page;
1575     unsigned int r_nid;
1576 
1577     if (hstate_is_gigantic(h))
1578         return NULL;
1579 
1580     /*
1581      * Make sure that anyone specifying 'nid' is not also specifying a VMA.
1582      * This makes sure the caller is picking _one_ of the modes with which
1583      * we can call this function, not both.
1584      */
1585     if (vma || (addr != -1)) {
1586         VM_WARN_ON_ONCE(addr == -1);
1587         VM_WARN_ON_ONCE(nid != NUMA_NO_NODE);
1588     }
1589     /*
1590      * Assume we will successfully allocate the surplus page to
1591      * prevent racing processes from causing the surplus to exceed
1592      * overcommit
1593      *
1594      * This however introduces a different race, where a process B
1595      * tries to grow the static hugepage pool while alloc_pages() is
1596      * called by process A. B will only examine the per-node
1597      * counters in determining if surplus huge pages can be
1598      * converted to normal huge pages in adjust_pool_surplus(). A
1599      * won't be able to increment the per-node counter, until the
1600      * lock is dropped by B, but B doesn't drop hugetlb_lock until
1601      * no more huge pages can be converted from surplus to normal
1602      * state (and doesn't try to convert again). Thus, we have a
1603      * case where a surplus huge page exists, the pool is grown, and
1604      * the surplus huge page still exists after, even though it
1605      * should just have been converted to a normal huge page. This
1606      * does not leak memory, though, as the hugepage will be freed
1607      * once it is out of use. It also does not allow the counters to
1608      * go out of whack in adjust_pool_surplus() as we don't modify
1609      * the node values until we've gotten the hugepage and only the
1610      * per-node value is checked there.
1611      */
1612     spin_lock(&hugetlb_lock);
1613     if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1614         spin_unlock(&hugetlb_lock);
1615         return NULL;
1616     } else {
1617         h->nr_huge_pages++;
1618         h->surplus_huge_pages++;
1619     }
1620     spin_unlock(&hugetlb_lock);
1621 
1622     page = __hugetlb_alloc_buddy_huge_page(h, vma, addr, nid);
1623 
1624     spin_lock(&hugetlb_lock);
1625     if (page) {
1626         INIT_LIST_HEAD(&page->lru);
1627         r_nid = page_to_nid(page);
1628         set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1629         set_hugetlb_cgroup(page, NULL);
1630         /*
1631          * We incremented the global counters already
1632          */
1633         h->nr_huge_pages_node[r_nid]++;
1634         h->surplus_huge_pages_node[r_nid]++;
1635         __count_vm_event(HTLB_BUDDY_PGALLOC);
1636     } else {
1637         h->nr_huge_pages--;
1638         h->surplus_huge_pages--;
1639         __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1640     }
1641     spin_unlock(&hugetlb_lock);
1642 
1643     return page;
1644 }
1645 
1646 /*
1647  * Allocate a huge page from 'nid'.  Note, 'nid' may be
1648  * NUMA_NO_NODE, which means that it may be allocated
1649  * anywhere.
1650  */
1651 static
1652 struct page *__alloc_buddy_huge_page_no_mpol(struct hstate *h, int nid)
1653 {
1654     unsigned long addr = -1;
1655 
1656     return __alloc_buddy_huge_page(h, NULL, addr, nid);
1657 }
1658 
1659 /*
1660  * Use the VMA's mpolicy to allocate a huge page from the buddy.
1661  */
1662 static
1663 struct page *__alloc_buddy_huge_page_with_mpol(struct hstate *h,
1664         struct vm_area_struct *vma, unsigned long addr)
1665 {
1666     return __alloc_buddy_huge_page(h, vma, addr, NUMA_NO_NODE);
1667 }
1668 
1669 /*
1670  * This allocation function is useful in the context where vma is irrelevant.
1671  * E.g. soft-offlining uses this function because it only cares physical
1672  * address of error page.
1673  */
1674 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1675 {
1676     struct page *page = NULL;
1677 
1678     spin_lock(&hugetlb_lock);
1679     if (h->free_huge_pages - h->resv_huge_pages > 0)
1680         page = dequeue_huge_page_node(h, nid);
1681     spin_unlock(&hugetlb_lock);
1682 
1683     if (!page)
1684         page = __alloc_buddy_huge_page_no_mpol(h, nid);
1685 
1686     return page;
1687 }
1688 
1689 /*
1690  * Increase the hugetlb pool such that it can accommodate a reservation
1691  * of size 'delta'.
1692  */
1693 static int gather_surplus_pages(struct hstate *h, int delta)
1694 {
1695     struct list_head surplus_list;
1696     struct page *page, *tmp;
1697     int ret, i;
1698     int needed, allocated;
1699     bool alloc_ok = true;
1700 
1701     needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1702     if (needed <= 0) {
1703         h->resv_huge_pages += delta;
1704         return 0;
1705     }
1706 
1707     allocated = 0;
1708     INIT_LIST_HEAD(&surplus_list);
1709 
1710     ret = -ENOMEM;
1711 retry:
1712     spin_unlock(&hugetlb_lock);
1713     for (i = 0; i < needed; i++) {
1714         page = __alloc_buddy_huge_page_no_mpol(h, NUMA_NO_NODE);
1715         if (!page) {
1716             alloc_ok = false;
1717             break;
1718         }
1719         list_add(&page->lru, &surplus_list);
1720     }
1721     allocated += i;
1722 
1723     /*
1724      * After retaking hugetlb_lock, we need to recalculate 'needed'
1725      * because either resv_huge_pages or free_huge_pages may have changed.
1726      */
1727     spin_lock(&hugetlb_lock);
1728     needed = (h->resv_huge_pages + delta) -
1729             (h->free_huge_pages + allocated);
1730     if (needed > 0) {
1731         if (alloc_ok)
1732             goto retry;
1733         /*
1734          * We were not able to allocate enough pages to
1735          * satisfy the entire reservation so we free what
1736          * we've allocated so far.
1737          */
1738         goto free;
1739     }
1740     /*
1741      * The surplus_list now contains _at_least_ the number of extra pages
1742      * needed to accommodate the reservation.  Add the appropriate number
1743      * of pages to the hugetlb pool and free the extras back to the buddy
1744      * allocator.  Commit the entire reservation here to prevent another
1745      * process from stealing the pages as they are added to the pool but
1746      * before they are reserved.
1747      */
1748     needed += allocated;
1749     h->resv_huge_pages += delta;
1750     ret = 0;
1751 
1752     /* Free the needed pages to the hugetlb pool */
1753     list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1754         if ((--needed) < 0)
1755             break;
1756         /*
1757          * This page is now managed by the hugetlb allocator and has
1758          * no users -- drop the buddy allocator's reference.
1759          */
1760         put_page_testzero(page);
1761         VM_BUG_ON_PAGE(page_count(page), page);
1762         enqueue_huge_page(h, page);
1763     }
1764 free:
1765     spin_unlock(&hugetlb_lock);
1766 
1767     /* Free unnecessary surplus pages to the buddy allocator */
1768     list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1769         put_page(page);
1770     spin_lock(&hugetlb_lock);
1771 
1772     return ret;
1773 }
1774 
1775 /*
1776  * This routine has two main purposes:
1777  * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1778  *    in unused_resv_pages.  This corresponds to the prior adjustments made
1779  *    to the associated reservation map.
1780  * 2) Free any unused surplus pages that may have been allocated to satisfy
1781  *    the reservation.  As many as unused_resv_pages may be freed.
1782  *
1783  * Called with hugetlb_lock held.  However, the lock could be dropped (and
1784  * reacquired) during calls to cond_resched_lock.  Whenever dropping the lock,
1785  * we must make sure nobody else can claim pages we are in the process of
1786  * freeing.  Do this by ensuring resv_huge_page always is greater than the
1787  * number of huge pages we plan to free when dropping the lock.
1788  */
1789 static void return_unused_surplus_pages(struct hstate *h,
1790                     unsigned long unused_resv_pages)
1791 {
1792     unsigned long nr_pages;
1793 
1794     /* Cannot return gigantic pages currently */
1795     if (hstate_is_gigantic(h))
1796         goto out;
1797 
1798     /*
1799      * Part (or even all) of the reservation could have been backed
1800      * by pre-allocated pages. Only free surplus pages.
1801      */
1802     nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1803 
1804     /*
1805      * We want to release as many surplus pages as possible, spread
1806      * evenly across all nodes with memory. Iterate across these nodes
1807      * until we can no longer free unreserved surplus pages. This occurs
1808      * when the nodes with surplus pages have no free pages.
1809      * free_pool_huge_page() will balance the the freed pages across the
1810      * on-line nodes with memory and will handle the hstate accounting.
1811      *
1812      * Note that we decrement resv_huge_pages as we free the pages.  If
1813      * we drop the lock, resv_huge_pages will still be sufficiently large
1814      * to cover subsequent pages we may free.
1815      */
1816     while (nr_pages--) {
1817         h->resv_huge_pages--;
1818         unused_resv_pages--;
1819         if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1820             goto out;
1821         cond_resched_lock(&hugetlb_lock);
1822     }
1823 
1824 out:
1825     /* Fully uncommit the reservation */
1826     h->resv_huge_pages -= unused_resv_pages;
1827 }
1828 
1829 
1830 /*
1831  * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1832  * are used by the huge page allocation routines to manage reservations.
1833  *
1834  * vma_needs_reservation is called to determine if the huge page at addr
1835  * within the vma has an associated reservation.  If a reservation is
1836  * needed, the value 1 is returned.  The caller is then responsible for
1837  * managing the global reservation and subpool usage counts.  After
1838  * the huge page has been allocated, vma_commit_reservation is called
1839  * to add the page to the reservation map.  If the page allocation fails,
1840  * the reservation must be ended instead of committed.  vma_end_reservation
1841  * is called in such cases.
1842  *
1843  * In the normal case, vma_commit_reservation returns the same value
1844  * as the preceding vma_needs_reservation call.  The only time this
1845  * is not the case is if a reserve map was changed between calls.  It
1846  * is the responsibility of the caller to notice the difference and
1847  * take appropriate action.
1848  *
1849  * vma_add_reservation is used in error paths where a reservation must
1850  * be restored when a newly allocated huge page must be freed.  It is
1851  * to be called after calling vma_needs_reservation to determine if a
1852  * reservation exists.
1853  */
1854 enum vma_resv_mode {
1855     VMA_NEEDS_RESV,
1856     VMA_COMMIT_RESV,
1857     VMA_END_RESV,
1858     VMA_ADD_RESV,
1859 };
1860 static long __vma_reservation_common(struct hstate *h,
1861                 struct vm_area_struct *vma, unsigned long addr,
1862                 enum vma_resv_mode mode)
1863 {
1864     struct resv_map *resv;
1865     pgoff_t idx;
1866     long ret;
1867 
1868     resv = vma_resv_map(vma);
1869     if (!resv)
1870         return 1;
1871 
1872     idx = vma_hugecache_offset(h, vma, addr);
1873     switch (mode) {
1874     case VMA_NEEDS_RESV:
1875         ret = region_chg(resv, idx, idx + 1);
1876         break;
1877     case VMA_COMMIT_RESV:
1878         ret = region_add(resv, idx, idx + 1);
1879         break;
1880     case VMA_END_RESV:
1881         region_abort(resv, idx, idx + 1);
1882         ret = 0;
1883         break;
1884     case VMA_ADD_RESV:
1885         if (vma->vm_flags & VM_MAYSHARE)
1886             ret = region_add(resv, idx, idx + 1);
1887         else {
1888             region_abort(resv, idx, idx + 1);
1889             ret = region_del(resv, idx, idx + 1);
1890         }
1891         break;
1892     default:
1893         BUG();
1894     }
1895 
1896     if (vma->vm_flags & VM_MAYSHARE)
1897         return ret;
1898     else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
1899         /*
1900          * In most cases, reserves always exist for private mappings.
1901          * However, a file associated with mapping could have been
1902          * hole punched or truncated after reserves were consumed.
1903          * As subsequent fault on such a range will not use reserves.
1904          * Subtle - The reserve map for private mappings has the
1905          * opposite meaning than that of shared mappings.  If NO
1906          * entry is in the reserve map, it means a reservation exists.
1907          * If an entry exists in the reserve map, it means the
1908          * reservation has already been consumed.  As a result, the
1909          * return value of this routine is the opposite of the
1910          * value returned from reserve map manipulation routines above.
1911          */
1912         if (ret)
1913             return 0;
1914         else
1915             return 1;
1916     }
1917     else
1918         return ret < 0 ? ret : 0;
1919 }
1920 
1921 static long vma_needs_reservation(struct hstate *h,
1922             struct vm_area_struct *vma, unsigned long addr)
1923 {
1924     return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
1925 }
1926 
1927 static long vma_commit_reservation(struct hstate *h,
1928             struct vm_area_struct *vma, unsigned long addr)
1929 {
1930     return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1931 }
1932 
1933 static void vma_end_reservation(struct hstate *h,
1934             struct vm_area_struct *vma, unsigned long addr)
1935 {
1936     (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
1937 }
1938 
1939 static long vma_add_reservation(struct hstate *h,
1940             struct vm_area_struct *vma, unsigned long addr)
1941 {
1942     return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
1943 }
1944 
1945 /*
1946  * This routine is called to restore a reservation on error paths.  In the
1947  * specific error paths, a huge page was allocated (via alloc_huge_page)
1948  * and is about to be freed.  If a reservation for the page existed,
1949  * alloc_huge_page would have consumed the reservation and set PagePrivate
1950  * in the newly allocated page.  When the page is freed via free_huge_page,
1951  * the global reservation count will be incremented if PagePrivate is set.
1952  * However, free_huge_page can not adjust the reserve map.  Adjust the
1953  * reserve map here to be consistent with global reserve count adjustments
1954  * to be made by free_huge_page.
1955  */
1956 static void restore_reserve_on_error(struct hstate *h,
1957             struct vm_area_struct *vma, unsigned long address,
1958             struct page *page)
1959 {
1960     if (unlikely(PagePrivate(page))) {
1961         long rc = vma_needs_reservation(h, vma, address);
1962 
1963         if (unlikely(rc < 0)) {
1964             /*
1965              * Rare out of memory condition in reserve map
1966              * manipulation.  Clear PagePrivate so that
1967              * global reserve count will not be incremented
1968              * by free_huge_page.  This will make it appear
1969              * as though the reservation for this page was
1970              * consumed.  This may prevent the task from
1971              * faulting in the page at a later time.  This
1972              * is better than inconsistent global huge page
1973              * accounting of reserve counts.
1974              */
1975             ClearPagePrivate(page);
1976         } else if (rc) {
1977             rc = vma_add_reservation(h, vma, address);
1978             if (unlikely(rc < 0))
1979                 /*
1980                  * See above comment about rare out of
1981                  * memory condition.
1982                  */
1983                 ClearPagePrivate(page);
1984         } else
1985             vma_end_reservation(h, vma, address);
1986     }
1987 }
1988 
1989 struct page *alloc_huge_page(struct vm_area_struct *vma,
1990                     unsigned long addr, int avoid_reserve)
1991 {
1992     struct hugepage_subpool *spool = subpool_vma(vma);
1993     struct hstate *h = hstate_vma(vma);
1994     struct page *page;
1995     long map_chg, map_commit;
1996     long gbl_chg;
1997     int ret, idx;
1998     struct hugetlb_cgroup *h_cg;
1999 
2000     idx = hstate_index(h);
2001     /*
2002      * Examine the region/reserve map to determine if the process
2003      * has a reservation for the page to be allocated.  A return
2004      * code of zero indicates a reservation exists (no change).
2005      */
2006     map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2007     if (map_chg < 0)
2008         return ERR_PTR(-ENOMEM);
2009 
2010     /*
2011      * Processes that did not create the mapping will have no
2012      * reserves as indicated by the region/reserve map. Check
2013      * that the allocation will not exceed the subpool limit.
2014      * Allocations for MAP_NORESERVE mappings also need to be
2015      * checked against any subpool limit.
2016      */
2017     if (map_chg || avoid_reserve) {
2018         gbl_chg = hugepage_subpool_get_pages(spool, 1);
2019         if (gbl_chg < 0) {
2020             vma_end_reservation(h, vma, addr);
2021             return ERR_PTR(-ENOSPC);
2022         }
2023 
2024         /*
2025          * Even though there was no reservation in the region/reserve
2026          * map, there could be reservations associated with the
2027          * subpool that can be used.  This would be indicated if the
2028          * return value of hugepage_subpool_get_pages() is zero.
2029          * However, if avoid_reserve is specified we still avoid even
2030          * the subpool reservations.
2031          */
2032         if (avoid_reserve)
2033             gbl_chg = 1;
2034     }
2035 
2036     ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2037     if (ret)
2038         goto out_subpool_put;
2039 
2040     spin_lock(&hugetlb_lock);
2041     /*
2042      * glb_chg is passed to indicate whether or not a page must be taken
2043      * from the global free pool (global change).  gbl_chg == 0 indicates
2044      * a reservation exists for the allocation.
2045      */
2046     page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2047     if (!page) {
2048         spin_unlock(&hugetlb_lock);
2049         page = __alloc_buddy_huge_page_with_mpol(h, vma, addr);
2050         if (!page)
2051             goto out_uncharge_cgroup;
2052         if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2053             SetPagePrivate(page);
2054             h->resv_huge_pages--;
2055         }
2056         spin_lock(&hugetlb_lock);
2057         list_move(&page->lru, &h->hugepage_activelist);
2058         /* Fall through */
2059     }
2060     hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2061     spin_unlock(&hugetlb_lock);
2062 
2063     set_page_private(page, (unsigned long)spool);
2064 
2065     map_commit = vma_commit_reservation(h, vma, addr);
2066     if (unlikely(map_chg > map_commit)) {
2067         /*
2068          * The page was added to the reservation map between
2069          * vma_needs_reservation and vma_commit_reservation.
2070          * This indicates a race with hugetlb_reserve_pages.
2071          * Adjust for the subpool count incremented above AND
2072          * in hugetlb_reserve_pages for the same page.  Also,
2073          * the reservation count added in hugetlb_reserve_pages
2074          * no longer applies.
2075          */
2076         long rsv_adjust;
2077 
2078         rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2079         hugetlb_acct_memory(h, -rsv_adjust);
2080     }
2081     return page;
2082 
2083 out_uncharge_cgroup:
2084     hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2085 out_subpool_put:
2086     if (map_chg || avoid_reserve)
2087         hugepage_subpool_put_pages(spool, 1);
2088     vma_end_reservation(h, vma, addr);
2089     return ERR_PTR(-ENOSPC);
2090 }
2091 
2092 /*
2093  * alloc_huge_page()'s wrapper which simply returns the page if allocation
2094  * succeeds, otherwise NULL. This function is called from new_vma_page(),
2095  * where no ERR_VALUE is expected to be returned.
2096  */
2097 struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
2098                 unsigned long addr, int avoid_reserve)
2099 {
2100     struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
2101     if (IS_ERR(page))
2102         page = NULL;
2103     return page;
2104 }
2105 
2106 int __weak alloc_bootmem_huge_page(struct hstate *h)
2107 {
2108     struct huge_bootmem_page *m;
2109     int nr_nodes, node;
2110 
2111     for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2112         void *addr;
2113 
2114         addr = memblock_virt_alloc_try_nid_nopanic(
2115                 huge_page_size(h), huge_page_size(h),
2116                 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
2117         if (addr) {
2118             /*
2119              * Use the beginning of the huge page to store the
2120              * huge_bootmem_page struct (until gather_bootmem
2121              * puts them into the mem_map).
2122              */
2123             m = addr;
2124             goto found;
2125         }
2126     }
2127     return 0;
2128 
2129 found:
2130     BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2131     /* Put them into a private list first because mem_map is not up yet */
2132     list_add(&m->list, &huge_boot_pages);
2133     m->hstate = h;
2134     return 1;
2135 }
2136 
2137 static void __init prep_compound_huge_page(struct page *page,
2138         unsigned int order)
2139 {
2140     if (unlikely(order > (MAX_ORDER - 1)))
2141         prep_compound_gigantic_page(page, order);
2142     else
2143         prep_compound_page(page, order);
2144 }
2145 
2146 /* Put bootmem huge pages into the standard lists after mem_map is up */
2147 static void __init gather_bootmem_prealloc(void)
2148 {
2149     struct huge_bootmem_page *m;
2150 
2151     list_for_each_entry(m, &huge_boot_pages, list) {
2152         struct hstate *h = m->hstate;
2153         struct page *page;
2154 
2155 #ifdef CONFIG_HIGHMEM
2156         page = pfn_to_page(m->phys >> PAGE_SHIFT);
2157         memblock_free_late(__pa(m),
2158                    sizeof(struct huge_bootmem_page));
2159 #else
2160         page = virt_to_page(m);
2161 #endif
2162         WARN_ON(page_count(page) != 1);
2163         prep_compound_huge_page(page, h->order);
2164         WARN_ON(PageReserved(page));
2165         prep_new_huge_page(h, page, page_to_nid(page));
2166         /*
2167          * If we had gigantic hugepages allocated at boot time, we need
2168          * to restore the 'stolen' pages to totalram_pages in order to
2169          * fix confusing memory reports from free(1) and another
2170          * side-effects, like CommitLimit going negative.
2171          */
2172         if (hstate_is_gigantic(h))
2173             adjust_managed_page_count(page, 1 << h->order);
2174     }
2175 }
2176 
2177 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2178 {
2179     unsigned long i;
2180 
2181     for (i = 0; i < h->max_huge_pages; ++i) {
2182         if (hstate_is_gigantic(h)) {
2183             if (!alloc_bootmem_huge_page(h))
2184                 break;
2185         } else if (!alloc_fresh_huge_page(h,
2186                      &node_states[N_MEMORY]))
2187             break;
2188     }
2189     h->max_huge_pages = i;
2190 }
2191 
2192 static void __init hugetlb_init_hstates(void)
2193 {
2194     struct hstate *h;
2195 
2196     for_each_hstate(h) {
2197         if (minimum_order > huge_page_order(h))
2198             minimum_order = huge_page_order(h);
2199 
2200         /* oversize hugepages were init'ed in early boot */
2201         if (!hstate_is_gigantic(h))
2202             hugetlb_hstate_alloc_pages(h);
2203     }
2204     VM_BUG_ON(minimum_order == UINT_MAX);
2205 }
2206 
2207 static char * __init memfmt(char *buf, unsigned long n)
2208 {
2209     if (n >= (1UL << 30))
2210         sprintf(buf, "%lu GB", n >> 30);
2211     else if (n >= (1UL << 20))
2212         sprintf(buf, "%lu MB", n >> 20);
2213     else
2214         sprintf(buf, "%lu KB", n >> 10);
2215     return buf;
2216 }
2217 
2218 static void __init report_hugepages(void)
2219 {
2220     struct hstate *h;
2221 
2222     for_each_hstate(h) {
2223         char buf[32];
2224         pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2225             memfmt(buf, huge_page_size(h)),
2226             h->free_huge_pages);
2227     }
2228 }
2229 
2230 #ifdef CONFIG_HIGHMEM
2231 static void try_to_free_low(struct hstate *h, unsigned long count,
2232                         nodemask_t *nodes_allowed)
2233 {
2234     int i;
2235 
2236     if (hstate_is_gigantic(h))
2237         return;
2238 
2239     for_each_node_mask(i, *nodes_allowed) {
2240         struct page *page, *next;
2241         struct list_head *freel = &h->hugepage_freelists[i];
2242         list_for_each_entry_safe(page, next, freel, lru) {
2243             if (count >= h->nr_huge_pages)
2244                 return;
2245             if (PageHighMem(page))
2246                 continue;
2247             list_del(&page->lru);
2248             update_and_free_page(h, page);
2249             h->free_huge_pages--;
2250             h->free_huge_pages_node[page_to_nid(page)]--;
2251         }
2252     }
2253 }
2254 #else
2255 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2256                         nodemask_t *nodes_allowed)
2257 {
2258 }
2259 #endif
2260 
2261 /*
2262  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
2263  * balanced by operating on them in a round-robin fashion.
2264  * Returns 1 if an adjustment was made.
2265  */
2266 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2267                 int delta)
2268 {
2269     int nr_nodes, node;
2270 
2271     VM_BUG_ON(delta != -1 && delta != 1);
2272 
2273     if (delta < 0) {
2274         for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2275             if (h->surplus_huge_pages_node[node])
2276                 goto found;
2277         }
2278     } else {
2279         for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2280             if (h->surplus_huge_pages_node[node] <
2281                     h->nr_huge_pages_node[node])
2282                 goto found;
2283         }
2284     }
2285     return 0;
2286 
2287 found:
2288     h->surplus_huge_pages += delta;
2289     h->surplus_huge_pages_node[node] += delta;
2290     return 1;
2291 }
2292 
2293 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2294 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
2295                         nodemask_t *nodes_allowed)
2296 {
2297     unsigned long min_count, ret;
2298 
2299     if (hstate_is_gigantic(h) && !gigantic_page_supported())
2300         return h->max_huge_pages;
2301 
2302     /*
2303      * Increase the pool size
2304      * First take pages out of surplus state.  Then make up the
2305      * remaining difference by allocating fresh huge pages.
2306      *
2307      * We might race with __alloc_buddy_huge_page() here and be unable
2308      * to convert a surplus huge page to a normal huge page. That is
2309      * not critical, though, it just means the overall size of the
2310      * pool might be one hugepage larger than it needs to be, but
2311      * within all the constraints specified by the sysctls.
2312      */
2313     spin_lock(&hugetlb_lock);
2314     while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2315         if (!adjust_pool_surplus(h, nodes_allowed, -1))
2316             break;
2317     }
2318 
2319     while (count > persistent_huge_pages(h)) {
2320         /*
2321          * If this allocation races such that we no longer need the
2322          * page, free_huge_page will handle it by freeing the page
2323          * and reducing the surplus.
2324          */
2325         spin_unlock(&hugetlb_lock);
2326 
2327         /* yield cpu to avoid soft lockup */
2328         cond_resched();
2329 
2330         if (hstate_is_gigantic(h))
2331             ret = alloc_fresh_gigantic_page(h, nodes_allowed);
2332         else
2333             ret = alloc_fresh_huge_page(h, nodes_allowed);
2334         spin_lock(&hugetlb_lock);
2335         if (!ret)
2336             goto out;
2337 
2338         /* Bail for signals. Probably ctrl-c from user */
2339         if (signal_pending(current))
2340             goto out;
2341     }
2342 
2343     /*
2344      * Decrease the pool size
2345      * First return free pages to the buddy allocator (being careful
2346      * to keep enough around to satisfy reservations).  Then place
2347      * pages into surplus state as needed so the pool will shrink
2348      * to the desired size as pages become free.
2349      *
2350      * By placing pages into the surplus state independent of the
2351      * overcommit value, we are allowing the surplus pool size to
2352      * exceed overcommit. There are few sane options here. Since
2353      * __alloc_buddy_huge_page() is checking the global counter,
2354      * though, we'll note that we're not allowed to exceed surplus
2355      * and won't grow the pool anywhere else. Not until one of the
2356      * sysctls are changed, or the surplus pages go out of use.
2357      */
2358     min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2359     min_count = max(count, min_count);
2360     try_to_free_low(h, min_count, nodes_allowed);
2361     while (min_count < persistent_huge_pages(h)) {
2362         if (!free_pool_huge_page(h, nodes_allowed, 0))
2363             break;
2364         cond_resched_lock(&hugetlb_lock);
2365     }
2366     while (count < persistent_huge_pages(h)) {
2367         if (!adjust_pool_surplus(h, nodes_allowed, 1))
2368             break;
2369     }
2370 out:
2371     ret = persistent_huge_pages(h);
2372     spin_unlock(&hugetlb_lock);
2373     return ret;
2374 }
2375 
2376 #define HSTATE_ATTR_RO(_name) \
2377     static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2378 
2379 #define HSTATE_ATTR(_name) \
2380     static struct kobj_attribute _name##_attr = \
2381         __ATTR(_name, 0644, _name##_show, _name##_store)
2382 
2383 static struct kobject *hugepages_kobj;
2384 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2385 
2386 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2387 
2388 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2389 {
2390     int i;
2391 
2392     for (i = 0; i < HUGE_MAX_HSTATE; i++)
2393         if (hstate_kobjs[i] == kobj) {
2394             if (nidp)
2395                 *nidp = NUMA_NO_NODE;
2396             return &hstates[i];
2397         }
2398 
2399     return kobj_to_node_hstate(kobj, nidp);
2400 }
2401 
2402 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2403                     struct kobj_attribute *attr, char *buf)
2404 {
2405     struct hstate *h;
2406     unsigned long nr_huge_pages;
2407     int nid;
2408 
2409     h = kobj_to_hstate(kobj, &nid);
2410     if (nid == NUMA_NO_NODE)
2411         nr_huge_pages = h->nr_huge_pages;
2412     else
2413         nr_huge_pages = h->nr_huge_pages_node[nid];
2414 
2415     return sprintf(buf, "%lu\n", nr_huge_pages);
2416 }
2417 
2418 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2419                        struct hstate *h, int nid,
2420                        unsigned long count, size_t len)
2421 {
2422     int err;
2423     NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
2424 
2425     if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
2426         err = -EINVAL;
2427         goto out;
2428     }
2429 
2430     if (nid == NUMA_NO_NODE) {
2431         /*
2432          * global hstate attribute
2433          */
2434         if (!(obey_mempolicy &&
2435                 init_nodemask_of_mempolicy(nodes_allowed))) {
2436             NODEMASK_FREE(nodes_allowed);
2437             nodes_allowed = &node_states[N_MEMORY];
2438         }
2439     } else if (nodes_allowed) {
2440         /*
2441          * per node hstate attribute: adjust count to global,
2442          * but restrict alloc/free to the specified node.
2443          */
2444         count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2445         init_nodemask_of_node(nodes_allowed, nid);
2446     } else
2447         nodes_allowed = &node_states[N_MEMORY];
2448 
2449     h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
2450 
2451     if (nodes_allowed != &node_states[N_MEMORY])
2452         NODEMASK_FREE(nodes_allowed);
2453 
2454     return len;
2455 out:
2456     NODEMASK_FREE(nodes_allowed);
2457     return err;
2458 }
2459 
2460 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2461                      struct kobject *kobj, const char *buf,
2462                      size_t len)
2463 {
2464     struct hstate *h;
2465     unsigned long count;
2466     int nid;
2467     int err;
2468 
2469     err = kstrtoul(buf, 10, &count);
2470     if (err)
2471         return err;
2472 
2473     h = kobj_to_hstate(kobj, &nid);
2474     return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2475 }
2476 
2477 static ssize_t nr_hugepages_show(struct kobject *kobj,
2478                        struct kobj_attribute *attr, char *buf)
2479 {
2480     return nr_hugepages_show_common(kobj, attr, buf);
2481 }
2482 
2483 static ssize_t nr_hugepages_store(struct kobject *kobj,
2484            struct kobj_attribute *attr, const char *buf, size_t len)
2485 {
2486     return nr_hugepages_store_common(false, kobj, buf, len);
2487 }
2488 HSTATE_ATTR(nr_hugepages);
2489 
2490 #ifdef CONFIG_NUMA
2491 
2492 /*
2493  * hstate attribute for optionally mempolicy-based constraint on persistent
2494  * huge page alloc/free.
2495  */
2496 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2497                        struct kobj_attribute *attr, char *buf)
2498 {
2499     return nr_hugepages_show_common(kobj, attr, buf);
2500 }
2501 
2502 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2503            struct kobj_attribute *attr, const char *buf, size_t len)
2504 {
2505     return nr_hugepages_store_common(true, kobj, buf, len);
2506 }
2507 HSTATE_ATTR(nr_hugepages_mempolicy);
2508 #endif
2509 
2510 
2511 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2512                     struct kobj_attribute *attr, char *buf)
2513 {
2514     struct hstate *h = kobj_to_hstate(kobj, NULL);
2515     return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2516 }
2517 
2518 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2519         struct kobj_attribute *attr, const char *buf, size_t count)
2520 {
2521     int err;
2522     unsigned long input;
2523     struct hstate *h = kobj_to_hstate(kobj, NULL);
2524 
2525     if (hstate_is_gigantic(h))
2526         return -EINVAL;
2527 
2528     err = kstrtoul(buf, 10, &input);
2529     if (err)
2530         return err;
2531 
2532     spin_lock(&hugetlb_lock);
2533     h->nr_overcommit_huge_pages = input;
2534     spin_unlock(&hugetlb_lock);
2535 
2536     return count;
2537 }
2538 HSTATE_ATTR(nr_overcommit_hugepages);
2539 
2540 static ssize_t free_hugepages_show(struct kobject *kobj,
2541                     struct kobj_attribute *attr, char *buf)
2542 {
2543     struct hstate *h;
2544     unsigned long free_huge_pages;
2545     int nid;
2546 
2547     h = kobj_to_hstate(kobj, &nid);
2548     if (nid == NUMA_NO_NODE)
2549         free_huge_pages = h->free_huge_pages;
2550     else
2551         free_huge_pages = h->free_huge_pages_node[nid];
2552 
2553     return sprintf(buf, "%lu\n", free_huge_pages);
2554 }
2555 HSTATE_ATTR_RO(free_hugepages);
2556 
2557 static ssize_t resv_hugepages_show(struct kobject *kobj,
2558                     struct kobj_attribute *attr, char *buf)
2559 {
2560     struct hstate *h = kobj_to_hstate(kobj, NULL);
2561     return sprintf(buf, "%lu\n", h->resv_huge_pages);
2562 }
2563 HSTATE_ATTR_RO(resv_hugepages);
2564 
2565 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2566                     struct kobj_attribute *attr, char *buf)
2567 {
2568     struct hstate *h;
2569     unsigned long surplus_huge_pages;
2570     int nid;
2571 
2572     h = kobj_to_hstate(kobj, &nid);
2573     if (nid == NUMA_NO_NODE)
2574         surplus_huge_pages = h->surplus_huge_pages;
2575     else
2576         surplus_huge_pages = h->surplus_huge_pages_node[nid];
2577 
2578     return sprintf(buf, "%lu\n", surplus_huge_pages);
2579 }
2580 HSTATE_ATTR_RO(surplus_hugepages);
2581 
2582 static struct attribute *hstate_attrs[] = {
2583     &nr_hugepages_attr.attr,
2584     &nr_overcommit_hugepages_attr.attr,
2585     &free_hugepages_attr.attr,
2586     &resv_hugepages_attr.attr,
2587     &surplus_hugepages_attr.attr,
2588 #ifdef CONFIG_NUMA
2589     &nr_hugepages_mempolicy_attr.attr,
2590 #endif
2591     NULL,
2592 };
2593 
2594 static struct attribute_group hstate_attr_group = {
2595     .attrs = hstate_attrs,
2596 };
2597 
2598 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2599                     struct kobject **hstate_kobjs,
2600                     struct attribute_group *hstate_attr_group)
2601 {
2602     int retval;
2603     int hi = hstate_index(h);
2604 
2605     hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2606     if (!hstate_kobjs[hi])
2607         return -ENOMEM;
2608 
2609     retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2610     if (retval)
2611         kobject_put(hstate_kobjs[hi]);
2612 
2613     return retval;
2614 }
2615 
2616 static void __init hugetlb_sysfs_init(void)
2617 {
2618     struct hstate *h;
2619     int err;
2620 
2621     hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2622     if (!hugepages_kobj)
2623         return;
2624 
2625     for_each_hstate(h) {
2626         err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2627                      hstate_kobjs, &hstate_attr_group);
2628         if (err)
2629             pr_err("Hugetlb: Unable to add hstate %s", h->name);
2630     }
2631 }
2632 
2633 #ifdef CONFIG_NUMA
2634 
2635 /*
2636  * node_hstate/s - associate per node hstate attributes, via their kobjects,
2637  * with node devices in node_devices[] using a parallel array.  The array
2638  * index of a node device or _hstate == node id.
2639  * This is here to avoid any static dependency of the node device driver, in
2640  * the base kernel, on the hugetlb module.
2641  */
2642 struct node_hstate {
2643     struct kobject      *hugepages_kobj;
2644     struct kobject      *hstate_kobjs[HUGE_MAX_HSTATE];
2645 };
2646 static struct node_hstate node_hstates[MAX_NUMNODES];
2647 
2648 /*
2649  * A subset of global hstate attributes for node devices
2650  */
2651 static struct attribute *per_node_hstate_attrs[] = {
2652     &nr_hugepages_attr.attr,
2653     &free_hugepages_attr.attr,
2654     &surplus_hugepages_attr.attr,
2655     NULL,
2656 };
2657 
2658 static struct attribute_group per_node_hstate_attr_group = {
2659     .attrs = per_node_hstate_attrs,
2660 };
2661 
2662 /*
2663  * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2664  * Returns node id via non-NULL nidp.
2665  */
2666 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2667 {
2668     int nid;
2669 
2670     for (nid = 0; nid < nr_node_ids; nid++) {
2671         struct node_hstate *nhs = &node_hstates[nid];
2672         int i;
2673         for (i = 0; i < HUGE_MAX_HSTATE; i++)
2674             if (nhs->hstate_kobjs[i] == kobj) {
2675                 if (nidp)
2676                     *nidp = nid;
2677                 return &hstates[i];
2678             }
2679     }
2680 
2681     BUG();
2682     return NULL;
2683 }
2684 
2685 /*
2686  * Unregister hstate attributes from a single node device.
2687  * No-op if no hstate attributes attached.
2688  */
2689 static void hugetlb_unregister_node(struct node *node)
2690 {
2691     struct hstate *h;
2692     struct node_hstate *nhs = &node_hstates[node->dev.id];
2693 
2694     if (!nhs->hugepages_kobj)
2695         return;     /* no hstate attributes */
2696 
2697     for_each_hstate(h) {
2698         int idx = hstate_index(h);
2699         if (nhs->hstate_kobjs[idx]) {
2700             kobject_put(nhs->hstate_kobjs[idx]);
2701             nhs->hstate_kobjs[idx] = NULL;
2702         }
2703     }
2704 
2705     kobject_put(nhs->hugepages_kobj);
2706     nhs->hugepages_kobj = NULL;
2707 }
2708 
2709 
2710 /*
2711  * Register hstate attributes for a single node device.
2712  * No-op if attributes already registered.
2713  */
2714 static void hugetlb_register_node(struct node *node)
2715 {
2716     struct hstate *h;
2717     struct node_hstate *nhs = &node_hstates[node->dev.id];
2718     int err;
2719 
2720     if (nhs->hugepages_kobj)
2721         return;     /* already allocated */
2722 
2723     nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2724                             &node->dev.kobj);
2725     if (!nhs->hugepages_kobj)
2726         return;
2727 
2728     for_each_hstate(h) {
2729         err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2730                         nhs->hstate_kobjs,
2731                         &per_node_hstate_attr_group);
2732         if (err) {
2733             pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2734                 h->name, node->dev.id);
2735             hugetlb_unregister_node(node);
2736             break;
2737         }
2738     }
2739 }
2740 
2741 /*
2742  * hugetlb init time:  register hstate attributes for all registered node
2743  * devices of nodes that have memory.  All on-line nodes should have
2744  * registered their associated device by this time.
2745  */
2746 static void __init hugetlb_register_all_nodes(void)
2747 {
2748     int nid;
2749 
2750     for_each_node_state(nid, N_MEMORY) {
2751         struct node *node = node_devices[nid];
2752         if (node->dev.id == nid)
2753             hugetlb_register_node(node);
2754     }
2755 
2756     /*
2757      * Let the node device driver know we're here so it can
2758      * [un]register hstate attributes on node hotplug.
2759      */
2760     register_hugetlbfs_with_node(hugetlb_register_node,
2761                      hugetlb_unregister_node);
2762 }
2763 #else   /* !CONFIG_NUMA */
2764 
2765 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2766 {
2767     BUG();
2768     if (nidp)
2769         *nidp = -1;
2770     return NULL;
2771 }
2772 
2773 static void hugetlb_register_all_nodes(void) { }
2774 
2775 #endif
2776 
2777 static int __init hugetlb_init(void)
2778 {
2779     int i;
2780 
2781     if (!hugepages_supported())
2782         return 0;
2783 
2784     if (!size_to_hstate(default_hstate_size)) {
2785         default_hstate_size = HPAGE_SIZE;
2786         if (!size_to_hstate(default_hstate_size))
2787             hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2788     }
2789     default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2790     if (default_hstate_max_huge_pages) {
2791         if (!default_hstate.max_huge_pages)
2792             default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2793     }
2794 
2795     hugetlb_init_hstates();
2796     gather_bootmem_prealloc();
2797     report_hugepages();
2798 
2799     hugetlb_sysfs_init();
2800     hugetlb_register_all_nodes();
2801     hugetlb_cgroup_file_init();
2802 
2803 #ifdef CONFIG_SMP
2804     num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2805 #else
2806     num_fault_mutexes = 1;
2807 #endif
2808     hugetlb_fault_mutex_table =
2809         kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
2810     BUG_ON(!hugetlb_fault_mutex_table);
2811 
2812     for (i = 0; i < num_fault_mutexes; i++)
2813         mutex_init(&hugetlb_fault_mutex_table[i]);
2814     return 0;
2815 }
2816 subsys_initcall(hugetlb_init);
2817 
2818 /* Should be called on processing a hugepagesz=... option */
2819 void __init hugetlb_bad_size(void)
2820 {
2821     parsed_valid_hugepagesz = false;
2822 }
2823 
2824 void __init hugetlb_add_hstate(unsigned int order)
2825 {
2826     struct hstate *h;
2827     unsigned long i;
2828 
2829     if (size_to_hstate(PAGE_SIZE << order)) {
2830         pr_warn("hugepagesz= specified twice, ignoring\n");
2831         return;
2832     }
2833     BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2834     BUG_ON(order == 0);
2835     h = &hstates[hugetlb_max_hstate++];
2836     h->order = order;
2837     h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2838     h->nr_huge_pages = 0;
2839     h->free_huge_pages = 0;
2840     for (i = 0; i < MAX_NUMNODES; ++i)
2841         INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2842     INIT_LIST_HEAD(&h->hugepage_activelist);
2843     h->next_nid_to_alloc = first_memory_node;
2844     h->next_nid_to_free = first_memory_node;
2845     snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2846                     huge_page_size(h)/1024);
2847 
2848     parsed_hstate = h;
2849 }
2850 
2851 static int __init hugetlb_nrpages_setup(char *s)
2852 {
2853     unsigned long *mhp;
2854     static unsigned long *last_mhp;
2855 
2856     if (!parsed_valid_hugepagesz) {
2857         pr_warn("hugepages = %s preceded by "
2858             "an unsupported hugepagesz, ignoring\n", s);
2859         parsed_valid_hugepagesz = true;
2860         return 1;
2861     }
2862     /*
2863      * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2864      * so this hugepages= parameter goes to the "default hstate".
2865      */
2866     else if (!hugetlb_max_hstate)
2867         mhp = &default_hstate_max_huge_pages;
2868     else
2869         mhp = &parsed_hstate->max_huge_pages;
2870 
2871     if (mhp == last_mhp) {
2872         pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
2873         return 1;
2874     }
2875 
2876     if (sscanf(s, "%lu", mhp) <= 0)
2877         *mhp = 0;
2878 
2879     /*
2880      * Global state is always initialized later in hugetlb_init.
2881      * But we need to allocate >= MAX_ORDER hstates here early to still
2882      * use the bootmem allocator.
2883      */
2884     if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2885         hugetlb_hstate_alloc_pages(parsed_hstate);
2886 
2887     last_mhp = mhp;
2888 
2889     return 1;
2890 }
2891 __setup("hugepages=", hugetlb_nrpages_setup);
2892 
2893 static int __init hugetlb_default_setup(char *s)
2894 {
2895     default_hstate_size = memparse(s, &s);
2896     return 1;
2897 }
2898 __setup("default_hugepagesz=", hugetlb_default_setup);
2899 
2900 static unsigned int cpuset_mems_nr(unsigned int *array)
2901 {
2902     int node;
2903     unsigned int nr = 0;
2904 
2905     for_each_node_mask(node, cpuset_current_mems_allowed)
2906         nr += array[node];
2907 
2908     return nr;
2909 }
2910 
2911 #ifdef CONFIG_SYSCTL
2912 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2913              struct ctl_table *table, int write,
2914              void __user *buffer, size_t *length, loff_t *ppos)
2915 {
2916     struct hstate *h = &default_hstate;
2917     unsigned long tmp = h->max_huge_pages;
2918     int ret;
2919 
2920     if (!hugepages_supported())
2921         return -EOPNOTSUPP;
2922 
2923     table->data = &tmp;
2924     table->maxlen = sizeof(unsigned long);
2925     ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2926     if (ret)
2927         goto out;
2928 
2929     if (write)
2930         ret = __nr_hugepages_store_common(obey_mempolicy, h,
2931                           NUMA_NO_NODE, tmp, *length);
2932 out:
2933     return ret;
2934 }
2935 
2936 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2937               void __user *buffer, size_t *length, loff_t *ppos)
2938 {
2939 
2940     return hugetlb_sysctl_handler_common(false, table, write,
2941                             buffer, length, ppos);
2942 }
2943 
2944 #ifdef CONFIG_NUMA
2945 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2946               void __user *buffer, size_t *length, loff_t *ppos)
2947 {
2948     return hugetlb_sysctl_handler_common(true, table, write,
2949                             buffer, length, ppos);
2950 }
2951 #endif /* CONFIG_NUMA */
2952 
2953 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2954             void __user *buffer,
2955             size_t *length, loff_t *ppos)
2956 {
2957     struct hstate *h = &default_hstate;
2958     unsigned long tmp;
2959     int ret;
2960 
2961     if (!hugepages_supported())
2962         return -EOPNOTSUPP;
2963 
2964     tmp = h->nr_overcommit_huge_pages;
2965 
2966     if (write && hstate_is_gigantic(h))
2967         return -EINVAL;
2968 
2969     table->data = &tmp;
2970     table->maxlen = sizeof(unsigned long);
2971     ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2972     if (ret)
2973         goto out;
2974 
2975     if (write) {
2976         spin_lock(&hugetlb_lock);
2977         h->nr_overcommit_huge_pages = tmp;
2978         spin_unlock(&hugetlb_lock);
2979     }
2980 out:
2981     return ret;
2982 }
2983 
2984 #endif /* CONFIG_SYSCTL */
2985 
2986 void hugetlb_report_meminfo(struct seq_file *m)
2987 {
2988     struct hstate *h = &default_hstate;
2989     if (!hugepages_supported())
2990         return;
2991     seq_printf(m,
2992             "HugePages_Total:   %5lu\n"
2993             "HugePages_Free:    %5lu\n"
2994             "HugePages_Rsvd:    %5lu\n"
2995             "HugePages_Surp:    %5lu\n"
2996             "Hugepagesize:   %8lu kB\n",
2997             h->nr_huge_pages,
2998             h->free_huge_pages,
2999             h->resv_huge_pages,
3000             h->surplus_huge_pages,
3001             1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3002 }
3003 
3004 int hugetlb_report_node_meminfo(int nid, char *buf)
3005 {
3006     struct hstate *h = &default_hstate;
3007     if (!hugepages_supported())
3008         return 0;
3009     return sprintf(buf,
3010         "Node %d HugePages_Total: %5u\n"
3011         "Node %d HugePages_Free:  %5u\n"
3012         "Node %d HugePages_Surp:  %5u\n",
3013         nid, h->nr_huge_pages_node[nid],
3014         nid, h->free_huge_pages_node[nid],
3015         nid, h->surplus_huge_pages_node[nid]);
3016 }
3017 
3018 void hugetlb_show_meminfo(void)
3019 {
3020     struct hstate *h;
3021     int nid;
3022 
3023     if (!hugepages_supported())
3024         return;
3025 
3026     for_each_node_state(nid, N_MEMORY)
3027         for_each_hstate(h)
3028             pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3029                 nid,
3030                 h->nr_huge_pages_node[nid],
3031                 h->free_huge_pages_node[nid],
3032                 h->surplus_huge_pages_node[nid],
3033                 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3034 }
3035 
3036 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3037 {
3038     seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3039            atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3040 }
3041 
3042 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3043 unsigned long hugetlb_total_pages(void)
3044 {
3045     struct hstate *h;
3046     unsigned long nr_total_pages = 0;
3047 
3048     for_each_hstate(h)
3049         nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3050     return nr_total_pages;
3051 }
3052 
3053 static int hugetlb_acct_memory(struct hstate *h, long delta)
3054 {
3055     int ret = -ENOMEM;
3056 
3057     spin_lock(&hugetlb_lock);
3058     /*
3059      * When cpuset is configured, it breaks the strict hugetlb page
3060      * reservation as the accounting is done on a global variable. Such
3061      * reservation is completely rubbish in the presence of cpuset because
3062      * the reservation is not checked against page availability for the
3063      * current cpuset. Application can still potentially OOM'ed by kernel
3064      * with lack of free htlb page in cpuset that the task is in.
3065      * Attempt to enforce strict accounting with cpuset is almost
3066      * impossible (or too ugly) because cpuset is too fluid that
3067      * task or memory node can be dynamically moved between cpusets.
3068      *
3069      * The change of semantics for shared hugetlb mapping with cpuset is
3070      * undesirable. However, in order to preserve some of the semantics,
3071      * we fall back to check against current free page availability as
3072      * a best attempt and hopefully to minimize the impact of changing
3073      * semantics that cpuset has.
3074      */
3075     if (delta > 0) {
3076         if (gather_surplus_pages(h, delta) < 0)
3077             goto out;
3078 
3079         if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
3080             return_unused_surplus_pages(h, delta);
3081             goto out;
3082         }
3083     }
3084 
3085     ret = 0;
3086     if (delta < 0)
3087         return_unused_surplus_pages(h, (unsigned long) -delta);
3088 
3089 out:
3090     spin_unlock(&hugetlb_lock);
3091     return ret;
3092 }
3093 
3094 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3095 {
3096     struct resv_map *resv = vma_resv_map(vma);
3097 
3098     /*
3099      * This new VMA should share its siblings reservation map if present.
3100      * The VMA will only ever have a valid reservation map pointer where
3101      * it is being copied for another still existing VMA.  As that VMA
3102      * has a reference to the reservation map it cannot disappear until
3103      * after this open call completes.  It is therefore safe to take a
3104      * new reference here without additional locking.
3105      */
3106     if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3107         kref_get(&resv->refs);
3108 }
3109 
3110 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3111 {
3112     struct hstate *h = hstate_vma(vma);
3113     struct resv_map *resv = vma_resv_map(vma);
3114     struct hugepage_subpool *spool = subpool_vma(vma);
3115     unsigned long reserve, start, end;
3116     long gbl_reserve;
3117 
3118     if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3119         return;
3120 
3121     start = vma_hugecache_offset(h, vma, vma->vm_start);
3122     end = vma_hugecache_offset(h, vma, vma->vm_end);
3123 
3124     reserve = (end - start) - region_count(resv, start, end);
3125 
3126     kref_put(&resv->refs, resv_map_release);
3127 
3128     if (reserve) {
3129         /*
3130          * Decrement reserve counts.  The global reserve count may be
3131          * adjusted if the subpool has a minimum size.
3132          */
3133         gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3134         hugetlb_acct_memory(h, -gbl_reserve);
3135     }
3136 }
3137 
3138 /*
3139  * We cannot handle pagefaults against hugetlb pages at all.  They cause
3140  * handle_mm_fault() to try to instantiate regular-sized pages in the
3141  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
3142  * this far.
3143  */
3144 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3145 {
3146     BUG();
3147     return 0;
3148 }
3149 
3150 const struct vm_operations_struct hugetlb_vm_ops = {
3151     .fault = hugetlb_vm_op_fault,
3152     .open = hugetlb_vm_op_open,
3153     .close = hugetlb_vm_op_close,
3154 };
3155 
3156 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3157                 int writable)
3158 {
3159     pte_t entry;
3160 
3161     if (writable) {
3162         entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3163                      vma->vm_page_prot)));
3164     } else {
3165         entry = huge_pte_wrprotect(mk_huge_pte(page,
3166                        vma->vm_page_prot));
3167     }
3168     entry = pte_mkyoung(entry);
3169     entry = pte_mkhuge(entry);
3170     entry = arch_make_huge_pte(entry, vma, page, writable);
3171 
3172     return entry;
3173 }
3174 
3175 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3176                    unsigned long address, pte_t *ptep)
3177 {
3178     pte_t entry;
3179 
3180     entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3181     if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3182         update_mmu_cache(vma, address, ptep);
3183 }
3184 
3185 static int is_hugetlb_entry_migration(pte_t pte)
3186 {
3187     swp_entry_t swp;
3188 
3189     if (huge_pte_none(pte) || pte_present(pte))
3190         return 0;
3191     swp = pte_to_swp_entry(pte);
3192     if (non_swap_entry(swp) && is_migration_entry(swp))
3193         return 1;
3194     else
3195         return 0;
3196 }
3197 
3198 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3199 {
3200     swp_entry_t swp;
3201 
3202     if (huge_pte_none(pte) || pte_present(pte))
3203         return 0;
3204     swp = pte_to_swp_entry(pte);
3205     if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3206         return 1;
3207     else
3208         return 0;
3209 }
3210 
3211 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3212                 struct vm_area_struct *vma)
3213 {
3214     pte_t *src_pte, *dst_pte, entry;
3215     struct page *ptepage;
3216     unsigned long addr;
3217     int cow;
3218     struct hstate *h = hstate_vma(vma);
3219     unsigned long sz = huge_page_size(h);
3220     unsigned long mmun_start;   /* For mmu_notifiers */
3221     unsigned long mmun_end;     /* For mmu_notifiers */
3222     int ret = 0;
3223 
3224     cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3225 
3226     mmun_start = vma->vm_start;
3227     mmun_end = vma->vm_end;
3228     if (cow)
3229         mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
3230 
3231     for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3232         spinlock_t *src_ptl, *dst_ptl;
3233         src_pte = huge_pte_offset(src, addr);
3234         if (!src_pte)
3235             continue;
3236         dst_pte = huge_pte_alloc(dst, addr, sz);
3237         if (!dst_pte) {
3238             ret = -ENOMEM;
3239             break;
3240         }
3241 
3242         /* If the pagetables are shared don't copy or take references */
3243         if (dst_pte == src_pte)
3244             continue;
3245 
3246         dst_ptl = huge_pte_lock(h, dst, dst_pte);
3247         src_ptl = huge_pte_lockptr(h, src, src_pte);
3248         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3249         entry = huge_ptep_get(src_pte);
3250         if (huge_pte_none(entry)) { /* skip none entry */
3251             ;
3252         } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3253                     is_hugetlb_entry_hwpoisoned(entry))) {
3254             swp_entry_t swp_entry = pte_to_swp_entry(entry);
3255 
3256             if (is_write_migration_entry(swp_entry) && cow) {
3257                 /*
3258                  * COW mappings require pages in both
3259                  * parent and child to be set to read.
3260                  */
3261                 make_migration_entry_read(&swp_entry);
3262                 entry = swp_entry_to_pte(swp_entry);
3263                 set_huge_pte_at(src, addr, src_pte, entry);
3264             }
3265             set_huge_pte_at(dst, addr, dst_pte, entry);
3266         } else {
3267             if (cow) {
3268                 huge_ptep_set_wrprotect(src, addr, src_pte);
3269                 mmu_notifier_invalidate_range(src, mmun_start,
3270                                    mmun_end);
3271             }
3272             entry = huge_ptep_get(src_pte);
3273             ptepage = pte_page(entry);
3274             get_page(ptepage);
3275             page_dup_rmap(ptepage, true);
3276             set_huge_pte_at(dst, addr, dst_pte, entry);
3277             hugetlb_count_add(pages_per_huge_page(h), dst);
3278         }
3279         spin_unlock(src_ptl);
3280         spin_unlock(dst_ptl);
3281     }
3282 
3283     if (cow)
3284         mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
3285 
3286     return ret;
3287 }
3288 
3289 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3290                 unsigned long start, unsigned long end,
3291                 struct page *ref_page)
3292 {
3293     struct mm_struct *mm = vma->vm_mm;
3294     unsigned long address;
3295     pte_t *ptep;
3296     pte_t pte;
3297     spinlock_t *ptl;
3298     struct page *page;
3299     struct hstate *h = hstate_vma(vma);
3300     unsigned long sz = huge_page_size(h);
3301     const unsigned long mmun_start = start; /* For mmu_notifiers */
3302     const unsigned long mmun_end   = end;   /* For mmu_notifiers */
3303 
3304     WARN_ON(!is_vm_hugetlb_page(vma));
3305     BUG_ON(start & ~huge_page_mask(h));
3306     BUG_ON(end & ~huge_page_mask(h));
3307 
3308     /*
3309      * This is a hugetlb vma, all the pte entries should point
3310      * to huge page.
3311      */
3312     tlb_remove_check_page_size_change(tlb, sz);
3313     tlb_start_vma(tlb, vma);
3314     mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3315     address = start;
3316     for (; address < end; address += sz) {
3317         ptep = huge_pte_offset(mm, address);
3318         if (!ptep)
3319             continue;
3320 
3321         ptl = huge_pte_lock(h, mm, ptep);
3322         if (huge_pmd_unshare(mm, &address, ptep)) {
3323             spin_unlock(ptl);
3324             continue;
3325         }
3326 
3327         pte = huge_ptep_get(ptep);
3328         if (huge_pte_none(pte)) {
3329             spin_unlock(ptl);
3330             continue;
3331         }
3332 
3333         /*
3334          * Migrating hugepage or HWPoisoned hugepage is already
3335          * unmapped and its refcount is dropped, so just clear pte here.
3336          */
3337         if (unlikely(!pte_present(pte))) {
3338             huge_pte_clear(mm, address, ptep);
3339             spin_unlock(ptl);
3340             continue;
3341         }
3342 
3343         page = pte_page(pte);
3344         /*
3345          * If a reference page is supplied, it is because a specific
3346          * page is being unmapped, not a range. Ensure the page we
3347          * are about to unmap is the actual page of interest.
3348          */
3349         if (ref_page) {
3350             if (page != ref_page) {
3351                 spin_unlock(ptl);
3352                 continue;
3353             }
3354             /*
3355              * Mark the VMA as having unmapped its page so that
3356              * future faults in this VMA will fail rather than
3357              * looking like data was lost
3358              */
3359             set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3360         }
3361 
3362         pte = huge_ptep_get_and_clear(mm, address, ptep);
3363         tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
3364         if (huge_pte_dirty(pte))
3365             set_page_dirty(page);
3366 
3367         hugetlb_count_sub(pages_per_huge_page(h), mm);
3368         page_remove_rmap(page, true);
3369 
3370         spin_unlock(ptl);
3371         tlb_remove_page_size(tlb, page, huge_page_size(h));
3372         /*
3373          * Bail out after unmapping reference page if supplied
3374          */
3375         if (ref_page)
3376             break;
3377     }
3378     mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3379     tlb_end_vma(tlb, vma);
3380 }
3381 
3382 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3383               struct vm_area_struct *vma, unsigned long start,
3384               unsigned long end, struct page *ref_page)
3385 {
3386     __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3387 
3388     /*
3389      * Clear this flag so that x86's huge_pmd_share page_table_shareable
3390      * test will fail on a vma being torn down, and not grab a page table
3391      * on its way out.  We're lucky that the flag has such an appropriate
3392      * name, and can in fact be safely cleared here. We could clear it
3393      * before the __unmap_hugepage_range above, but all that's necessary
3394      * is to clear it before releasing the i_mmap_rwsem. This works
3395      * because in the context this is called, the VMA is about to be
3396      * destroyed and the i_mmap_rwsem is held.
3397      */
3398     vma->vm_flags &= ~VM_MAYSHARE;
3399 }
3400 
3401 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3402               unsigned long end, struct page *ref_page)
3403 {
3404     struct mm_struct *mm;
3405     struct mmu_gather tlb;
3406 
3407     mm = vma->vm_mm;
3408 
3409     tlb_gather_mmu(&tlb, mm, start, end);
3410     __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3411     tlb_finish_mmu(&tlb, start, end);
3412 }
3413 
3414 /*
3415  * This is called when the original mapper is failing to COW a MAP_PRIVATE
3416  * mappping it owns the reserve page for. The intention is to unmap the page
3417  * from other VMAs and let the children be SIGKILLed if they are faulting the
3418  * same region.
3419  */
3420 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3421                   struct page *page, unsigned long address)
3422 {
3423     struct hstate *h = hstate_vma(vma);
3424     struct vm_area_struct *iter_vma;
3425     struct address_space *mapping;
3426     pgoff_t pgoff;
3427 
3428     /*
3429      * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3430      * from page cache lookup which is in HPAGE_SIZE units.
3431      */
3432     address = address & huge_page_mask(h);
3433     pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3434             vma->vm_pgoff;
3435     mapping = vma->vm_file->f_mapping;
3436 
3437     /*
3438      * Take the mapping lock for the duration of the table walk. As
3439      * this mapping should be shared between all the VMAs,
3440      * __unmap_hugepage_range() is called as the lock is already held
3441      */
3442     i_mmap_lock_write(mapping);
3443     vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3444         /* Do not unmap the current VMA */
3445         if (iter_vma == vma)
3446             continue;
3447 
3448         /*
3449          * Shared VMAs have their own reserves and do not affect
3450          * MAP_PRIVATE accounting but it is possible that a shared
3451          * VMA is using the same page so check and skip such VMAs.
3452          */
3453         if (iter_vma->vm_flags & VM_MAYSHARE)
3454             continue;
3455 
3456         /*
3457          * Unmap the page from other VMAs without their own reserves.
3458          * They get marked to be SIGKILLed if they fault in these
3459          * areas. This is because a future no-page fault on this VMA
3460          * could insert a zeroed page instead of the data existing
3461          * from the time of fork. This would look like data corruption
3462          */
3463         if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3464             unmap_hugepage_range(iter_vma, address,
3465                          address + huge_page_size(h), page);
3466     }
3467     i_mmap_unlock_write(mapping);
3468 }
3469 
3470 /*
3471  * Hugetlb_cow() should be called with page lock of the original hugepage held.
3472  * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3473  * cannot race with other handlers or page migration.
3474  * Keep the pte_same checks anyway to make transition from the mutex easier.
3475  */
3476 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3477                unsigned long address, pte_t *ptep,
3478                struct page *pagecache_page, spinlock_t *ptl)
3479 {
3480     pte_t pte;
3481     struct hstate *h = hstate_vma(vma);
3482     struct page *old_page, *new_page;
3483     int ret = 0, outside_reserve = 0;
3484     unsigned long mmun_start;   /* For mmu_notifiers */
3485     unsigned long mmun_end;     /* For mmu_notifiers */
3486 
3487     pte = huge_ptep_get(ptep);
3488     old_page = pte_page(pte);
3489 
3490 retry_avoidcopy:
3491     /* If no-one else is actually using this page, avoid the copy
3492      * and just make the page writable */
3493     if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3494         page_move_anon_rmap(old_page, vma);
3495         set_huge_ptep_writable(vma, address, ptep);
3496         return 0;
3497     }
3498 
3499     /*
3500      * If the process that created a MAP_PRIVATE mapping is about to
3501      * perform a COW due to a shared page count, attempt to satisfy
3502      * the allocation without using the existing reserves. The pagecache
3503      * page is used to determine if the reserve at this address was
3504      * consumed or not. If reserves were used, a partial faulted mapping
3505      * at the time of fork() could consume its reserves on COW instead
3506      * of the full address range.
3507      */
3508     if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3509             old_page != pagecache_page)
3510         outside_reserve = 1;
3511 
3512     get_page(old_page);
3513 
3514     /*
3515      * Drop page table lock as buddy allocator may be called. It will
3516      * be acquired again before returning to the caller, as expected.
3517      */
3518     spin_unlock(ptl);
3519     new_page = alloc_huge_page(vma, address, outside_reserve);
3520 
3521     if (IS_ERR(new_page)) {
3522         /*
3523          * If a process owning a MAP_PRIVATE mapping fails to COW,
3524          * it is due to references held by a child and an insufficient
3525          * huge page pool. To guarantee the original mappers
3526          * reliability, unmap the page from child processes. The child
3527          * may get SIGKILLed if it later faults.
3528          */
3529         if (outside_reserve) {
3530             put_page(old_page);
3531             BUG_ON(huge_pte_none(pte));
3532             unmap_ref_private(mm, vma, old_page, address);
3533             BUG_ON(huge_pte_none(pte));
3534             spin_lock(ptl);
3535             ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3536             if (likely(ptep &&
3537                    pte_same(huge_ptep_get(ptep), pte)))
3538                 goto retry_avoidcopy;
3539             /*
3540              * race occurs while re-acquiring page table
3541              * lock, and our job is done.
3542              */
3543             return 0;
3544         }
3545 
3546         ret = (PTR_ERR(new_page) == -ENOMEM) ?
3547             VM_FAULT_OOM : VM_FAULT_SIGBUS;
3548         goto out_release_old;
3549     }
3550 
3551     /*
3552      * When the original hugepage is shared one, it does not have
3553      * anon_vma prepared.
3554      */
3555     if (unlikely(anon_vma_prepare(vma))) {
3556         ret = VM_FAULT_OOM;
3557         goto out_release_all;
3558     }
3559 
3560     copy_user_huge_page(new_page, old_page, address, vma,
3561                 pages_per_huge_page(h));
3562     __SetPageUptodate(new_page);
3563     set_page_huge_active(new_page);
3564 
3565     mmun_start = address & huge_page_mask(h);
3566     mmun_end = mmun_start + huge_page_size(h);
3567     mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3568 
3569     /*
3570      * Retake the page table lock to check for racing updates
3571      * before the page tables are altered
3572      */
3573     spin_lock(ptl);
3574     ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3575     if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3576         ClearPagePrivate(new_page);
3577 
3578         /* Break COW */
3579         huge_ptep_clear_flush(vma, address, ptep);
3580         mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
3581         set_huge_pte_at(mm, address, ptep,
3582                 make_huge_pte(vma, new_page, 1));
3583         page_remove_rmap(old_page, true);
3584         hugepage_add_new_anon_rmap(new_page, vma, address);
3585         /* Make the old page be freed below */
3586         new_page = old_page;
3587     }
3588     spin_unlock(ptl);
3589     mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3590 out_release_all:
3591     restore_reserve_on_error(h, vma, address, new_page);
3592     put_page(new_page);
3593 out_release_old:
3594     put_page(old_page);
3595 
3596     spin_lock(ptl); /* Caller expects lock to be held */
3597     return ret;
3598 }
3599 
3600 /* Return the pagecache page at a given address within a VMA */
3601 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3602             struct vm_area_struct *vma, unsigned long address)
3603 {
3604     struct address_space *mapping;
3605     pgoff_t idx;
3606 
3607     mapping = vma->vm_file->f_mapping;
3608     idx = vma_hugecache_offset(h, vma, address);
3609 
3610     return find_lock_page(mapping, idx);
3611 }
3612 
3613 /*
3614  * Return whether there is a pagecache page to back given address within VMA.
3615  * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3616  */
3617 static bool hugetlbfs_pagecache_present(struct hstate *h,
3618             struct vm_area_struct *vma, unsigned long address)
3619 {
3620     struct address_space *mapping;
3621     pgoff_t idx;
3622     struct page *page;
3623 
3624     mapping = vma->vm_file->f_mapping;
3625     idx = vma_hugecache_offset(h, vma, address);
3626 
3627     page = find_get_page(mapping, idx);
3628     if (page)
3629         put_page(page);
3630     return page != NULL;
3631 }
3632 
3633 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3634                pgoff_t idx)
3635 {
3636     struct inode *inode = mapping->host;
3637     struct hstate *h = hstate_inode(inode);
3638     int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3639 
3640     if (err)
3641         return err;
3642     ClearPagePrivate(page);
3643 
3644     spin_lock(&inode->i_lock);
3645     inode->i_blocks += blocks_per_huge_page(h);
3646     spin_unlock(&inode->i_lock);
3647     return 0;
3648 }
3649 
3650 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
3651                struct address_space *mapping, pgoff_t idx,
3652                unsigned long address, pte_t *ptep, unsigned int flags)
3653 {
3654     struct hstate *h = hstate_vma(vma);
3655     int ret = VM_FAULT_SIGBUS;
3656     int anon_rmap = 0;
3657     unsigned long size;
3658     struct page *page;
3659     pte_t new_pte;
3660     spinlock_t *ptl;
3661 
3662     /*
3663      * Currently, we are forced to kill the process in the event the
3664      * original mapper has unmapped pages from the child due to a failed
3665      * COW. Warn that such a situation has occurred as it may not be obvious
3666      */
3667     if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3668         pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
3669                current->pid);
3670         return ret;
3671     }
3672 
3673     /*
3674      * Use page lock to guard against racing truncation
3675      * before we get page_table_lock.
3676      */
3677 retry:
3678     page = find_lock_page(mapping, idx);
3679     if (!page) {
3680         size = i_size_read(mapping->host) >> huge_page_shift(h);
3681         if (idx >= size)
3682             goto out;
3683         page = alloc_huge_page(vma, address, 0);
3684         if (IS_ERR(page)) {
3685             ret = PTR_ERR(page);
3686             if (ret == -ENOMEM)
3687                 ret = VM_FAULT_OOM;
3688             else
3689                 ret = VM_FAULT_SIGBUS;
3690             goto out;
3691         }
3692         clear_huge_page(page, address, pages_per_huge_page(h));
3693         __SetPageUptodate(page);
3694         set_page_huge_active(page);
3695 
3696         if (vma->vm_flags & VM_MAYSHARE) {
3697             int err = huge_add_to_page_cache(page, mapping, idx);
3698             if (err) {
3699                 put_page(page);
3700                 if (err == -EEXIST)
3701                     goto retry;
3702                 goto out;
3703             }
3704         } else {
3705             lock_page(page);
3706             if (unlikely(anon_vma_prepare(vma))) {
3707                 ret = VM_FAULT_OOM;
3708                 goto backout_unlocked;
3709             }
3710             anon_rmap = 1;
3711         }
3712     } else {
3713         /*
3714          * If memory error occurs between mmap() and fault, some process
3715          * don't have hwpoisoned swap entry for errored virtual address.
3716          * So we need to block hugepage fault by PG_hwpoison bit check.
3717          */
3718         if (unlikely(PageHWPoison(page))) {
3719             ret = VM_FAULT_HWPOISON |
3720                 VM_FAULT_SET_HINDEX(hstate_index(h));
3721             goto backout_unlocked;
3722         }
3723     }
3724 
3725     /*
3726      * If we are going to COW a private mapping later, we examine the
3727      * pending reservations for this page now. This will ensure that
3728      * any allocations necessary to record that reservation occur outside
3729      * the spinlock.
3730      */
3731     if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3732         if (vma_needs_reservation(h, vma, address) < 0) {
3733             ret = VM_FAULT_OOM;
3734             goto backout_unlocked;
3735         }
3736         /* Just decrements count, does not deallocate */
3737         vma_end_reservation(h, vma, address);
3738     }
3739 
3740     ptl = huge_pte_lock(h, mm, ptep);
3741     size = i_size_read(mapping->host) >> huge_page_shift(h);
3742     if (idx >= size)
3743         goto backout;
3744 
3745     ret = 0;
3746     if (!huge_pte_none(huge_ptep_get(ptep)))
3747         goto backout;
3748 
3749     if (anon_rmap) {
3750         ClearPagePrivate(page);
3751         hugepage_add_new_anon_rmap(page, vma, address);
3752     } else
3753         page_dup_rmap(page, true);
3754     new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3755                 && (vma->vm_flags & VM_SHARED)));
3756     set_huge_pte_at(mm, address, ptep, new_pte);
3757 
3758     hugetlb_count_add(pages_per_huge_page(h), mm);
3759     if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3760         /* Optimization, do the COW without a second fault */
3761         ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
3762     }
3763 
3764     spin_unlock(ptl);
3765     unlock_page(page);
3766 out:
3767     return ret;
3768 
3769 backout:
3770     spin_unlock(ptl);
3771 backout_unlocked:
3772     unlock_page(page);
3773     restore_reserve_on_error(h, vma, address, page);
3774     put_page(page);
3775     goto out;
3776 }
3777 
3778 #ifdef CONFIG_SMP
3779 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3780                 struct vm_area_struct *vma,
3781                 struct address_space *mapping,
3782                 pgoff_t idx, unsigned long address)
3783 {
3784     unsigned long key[2];
3785     u32 hash;
3786 
3787     if (vma->vm_flags & VM_SHARED) {
3788         key[0] = (unsigned long) mapping;
3789         key[1] = idx;
3790     } else {
3791         key[0] = (unsigned long) mm;
3792         key[1] = address >> huge_page_shift(h);
3793     }
3794 
3795     hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3796 
3797     return hash & (num_fault_mutexes - 1);
3798 }
3799 #else
3800 /*
3801  * For uniprocesor systems we always use a single mutex, so just
3802  * return 0 and avoid the hashing overhead.
3803  */
3804 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3805                 struct vm_area_struct *vma,
3806                 struct address_space *mapping,
3807                 pgoff_t idx, unsigned long address)
3808 {
3809     return 0;
3810 }
3811 #endif
3812 
3813 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3814             unsigned long address, unsigned int flags)
3815 {
3816     pte_t *ptep, entry;
3817     spinlock_t *ptl;
3818     int ret;
3819     u32 hash;
3820     pgoff_t idx;
3821     struct page *page = NULL;
3822     struct page *pagecache_page = NULL;
3823     struct hstate *h = hstate_vma(vma);
3824     struct address_space *mapping;
3825     int need_wait_lock = 0;
3826 
3827     address &= huge_page_mask(h);
3828 
3829     ptep = huge_pte_offset(mm, address);
3830     if (ptep) {
3831         entry = huge_ptep_get(ptep);
3832         if (unlikely(is_hugetlb_entry_migration(entry))) {
3833             migration_entry_wait_huge(vma, mm, ptep);
3834             return 0;
3835         } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
3836             return VM_FAULT_HWPOISON_LARGE |
3837                 VM_FAULT_SET_HINDEX(hstate_index(h));
3838     } else {
3839         ptep = huge_pte_alloc(mm, address, huge_page_size(h));
3840         if (!ptep)
3841             return VM_FAULT_OOM;
3842     }
3843 
3844     mapping = vma->vm_file->f_mapping;
3845     idx = vma_hugecache_offset(h, vma, address);
3846 
3847     /*
3848      * Serialize hugepage allocation and instantiation, so that we don't
3849      * get spurious allocation failures if two CPUs race to instantiate
3850      * the same page in the page cache.
3851      */
3852     hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address);
3853     mutex_lock(&hugetlb_fault_mutex_table[hash]);
3854 
3855     entry = huge_ptep_get(ptep);
3856     if (huge_pte_none(entry)) {
3857         ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
3858         goto out_mutex;
3859     }
3860 
3861     ret = 0;
3862 
3863     /*
3864      * entry could be a migration/hwpoison entry at this point, so this
3865      * check prevents the kernel from going below assuming that we have
3866      * a active hugepage in pagecache. This goto expects the 2nd page fault,
3867      * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3868      * handle it.
3869      */
3870     if (!pte_present(entry))
3871         goto out_mutex;
3872 
3873     /*
3874      * If we are going to COW the mapping later, we examine the pending
3875      * reservations for this page now. This will ensure that any
3876      * allocations necessary to record that reservation occur outside the
3877      * spinlock. For private mappings, we also lookup the pagecache
3878      * page now as it is used to determine if a reservation has been
3879      * consumed.
3880      */
3881     if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
3882         if (vma_needs_reservation(h, vma, address) < 0) {
3883             ret = VM_FAULT_OOM;
3884             goto out_mutex;
3885         }
3886         /* Just decrements count, does not deallocate */
3887         vma_end_reservation(h, vma, address);
3888 
3889         if (!(vma->vm_flags & VM_MAYSHARE))
3890             pagecache_page = hugetlbfs_pagecache_page(h,
3891                                 vma, address);
3892     }
3893 
3894     ptl = huge_pte_lock(h, mm, ptep);
3895 
3896     /* Check for a racing update before calling hugetlb_cow */
3897     if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
3898         goto out_ptl;
3899 
3900     /*
3901      * hugetlb_cow() requires page locks of pte_page(entry) and
3902      * pagecache_page, so here we need take the former one
3903      * when page != pagecache_page or !pagecache_page.
3904      */
3905     page = pte_page(entry);
3906     if (page != pagecache_page)
3907         if (!trylock_page(page)) {
3908             need_wait_lock = 1;
3909             goto out_ptl;
3910         }
3911 
3912     get_page(page);
3913 
3914     if (flags & FAULT_FLAG_WRITE) {
3915         if (!huge_pte_write(entry)) {
3916             ret = hugetlb_cow(mm, vma, address, ptep,
3917                       pagecache_page, ptl);
3918             goto out_put_page;
3919         }
3920         entry = huge_pte_mkdirty(entry);
3921     }
3922     entry = pte_mkyoung(entry);
3923     if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3924                         flags & FAULT_FLAG_WRITE))
3925         update_mmu_cache(vma, address, ptep);
3926 out_put_page:
3927     if (page != pagecache_page)
3928         unlock_page(page);
3929     put_page(page);
3930 out_ptl:
3931     spin_unlock(ptl);
3932 
3933     if (pagecache_page) {
3934         unlock_page(pagecache_page);
3935         put_page(pagecache_page);
3936     }
3937 out_mutex:
3938     mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3939     /*
3940      * Generally it's safe to hold refcount during waiting page lock. But
3941      * here we just wait to defer the next page fault to avoid busy loop and
3942      * the page is not used after unlocked before returning from the current
3943      * page fault. So we are safe from accessing freed page, even if we wait
3944      * here without taking refcount.
3945      */
3946     if (need_wait_lock)
3947         wait_on_page_locked(page);
3948     return ret;
3949 }
3950 
3951 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3952              struct page **pages, struct vm_area_struct **vmas,
3953              unsigned long *position, unsigned long *nr_pages,
3954              long i, unsigned int flags)
3955 {
3956     unsigned long pfn_offset;
3957     unsigned long vaddr = *position;
3958     unsigned long remainder = *nr_pages;
3959     struct hstate *h = hstate_vma(vma);
3960 
3961     while (vaddr < vma->vm_end && remainder) {
3962         pte_t *pte;
3963         spinlock_t *ptl = NULL;
3964         int absent;
3965         struct page *page;
3966 
3967         /*
3968          * If we have a pending SIGKILL, don't keep faulting pages and
3969          * potentially allocating memory.
3970          */
3971         if (unlikely(fatal_signal_pending(current))) {
3972             remainder = 0;
3973             break;
3974         }
3975 
3976         /*
3977          * Some archs (sparc64, sh*) have multiple pte_ts to
3978          * each hugepage.  We have to make sure we get the
3979          * first, for the page indexing below to work.
3980          *
3981          * Note that page table lock is not held when pte is null.
3982          */
3983         pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
3984         if (pte)
3985             ptl = huge_pte_lock(h, mm, pte);
3986         absent = !pte || huge_pte_none(huge_ptep_get(pte));
3987 
3988         /*
3989          * When coredumping, it suits get_dump_page if we just return
3990          * an error where there's an empty slot with no huge pagecache
3991          * to back it.  This way, we avoid allocating a hugepage, and
3992          * the sparse dumpfile avoids allocating disk blocks, but its
3993          * huge holes still show up with zeroes where they need to be.
3994          */
3995         if (absent && (flags & FOLL_DUMP) &&
3996             !hugetlbfs_pagecache_present(h, vma, vaddr)) {
3997             if (pte)
3998                 spin_unlock(ptl);
3999             remainder = 0;
4000             break;
4001         }
4002 
4003         /*
4004          * We need call hugetlb_fault for both hugepages under migration
4005          * (in which case hugetlb_fault waits for the migration,) and
4006          * hwpoisoned hugepages (in which case we need to prevent the
4007          * caller from accessing to them.) In order to do this, we use
4008          * here is_swap_pte instead of is_hugetlb_entry_migration and
4009          * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4010          * both cases, and because we can't follow correct pages
4011          * directly from any kind of swap entries.
4012          */
4013         if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4014             ((flags & FOLL_WRITE) &&
4015               !huge_pte_write(huge_ptep_get(pte)))) {
4016             int ret;
4017 
4018             if (pte)
4019                 spin_unlock(ptl);
4020             ret = hugetlb_fault(mm, vma, vaddr,
4021                 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
4022             if (!(ret & VM_FAULT_ERROR))
4023                 continue;
4024 
4025             remainder = 0;
4026             break;
4027         }
4028 
4029         pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4030         page = pte_page(huge_ptep_get(pte));
4031 same_page:
4032         if (pages) {
4033             pages[i] = mem_map_offset(page, pfn_offset);
4034             get_page(pages[i]);
4035         }
4036 
4037         if (vmas)
4038             vmas[i] = vma;
4039 
4040         vaddr += PAGE_SIZE;
4041         ++pfn_offset;
4042         --remainder;
4043         ++i;
4044         if (vaddr < vma->vm_end && remainder &&
4045                 pfn_offset < pages_per_huge_page(h)) {
4046             /*
4047              * We use pfn_offset to avoid touching the pageframes
4048              * of this compound page.
4049              */
4050             goto same_page;
4051         }
4052         spin_unlock(ptl);
4053     }
4054     *nr_pages = remainder;
4055     *position = vaddr;
4056 
4057     return i ? i : -EFAULT;
4058 }
4059 
4060 #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4061 /*
4062  * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4063  * implement this.
4064  */
4065 #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4066 #endif
4067 
4068 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
4069         unsigned long address, unsigned long end, pgprot_t newprot)
4070 {
4071     struct mm_struct *mm = vma->vm_mm;
4072     unsigned long start = address;
4073     pte_t *ptep;
4074     pte_t pte;
4075     struct hstate *h = hstate_vma(vma);
4076     unsigned long pages = 0;
4077 
4078     BUG_ON(address >= end);
4079     flush_cache_range(vma, address, end);
4080 
4081     mmu_notifier_invalidate_range_start(mm, start, end);
4082     i_mmap_lock_write(vma->vm_file->f_mapping);
4083     for (; address < end; address += huge_page_size(h)) {
4084         spinlock_t *ptl;
4085         ptep = huge_pte_offset(mm, address);
4086         if (!ptep)
4087             continue;
4088         ptl = huge_pte_lock(h, mm, ptep);
4089         if (huge_pmd_unshare(mm, &address, ptep)) {
4090             pages++;
4091             spin_unlock(ptl);
4092             continue;
4093         }
4094         pte = huge_ptep_get(ptep);
4095         if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
4096             spin_unlock(ptl);
4097             continue;
4098         }
4099         if (unlikely(is_hugetlb_entry_migration(pte))) {
4100             swp_entry_t entry = pte_to_swp_entry(pte);
4101 
4102             if (is_write_migration_entry(entry)) {
4103                 pte_t newpte;
4104 
4105                 make_migration_entry_read(&entry);
4106                 newpte = swp_entry_to_pte(entry);
4107                 set_huge_pte_at(mm, address, ptep, newpte);
4108                 pages++;
4109             }
4110             spin_unlock(ptl);
4111             continue;
4112         }
4113         if (!huge_pte_none(pte)) {
4114             pte = huge_ptep_get_and_clear(mm, address, ptep);
4115             pte = pte_mkhuge(huge_pte_modify(pte, newprot));
4116             pte = arch_make_huge_pte(pte, vma, NULL, 0);
4117             set_huge_pte_at(mm, address, ptep, pte);
4118             pages++;
4119         }
4120         spin_unlock(ptl);
4121     }
4122     /*
4123      * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
4124      * may have cleared our pud entry and done put_page on the page table:
4125      * once we release i_mmap_rwsem, another task can do the final put_page
4126      * and that page table be reused and filled with junk.
4127      */
4128     flush_hugetlb_tlb_range(vma, start, end);
4129     mmu_notifier_invalidate_range(mm, start, end);
4130     i_mmap_unlock_write(vma->vm_file->f_mapping);
4131     mmu_notifier_invalidate_range_end(mm, start, end);
4132 
4133     return pages << h->order;
4134 }
4135 
4136 int hugetlb_reserve_pages(struct inode *inode,
4137                     long from, long to,
4138                     struct vm_area_struct *vma,
4139                     vm_flags_t vm_flags)
4140 {
4141     long ret, chg;
4142     struct hstate *h = hstate_inode(inode);
4143     struct hugepage_subpool *spool = subpool_inode(inode);
4144     struct resv_map *resv_map;
4145     long gbl_reserve;
4146 
4147     /*
4148      * Only apply hugepage reservation if asked. At fault time, an
4149      * attempt will be made for VM_NORESERVE to allocate a page
4150      * without using reserves
4151      */
4152     if (vm_flags & VM_NORESERVE)
4153         return 0;
4154 
4155     /*
4156      * Shared mappings base their reservation on the number of pages that
4157      * are already allocated on behalf of the file. Private mappings need
4158      * to reserve the full area even if read-only as mprotect() may be
4159      * called to make the mapping read-write. Assume !vma is a shm mapping
4160      */
4161     if (!vma || vma->vm_flags & VM_MAYSHARE) {
4162         resv_map = inode_resv_map(inode);
4163 
4164         chg = region_chg(resv_map, from, to);
4165 
4166     } else {
4167         resv_map = resv_map_alloc();
4168         if (!resv_map)
4169             return -ENOMEM;
4170 
4171         chg = to - from;
4172 
4173         set_vma_resv_map(vma, resv_map);
4174         set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4175     }
4176 
4177     if (chg < 0) {
4178         ret = chg;
4179         goto out_err;
4180     }
4181 
4182     /*
4183      * There must be enough pages in the subpool for the mapping. If
4184      * the subpool has a minimum size, there may be some global
4185      * reservations already in place (gbl_reserve).
4186      */
4187     gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4188     if (gbl_reserve < 0) {
4189         ret = -ENOSPC;
4190         goto out_err;
4191     }
4192 
4193     /*
4194      * Check enough hugepages are available for the reservation.
4195      * Hand the pages back to the subpool if there are not
4196      */
4197     ret = hugetlb_acct_memory(h, gbl_reserve);
4198     if (ret < 0) {
4199         /* put back original number of pages, chg */
4200         (void)hugepage_subpool_put_pages(spool, chg);
4201         goto out_err;
4202     }
4203 
4204     /*
4205      * Account for the reservations made. Shared mappings record regions
4206      * that have reservations as they are shared by multiple VMAs.
4207      * When the last VMA disappears, the region map says how much
4208      * the reservation was and the page cache tells how much of
4209      * the reservation was consumed. Private mappings are per-VMA and
4210      * only the consumed reservations are tracked. When the VMA
4211      * disappears, the original reservation is the VMA size and the
4212      * consumed reservations are stored in the map. Hence, nothing
4213      * else has to be done for private mappings here
4214      */
4215     if (!vma || vma->vm_flags & VM_MAYSHARE) {
4216         long add = region_add(resv_map, from, to);
4217 
4218         if (unlikely(chg > add)) {
4219             /*
4220              * pages in this range were added to the reserve
4221              * map between region_chg and region_add.  This
4222              * indicates a race with alloc_huge_page.  Adjust
4223              * the subpool and reserve counts modified above
4224              * based on the difference.
4225              */
4226             long rsv_adjust;
4227 
4228             rsv_adjust = hugepage_subpool_put_pages(spool,
4229                                 chg - add);
4230             hugetlb_acct_memory(h, -rsv_adjust);
4231         }
4232     }
4233     return 0;
4234 out_err:
4235     if (!vma || vma->vm_flags & VM_MAYSHARE)
4236         region_abort(resv_map, from, to);
4237     if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4238         kref_put(&resv_map->refs, resv_map_release);
4239     return ret;
4240 }
4241 
4242 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4243                                 long freed)
4244 {
4245     struct hstate *h = hstate_inode(inode);
4246     struct resv_map *resv_map = inode_resv_map(inode);
4247     long chg = 0;
4248     struct hugepage_subpool *spool = subpool_inode(inode);
4249     long gbl_reserve;
4250 
4251     if (resv_map) {
4252         chg = region_del(resv_map, start, end);
4253         /*
4254          * region_del() can fail in the rare case where a region
4255          * must be split and another region descriptor can not be
4256          * allocated.  If end == LONG_MAX, it will not fail.
4257          */
4258         if (chg < 0)
4259             return chg;
4260     }
4261 
4262     spin_lock(&inode->i_lock);
4263     inode->i_blocks -= (blocks_per_huge_page(h) * freed);
4264     spin_unlock(&inode->i_lock);
4265 
4266     /*
4267      * If the subpool has a minimum size, the number of global
4268      * reservations to be released may be adjusted.
4269      */
4270     gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4271     hugetlb_acct_memory(h, -gbl_reserve);
4272 
4273     return 0;
4274 }
4275 
4276 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4277 static unsigned long page_table_shareable(struct vm_area_struct *svma,
4278                 struct vm_area_struct *vma,
4279                 unsigned long addr, pgoff_t idx)
4280 {
4281     unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4282                 svma->vm_start;
4283     unsigned long sbase = saddr & PUD_MASK;
4284     unsigned long s_end = sbase + PUD_SIZE;
4285 
4286     /* Allow segments to share if only one is marked locked */
4287     unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4288     unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
4289 
4290     /*
4291      * match the virtual addresses, permission and the alignment of the
4292      * page table page.
4293      */
4294     if (pmd_index(addr) != pmd_index(saddr) ||
4295         vm_flags != svm_flags ||
4296         sbase < svma->vm_start || svma->vm_end < s_end)
4297         return 0;
4298 
4299     return saddr;
4300 }
4301 
4302 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
4303 {
4304     unsigned long base = addr & PUD_MASK;
4305     unsigned long end = base + PUD_SIZE;
4306 
4307     /*
4308      * check on proper vm_flags and page table alignment
4309      */
4310     if (vma->vm_flags & VM_MAYSHARE &&
4311         vma->vm_start <= base && end <= vma->vm_end)
4312         return true;
4313     return false;
4314 }
4315 
4316 /*
4317  * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4318  * and returns the corresponding pte. While this is not necessary for the
4319  * !shared pmd case because we can allocate the pmd later as well, it makes the
4320  * code much cleaner. pmd allocation is essential for the shared case because
4321  * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4322  * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4323  * bad pmd for sharing.
4324  */
4325 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4326 {
4327     struct vm_area_struct *vma = find_vma(mm, addr);
4328     struct address_space *mapping = vma->vm_file->f_mapping;
4329     pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4330             vma->vm_pgoff;
4331     struct vm_area_struct *svma;
4332     unsigned long saddr;
4333     pte_t *spte = NULL;
4334     pte_t *pte;
4335     spinlock_t *ptl;
4336 
4337     if (!vma_shareable(vma, addr))
4338         return (pte_t *)pmd_alloc(mm, pud, addr);
4339 
4340     i_mmap_lock_write(mapping);
4341     vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4342         if (svma == vma)
4343             continue;
4344 
4345         saddr = page_table_shareable(svma, vma, addr, idx);
4346         if (saddr) {
4347             spte = huge_pte_offset(svma->vm_mm, saddr);
4348             if (spte) {
4349                 get_page(virt_to_page(spte));
4350                 break;
4351             }
4352         }
4353     }
4354 
4355     if (!spte)
4356         goto out;
4357 
4358     ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
4359     if (pud_none(*pud)) {
4360         pud_populate(mm, pud,
4361                 (pmd_t *)((unsigned long)spte & PAGE_MASK));
4362         mm_inc_nr_pmds(mm);
4363     } else {
4364         put_page(virt_to_page(spte));
4365     }
4366     spin_unlock(ptl);
4367 out:
4368     pte = (pte_t *)pmd_alloc(mm, pud, addr);
4369     i_mmap_unlock_write(mapping);
4370     return pte;
4371 }
4372 
4373 /*
4374  * unmap huge page backed by shared pte.
4375  *
4376  * Hugetlb pte page is ref counted at the time of mapping.  If pte is shared
4377  * indicated by page_count > 1, unmap is achieved by clearing pud and
4378  * decrementing the ref count. If count == 1, the pte page is not shared.
4379  *
4380  * called with page table lock held.
4381  *
4382  * returns: 1 successfully unmapped a shared pte page
4383  *      0 the underlying pte page is not shared, or it is the last user
4384  */
4385 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4386 {
4387     pgd_t *pgd = pgd_offset(mm, *addr);
4388     pud_t *pud = pud_offset(pgd, *addr);
4389 
4390     BUG_ON(page_count(virt_to_page(ptep)) == 0);
4391     if (page_count(virt_to_page(ptep)) == 1)
4392         return 0;
4393 
4394     pud_clear(pud);
4395     put_page(virt_to_page(ptep));
4396     mm_dec_nr_pmds(mm);
4397     *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4398     return 1;
4399 }
4400 #define want_pmd_share()    (1)
4401 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4402 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4403 {
4404     return NULL;
4405 }
4406 
4407 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4408 {
4409     return 0;
4410 }
4411 #define want_pmd_share()    (0)
4412 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4413 
4414 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4415 pte_t *huge_pte_alloc(struct mm_struct *mm,
4416             unsigned long addr, unsigned long sz)
4417 {
4418     pgd_t *pgd;
4419     pud_t *pud;
4420     pte_t *pte = NULL;
4421 
4422     pgd = pgd_offset(mm, addr);
4423     pud = pud_alloc(mm, pgd, addr);
4424     if (pud) {
4425         if (sz == PUD_SIZE) {
4426             pte = (pte_t *)pud;
4427         } else {
4428             BUG_ON(sz != PMD_SIZE);
4429             if (want_pmd_share() && pud_none(*pud))
4430                 pte = huge_pmd_share(mm, addr, pud);
4431             else
4432                 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4433         }
4434     }
4435     BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
4436 
4437     return pte;
4438 }
4439 
4440 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
4441 {
4442     pgd_t *pgd;
4443     pud_t *pud;
4444     pmd_t *pmd = NULL;
4445 
4446     pgd = pgd_offset(mm, addr);
4447     if (pgd_present(*pgd)) {
4448         pud = pud_offset(pgd, addr);
4449         if (pud_present(*pud)) {
4450             if (pud_huge(*pud))
4451                 return (pte_t *)pud;
4452             pmd = pmd_offset(pud, addr);
4453         }
4454     }
4455     return (pte_t *) pmd;
4456 }
4457 
4458 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4459 
4460 /*
4461  * These functions are overwritable if your architecture needs its own
4462  * behavior.
4463  */
4464 struct page * __weak
4465 follow_huge_addr(struct mm_struct *mm, unsigned long address,
4466                   int write)
4467 {
4468     return ERR_PTR(-EINVAL);
4469 }
4470 
4471 struct page * __weak
4472 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
4473         pmd_t *pmd, int flags)
4474 {
4475     struct page *page = NULL;
4476     spinlock_t *ptl;
4477 retry:
4478     ptl = pmd_lockptr(mm, pmd);
4479     spin_lock(ptl);
4480     /*
4481      * make sure that the address range covered by this pmd is not
4482      * unmapped from other threads.
4483      */
4484     if (!pmd_huge(*pmd))
4485         goto out;
4486     if (pmd_present(*pmd)) {
4487         page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
4488         if (flags & FOLL_GET)
4489             get_page(page);
4490     } else {
4491         if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) {
4492             spin_unlock(ptl);
4493             __migration_entry_wait(mm, (pte_t *)pmd, ptl);
4494             goto retry;
4495         }
4496         /*
4497          * hwpoisoned entry is treated as no_page_table in
4498          * follow_page_mask().
4499          */
4500     }
4501 out:
4502     spin_unlock(ptl);
4503     return page;
4504 }
4505 
4506 struct page * __weak
4507 follow_huge_pud(struct mm_struct *mm, unsigned long address,
4508         pud_t *pud, int flags)
4509 {
4510     if (flags & FOLL_GET)
4511         return NULL;
4512 
4513     return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
4514 }
4515 
4516 #ifdef CONFIG_MEMORY_FAILURE
4517 
4518 /*
4519  * This function is called from memory failure code.
4520  */
4521 int dequeue_hwpoisoned_huge_page(struct page *hpage)
4522 {
4523     struct hstate *h = page_hstate(hpage);
4524     int nid = page_to_nid(hpage);
4525     int ret = -EBUSY;
4526 
4527     spin_lock(&hugetlb_lock);
4528     /*
4529      * Just checking !page_huge_active is not enough, because that could be
4530      * an isolated/hwpoisoned hugepage (which have >0 refcount).
4531      */
4532     if (!page_huge_active(hpage) && !page_count(hpage)) {
4533         /*
4534          * Hwpoisoned hugepage isn't linked to activelist or freelist,
4535          * but dangling hpage->lru can trigger list-debug warnings
4536          * (this happens when we call unpoison_memory() on it),
4537          * so let it point to itself with list_del_init().
4538          */
4539         list_del_init(&hpage->lru);
4540         set_page_refcounted(hpage);
4541         h->free_huge_pages--;
4542         h->free_huge_pages_node[nid]--;
4543         ret = 0;
4544     }
4545     spin_unlock(&hugetlb_lock);
4546     return ret;
4547 }
4548 #endif
4549 
4550 bool isolate_huge_page(struct page *page, struct list_head *list)
4551 {
4552     bool ret = true;
4553 
4554     VM_BUG_ON_PAGE(!PageHead(page), page);
4555     spin_lock(&hugetlb_lock);
4556     if (!page_huge_active(page) || !get_page_unless_zero(page)) {
4557         ret = false;
4558         goto unlock;
4559     }
4560     clear_page_huge_active(page);
4561     list_move_tail(&page->lru, list);
4562 unlock:
4563     spin_unlock(&hugetlb_lock);
4564     return ret;
4565 }
4566 
4567 void putback_active_hugepage(struct page *page)
4568 {
4569     VM_BUG_ON_PAGE(!PageHead(page), page);
4570     spin_lock(&hugetlb_lock);
4571     set_page_huge_active(page);
4572     list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
4573     spin_unlock(&hugetlb_lock);
4574     put_page(page);
4575 }