Back to home page

LXR

 
 

    


0001 /*
0002  *  Copyright (C) 2009  Red Hat, Inc.
0003  *
0004  *  This work is licensed under the terms of the GNU GPL, version 2. See
0005  *  the COPYING file in the top-level directory.
0006  */
0007 
0008 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
0009 
0010 #include <linux/mm.h>
0011 #include <linux/sched.h>
0012 #include <linux/highmem.h>
0013 #include <linux/hugetlb.h>
0014 #include <linux/mmu_notifier.h>
0015 #include <linux/rmap.h>
0016 #include <linux/swap.h>
0017 #include <linux/shrinker.h>
0018 #include <linux/mm_inline.h>
0019 #include <linux/swapops.h>
0020 #include <linux/dax.h>
0021 #include <linux/khugepaged.h>
0022 #include <linux/freezer.h>
0023 #include <linux/pfn_t.h>
0024 #include <linux/mman.h>
0025 #include <linux/memremap.h>
0026 #include <linux/pagemap.h>
0027 #include <linux/debugfs.h>
0028 #include <linux/migrate.h>
0029 #include <linux/hashtable.h>
0030 #include <linux/userfaultfd_k.h>
0031 #include <linux/page_idle.h>
0032 #include <linux/shmem_fs.h>
0033 
0034 #include <asm/tlb.h>
0035 #include <asm/pgalloc.h>
0036 #include "internal.h"
0037 
0038 /*
0039  * By default transparent hugepage support is disabled in order that avoid
0040  * to risk increase the memory footprint of applications without a guaranteed
0041  * benefit. When transparent hugepage support is enabled, is for all mappings,
0042  * and khugepaged scans all mappings.
0043  * Defrag is invoked by khugepaged hugepage allocations and by page faults
0044  * for all hugepage allocations.
0045  */
0046 unsigned long transparent_hugepage_flags __read_mostly =
0047 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
0048     (1<<TRANSPARENT_HUGEPAGE_FLAG)|
0049 #endif
0050 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
0051     (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
0052 #endif
0053     (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
0054     (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
0055     (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
0056 
0057 static struct shrinker deferred_split_shrinker;
0058 
0059 static atomic_t huge_zero_refcount;
0060 struct page *huge_zero_page __read_mostly;
0061 
0062 static struct page *get_huge_zero_page(void)
0063 {
0064     struct page *zero_page;
0065 retry:
0066     if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
0067         return READ_ONCE(huge_zero_page);
0068 
0069     zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
0070             HPAGE_PMD_ORDER);
0071     if (!zero_page) {
0072         count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
0073         return NULL;
0074     }
0075     count_vm_event(THP_ZERO_PAGE_ALLOC);
0076     preempt_disable();
0077     if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
0078         preempt_enable();
0079         __free_pages(zero_page, compound_order(zero_page));
0080         goto retry;
0081     }
0082 
0083     /* We take additional reference here. It will be put back by shrinker */
0084     atomic_set(&huge_zero_refcount, 2);
0085     preempt_enable();
0086     return READ_ONCE(huge_zero_page);
0087 }
0088 
0089 static void put_huge_zero_page(void)
0090 {
0091     /*
0092      * Counter should never go to zero here. Only shrinker can put
0093      * last reference.
0094      */
0095     BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
0096 }
0097 
0098 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
0099 {
0100     if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
0101         return READ_ONCE(huge_zero_page);
0102 
0103     if (!get_huge_zero_page())
0104         return NULL;
0105 
0106     if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
0107         put_huge_zero_page();
0108 
0109     return READ_ONCE(huge_zero_page);
0110 }
0111 
0112 void mm_put_huge_zero_page(struct mm_struct *mm)
0113 {
0114     if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
0115         put_huge_zero_page();
0116 }
0117 
0118 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
0119                     struct shrink_control *sc)
0120 {
0121     /* we can free zero page only if last reference remains */
0122     return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
0123 }
0124 
0125 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
0126                        struct shrink_control *sc)
0127 {
0128     if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
0129         struct page *zero_page = xchg(&huge_zero_page, NULL);
0130         BUG_ON(zero_page == NULL);
0131         __free_pages(zero_page, compound_order(zero_page));
0132         return HPAGE_PMD_NR;
0133     }
0134 
0135     return 0;
0136 }
0137 
0138 static struct shrinker huge_zero_page_shrinker = {
0139     .count_objects = shrink_huge_zero_page_count,
0140     .scan_objects = shrink_huge_zero_page_scan,
0141     .seeks = DEFAULT_SEEKS,
0142 };
0143 
0144 #ifdef CONFIG_SYSFS
0145 
0146 static ssize_t triple_flag_store(struct kobject *kobj,
0147                  struct kobj_attribute *attr,
0148                  const char *buf, size_t count,
0149                  enum transparent_hugepage_flag enabled,
0150                  enum transparent_hugepage_flag deferred,
0151                  enum transparent_hugepage_flag req_madv)
0152 {
0153     if (!memcmp("defer", buf,
0154             min(sizeof("defer")-1, count))) {
0155         if (enabled == deferred)
0156             return -EINVAL;
0157         clear_bit(enabled, &transparent_hugepage_flags);
0158         clear_bit(req_madv, &transparent_hugepage_flags);
0159         set_bit(deferred, &transparent_hugepage_flags);
0160     } else if (!memcmp("always", buf,
0161             min(sizeof("always")-1, count))) {
0162         clear_bit(deferred, &transparent_hugepage_flags);
0163         clear_bit(req_madv, &transparent_hugepage_flags);
0164         set_bit(enabled, &transparent_hugepage_flags);
0165     } else if (!memcmp("madvise", buf,
0166                min(sizeof("madvise")-1, count))) {
0167         clear_bit(enabled, &transparent_hugepage_flags);
0168         clear_bit(deferred, &transparent_hugepage_flags);
0169         set_bit(req_madv, &transparent_hugepage_flags);
0170     } else if (!memcmp("never", buf,
0171                min(sizeof("never")-1, count))) {
0172         clear_bit(enabled, &transparent_hugepage_flags);
0173         clear_bit(req_madv, &transparent_hugepage_flags);
0174         clear_bit(deferred, &transparent_hugepage_flags);
0175     } else
0176         return -EINVAL;
0177 
0178     return count;
0179 }
0180 
0181 static ssize_t enabled_show(struct kobject *kobj,
0182                 struct kobj_attribute *attr, char *buf)
0183 {
0184     if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
0185         return sprintf(buf, "[always] madvise never\n");
0186     else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
0187         return sprintf(buf, "always [madvise] never\n");
0188     else
0189         return sprintf(buf, "always madvise [never]\n");
0190 }
0191 
0192 static ssize_t enabled_store(struct kobject *kobj,
0193                  struct kobj_attribute *attr,
0194                  const char *buf, size_t count)
0195 {
0196     ssize_t ret;
0197 
0198     ret = triple_flag_store(kobj, attr, buf, count,
0199                 TRANSPARENT_HUGEPAGE_FLAG,
0200                 TRANSPARENT_HUGEPAGE_FLAG,
0201                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
0202 
0203     if (ret > 0) {
0204         int err = start_stop_khugepaged();
0205         if (err)
0206             ret = err;
0207     }
0208 
0209     return ret;
0210 }
0211 static struct kobj_attribute enabled_attr =
0212     __ATTR(enabled, 0644, enabled_show, enabled_store);
0213 
0214 ssize_t single_hugepage_flag_show(struct kobject *kobj,
0215                 struct kobj_attribute *attr, char *buf,
0216                 enum transparent_hugepage_flag flag)
0217 {
0218     return sprintf(buf, "%d\n",
0219                !!test_bit(flag, &transparent_hugepage_flags));
0220 }
0221 
0222 ssize_t single_hugepage_flag_store(struct kobject *kobj,
0223                  struct kobj_attribute *attr,
0224                  const char *buf, size_t count,
0225                  enum transparent_hugepage_flag flag)
0226 {
0227     unsigned long value;
0228     int ret;
0229 
0230     ret = kstrtoul(buf, 10, &value);
0231     if (ret < 0)
0232         return ret;
0233     if (value > 1)
0234         return -EINVAL;
0235 
0236     if (value)
0237         set_bit(flag, &transparent_hugepage_flags);
0238     else
0239         clear_bit(flag, &transparent_hugepage_flags);
0240 
0241     return count;
0242 }
0243 
0244 /*
0245  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
0246  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
0247  * memory just to allocate one more hugepage.
0248  */
0249 static ssize_t defrag_show(struct kobject *kobj,
0250                struct kobj_attribute *attr, char *buf)
0251 {
0252     if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
0253         return sprintf(buf, "[always] defer madvise never\n");
0254     if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
0255         return sprintf(buf, "always [defer] madvise never\n");
0256     else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
0257         return sprintf(buf, "always defer [madvise] never\n");
0258     else
0259         return sprintf(buf, "always defer madvise [never]\n");
0260 
0261 }
0262 static ssize_t defrag_store(struct kobject *kobj,
0263                 struct kobj_attribute *attr,
0264                 const char *buf, size_t count)
0265 {
0266     return triple_flag_store(kobj, attr, buf, count,
0267                  TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
0268                  TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
0269                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
0270 }
0271 static struct kobj_attribute defrag_attr =
0272     __ATTR(defrag, 0644, defrag_show, defrag_store);
0273 
0274 static ssize_t use_zero_page_show(struct kobject *kobj,
0275         struct kobj_attribute *attr, char *buf)
0276 {
0277     return single_hugepage_flag_show(kobj, attr, buf,
0278                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
0279 }
0280 static ssize_t use_zero_page_store(struct kobject *kobj,
0281         struct kobj_attribute *attr, const char *buf, size_t count)
0282 {
0283     return single_hugepage_flag_store(kobj, attr, buf, count,
0284                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
0285 }
0286 static struct kobj_attribute use_zero_page_attr =
0287     __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
0288 
0289 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
0290         struct kobj_attribute *attr, char *buf)
0291 {
0292     return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE);
0293 }
0294 static struct kobj_attribute hpage_pmd_size_attr =
0295     __ATTR_RO(hpage_pmd_size);
0296 
0297 #ifdef CONFIG_DEBUG_VM
0298 static ssize_t debug_cow_show(struct kobject *kobj,
0299                 struct kobj_attribute *attr, char *buf)
0300 {
0301     return single_hugepage_flag_show(kobj, attr, buf,
0302                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
0303 }
0304 static ssize_t debug_cow_store(struct kobject *kobj,
0305                    struct kobj_attribute *attr,
0306                    const char *buf, size_t count)
0307 {
0308     return single_hugepage_flag_store(kobj, attr, buf, count,
0309                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
0310 }
0311 static struct kobj_attribute debug_cow_attr =
0312     __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
0313 #endif /* CONFIG_DEBUG_VM */
0314 
0315 static struct attribute *hugepage_attr[] = {
0316     &enabled_attr.attr,
0317     &defrag_attr.attr,
0318     &use_zero_page_attr.attr,
0319     &hpage_pmd_size_attr.attr,
0320 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
0321     &shmem_enabled_attr.attr,
0322 #endif
0323 #ifdef CONFIG_DEBUG_VM
0324     &debug_cow_attr.attr,
0325 #endif
0326     NULL,
0327 };
0328 
0329 static struct attribute_group hugepage_attr_group = {
0330     .attrs = hugepage_attr,
0331 };
0332 
0333 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
0334 {
0335     int err;
0336 
0337     *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
0338     if (unlikely(!*hugepage_kobj)) {
0339         pr_err("failed to create transparent hugepage kobject\n");
0340         return -ENOMEM;
0341     }
0342 
0343     err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
0344     if (err) {
0345         pr_err("failed to register transparent hugepage group\n");
0346         goto delete_obj;
0347     }
0348 
0349     err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
0350     if (err) {
0351         pr_err("failed to register transparent hugepage group\n");
0352         goto remove_hp_group;
0353     }
0354 
0355     return 0;
0356 
0357 remove_hp_group:
0358     sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
0359 delete_obj:
0360     kobject_put(*hugepage_kobj);
0361     return err;
0362 }
0363 
0364 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
0365 {
0366     sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
0367     sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
0368     kobject_put(hugepage_kobj);
0369 }
0370 #else
0371 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
0372 {
0373     return 0;
0374 }
0375 
0376 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
0377 {
0378 }
0379 #endif /* CONFIG_SYSFS */
0380 
0381 static int __init hugepage_init(void)
0382 {
0383     int err;
0384     struct kobject *hugepage_kobj;
0385 
0386     if (!has_transparent_hugepage()) {
0387         transparent_hugepage_flags = 0;
0388         return -EINVAL;
0389     }
0390 
0391     /*
0392      * hugepages can't be allocated by the buddy allocator
0393      */
0394     MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
0395     /*
0396      * we use page->mapping and page->index in second tail page
0397      * as list_head: assuming THP order >= 2
0398      */
0399     MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
0400 
0401     err = hugepage_init_sysfs(&hugepage_kobj);
0402     if (err)
0403         goto err_sysfs;
0404 
0405     err = khugepaged_init();
0406     if (err)
0407         goto err_slab;
0408 
0409     err = register_shrinker(&huge_zero_page_shrinker);
0410     if (err)
0411         goto err_hzp_shrinker;
0412     err = register_shrinker(&deferred_split_shrinker);
0413     if (err)
0414         goto err_split_shrinker;
0415 
0416     /*
0417      * By default disable transparent hugepages on smaller systems,
0418      * where the extra memory used could hurt more than TLB overhead
0419      * is likely to save.  The admin can still enable it through /sys.
0420      */
0421     if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
0422         transparent_hugepage_flags = 0;
0423         return 0;
0424     }
0425 
0426     err = start_stop_khugepaged();
0427     if (err)
0428         goto err_khugepaged;
0429 
0430     return 0;
0431 err_khugepaged:
0432     unregister_shrinker(&deferred_split_shrinker);
0433 err_split_shrinker:
0434     unregister_shrinker(&huge_zero_page_shrinker);
0435 err_hzp_shrinker:
0436     khugepaged_destroy();
0437 err_slab:
0438     hugepage_exit_sysfs(hugepage_kobj);
0439 err_sysfs:
0440     return err;
0441 }
0442 subsys_initcall(hugepage_init);
0443 
0444 static int __init setup_transparent_hugepage(char *str)
0445 {
0446     int ret = 0;
0447     if (!str)
0448         goto out;
0449     if (!strcmp(str, "always")) {
0450         set_bit(TRANSPARENT_HUGEPAGE_FLAG,
0451             &transparent_hugepage_flags);
0452         clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
0453               &transparent_hugepage_flags);
0454         ret = 1;
0455     } else if (!strcmp(str, "madvise")) {
0456         clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
0457               &transparent_hugepage_flags);
0458         set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
0459             &transparent_hugepage_flags);
0460         ret = 1;
0461     } else if (!strcmp(str, "never")) {
0462         clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
0463               &transparent_hugepage_flags);
0464         clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
0465               &transparent_hugepage_flags);
0466         ret = 1;
0467     }
0468 out:
0469     if (!ret)
0470         pr_warn("transparent_hugepage= cannot parse, ignored\n");
0471     return ret;
0472 }
0473 __setup("transparent_hugepage=", setup_transparent_hugepage);
0474 
0475 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
0476 {
0477     if (likely(vma->vm_flags & VM_WRITE))
0478         pmd = pmd_mkwrite(pmd);
0479     return pmd;
0480 }
0481 
0482 static inline struct list_head *page_deferred_list(struct page *page)
0483 {
0484     /*
0485      * ->lru in the tail pages is occupied by compound_head.
0486      * Let's use ->mapping + ->index in the second tail page as list_head.
0487      */
0488     return (struct list_head *)&page[2].mapping;
0489 }
0490 
0491 void prep_transhuge_page(struct page *page)
0492 {
0493     /*
0494      * we use page->mapping and page->indexlru in second tail page
0495      * as list_head: assuming THP order >= 2
0496      */
0497 
0498     INIT_LIST_HEAD(page_deferred_list(page));
0499     set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
0500 }
0501 
0502 unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len,
0503         loff_t off, unsigned long flags, unsigned long size)
0504 {
0505     unsigned long addr;
0506     loff_t off_end = off + len;
0507     loff_t off_align = round_up(off, size);
0508     unsigned long len_pad;
0509 
0510     if (off_end <= off_align || (off_end - off_align) < size)
0511         return 0;
0512 
0513     len_pad = len + size;
0514     if (len_pad < len || (off + len_pad) < off)
0515         return 0;
0516 
0517     addr = current->mm->get_unmapped_area(filp, 0, len_pad,
0518                           off >> PAGE_SHIFT, flags);
0519     if (IS_ERR_VALUE(addr))
0520         return 0;
0521 
0522     addr += (off - addr) & (size - 1);
0523     return addr;
0524 }
0525 
0526 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
0527         unsigned long len, unsigned long pgoff, unsigned long flags)
0528 {
0529     loff_t off = (loff_t)pgoff << PAGE_SHIFT;
0530 
0531     if (addr)
0532         goto out;
0533     if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
0534         goto out;
0535 
0536     addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE);
0537     if (addr)
0538         return addr;
0539 
0540  out:
0541     return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
0542 }
0543 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
0544 
0545 static int __do_huge_pmd_anonymous_page(struct vm_fault *vmf, struct page *page,
0546         gfp_t gfp)
0547 {
0548     struct vm_area_struct *vma = vmf->vma;
0549     struct mem_cgroup *memcg;
0550     pgtable_t pgtable;
0551     unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
0552 
0553     VM_BUG_ON_PAGE(!PageCompound(page), page);
0554 
0555     if (mem_cgroup_try_charge(page, vma->vm_mm, gfp, &memcg, true)) {
0556         put_page(page);
0557         count_vm_event(THP_FAULT_FALLBACK);
0558         return VM_FAULT_FALLBACK;
0559     }
0560 
0561     pgtable = pte_alloc_one(vma->vm_mm, haddr);
0562     if (unlikely(!pgtable)) {
0563         mem_cgroup_cancel_charge(page, memcg, true);
0564         put_page(page);
0565         return VM_FAULT_OOM;
0566     }
0567 
0568     clear_huge_page(page, haddr, HPAGE_PMD_NR);
0569     /*
0570      * The memory barrier inside __SetPageUptodate makes sure that
0571      * clear_huge_page writes become visible before the set_pmd_at()
0572      * write.
0573      */
0574     __SetPageUptodate(page);
0575 
0576     vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
0577     if (unlikely(!pmd_none(*vmf->pmd))) {
0578         spin_unlock(vmf->ptl);
0579         mem_cgroup_cancel_charge(page, memcg, true);
0580         put_page(page);
0581         pte_free(vma->vm_mm, pgtable);
0582     } else {
0583         pmd_t entry;
0584 
0585         /* Deliver the page fault to userland */
0586         if (userfaultfd_missing(vma)) {
0587             int ret;
0588 
0589             spin_unlock(vmf->ptl);
0590             mem_cgroup_cancel_charge(page, memcg, true);
0591             put_page(page);
0592             pte_free(vma->vm_mm, pgtable);
0593             ret = handle_userfault(vmf, VM_UFFD_MISSING);
0594             VM_BUG_ON(ret & VM_FAULT_FALLBACK);
0595             return ret;
0596         }
0597 
0598         entry = mk_huge_pmd(page, vma->vm_page_prot);
0599         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
0600         page_add_new_anon_rmap(page, vma, haddr, true);
0601         mem_cgroup_commit_charge(page, memcg, false, true);
0602         lru_cache_add_active_or_unevictable(page, vma);
0603         pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
0604         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
0605         add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
0606         atomic_long_inc(&vma->vm_mm->nr_ptes);
0607         spin_unlock(vmf->ptl);
0608         count_vm_event(THP_FAULT_ALLOC);
0609     }
0610 
0611     return 0;
0612 }
0613 
0614 /*
0615  * If THP defrag is set to always then directly reclaim/compact as necessary
0616  * If set to defer then do only background reclaim/compact and defer to khugepaged
0617  * If set to madvise and the VMA is flagged then directly reclaim/compact
0618  * When direct reclaim/compact is allowed, don't retry except for flagged VMA's
0619  */
0620 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
0621 {
0622     bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
0623 
0624     if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
0625                 &transparent_hugepage_flags) && vma_madvised)
0626         return GFP_TRANSHUGE;
0627     else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
0628                         &transparent_hugepage_flags))
0629         return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
0630     else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
0631                         &transparent_hugepage_flags))
0632         return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
0633 
0634     return GFP_TRANSHUGE_LIGHT;
0635 }
0636 
0637 /* Caller must hold page table lock. */
0638 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
0639         struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
0640         struct page *zero_page)
0641 {
0642     pmd_t entry;
0643     if (!pmd_none(*pmd))
0644         return false;
0645     entry = mk_pmd(zero_page, vma->vm_page_prot);
0646     entry = pmd_mkhuge(entry);
0647     if (pgtable)
0648         pgtable_trans_huge_deposit(mm, pmd, pgtable);
0649     set_pmd_at(mm, haddr, pmd, entry);
0650     atomic_long_inc(&mm->nr_ptes);
0651     return true;
0652 }
0653 
0654 int do_huge_pmd_anonymous_page(struct vm_fault *vmf)
0655 {
0656     struct vm_area_struct *vma = vmf->vma;
0657     gfp_t gfp;
0658     struct page *page;
0659     unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
0660 
0661     if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
0662         return VM_FAULT_FALLBACK;
0663     if (unlikely(anon_vma_prepare(vma)))
0664         return VM_FAULT_OOM;
0665     if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
0666         return VM_FAULT_OOM;
0667     if (!(vmf->flags & FAULT_FLAG_WRITE) &&
0668             !mm_forbids_zeropage(vma->vm_mm) &&
0669             transparent_hugepage_use_zero_page()) {
0670         pgtable_t pgtable;
0671         struct page *zero_page;
0672         bool set;
0673         int ret;
0674         pgtable = pte_alloc_one(vma->vm_mm, haddr);
0675         if (unlikely(!pgtable))
0676             return VM_FAULT_OOM;
0677         zero_page = mm_get_huge_zero_page(vma->vm_mm);
0678         if (unlikely(!zero_page)) {
0679             pte_free(vma->vm_mm, pgtable);
0680             count_vm_event(THP_FAULT_FALLBACK);
0681             return VM_FAULT_FALLBACK;
0682         }
0683         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
0684         ret = 0;
0685         set = false;
0686         if (pmd_none(*vmf->pmd)) {
0687             if (userfaultfd_missing(vma)) {
0688                 spin_unlock(vmf->ptl);
0689                 ret = handle_userfault(vmf, VM_UFFD_MISSING);
0690                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
0691             } else {
0692                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
0693                            haddr, vmf->pmd, zero_page);
0694                 spin_unlock(vmf->ptl);
0695                 set = true;
0696             }
0697         } else
0698             spin_unlock(vmf->ptl);
0699         if (!set)
0700             pte_free(vma->vm_mm, pgtable);
0701         return ret;
0702     }
0703     gfp = alloc_hugepage_direct_gfpmask(vma);
0704     page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
0705     if (unlikely(!page)) {
0706         count_vm_event(THP_FAULT_FALLBACK);
0707         return VM_FAULT_FALLBACK;
0708     }
0709     prep_transhuge_page(page);
0710     return __do_huge_pmd_anonymous_page(vmf, page, gfp);
0711 }
0712 
0713 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
0714         pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write)
0715 {
0716     struct mm_struct *mm = vma->vm_mm;
0717     pmd_t entry;
0718     spinlock_t *ptl;
0719 
0720     ptl = pmd_lock(mm, pmd);
0721     entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
0722     if (pfn_t_devmap(pfn))
0723         entry = pmd_mkdevmap(entry);
0724     if (write) {
0725         entry = pmd_mkyoung(pmd_mkdirty(entry));
0726         entry = maybe_pmd_mkwrite(entry, vma);
0727     }
0728     set_pmd_at(mm, addr, pmd, entry);
0729     update_mmu_cache_pmd(vma, addr, pmd);
0730     spin_unlock(ptl);
0731 }
0732 
0733 int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
0734             pmd_t *pmd, pfn_t pfn, bool write)
0735 {
0736     pgprot_t pgprot = vma->vm_page_prot;
0737     /*
0738      * If we had pmd_special, we could avoid all these restrictions,
0739      * but we need to be consistent with PTEs and architectures that
0740      * can't support a 'special' bit.
0741      */
0742     BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
0743     BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
0744                         (VM_PFNMAP|VM_MIXEDMAP));
0745     BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
0746     BUG_ON(!pfn_t_devmap(pfn));
0747 
0748     if (addr < vma->vm_start || addr >= vma->vm_end)
0749         return VM_FAULT_SIGBUS;
0750 
0751     track_pfn_insert(vma, &pgprot, pfn);
0752 
0753     insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
0754     return VM_FAULT_NOPAGE;
0755 }
0756 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
0757 
0758 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
0759         pmd_t *pmd)
0760 {
0761     pmd_t _pmd;
0762 
0763     /*
0764      * We should set the dirty bit only for FOLL_WRITE but for now
0765      * the dirty bit in the pmd is meaningless.  And if the dirty
0766      * bit will become meaningful and we'll only set it with
0767      * FOLL_WRITE, an atomic set_bit will be required on the pmd to
0768      * set the young bit, instead of the current set_pmd_at.
0769      */
0770     _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
0771     if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
0772                 pmd, _pmd,  1))
0773         update_mmu_cache_pmd(vma, addr, pmd);
0774 }
0775 
0776 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
0777         pmd_t *pmd, int flags)
0778 {
0779     unsigned long pfn = pmd_pfn(*pmd);
0780     struct mm_struct *mm = vma->vm_mm;
0781     struct dev_pagemap *pgmap;
0782     struct page *page;
0783 
0784     assert_spin_locked(pmd_lockptr(mm, pmd));
0785 
0786     /*
0787      * When we COW a devmap PMD entry, we split it into PTEs, so we should
0788      * not be in this function with `flags & FOLL_COW` set.
0789      */
0790     WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
0791 
0792     if (flags & FOLL_WRITE && !pmd_write(*pmd))
0793         return NULL;
0794 
0795     if (pmd_present(*pmd) && pmd_devmap(*pmd))
0796         /* pass */;
0797     else
0798         return NULL;
0799 
0800     if (flags & FOLL_TOUCH)
0801         touch_pmd(vma, addr, pmd);
0802 
0803     /*
0804      * device mapped pages can only be returned if the
0805      * caller will manage the page reference count.
0806      */
0807     if (!(flags & FOLL_GET))
0808         return ERR_PTR(-EEXIST);
0809 
0810     pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
0811     pgmap = get_dev_pagemap(pfn, NULL);
0812     if (!pgmap)
0813         return ERR_PTR(-EFAULT);
0814     page = pfn_to_page(pfn);
0815     get_page(page);
0816     put_dev_pagemap(pgmap);
0817 
0818     return page;
0819 }
0820 
0821 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
0822           pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
0823           struct vm_area_struct *vma)
0824 {
0825     spinlock_t *dst_ptl, *src_ptl;
0826     struct page *src_page;
0827     pmd_t pmd;
0828     pgtable_t pgtable = NULL;
0829     int ret = -ENOMEM;
0830 
0831     /* Skip if can be re-fill on fault */
0832     if (!vma_is_anonymous(vma))
0833         return 0;
0834 
0835     pgtable = pte_alloc_one(dst_mm, addr);
0836     if (unlikely(!pgtable))
0837         goto out;
0838 
0839     dst_ptl = pmd_lock(dst_mm, dst_pmd);
0840     src_ptl = pmd_lockptr(src_mm, src_pmd);
0841     spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
0842 
0843     ret = -EAGAIN;
0844     pmd = *src_pmd;
0845     if (unlikely(!pmd_trans_huge(pmd))) {
0846         pte_free(dst_mm, pgtable);
0847         goto out_unlock;
0848     }
0849     /*
0850      * When page table lock is held, the huge zero pmd should not be
0851      * under splitting since we don't split the page itself, only pmd to
0852      * a page table.
0853      */
0854     if (is_huge_zero_pmd(pmd)) {
0855         struct page *zero_page;
0856         /*
0857          * get_huge_zero_page() will never allocate a new page here,
0858          * since we already have a zero page to copy. It just takes a
0859          * reference.
0860          */
0861         zero_page = mm_get_huge_zero_page(dst_mm);
0862         set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
0863                 zero_page);
0864         ret = 0;
0865         goto out_unlock;
0866     }
0867 
0868     src_page = pmd_page(pmd);
0869     VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
0870     get_page(src_page);
0871     page_dup_rmap(src_page, true);
0872     add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
0873     atomic_long_inc(&dst_mm->nr_ptes);
0874     pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
0875 
0876     pmdp_set_wrprotect(src_mm, addr, src_pmd);
0877     pmd = pmd_mkold(pmd_wrprotect(pmd));
0878     set_pmd_at(dst_mm, addr, dst_pmd, pmd);
0879 
0880     ret = 0;
0881 out_unlock:
0882     spin_unlock(src_ptl);
0883     spin_unlock(dst_ptl);
0884 out:
0885     return ret;
0886 }
0887 
0888 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
0889 {
0890     pmd_t entry;
0891     unsigned long haddr;
0892     bool write = vmf->flags & FAULT_FLAG_WRITE;
0893 
0894     vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
0895     if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
0896         goto unlock;
0897 
0898     entry = pmd_mkyoung(orig_pmd);
0899     if (write)
0900         entry = pmd_mkdirty(entry);
0901     haddr = vmf->address & HPAGE_PMD_MASK;
0902     if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
0903         update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
0904 
0905 unlock:
0906     spin_unlock(vmf->ptl);
0907 }
0908 
0909 static int do_huge_pmd_wp_page_fallback(struct vm_fault *vmf, pmd_t orig_pmd,
0910         struct page *page)
0911 {
0912     struct vm_area_struct *vma = vmf->vma;
0913     unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
0914     struct mem_cgroup *memcg;
0915     pgtable_t pgtable;
0916     pmd_t _pmd;
0917     int ret = 0, i;
0918     struct page **pages;
0919     unsigned long mmun_start;   /* For mmu_notifiers */
0920     unsigned long mmun_end;     /* For mmu_notifiers */
0921 
0922     pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
0923             GFP_KERNEL);
0924     if (unlikely(!pages)) {
0925         ret |= VM_FAULT_OOM;
0926         goto out;
0927     }
0928 
0929     for (i = 0; i < HPAGE_PMD_NR; i++) {
0930         pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
0931                            vmf->address, page_to_nid(page));
0932         if (unlikely(!pages[i] ||
0933                  mem_cgroup_try_charge(pages[i], vma->vm_mm,
0934                      GFP_KERNEL, &memcg, false))) {
0935             if (pages[i])
0936                 put_page(pages[i]);
0937             while (--i >= 0) {
0938                 memcg = (void *)page_private(pages[i]);
0939                 set_page_private(pages[i], 0);
0940                 mem_cgroup_cancel_charge(pages[i], memcg,
0941                         false);
0942                 put_page(pages[i]);
0943             }
0944             kfree(pages);
0945             ret |= VM_FAULT_OOM;
0946             goto out;
0947         }
0948         set_page_private(pages[i], (unsigned long)memcg);
0949     }
0950 
0951     for (i = 0; i < HPAGE_PMD_NR; i++) {
0952         copy_user_highpage(pages[i], page + i,
0953                    haddr + PAGE_SIZE * i, vma);
0954         __SetPageUptodate(pages[i]);
0955         cond_resched();
0956     }
0957 
0958     mmun_start = haddr;
0959     mmun_end   = haddr + HPAGE_PMD_SIZE;
0960     mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
0961 
0962     vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
0963     if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
0964         goto out_free_pages;
0965     VM_BUG_ON_PAGE(!PageHead(page), page);
0966 
0967     pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
0968     /* leave pmd empty until pte is filled */
0969 
0970     pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
0971     pmd_populate(vma->vm_mm, &_pmd, pgtable);
0972 
0973     for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
0974         pte_t entry;
0975         entry = mk_pte(pages[i], vma->vm_page_prot);
0976         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
0977         memcg = (void *)page_private(pages[i]);
0978         set_page_private(pages[i], 0);
0979         page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
0980         mem_cgroup_commit_charge(pages[i], memcg, false, false);
0981         lru_cache_add_active_or_unevictable(pages[i], vma);
0982         vmf->pte = pte_offset_map(&_pmd, haddr);
0983         VM_BUG_ON(!pte_none(*vmf->pte));
0984         set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
0985         pte_unmap(vmf->pte);
0986     }
0987     kfree(pages);
0988 
0989     smp_wmb(); /* make pte visible before pmd */
0990     pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
0991     page_remove_rmap(page, true);
0992     spin_unlock(vmf->ptl);
0993 
0994     mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
0995 
0996     ret |= VM_FAULT_WRITE;
0997     put_page(page);
0998 
0999 out:
1000     return ret;
1001 
1002 out_free_pages:
1003     spin_unlock(vmf->ptl);
1004     mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1005     for (i = 0; i < HPAGE_PMD_NR; i++) {
1006         memcg = (void *)page_private(pages[i]);
1007         set_page_private(pages[i], 0);
1008         mem_cgroup_cancel_charge(pages[i], memcg, false);
1009         put_page(pages[i]);
1010     }
1011     kfree(pages);
1012     goto out;
1013 }
1014 
1015 int do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1016 {
1017     struct vm_area_struct *vma = vmf->vma;
1018     struct page *page = NULL, *new_page;
1019     struct mem_cgroup *memcg;
1020     unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1021     unsigned long mmun_start;   /* For mmu_notifiers */
1022     unsigned long mmun_end;     /* For mmu_notifiers */
1023     gfp_t huge_gfp;         /* for allocation and charge */
1024     int ret = 0;
1025 
1026     vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1027     VM_BUG_ON_VMA(!vma->anon_vma, vma);
1028     if (is_huge_zero_pmd(orig_pmd))
1029         goto alloc;
1030     spin_lock(vmf->ptl);
1031     if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1032         goto out_unlock;
1033 
1034     page = pmd_page(orig_pmd);
1035     VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1036     /*
1037      * We can only reuse the page if nobody else maps the huge page or it's
1038      * part.
1039      */
1040     if (page_trans_huge_mapcount(page, NULL) == 1) {
1041         pmd_t entry;
1042         entry = pmd_mkyoung(orig_pmd);
1043         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1044         if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry,  1))
1045             update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1046         ret |= VM_FAULT_WRITE;
1047         goto out_unlock;
1048     }
1049     get_page(page);
1050     spin_unlock(vmf->ptl);
1051 alloc:
1052     if (transparent_hugepage_enabled(vma) &&
1053         !transparent_hugepage_debug_cow()) {
1054         huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1055         new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1056     } else
1057         new_page = NULL;
1058 
1059     if (likely(new_page)) {
1060         prep_transhuge_page(new_page);
1061     } else {
1062         if (!page) {
1063             split_huge_pmd(vma, vmf->pmd, vmf->address);
1064             ret |= VM_FAULT_FALLBACK;
1065         } else {
1066             ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
1067             if (ret & VM_FAULT_OOM) {
1068                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1069                 ret |= VM_FAULT_FALLBACK;
1070             }
1071             put_page(page);
1072         }
1073         count_vm_event(THP_FAULT_FALLBACK);
1074         goto out;
1075     }
1076 
1077     if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm,
1078                     huge_gfp, &memcg, true))) {
1079         put_page(new_page);
1080         split_huge_pmd(vma, vmf->pmd, vmf->address);
1081         if (page)
1082             put_page(page);
1083         ret |= VM_FAULT_FALLBACK;
1084         count_vm_event(THP_FAULT_FALLBACK);
1085         goto out;
1086     }
1087 
1088     count_vm_event(THP_FAULT_ALLOC);
1089 
1090     if (!page)
1091         clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1092     else
1093         copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1094     __SetPageUptodate(new_page);
1095 
1096     mmun_start = haddr;
1097     mmun_end   = haddr + HPAGE_PMD_SIZE;
1098     mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
1099 
1100     spin_lock(vmf->ptl);
1101     if (page)
1102         put_page(page);
1103     if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1104         spin_unlock(vmf->ptl);
1105         mem_cgroup_cancel_charge(new_page, memcg, true);
1106         put_page(new_page);
1107         goto out_mn;
1108     } else {
1109         pmd_t entry;
1110         entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1111         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1112         pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1113         page_add_new_anon_rmap(new_page, vma, haddr, true);
1114         mem_cgroup_commit_charge(new_page, memcg, false, true);
1115         lru_cache_add_active_or_unevictable(new_page, vma);
1116         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1117         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1118         if (!page) {
1119             add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1120         } else {
1121             VM_BUG_ON_PAGE(!PageHead(page), page);
1122             page_remove_rmap(page, true);
1123             put_page(page);
1124         }
1125         ret |= VM_FAULT_WRITE;
1126     }
1127     spin_unlock(vmf->ptl);
1128 out_mn:
1129     mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1130 out:
1131     return ret;
1132 out_unlock:
1133     spin_unlock(vmf->ptl);
1134     return ret;
1135 }
1136 
1137 /*
1138  * FOLL_FORCE can write to even unwritable pmd's, but only
1139  * after we've gone through a COW cycle and they are dirty.
1140  */
1141 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1142 {
1143     return pmd_write(pmd) ||
1144            ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1145 }
1146 
1147 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1148                    unsigned long addr,
1149                    pmd_t *pmd,
1150                    unsigned int flags)
1151 {
1152     struct mm_struct *mm = vma->vm_mm;
1153     struct page *page = NULL;
1154 
1155     assert_spin_locked(pmd_lockptr(mm, pmd));
1156 
1157     if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1158         goto out;
1159 
1160     /* Avoid dumping huge zero page */
1161     if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1162         return ERR_PTR(-EFAULT);
1163 
1164     /* Full NUMA hinting faults to serialise migration in fault paths */
1165     if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1166         goto out;
1167 
1168     page = pmd_page(*pmd);
1169     VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1170     if (flags & FOLL_TOUCH)
1171         touch_pmd(vma, addr, pmd);
1172     if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1173         /*
1174          * We don't mlock() pte-mapped THPs. This way we can avoid
1175          * leaking mlocked pages into non-VM_LOCKED VMAs.
1176          *
1177          * For anon THP:
1178          *
1179          * In most cases the pmd is the only mapping of the page as we
1180          * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1181          * writable private mappings in populate_vma_page_range().
1182          *
1183          * The only scenario when we have the page shared here is if we
1184          * mlocking read-only mapping shared over fork(). We skip
1185          * mlocking such pages.
1186          *
1187          * For file THP:
1188          *
1189          * We can expect PageDoubleMap() to be stable under page lock:
1190          * for file pages we set it in page_add_file_rmap(), which
1191          * requires page to be locked.
1192          */
1193 
1194         if (PageAnon(page) && compound_mapcount(page) != 1)
1195             goto skip_mlock;
1196         if (PageDoubleMap(page) || !page->mapping)
1197             goto skip_mlock;
1198         if (!trylock_page(page))
1199             goto skip_mlock;
1200         lru_add_drain();
1201         if (page->mapping && !PageDoubleMap(page))
1202             mlock_vma_page(page);
1203         unlock_page(page);
1204     }
1205 skip_mlock:
1206     page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1207     VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1208     if (flags & FOLL_GET)
1209         get_page(page);
1210 
1211 out:
1212     return page;
1213 }
1214 
1215 /* NUMA hinting page fault entry point for trans huge pmds */
1216 int do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1217 {
1218     struct vm_area_struct *vma = vmf->vma;
1219     struct anon_vma *anon_vma = NULL;
1220     struct page *page;
1221     unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1222     int page_nid = -1, this_nid = numa_node_id();
1223     int target_nid, last_cpupid = -1;
1224     bool page_locked;
1225     bool migrated = false;
1226     bool was_writable;
1227     int flags = 0;
1228 
1229     vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1230     if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1231         goto out_unlock;
1232 
1233     /*
1234      * If there are potential migrations, wait for completion and retry
1235      * without disrupting NUMA hinting information. Do not relock and
1236      * check_same as the page may no longer be mapped.
1237      */
1238     if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1239         page = pmd_page(*vmf->pmd);
1240         spin_unlock(vmf->ptl);
1241         wait_on_page_locked(page);
1242         goto out;
1243     }
1244 
1245     page = pmd_page(pmd);
1246     BUG_ON(is_huge_zero_page(page));
1247     page_nid = page_to_nid(page);
1248     last_cpupid = page_cpupid_last(page);
1249     count_vm_numa_event(NUMA_HINT_FAULTS);
1250     if (page_nid == this_nid) {
1251         count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1252         flags |= TNF_FAULT_LOCAL;
1253     }
1254 
1255     /* See similar comment in do_numa_page for explanation */
1256     if (!pmd_write(pmd))
1257         flags |= TNF_NO_GROUP;
1258 
1259     /*
1260      * Acquire the page lock to serialise THP migrations but avoid dropping
1261      * page_table_lock if at all possible
1262      */
1263     page_locked = trylock_page(page);
1264     target_nid = mpol_misplaced(page, vma, haddr);
1265     if (target_nid == -1) {
1266         /* If the page was locked, there are no parallel migrations */
1267         if (page_locked)
1268             goto clear_pmdnuma;
1269     }
1270 
1271     /* Migration could have started since the pmd_trans_migrating check */
1272     if (!page_locked) {
1273         spin_unlock(vmf->ptl);
1274         wait_on_page_locked(page);
1275         page_nid = -1;
1276         goto out;
1277     }
1278 
1279     /*
1280      * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1281      * to serialises splits
1282      */
1283     get_page(page);
1284     spin_unlock(vmf->ptl);
1285     anon_vma = page_lock_anon_vma_read(page);
1286 
1287     /* Confirm the PMD did not change while page_table_lock was released */
1288     spin_lock(vmf->ptl);
1289     if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1290         unlock_page(page);
1291         put_page(page);
1292         page_nid = -1;
1293         goto out_unlock;
1294     }
1295 
1296     /* Bail if we fail to protect against THP splits for any reason */
1297     if (unlikely(!anon_vma)) {
1298         put_page(page);
1299         page_nid = -1;
1300         goto clear_pmdnuma;
1301     }
1302 
1303     /*
1304      * Migrate the THP to the requested node, returns with page unlocked
1305      * and access rights restored.
1306      */
1307     spin_unlock(vmf->ptl);
1308     migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1309                 vmf->pmd, pmd, vmf->address, page, target_nid);
1310     if (migrated) {
1311         flags |= TNF_MIGRATED;
1312         page_nid = target_nid;
1313     } else
1314         flags |= TNF_MIGRATE_FAIL;
1315 
1316     goto out;
1317 clear_pmdnuma:
1318     BUG_ON(!PageLocked(page));
1319     was_writable = pmd_write(pmd);
1320     pmd = pmd_modify(pmd, vma->vm_page_prot);
1321     pmd = pmd_mkyoung(pmd);
1322     if (was_writable)
1323         pmd = pmd_mkwrite(pmd);
1324     set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1325     update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1326     unlock_page(page);
1327 out_unlock:
1328     spin_unlock(vmf->ptl);
1329 
1330 out:
1331     if (anon_vma)
1332         page_unlock_anon_vma_read(anon_vma);
1333 
1334     if (page_nid != -1)
1335         task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1336                 vmf->flags);
1337 
1338     return 0;
1339 }
1340 
1341 /*
1342  * Return true if we do MADV_FREE successfully on entire pmd page.
1343  * Otherwise, return false.
1344  */
1345 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1346         pmd_t *pmd, unsigned long addr, unsigned long next)
1347 {
1348     spinlock_t *ptl;
1349     pmd_t orig_pmd;
1350     struct page *page;
1351     struct mm_struct *mm = tlb->mm;
1352     bool ret = false;
1353 
1354     tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1355 
1356     ptl = pmd_trans_huge_lock(pmd, vma);
1357     if (!ptl)
1358         goto out_unlocked;
1359 
1360     orig_pmd = *pmd;
1361     if (is_huge_zero_pmd(orig_pmd))
1362         goto out;
1363 
1364     page = pmd_page(orig_pmd);
1365     /*
1366      * If other processes are mapping this page, we couldn't discard
1367      * the page unless they all do MADV_FREE so let's skip the page.
1368      */
1369     if (page_mapcount(page) != 1)
1370         goto out;
1371 
1372     if (!trylock_page(page))
1373         goto out;
1374 
1375     /*
1376      * If user want to discard part-pages of THP, split it so MADV_FREE
1377      * will deactivate only them.
1378      */
1379     if (next - addr != HPAGE_PMD_SIZE) {
1380         get_page(page);
1381         spin_unlock(ptl);
1382         split_huge_page(page);
1383         put_page(page);
1384         unlock_page(page);
1385         goto out_unlocked;
1386     }
1387 
1388     if (PageDirty(page))
1389         ClearPageDirty(page);
1390     unlock_page(page);
1391 
1392     if (PageActive(page))
1393         deactivate_page(page);
1394 
1395     if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1396         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1397             tlb->fullmm);
1398         orig_pmd = pmd_mkold(orig_pmd);
1399         orig_pmd = pmd_mkclean(orig_pmd);
1400 
1401         set_pmd_at(mm, addr, pmd, orig_pmd);
1402         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1403     }
1404     ret = true;
1405 out:
1406     spin_unlock(ptl);
1407 out_unlocked:
1408     return ret;
1409 }
1410 
1411 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1412 {
1413     pgtable_t pgtable;
1414 
1415     pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1416     pte_free(mm, pgtable);
1417     atomic_long_dec(&mm->nr_ptes);
1418 }
1419 
1420 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1421          pmd_t *pmd, unsigned long addr)
1422 {
1423     pmd_t orig_pmd;
1424     spinlock_t *ptl;
1425 
1426     tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1427 
1428     ptl = __pmd_trans_huge_lock(pmd, vma);
1429     if (!ptl)
1430         return 0;
1431     /*
1432      * For architectures like ppc64 we look at deposited pgtable
1433      * when calling pmdp_huge_get_and_clear. So do the
1434      * pgtable_trans_huge_withdraw after finishing pmdp related
1435      * operations.
1436      */
1437     orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1438             tlb->fullmm);
1439     tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1440     if (vma_is_dax(vma)) {
1441         spin_unlock(ptl);
1442         if (is_huge_zero_pmd(orig_pmd))
1443             tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1444     } else if (is_huge_zero_pmd(orig_pmd)) {
1445         pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1446         atomic_long_dec(&tlb->mm->nr_ptes);
1447         spin_unlock(ptl);
1448         tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1449     } else {
1450         struct page *page = pmd_page(orig_pmd);
1451         page_remove_rmap(page, true);
1452         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1453         VM_BUG_ON_PAGE(!PageHead(page), page);
1454         if (PageAnon(page)) {
1455             pgtable_t pgtable;
1456             pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1457             pte_free(tlb->mm, pgtable);
1458             atomic_long_dec(&tlb->mm->nr_ptes);
1459             add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1460         } else {
1461             if (arch_needs_pgtable_deposit())
1462                 zap_deposited_table(tlb->mm, pmd);
1463             add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1464         }
1465         spin_unlock(ptl);
1466         tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1467     }
1468     return 1;
1469 }
1470 
1471 #ifndef pmd_move_must_withdraw
1472 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1473                      spinlock_t *old_pmd_ptl,
1474                      struct vm_area_struct *vma)
1475 {
1476     /*
1477      * With split pmd lock we also need to move preallocated
1478      * PTE page table if new_pmd is on different PMD page table.
1479      *
1480      * We also don't deposit and withdraw tables for file pages.
1481      */
1482     return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1483 }
1484 #endif
1485 
1486 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1487           unsigned long new_addr, unsigned long old_end,
1488           pmd_t *old_pmd, pmd_t *new_pmd, bool *need_flush)
1489 {
1490     spinlock_t *old_ptl, *new_ptl;
1491     pmd_t pmd;
1492     struct mm_struct *mm = vma->vm_mm;
1493     bool force_flush = false;
1494 
1495     if ((old_addr & ~HPAGE_PMD_MASK) ||
1496         (new_addr & ~HPAGE_PMD_MASK) ||
1497         old_end - old_addr < HPAGE_PMD_SIZE)
1498         return false;
1499 
1500     /*
1501      * The destination pmd shouldn't be established, free_pgtables()
1502      * should have release it.
1503      */
1504     if (WARN_ON(!pmd_none(*new_pmd))) {
1505         VM_BUG_ON(pmd_trans_huge(*new_pmd));
1506         return false;
1507     }
1508 
1509     /*
1510      * We don't have to worry about the ordering of src and dst
1511      * ptlocks because exclusive mmap_sem prevents deadlock.
1512      */
1513     old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1514     if (old_ptl) {
1515         new_ptl = pmd_lockptr(mm, new_pmd);
1516         if (new_ptl != old_ptl)
1517             spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1518         pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1519         if (pmd_present(pmd) && pmd_dirty(pmd))
1520             force_flush = true;
1521         VM_BUG_ON(!pmd_none(*new_pmd));
1522 
1523         if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1524             pgtable_t pgtable;
1525             pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1526             pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1527         }
1528         set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1529         if (new_ptl != old_ptl)
1530             spin_unlock(new_ptl);
1531         if (force_flush)
1532             flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1533         else
1534             *need_flush = true;
1535         spin_unlock(old_ptl);
1536         return true;
1537     }
1538     return false;
1539 }
1540 
1541 /*
1542  * Returns
1543  *  - 0 if PMD could not be locked
1544  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1545  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1546  */
1547 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1548         unsigned long addr, pgprot_t newprot, int prot_numa)
1549 {
1550     struct mm_struct *mm = vma->vm_mm;
1551     spinlock_t *ptl;
1552     int ret = 0;
1553 
1554     ptl = __pmd_trans_huge_lock(pmd, vma);
1555     if (ptl) {
1556         pmd_t entry;
1557         bool preserve_write = prot_numa && pmd_write(*pmd);
1558         ret = 1;
1559 
1560         /*
1561          * Avoid trapping faults against the zero page. The read-only
1562          * data is likely to be read-cached on the local CPU and
1563          * local/remote hits to the zero page are not interesting.
1564          */
1565         if (prot_numa && is_huge_zero_pmd(*pmd)) {
1566             spin_unlock(ptl);
1567             return ret;
1568         }
1569 
1570         if (!prot_numa || !pmd_protnone(*pmd)) {
1571             entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
1572             entry = pmd_modify(entry, newprot);
1573             if (preserve_write)
1574                 entry = pmd_mkwrite(entry);
1575             ret = HPAGE_PMD_NR;
1576             set_pmd_at(mm, addr, pmd, entry);
1577             BUG_ON(vma_is_anonymous(vma) && !preserve_write &&
1578                     pmd_write(entry));
1579         }
1580         spin_unlock(ptl);
1581     }
1582 
1583     return ret;
1584 }
1585 
1586 /*
1587  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1588  *
1589  * Note that if it returns page table lock pointer, this routine returns without
1590  * unlocking page table lock. So callers must unlock it.
1591  */
1592 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1593 {
1594     spinlock_t *ptl;
1595     ptl = pmd_lock(vma->vm_mm, pmd);
1596     if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd)))
1597         return ptl;
1598     spin_unlock(ptl);
1599     return NULL;
1600 }
1601 
1602 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1603         unsigned long haddr, pmd_t *pmd)
1604 {
1605     struct mm_struct *mm = vma->vm_mm;
1606     pgtable_t pgtable;
1607     pmd_t _pmd;
1608     int i;
1609 
1610     /* leave pmd empty until pte is filled */
1611     pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1612 
1613     pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1614     pmd_populate(mm, &_pmd, pgtable);
1615 
1616     for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1617         pte_t *pte, entry;
1618         entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
1619         entry = pte_mkspecial(entry);
1620         pte = pte_offset_map(&_pmd, haddr);
1621         VM_BUG_ON(!pte_none(*pte));
1622         set_pte_at(mm, haddr, pte, entry);
1623         pte_unmap(pte);
1624     }
1625     smp_wmb(); /* make pte visible before pmd */
1626     pmd_populate(mm, pmd, pgtable);
1627 }
1628 
1629 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
1630         unsigned long haddr, bool freeze)
1631 {
1632     struct mm_struct *mm = vma->vm_mm;
1633     struct page *page;
1634     pgtable_t pgtable;
1635     pmd_t _pmd;
1636     bool young, write, dirty, soft_dirty;
1637     unsigned long addr;
1638     int i;
1639 
1640     VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
1641     VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1642     VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
1643     VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd));
1644 
1645     count_vm_event(THP_SPLIT_PMD);
1646 
1647     if (!vma_is_anonymous(vma)) {
1648         _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1649         /*
1650          * We are going to unmap this huge page. So
1651          * just go ahead and zap it
1652          */
1653         if (arch_needs_pgtable_deposit())
1654             zap_deposited_table(mm, pmd);
1655         if (vma_is_dax(vma))
1656             return;
1657         page = pmd_page(_pmd);
1658         if (!PageReferenced(page) && pmd_young(_pmd))
1659             SetPageReferenced(page);
1660         page_remove_rmap(page, true);
1661         put_page(page);
1662         add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1663         return;
1664     } else if (is_huge_zero_pmd(*pmd)) {
1665         return __split_huge_zero_page_pmd(vma, haddr, pmd);
1666     }
1667 
1668     page = pmd_page(*pmd);
1669     VM_BUG_ON_PAGE(!page_count(page), page);
1670     page_ref_add(page, HPAGE_PMD_NR - 1);
1671     write = pmd_write(*pmd);
1672     young = pmd_young(*pmd);
1673     dirty = pmd_dirty(*pmd);
1674     soft_dirty = pmd_soft_dirty(*pmd);
1675 
1676     pmdp_huge_split_prepare(vma, haddr, pmd);
1677     pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1678     pmd_populate(mm, &_pmd, pgtable);
1679 
1680     for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
1681         pte_t entry, *pte;
1682         /*
1683          * Note that NUMA hinting access restrictions are not
1684          * transferred to avoid any possibility of altering
1685          * permissions across VMAs.
1686          */
1687         if (freeze) {
1688             swp_entry_t swp_entry;
1689             swp_entry = make_migration_entry(page + i, write);
1690             entry = swp_entry_to_pte(swp_entry);
1691             if (soft_dirty)
1692                 entry = pte_swp_mksoft_dirty(entry);
1693         } else {
1694             entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
1695             entry = maybe_mkwrite(entry, vma);
1696             if (!write)
1697                 entry = pte_wrprotect(entry);
1698             if (!young)
1699                 entry = pte_mkold(entry);
1700             if (soft_dirty)
1701                 entry = pte_mksoft_dirty(entry);
1702         }
1703         if (dirty)
1704             SetPageDirty(page + i);
1705         pte = pte_offset_map(&_pmd, addr);
1706         BUG_ON(!pte_none(*pte));
1707         set_pte_at(mm, addr, pte, entry);
1708         atomic_inc(&page[i]._mapcount);
1709         pte_unmap(pte);
1710     }
1711 
1712     /*
1713      * Set PG_double_map before dropping compound_mapcount to avoid
1714      * false-negative page_mapped().
1715      */
1716     if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
1717         for (i = 0; i < HPAGE_PMD_NR; i++)
1718             atomic_inc(&page[i]._mapcount);
1719     }
1720 
1721     if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
1722         /* Last compound_mapcount is gone. */
1723         __dec_node_page_state(page, NR_ANON_THPS);
1724         if (TestClearPageDoubleMap(page)) {
1725             /* No need in mapcount reference anymore */
1726             for (i = 0; i < HPAGE_PMD_NR; i++)
1727                 atomic_dec(&page[i]._mapcount);
1728         }
1729     }
1730 
1731     smp_wmb(); /* make pte visible before pmd */
1732     /*
1733      * Up to this point the pmd is present and huge and userland has the
1734      * whole access to the hugepage during the split (which happens in
1735      * place). If we overwrite the pmd with the not-huge version pointing
1736      * to the pte here (which of course we could if all CPUs were bug
1737      * free), userland could trigger a small page size TLB miss on the
1738      * small sized TLB while the hugepage TLB entry is still established in
1739      * the huge TLB. Some CPU doesn't like that.
1740      * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
1741      * 383 on page 93. Intel should be safe but is also warns that it's
1742      * only safe if the permission and cache attributes of the two entries
1743      * loaded in the two TLB is identical (which should be the case here).
1744      * But it is generally safer to never allow small and huge TLB entries
1745      * for the same virtual address to be loaded simultaneously. So instead
1746      * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
1747      * current pmd notpresent (atomically because here the pmd_trans_huge
1748      * and pmd_trans_splitting must remain set at all times on the pmd
1749      * until the split is complete for this pmd), then we flush the SMP TLB
1750      * and finally we write the non-huge version of the pmd entry with
1751      * pmd_populate.
1752      */
1753     pmdp_invalidate(vma, haddr, pmd);
1754     pmd_populate(mm, pmd, pgtable);
1755 
1756     if (freeze) {
1757         for (i = 0; i < HPAGE_PMD_NR; i++) {
1758             page_remove_rmap(page + i, false);
1759             put_page(page + i);
1760         }
1761     }
1762 }
1763 
1764 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1765         unsigned long address, bool freeze, struct page *page)
1766 {
1767     spinlock_t *ptl;
1768     struct mm_struct *mm = vma->vm_mm;
1769     unsigned long haddr = address & HPAGE_PMD_MASK;
1770 
1771     mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
1772     ptl = pmd_lock(mm, pmd);
1773 
1774     /*
1775      * If caller asks to setup a migration entries, we need a page to check
1776      * pmd against. Otherwise we can end up replacing wrong page.
1777      */
1778     VM_BUG_ON(freeze && !page);
1779     if (page && page != pmd_page(*pmd))
1780             goto out;
1781 
1782     if (pmd_trans_huge(*pmd)) {
1783         page = pmd_page(*pmd);
1784         if (PageMlocked(page))
1785             clear_page_mlock(page);
1786     } else if (!pmd_devmap(*pmd))
1787         goto out;
1788     __split_huge_pmd_locked(vma, pmd, haddr, freeze);
1789 out:
1790     spin_unlock(ptl);
1791     mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
1792 }
1793 
1794 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
1795         bool freeze, struct page *page)
1796 {
1797     pgd_t *pgd;
1798     pud_t *pud;
1799     pmd_t *pmd;
1800 
1801     pgd = pgd_offset(vma->vm_mm, address);
1802     if (!pgd_present(*pgd))
1803         return;
1804 
1805     pud = pud_offset(pgd, address);
1806     if (!pud_present(*pud))
1807         return;
1808 
1809     pmd = pmd_offset(pud, address);
1810 
1811     __split_huge_pmd(vma, pmd, address, freeze, page);
1812 }
1813 
1814 void vma_adjust_trans_huge(struct vm_area_struct *vma,
1815                  unsigned long start,
1816                  unsigned long end,
1817                  long adjust_next)
1818 {
1819     /*
1820      * If the new start address isn't hpage aligned and it could
1821      * previously contain an hugepage: check if we need to split
1822      * an huge pmd.
1823      */
1824     if (start & ~HPAGE_PMD_MASK &&
1825         (start & HPAGE_PMD_MASK) >= vma->vm_start &&
1826         (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
1827         split_huge_pmd_address(vma, start, false, NULL);
1828 
1829     /*
1830      * If the new end address isn't hpage aligned and it could
1831      * previously contain an hugepage: check if we need to split
1832      * an huge pmd.
1833      */
1834     if (end & ~HPAGE_PMD_MASK &&
1835         (end & HPAGE_PMD_MASK) >= vma->vm_start &&
1836         (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
1837         split_huge_pmd_address(vma, end, false, NULL);
1838 
1839     /*
1840      * If we're also updating the vma->vm_next->vm_start, if the new
1841      * vm_next->vm_start isn't page aligned and it could previously
1842      * contain an hugepage: check if we need to split an huge pmd.
1843      */
1844     if (adjust_next > 0) {
1845         struct vm_area_struct *next = vma->vm_next;
1846         unsigned long nstart = next->vm_start;
1847         nstart += adjust_next << PAGE_SHIFT;
1848         if (nstart & ~HPAGE_PMD_MASK &&
1849             (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
1850             (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
1851             split_huge_pmd_address(next, nstart, false, NULL);
1852     }
1853 }
1854 
1855 static void freeze_page(struct page *page)
1856 {
1857     enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
1858         TTU_RMAP_LOCKED;
1859     int i, ret;
1860 
1861     VM_BUG_ON_PAGE(!PageHead(page), page);
1862 
1863     if (PageAnon(page))
1864         ttu_flags |= TTU_MIGRATION;
1865 
1866     /* We only need TTU_SPLIT_HUGE_PMD once */
1867     ret = try_to_unmap(page, ttu_flags | TTU_SPLIT_HUGE_PMD);
1868     for (i = 1; !ret && i < HPAGE_PMD_NR; i++) {
1869         /* Cut short if the page is unmapped */
1870         if (page_count(page) == 1)
1871             return;
1872 
1873         ret = try_to_unmap(page + i, ttu_flags);
1874     }
1875     VM_BUG_ON_PAGE(ret, page + i - 1);
1876 }
1877 
1878 static void unfreeze_page(struct page *page)
1879 {
1880     int i;
1881 
1882     for (i = 0; i < HPAGE_PMD_NR; i++)
1883         remove_migration_ptes(page + i, page + i, true);
1884 }
1885 
1886 static void __split_huge_page_tail(struct page *head, int tail,
1887         struct lruvec *lruvec, struct list_head *list)
1888 {
1889     struct page *page_tail = head + tail;
1890 
1891     VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
1892     VM_BUG_ON_PAGE(page_ref_count(page_tail) != 0, page_tail);
1893 
1894     /*
1895      * tail_page->_refcount is zero and not changing from under us. But
1896      * get_page_unless_zero() may be running from under us on the
1897      * tail_page. If we used atomic_set() below instead of atomic_inc() or
1898      * atomic_add(), we would then run atomic_set() concurrently with
1899      * get_page_unless_zero(), and atomic_set() is implemented in C not
1900      * using locked ops. spin_unlock on x86 sometime uses locked ops
1901      * because of PPro errata 66, 92, so unless somebody can guarantee
1902      * atomic_set() here would be safe on all archs (and not only on x86),
1903      * it's safer to use atomic_inc()/atomic_add().
1904      */
1905     if (PageAnon(head)) {
1906         page_ref_inc(page_tail);
1907     } else {
1908         /* Additional pin to radix tree */
1909         page_ref_add(page_tail, 2);
1910     }
1911 
1912     page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1913     page_tail->flags |= (head->flags &
1914             ((1L << PG_referenced) |
1915              (1L << PG_swapbacked) |
1916              (1L << PG_mlocked) |
1917              (1L << PG_uptodate) |
1918              (1L << PG_active) |
1919              (1L << PG_locked) |
1920              (1L << PG_unevictable) |
1921              (1L << PG_dirty)));
1922 
1923     /*
1924      * After clearing PageTail the gup refcount can be released.
1925      * Page flags also must be visible before we make the page non-compound.
1926      */
1927     smp_wmb();
1928 
1929     clear_compound_head(page_tail);
1930 
1931     if (page_is_young(head))
1932         set_page_young(page_tail);
1933     if (page_is_idle(head))
1934         set_page_idle(page_tail);
1935 
1936     /* ->mapping in first tail page is compound_mapcount */
1937     VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
1938             page_tail);
1939     page_tail->mapping = head->mapping;
1940 
1941     page_tail->index = head->index + tail;
1942     page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
1943     lru_add_page_tail(head, page_tail, lruvec, list);
1944 }
1945 
1946 static void __split_huge_page(struct page *page, struct list_head *list,
1947         unsigned long flags)
1948 {
1949     struct page *head = compound_head(page);
1950     struct zone *zone = page_zone(head);
1951     struct lruvec *lruvec;
1952     pgoff_t end = -1;
1953     int i;
1954 
1955     lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat);
1956 
1957     /* complete memcg works before add pages to LRU */
1958     mem_cgroup_split_huge_fixup(head);
1959 
1960     if (!PageAnon(page))
1961         end = DIV_ROUND_UP(i_size_read(head->mapping->host), PAGE_SIZE);
1962 
1963     for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1964         __split_huge_page_tail(head, i, lruvec, list);
1965         /* Some pages can be beyond i_size: drop them from page cache */
1966         if (head[i].index >= end) {
1967             __ClearPageDirty(head + i);
1968             __delete_from_page_cache(head + i, NULL);
1969             if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
1970                 shmem_uncharge(head->mapping->host, 1);
1971             put_page(head + i);
1972         }
1973     }
1974 
1975     ClearPageCompound(head);
1976     /* See comment in __split_huge_page_tail() */
1977     if (PageAnon(head)) {
1978         page_ref_inc(head);
1979     } else {
1980         /* Additional pin to radix tree */
1981         page_ref_add(head, 2);
1982         spin_unlock(&head->mapping->tree_lock);
1983     }
1984 
1985     spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
1986 
1987     unfreeze_page(head);
1988 
1989     for (i = 0; i < HPAGE_PMD_NR; i++) {
1990         struct page *subpage = head + i;
1991         if (subpage == page)
1992             continue;
1993         unlock_page(subpage);
1994 
1995         /*
1996          * Subpages may be freed if there wasn't any mapping
1997          * like if add_to_swap() is running on a lru page that
1998          * had its mapping zapped. And freeing these pages
1999          * requires taking the lru_lock so we do the put_page
2000          * of the tail pages after the split is complete.
2001          */
2002         put_page(subpage);
2003     }
2004 }
2005 
2006 int total_mapcount(struct page *page)
2007 {
2008     int i, compound, ret;
2009 
2010     VM_BUG_ON_PAGE(PageTail(page), page);
2011 
2012     if (likely(!PageCompound(page)))
2013         return atomic_read(&page->_mapcount) + 1;
2014 
2015     compound = compound_mapcount(page);
2016     if (PageHuge(page))
2017         return compound;
2018     ret = compound;
2019     for (i = 0; i < HPAGE_PMD_NR; i++)
2020         ret += atomic_read(&page[i]._mapcount) + 1;
2021     /* File pages has compound_mapcount included in _mapcount */
2022     if (!PageAnon(page))
2023         return ret - compound * HPAGE_PMD_NR;
2024     if (PageDoubleMap(page))
2025         ret -= HPAGE_PMD_NR;
2026     return ret;
2027 }
2028 
2029 /*
2030  * This calculates accurately how many mappings a transparent hugepage
2031  * has (unlike page_mapcount() which isn't fully accurate). This full
2032  * accuracy is primarily needed to know if copy-on-write faults can
2033  * reuse the page and change the mapping to read-write instead of
2034  * copying them. At the same time this returns the total_mapcount too.
2035  *
2036  * The function returns the highest mapcount any one of the subpages
2037  * has. If the return value is one, even if different processes are
2038  * mapping different subpages of the transparent hugepage, they can
2039  * all reuse it, because each process is reusing a different subpage.
2040  *
2041  * The total_mapcount is instead counting all virtual mappings of the
2042  * subpages. If the total_mapcount is equal to "one", it tells the
2043  * caller all mappings belong to the same "mm" and in turn the
2044  * anon_vma of the transparent hugepage can become the vma->anon_vma
2045  * local one as no other process may be mapping any of the subpages.
2046  *
2047  * It would be more accurate to replace page_mapcount() with
2048  * page_trans_huge_mapcount(), however we only use
2049  * page_trans_huge_mapcount() in the copy-on-write faults where we
2050  * need full accuracy to avoid breaking page pinning, because
2051  * page_trans_huge_mapcount() is slower than page_mapcount().
2052  */
2053 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2054 {
2055     int i, ret, _total_mapcount, mapcount;
2056 
2057     /* hugetlbfs shouldn't call it */
2058     VM_BUG_ON_PAGE(PageHuge(page), page);
2059 
2060     if (likely(!PageTransCompound(page))) {
2061         mapcount = atomic_read(&page->_mapcount) + 1;
2062         if (total_mapcount)
2063             *total_mapcount = mapcount;
2064         return mapcount;
2065     }
2066 
2067     page = compound_head(page);
2068 
2069     _total_mapcount = ret = 0;
2070     for (i = 0; i < HPAGE_PMD_NR; i++) {
2071         mapcount = atomic_read(&page[i]._mapcount) + 1;
2072         ret = max(ret, mapcount);
2073         _total_mapcount += mapcount;
2074     }
2075     if (PageDoubleMap(page)) {
2076         ret -= 1;
2077         _total_mapcount -= HPAGE_PMD_NR;
2078     }
2079     mapcount = compound_mapcount(page);
2080     ret += mapcount;
2081     _total_mapcount += mapcount;
2082     if (total_mapcount)
2083         *total_mapcount = _total_mapcount;
2084     return ret;
2085 }
2086 
2087 /*
2088  * This function splits huge page into normal pages. @page can point to any
2089  * subpage of huge page to split. Split doesn't change the position of @page.
2090  *
2091  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2092  * The huge page must be locked.
2093  *
2094  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2095  *
2096  * Both head page and tail pages will inherit mapping, flags, and so on from
2097  * the hugepage.
2098  *
2099  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2100  * they are not mapped.
2101  *
2102  * Returns 0 if the hugepage is split successfully.
2103  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2104  * us.
2105  */
2106 int split_huge_page_to_list(struct page *page, struct list_head *list)
2107 {
2108     struct page *head = compound_head(page);
2109     struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2110     struct anon_vma *anon_vma = NULL;
2111     struct address_space *mapping = NULL;
2112     int count, mapcount, extra_pins, ret;
2113     bool mlocked;
2114     unsigned long flags;
2115 
2116     VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
2117     VM_BUG_ON_PAGE(!PageLocked(page), page);
2118     VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2119     VM_BUG_ON_PAGE(!PageCompound(page), page);
2120 
2121     if (PageAnon(head)) {
2122         /*
2123          * The caller does not necessarily hold an mmap_sem that would
2124          * prevent the anon_vma disappearing so we first we take a
2125          * reference to it and then lock the anon_vma for write. This
2126          * is similar to page_lock_anon_vma_read except the write lock
2127          * is taken to serialise against parallel split or collapse
2128          * operations.
2129          */
2130         anon_vma = page_get_anon_vma(head);
2131         if (!anon_vma) {
2132             ret = -EBUSY;
2133             goto out;
2134         }
2135         extra_pins = 0;
2136         mapping = NULL;
2137         anon_vma_lock_write(anon_vma);
2138     } else {
2139         mapping = head->mapping;
2140 
2141         /* Truncated ? */
2142         if (!mapping) {
2143             ret = -EBUSY;
2144             goto out;
2145         }
2146 
2147         /* Addidional pins from radix tree */
2148         extra_pins = HPAGE_PMD_NR;
2149         anon_vma = NULL;
2150         i_mmap_lock_read(mapping);
2151     }
2152 
2153     /*
2154      * Racy check if we can split the page, before freeze_page() will
2155      * split PMDs
2156      */
2157     if (total_mapcount(head) != page_count(head) - extra_pins - 1) {
2158         ret = -EBUSY;
2159         goto out_unlock;
2160     }
2161 
2162     mlocked = PageMlocked(page);
2163     freeze_page(head);
2164     VM_BUG_ON_PAGE(compound_mapcount(head), head);
2165 
2166     /* Make sure the page is not on per-CPU pagevec as it takes pin */
2167     if (mlocked)
2168         lru_add_drain();
2169 
2170     /* prevent PageLRU to go away from under us, and freeze lru stats */
2171     spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags);
2172 
2173     if (mapping) {
2174         void **pslot;
2175 
2176         spin_lock(&mapping->tree_lock);
2177         pslot = radix_tree_lookup_slot(&mapping->page_tree,
2178                 page_index(head));
2179         /*
2180          * Check if the head page is present in radix tree.
2181          * We assume all tail are present too, if head is there.
2182          */
2183         if (radix_tree_deref_slot_protected(pslot,
2184                     &mapping->tree_lock) != head)
2185             goto fail;
2186     }
2187 
2188     /* Prevent deferred_split_scan() touching ->_refcount */
2189     spin_lock(&pgdata->split_queue_lock);
2190     count = page_count(head);
2191     mapcount = total_mapcount(head);
2192     if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2193         if (!list_empty(page_deferred_list(head))) {
2194             pgdata->split_queue_len--;
2195             list_del(page_deferred_list(head));
2196         }
2197         if (mapping)
2198             __dec_node_page_state(page, NR_SHMEM_THPS);
2199         spin_unlock(&pgdata->split_queue_lock);
2200         __split_huge_page(page, list, flags);
2201         ret = 0;
2202     } else {
2203         if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2204             pr_alert("total_mapcount: %u, page_count(): %u\n",
2205                     mapcount, count);
2206             if (PageTail(page))
2207                 dump_page(head, NULL);
2208             dump_page(page, "total_mapcount(head) > 0");
2209             BUG();
2210         }
2211         spin_unlock(&pgdata->split_queue_lock);
2212 fail:       if (mapping)
2213             spin_unlock(&mapping->tree_lock);
2214         spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
2215         unfreeze_page(head);
2216         ret = -EBUSY;
2217     }
2218 
2219 out_unlock:
2220     if (anon_vma) {
2221         anon_vma_unlock_write(anon_vma);
2222         put_anon_vma(anon_vma);
2223     }
2224     if (mapping)
2225         i_mmap_unlock_read(mapping);
2226 out:
2227     count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2228     return ret;
2229 }
2230 
2231 void free_transhuge_page(struct page *page)
2232 {
2233     struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2234     unsigned long flags;
2235 
2236     spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2237     if (!list_empty(page_deferred_list(page))) {
2238         pgdata->split_queue_len--;
2239         list_del(page_deferred_list(page));
2240     }
2241     spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2242     free_compound_page(page);
2243 }
2244 
2245 void deferred_split_huge_page(struct page *page)
2246 {
2247     struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2248     unsigned long flags;
2249 
2250     VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2251 
2252     spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2253     if (list_empty(page_deferred_list(page))) {
2254         count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2255         list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2256         pgdata->split_queue_len++;
2257     }
2258     spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2259 }
2260 
2261 static unsigned long deferred_split_count(struct shrinker *shrink,
2262         struct shrink_control *sc)
2263 {
2264     struct pglist_data *pgdata = NODE_DATA(sc->nid);
2265     return ACCESS_ONCE(pgdata->split_queue_len);
2266 }
2267 
2268 static unsigned long deferred_split_scan(struct shrinker *shrink,
2269         struct shrink_control *sc)
2270 {
2271     struct pglist_data *pgdata = NODE_DATA(sc->nid);
2272     unsigned long flags;
2273     LIST_HEAD(list), *pos, *next;
2274     struct page *page;
2275     int split = 0;
2276 
2277     spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2278     /* Take pin on all head pages to avoid freeing them under us */
2279     list_for_each_safe(pos, next, &pgdata->split_queue) {
2280         page = list_entry((void *)pos, struct page, mapping);
2281         page = compound_head(page);
2282         if (get_page_unless_zero(page)) {
2283             list_move(page_deferred_list(page), &list);
2284         } else {
2285             /* We lost race with put_compound_page() */
2286             list_del_init(page_deferred_list(page));
2287             pgdata->split_queue_len--;
2288         }
2289         if (!--sc->nr_to_scan)
2290             break;
2291     }
2292     spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2293 
2294     list_for_each_safe(pos, next, &list) {
2295         page = list_entry((void *)pos, struct page, mapping);
2296         lock_page(page);
2297         /* split_huge_page() removes page from list on success */
2298         if (!split_huge_page(page))
2299             split++;
2300         unlock_page(page);
2301         put_page(page);
2302     }
2303 
2304     spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2305     list_splice_tail(&list, &pgdata->split_queue);
2306     spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2307 
2308     /*
2309      * Stop shrinker if we didn't split any page, but the queue is empty.
2310      * This can happen if pages were freed under us.
2311      */
2312     if (!split && list_empty(&pgdata->split_queue))
2313         return SHRINK_STOP;
2314     return split;
2315 }
2316 
2317 static struct shrinker deferred_split_shrinker = {
2318     .count_objects = deferred_split_count,
2319     .scan_objects = deferred_split_scan,
2320     .seeks = DEFAULT_SEEKS,
2321     .flags = SHRINKER_NUMA_AWARE,
2322 };
2323 
2324 #ifdef CONFIG_DEBUG_FS
2325 static int split_huge_pages_set(void *data, u64 val)
2326 {
2327     struct zone *zone;
2328     struct page *page;
2329     unsigned long pfn, max_zone_pfn;
2330     unsigned long total = 0, split = 0;
2331 
2332     if (val != 1)
2333         return -EINVAL;
2334 
2335     for_each_populated_zone(zone) {
2336         max_zone_pfn = zone_end_pfn(zone);
2337         for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2338             if (!pfn_valid(pfn))
2339                 continue;
2340 
2341             page = pfn_to_page(pfn);
2342             if (!get_page_unless_zero(page))
2343                 continue;
2344 
2345             if (zone != page_zone(page))
2346                 goto next;
2347 
2348             if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2349                 goto next;
2350 
2351             total++;
2352             lock_page(page);
2353             if (!split_huge_page(page))
2354                 split++;
2355             unlock_page(page);
2356 next:
2357             put_page(page);
2358         }
2359     }
2360 
2361     pr_info("%lu of %lu THP split\n", split, total);
2362 
2363     return 0;
2364 }
2365 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2366         "%llu\n");
2367 
2368 static int __init split_huge_pages_debugfs(void)
2369 {
2370     void *ret;
2371 
2372     ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2373             &split_huge_pages_fops);
2374     if (!ret)
2375         pr_warn("Failed to create split_huge_pages in debugfs");
2376     return 0;
2377 }
2378 late_initcall(split_huge_pages_debugfs);
2379 #endif