mm/huge_memory.c

   1 /*
   2  *  Copyright (C) 2009  Red Hat, Inc.
   3  *
   4  *  This work is licensed under the terms of the GNU GPL, version 2. See
   5  *  the COPYING file in the top-level directory.
   6  */
   7
   8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   9
  10 #include <linux/mm.h>
  11 #include <linux/sched.h>
  12 #include <linux/highmem.h>
  13 #include <linux/hugetlb.h>
  14 #include <linux/mmu_notifier.h>
  15 #include <linux/rmap.h>
  16 #include <linux/swap.h>
  17 #include <linux/shrinker.h>
  18 #include <linux/mm_inline.h>
  19 #include <linux/swapops.h>
  20 #include <linux/dax.h>
  21 #include <linux/khugepaged.h>
  22 #include <linux/freezer.h>
  23 #include <linux/pfn_t.h>
  24 #include <linux/mman.h>
  25 #include <linux/memremap.h>
  26 #include <linux/pagemap.h>
  27 #include <linux/debugfs.h>
  28 #include <linux/migrate.h>
  29 #include <linux/hashtable.h>
  30 #include <linux/userfaultfd_k.h>
  31 #include <linux/page_idle.h>
  32 #include <linux/shmem_fs.h>
  33
  34 #include <asm/tlb.h>
  35 #include <asm/pgalloc.h>
  36 #include "internal.h"
  37
  38 /*
  39  * By default transparent hugepage support is disabled in order that avoid
  40  * to risk increase the memory footprint of applications without a guaranteed
  41  * benefit. When transparent hugepage support is enabled, is for all mappings,
  42  * and khugepaged scans all mappings.
  43  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  44  * for all hugepage allocations.
  45  */
  46 unsigned long transparent_hugepage_flags __read_mostly =
  47 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  48         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  49 #endif
  50 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  51         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  52 #endif
  53         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
  54         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  55         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  56
  57 static struct shrinker deferred_split_shrinker;
  58
  59 static atomic_t huge_zero_refcount;
  60 struct page *huge_zero_page __read_mostly;
  61
  62 static struct page *get_huge_zero_page(void)
  63 {
  64         struct page *zero_page;
  65 retry:
  66         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
  67                 return READ_ONCE(huge_zero_page);
  68
  69         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
  70                         HPAGE_PMD_ORDER);
  71         if (!zero_page) {
  72                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
  73                 return NULL;
  74         }
  75         count_vm_event(THP_ZERO_PAGE_ALLOC);
  76         preempt_disable();
  77         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
  78                 preempt_enable();
  79                 __free_pages(zero_page, compound_order(zero_page));
  80                 goto retry;
  81         }
  82
  83         /* We take additional reference here. It will be put back by shrinker */
  84         atomic_set(&huge_zero_refcount, 2);
  85         preempt_enable();
  86         return READ_ONCE(huge_zero_page);
  87 }
  88
  89 static void put_huge_zero_page(void)
  90 {
  91         /*
  92          * Counter should never go to zero here. Only shrinker can put
  93          * last reference.
  94          */
  95         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
  96 }
  97
  98 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
  99 {
 100         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 101                 return READ_ONCE(huge_zero_page);
 102
 103         if (!get_huge_zero_page())
 104                 return NULL;
 105
 106         if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 107                 put_huge_zero_page();
 108
 109         return READ_ONCE(huge_zero_page);
 110 }
 111
 112 void mm_put_huge_zero_page(struct mm_struct *mm)
 113 {
 114         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 115                 put_huge_zero_page();
 116 }
 117
 118 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 119                                         struct shrink_control *sc)
 120 {
 121         /* we can free zero page only if last reference remains */
 122         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 123 }
 124
 125 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 126                                        struct shrink_control *sc)
 127 {
 128         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 129                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 130                 BUG_ON(zero_page == NULL);
 131                 __free_pages(zero_page, compound_order(zero_page));
 132                 return HPAGE_PMD_NR;
 133         }
 134
 135         return 0;
 136 }
 137
 138 static struct shrinker huge_zero_page_shrinker = {
 139         .count_objects = shrink_huge_zero_page_count,
 140         .scan_objects = shrink_huge_zero_page_scan,
 141         .seeks = DEFAULT_SEEKS,
 142 };
 143
 144 #ifdef CONFIG_SYSFS
 145
 146 static ssize_t triple_flag_store(struct kobject *kobj,
 147                                  struct kobj_attribute *attr,
 148                                  const char *buf, size_t count,
 149                                  enum transparent_hugepage_flag enabled,
 150                                  enum transparent_hugepage_flag deferred,
 151                                  enum transparent_hugepage_flag req_madv)
 152 {
 153         if (!memcmp("defer", buf,
 154                     min(sizeof("defer")-1, count))) {
 155                 if (enabled == deferred)
 156                         return -EINVAL;
 157                 clear_bit(enabled, &transparent_hugepage_flags);
 158                 clear_bit(req_madv, &transparent_hugepage_flags);
 159                 set_bit(deferred, &transparent_hugepage_flags);
 160         } else if (!memcmp("always", buf,
 161                     min(sizeof("always")-1, count))) {
 162                 clear_bit(deferred, &transparent_hugepage_flags);
 163                 clear_bit(req_madv, &transparent_hugepage_flags);
 164                 set_bit(enabled, &transparent_hugepage_flags);
 165         } else if (!memcmp("madvise", buf,
 166                            min(sizeof("madvise")-1, count))) {
 167                 clear_bit(enabled, &transparent_hugepage_flags);
 168                 clear_bit(deferred, &transparent_hugepage_flags);
 169                 set_bit(req_madv, &transparent_hugepage_flags);
 170         } else if (!memcmp("never", buf,
 171                            min(sizeof("never")-1, count))) {
 172                 clear_bit(enabled, &transparent_hugepage_flags);
 173                 clear_bit(req_madv, &transparent_hugepage_flags);
 174                 clear_bit(deferred, &transparent_hugepage_flags);
 175         } else
 176                 return -EINVAL;
 177
 178         return count;
 179 }
 180
 181 static ssize_t enabled_show(struct kobject *kobj,
 182                             struct kobj_attribute *attr, char *buf)
 183 {
 184         if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
 185                 return sprintf(buf, "[always] madvise never\n");
 186         else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
 187                 return sprintf(buf, "always [madvise] never\n");
 188         else
 189                 return sprintf(buf, "always madvise [never]\n");
 190 }
 191
 192 static ssize_t enabled_store(struct kobject *kobj,
 193                              struct kobj_attribute *attr,
 194                              const char *buf, size_t count)
 195 {
 196         ssize_t ret;
 197
 198         ret = triple_flag_store(kobj, attr, buf, count,
 199                                 TRANSPARENT_HUGEPAGE_FLAG,
 200                                 TRANSPARENT_HUGEPAGE_FLAG,
 201                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 202
 203         if (ret > 0) {
 204                 int err = start_stop_khugepaged();
 205                 if (err)
 206                         ret = err;
 207         }
 208
 209         return ret;
 210 }
 211 static struct kobj_attribute enabled_attr =
 212         __ATTR(enabled, 0644, enabled_show, enabled_store);
 213
 214 ssize_t single_hugepage_flag_show(struct kobject *kobj,
 215                                 struct kobj_attribute *attr, char *buf,
 216                                 enum transparent_hugepage_flag flag)
 217 {
 218         return sprintf(buf, "%d\n",
 219                        !!test_bit(flag, &transparent_hugepage_flags));
 220 }
 221
 222 ssize_t single_hugepage_flag_store(struct kobject *kobj,
 223                                  struct kobj_attribute *attr,
 224                                  const char *buf, size_t count,
 225                                  enum transparent_hugepage_flag flag)
 226 {
 227         unsigned long value;
 228         int ret;
 229
 230         ret = kstrtoul(buf, 10, &value);
 231         if (ret < 0)
 232                 return ret;
 233         if (value > 1)
 234                 return -EINVAL;
 235
 236         if (value)
 237                 set_bit(flag, &transparent_hugepage_flags);
 238         else
 239                 clear_bit(flag, &transparent_hugepage_flags);
 240
 241         return count;
 242 }
 243
 244 /*
 245  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 246  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 247  * memory just to allocate one more hugepage.
 248  */
 249 static ssize_t defrag_show(struct kobject *kobj,
 250                            struct kobj_attribute *attr, char *buf)
 251 {
 252         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 253                 return sprintf(buf, "[always] defer madvise never\n");
 254         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 255                 return sprintf(buf, "always [defer] madvise never\n");
 256         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 257                 return sprintf(buf, "always defer [madvise] never\n");
 258         else
 259                 return sprintf(buf, "always defer madvise [never]\n");
 260
 261 }
 262 static ssize_t defrag_store(struct kobject *kobj,
 263                             struct kobj_attribute *attr,
 264                             const char *buf, size_t count)
 265 {
 266         return triple_flag_store(kobj, attr, buf, count,
 267                                  TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
 268                                  TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
 269                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 270 }
 271 static struct kobj_attribute defrag_attr =
 272         __ATTR(defrag, 0644, defrag_show, defrag_store);
 273
 274 static ssize_t use_zero_page_show(struct kobject *kobj,
 275                 struct kobj_attribute *attr, char *buf)
 276 {
 277         return single_hugepage_flag_show(kobj, attr, buf,
 278                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 279 }
 280 static ssize_t use_zero_page_store(struct kobject *kobj,
 281                 struct kobj_attribute *attr, const char *buf, size_t count)
 282 {
 283         return single_hugepage_flag_store(kobj, attr, buf, count,
 284                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 285 }
 286 static struct kobj_attribute use_zero_page_attr =
 287         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 288 #ifdef CONFIG_DEBUG_VM
 289 static ssize_t debug_cow_show(struct kobject *kobj,
 290                                 struct kobj_attribute *attr, char *buf)
 291 {
 292         return single_hugepage_flag_show(kobj, attr, buf,
 293                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 294 }
 295 static ssize_t debug_cow_store(struct kobject *kobj,
 296                                struct kobj_attribute *attr,
 297                                const char *buf, size_t count)
 298 {
 299         return single_hugepage_flag_store(kobj, attr, buf, count,
 300                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 301 }
 302 static struct kobj_attribute debug_cow_attr =
 303         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 304 #endif /* CONFIG_DEBUG_VM */
 305
 306 static struct attribute *hugepage_attr[] = {
 307         &enabled_attr.attr,
 308         &defrag_attr.attr,
 309         &use_zero_page_attr.attr,
 310 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
 311         &shmem_enabled_attr.attr,
 312 #endif
 313 #ifdef CONFIG_DEBUG_VM
 314         &debug_cow_attr.attr,
 315 #endif
 316         NULL,
 317 };
 318
 319 static struct attribute_group hugepage_attr_group = {
 320         .attrs = hugepage_attr,
 321 };
 322
 323 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 324 {
 325         int err;
 326
 327         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 328         if (unlikely(!*hugepage_kobj)) {
 329                 pr_err("failed to create transparent hugepage kobject\n");
 330                 return -ENOMEM;
 331         }
 332
 333         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 334         if (err) {
 335                 pr_err("failed to register transparent hugepage group\n");
 336                 goto delete_obj;
 337         }
 338
 339         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 340         if (err) {
 341                 pr_err("failed to register transparent hugepage group\n");
 342                 goto remove_hp_group;
 343         }
 344
 345         return 0;
 346
 347 remove_hp_group:
 348         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 349 delete_obj:
 350         kobject_put(*hugepage_kobj);
 351         return err;
 352 }
 353
 354 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 355 {
 356         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 357         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 358         kobject_put(hugepage_kobj);
 359 }
 360 #else
 361 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 362 {
 363         return 0;
 364 }
 365
 366 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 367 {
 368 }
 369 #endif /* CONFIG_SYSFS */
 370
 371 static int __init hugepage_init(void)
 372 {
 373         int err;
 374         struct kobject *hugepage_kobj;
 375
 376         if (!has_transparent_hugepage()) {
 377                 transparent_hugepage_flags = 0;
 378                 return -EINVAL;
 379         }
 380
 381         /*
 382          * hugepages can't be allocated by the buddy allocator
 383          */
 384         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
 385         /*
 386          * we use page->mapping and page->index in second tail page
 387          * as list_head: assuming THP order >= 2
 388          */
 389         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 390
 391         err = hugepage_init_sysfs(&hugepage_kobj);
 392         if (err)
 393                 goto err_sysfs;
 394
 395         err = khugepaged_init();
 396         if (err)
 397                 goto err_slab;
 398
 399         err = register_shrinker(&huge_zero_page_shrinker);
 400         if (err)
 401                 goto err_hzp_shrinker;
 402         err = register_shrinker(&deferred_split_shrinker);
 403         if (err)
 404                 goto err_split_shrinker;
 405
 406         /*
 407          * By default disable transparent hugepages on smaller systems,
 408          * where the extra memory used could hurt more than TLB overhead
 409          * is likely to save.  The admin can still enable it through /sys.
 410          */
 411         if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
 412                 transparent_hugepage_flags = 0;
 413                 return 0;
 414         }
 415
 416         err = start_stop_khugepaged();
 417         if (err)
 418                 goto err_khugepaged;
 419
 420         return 0;
 421 err_khugepaged:
 422         unregister_shrinker(&deferred_split_shrinker);
 423 err_split_shrinker:
 424         unregister_shrinker(&huge_zero_page_shrinker);
 425 err_hzp_shrinker:
 426         khugepaged_destroy();
 427 err_slab:
 428         hugepage_exit_sysfs(hugepage_kobj);
 429 err_sysfs:
 430         return err;
 431 }
 432 subsys_initcall(hugepage_init);
 433
 434 static int __init setup_transparent_hugepage(char *str)
 435 {
 436         int ret = 0;
 437         if (!str)
 438                 goto out;
 439         if (!strcmp(str, "always")) {
 440                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 441                         &transparent_hugepage_flags);
 442                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 443                           &transparent_hugepage_flags);
 444                 ret = 1;
 445         } else if (!strcmp(str, "madvise")) {
 446                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 447                           &transparent_hugepage_flags);
 448                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 449                         &transparent_hugepage_flags);
 450                 ret = 1;
 451         } else if (!strcmp(str, "never")) {
 452                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 453                           &transparent_hugepage_flags);
 454                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 455                           &transparent_hugepage_flags);
 456                 ret = 1;
 457         }
 458 out:
 459         if (!ret)
 460                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 461         return ret;
 462 }
 463 __setup("transparent_hugepage=", setup_transparent_hugepage);
 464
 465 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 466 {
 467         if (likely(vma->vm_flags & VM_WRITE))
 468                 pmd = pmd_mkwrite(pmd);
 469         return pmd;
 470 }
 471
 472 static inline struct list_head *page_deferred_list(struct page *page)
 473 {
 474         /*
 475          * ->lru in the tail pages is occupied by compound_head.
 476          * Let's use ->mapping + ->index in the second tail page as list_head.
 477          */
 478         return (struct list_head *)&page[2].mapping;
 479 }
 480
 481 void prep_transhuge_page(struct page *page)
 482 {
 483         /*
 484          * we use page->mapping and page->indexlru in second tail page
 485          * as list_head: assuming THP order >= 2
 486          */
 487
 488         INIT_LIST_HEAD(page_deferred_list(page));
 489         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 490 }
 491
 492 unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len,
 493                 loff_t off, unsigned long flags, unsigned long size)
 494 {
 495         unsigned long addr;
 496         loff_t off_end = off + len;
 497         loff_t off_align = round_up(off, size);
 498         unsigned long len_pad;
 499
 500         if (off_end <= off_align || (off_end - off_align) < size)
 501                 return 0;
 502
 503         len_pad = len + size;
 504         if (len_pad < len || (off + len_pad) < off)
 505                 return 0;
 506
 507         addr = current->mm->get_unmapped_area(filp, 0, len_pad,
 508                                               off >> PAGE_SHIFT, flags);
 509         if (IS_ERR_VALUE(addr))
 510                 return 0;
 511
 512         addr += (off - addr) & (size - 1);
 513         return addr;
 514 }
 515
 516 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 517                 unsigned long len, unsigned long pgoff, unsigned long flags)
 518 {
 519         loff_t off = (loff_t)pgoff << PAGE_SHIFT;
 520
 521         if (addr)
 522                 goto out;
 523         if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
 524                 goto out;
 525
 526         addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE);
 527         if (addr)
 528                 return addr;
 529
 530  out:
 531         return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
 532 }
 533 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
 534
 535 static int __do_huge_pmd_anonymous_page(struct fault_env *fe, struct page *page,
 536                 gfp_t gfp)
 537 {
 538         struct vm_area_struct *vma = fe->vma;
 539         struct mem_cgroup *memcg;
 540         pgtable_t pgtable;
 541         unsigned long haddr = fe->address & HPAGE_PMD_MASK;
 542
 543         VM_BUG_ON_PAGE(!PageCompound(page), page);
 544
 545         if (mem_cgroup_try_charge(page, vma->vm_mm, gfp, &memcg, true)) {
 546                 put_page(page);
 547                 count_vm_event(THP_FAULT_FALLBACK);
 548                 return VM_FAULT_FALLBACK;
 549         }
 550
 551         pgtable = pte_alloc_one(vma->vm_mm, haddr);
 552         if (unlikely(!pgtable)) {
 553                 mem_cgroup_cancel_charge(page, memcg, true);
 554                 put_page(page);
 555                 return VM_FAULT_OOM;
 556         }
 557
 558         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 559         /*
 560          * The memory barrier inside __SetPageUptodate makes sure that
 561          * clear_huge_page writes become visible before the set_pmd_at()
 562          * write.
 563          */
 564         __SetPageUptodate(page);
 565
 566         fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
 567         if (unlikely(!pmd_none(*fe->pmd))) {
 568                 spin_unlock(fe->ptl);
 569                 mem_cgroup_cancel_charge(page, memcg, true);
 570                 put_page(page);
 571                 pte_free(vma->vm_mm, pgtable);
 572         } else {
 573                 pmd_t entry;
 574
 575                 /* Deliver the page fault to userland */
 576                 if (userfaultfd_missing(vma)) {
 577                         int ret;
 578
 579                         spin_unlock(fe->ptl);
 580                         mem_cgroup_cancel_charge(page, memcg, true);
 581                         put_page(page);
 582                         pte_free(vma->vm_mm, pgtable);
 583                         ret = handle_userfault(fe, VM_UFFD_MISSING);
 584                         VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 585                         return ret;
 586                 }
 587
 588                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 589                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 590                 page_add_new_anon_rmap(page, vma, haddr, true);
 591                 mem_cgroup_commit_charge(page, memcg, false, true);
 592                 lru_cache_add_active_or_unevictable(page, vma);
 593                 pgtable_trans_huge_deposit(vma->vm_mm, fe->pmd, pgtable);
 594                 set_pmd_at(vma->vm_mm, haddr, fe->pmd, entry);
 595                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 596                 atomic_long_inc(&vma->vm_mm->nr_ptes);
 597                 spin_unlock(fe->ptl);
 598                 count_vm_event(THP_FAULT_ALLOC);
 599         }
 600
 601         return 0;
 602 }
 603
 604 /*
 605  * If THP defrag is set to always then directly reclaim/compact as necessary
 606  * If set to defer then do only background reclaim/compact and defer to khugepaged
 607  * If set to madvise and the VMA is flagged then directly reclaim/compact
 608  * When direct reclaim/compact is allowed, don't retry except for flagged VMA's
 609  */
 610 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
 611 {
 612         bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
 613
 614         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
 615                                 &transparent_hugepage_flags) && vma_madvised)
 616                 return GFP_TRANSHUGE;
 617         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
 618                                                 &transparent_hugepage_flags))
 619                 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
 620         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
 621                                                 &transparent_hugepage_flags))
 622                 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
 623
 624         return GFP_TRANSHUGE_LIGHT;
 625 }
 626
 627 /* Caller must hold page table lock. */
 628 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 629                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 630                 struct page *zero_page)
 631 {
 632         pmd_t entry;
 633         if (!pmd_none(*pmd))
 634                 return false;
 635         entry = mk_pmd(zero_page, vma->vm_page_prot);
 636         entry = pmd_mkhuge(entry);
 637         if (pgtable)
 638                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 639         set_pmd_at(mm, haddr, pmd, entry);
 640         atomic_long_inc(&mm->nr_ptes);
 641         return true;
 642 }
 643
 644 int do_huge_pmd_anonymous_page(struct fault_env *fe)
 645 {
 646         struct vm_area_struct *vma = fe->vma;
 647         gfp_t gfp;
 648         struct page *page;
 649         unsigned long haddr = fe->address & HPAGE_PMD_MASK;
 650
 651         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 652                 return VM_FAULT_FALLBACK;
 653         if (unlikely(anon_vma_prepare(vma)))
 654                 return VM_FAULT_OOM;
 655         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 656                 return VM_FAULT_OOM;
 657         if (!(fe->flags & FAULT_FLAG_WRITE) &&
 658                         !mm_forbids_zeropage(vma->vm_mm) &&
 659                         transparent_hugepage_use_zero_page()) {
 660                 pgtable_t pgtable;
 661                 struct page *zero_page;
 662                 bool set;
 663                 int ret;
 664                 pgtable = pte_alloc_one(vma->vm_mm, haddr);
 665                 if (unlikely(!pgtable))
 666                         return VM_FAULT_OOM;
 667                 zero_page = mm_get_huge_zero_page(vma->vm_mm);
 668                 if (unlikely(!zero_page)) {
 669                         pte_free(vma->vm_mm, pgtable);
 670                         count_vm_event(THP_FAULT_FALLBACK);
 671                         return VM_FAULT_FALLBACK;
 672                 }
 673                 fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
 674                 ret = 0;
 675                 set = false;
 676                 if (pmd_none(*fe->pmd)) {
 677                         if (userfaultfd_missing(vma)) {
 678                                 spin_unlock(fe->ptl);
 679                                 ret = handle_userfault(fe, VM_UFFD_MISSING);
 680                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 681                         } else {
 682                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
 683                                                    haddr, fe->pmd, zero_page);
 684                                 spin_unlock(fe->ptl);
 685                                 set = true;
 686                         }
 687                 } else
 688                         spin_unlock(fe->ptl);
 689                 if (!set)
 690                         pte_free(vma->vm_mm, pgtable);
 691                 return ret;
 692         }
 693         gfp = alloc_hugepage_direct_gfpmask(vma);
 694         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 695         if (unlikely(!page)) {
 696                 count_vm_event(THP_FAULT_FALLBACK);
 697                 return VM_FAULT_FALLBACK;
 698         }
 699         prep_transhuge_page(page);
 700         return __do_huge_pmd_anonymous_page(fe, page, gfp);
 701 }
 702
 703 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 704                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write)
 705 {
 706         struct mm_struct *mm = vma->vm_mm;
 707         pmd_t entry;
 708         spinlock_t *ptl;
 709
 710         ptl = pmd_lock(mm, pmd);
 711         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 712         if (pfn_t_devmap(pfn))
 713                 entry = pmd_mkdevmap(entry);
 714         if (write) {
 715                 entry = pmd_mkyoung(pmd_mkdirty(entry));
 716                 entry = maybe_pmd_mkwrite(entry, vma);
 717         }
 718         set_pmd_at(mm, addr, pmd, entry);
 719         update_mmu_cache_pmd(vma, addr, pmd);
 720         spin_unlock(ptl);
 721 }
 722
 723 int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 724                         pmd_t *pmd, pfn_t pfn, bool write)
 725 {
 726         pgprot_t pgprot = vma->vm_page_prot;
 727         /*
 728          * If we had pmd_special, we could avoid all these restrictions,
 729          * but we need to be consistent with PTEs and architectures that
 730          * can't support a 'special' bit.
 731          */
 732         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
 733         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 734                                                 (VM_PFNMAP|VM_MIXEDMAP));
 735         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 736         BUG_ON(!pfn_t_devmap(pfn));
 737
 738         if (addr < vma->vm_start || addr >= vma->vm_end)
 739                 return VM_FAULT_SIGBUS;
 740         if (track_pfn_insert(vma, &pgprot, pfn))
 741                 return VM_FAULT_SIGBUS;
 742         insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
 743         return VM_FAULT_NOPAGE;
 744 }
 745 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
 746
 747 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
 748                 pmd_t *pmd, int flags)
 749 {
 750         pmd_t _pmd;
 751
 752         _pmd = pmd_mkyoung(*pmd);
 753         if (flags & FOLL_WRITE)
 754                 _pmd = pmd_mkdirty(_pmd);
 755         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
 756                                 pmd, _pmd, flags & FOLL_WRITE))
 757                 update_mmu_cache_pmd(vma, addr, pmd);
 758 }
 759
 760 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 761                 pmd_t *pmd, int flags)
 762 {
 763         unsigned long pfn = pmd_pfn(*pmd);
 764         struct mm_struct *mm = vma->vm_mm;
 765         struct dev_pagemap *pgmap;
 766         struct page *page;
 767
 768         assert_spin_locked(pmd_lockptr(mm, pmd));
 769
 770         /*
 771          * When we COW a devmap PMD entry, we split it into PTEs, so we should
 772          * not be in this function with `flags & FOLL_COW` set.
 773          */
 774         WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
 775
 776         if (flags & FOLL_WRITE && !pmd_write(*pmd))
 777                 return NULL;
 778
 779         if (pmd_present(*pmd) && pmd_devmap(*pmd))
 780                 /* pass */;
 781         else
 782                 return NULL;
 783
 784         if (flags & FOLL_TOUCH)
 785                 touch_pmd(vma, addr, pmd, flags);
 786
 787         /*
 788          * device mapped pages can only be returned if the
 789          * caller will manage the page reference count.
 790          */
 791         if (!(flags & FOLL_GET))
 792                 return ERR_PTR(-EEXIST);
 793
 794         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
 795         pgmap = get_dev_pagemap(pfn, NULL);
 796         if (!pgmap)
 797                 return ERR_PTR(-EFAULT);
 798         page = pfn_to_page(pfn);
 799         get_page(page);
 800         put_dev_pagemap(pgmap);
 801
 802         return page;
 803 }
 804
 805 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 806                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 807                   struct vm_area_struct *vma)
 808 {
 809         spinlock_t *dst_ptl, *src_ptl;
 810         struct page *src_page;
 811         pmd_t pmd;
 812         pgtable_t pgtable = NULL;
 813         int ret = -ENOMEM;
 814
 815         /* Skip if can be re-fill on fault */
 816         if (!vma_is_anonymous(vma))
 817                 return 0;
 818
 819         pgtable = pte_alloc_one(dst_mm, addr);
 820         if (unlikely(!pgtable))
 821                 goto out;
 822
 823         dst_ptl = pmd_lock(dst_mm, dst_pmd);
 824         src_ptl = pmd_lockptr(src_mm, src_pmd);
 825         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 826
 827         ret = -EAGAIN;
 828         pmd = *src_pmd;
 829         if (unlikely(!pmd_trans_huge(pmd))) {
 830                 pte_free(dst_mm, pgtable);
 831                 goto out_unlock;
 832         }
 833         /*
 834          * When page table lock is held, the huge zero pmd should not be
 835          * under splitting since we don't split the page itself, only pmd to
 836          * a page table.
 837          */
 838         if (is_huge_zero_pmd(pmd)) {
 839                 struct page *zero_page;
 840                 /*
 841                  * get_huge_zero_page() will never allocate a new page here,
 842                  * since we already have a zero page to copy. It just takes a
 843                  * reference.
 844                  */
 845                 zero_page = mm_get_huge_zero_page(dst_mm);
 846                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
 847                                 zero_page);
 848                 ret = 0;
 849                 goto out_unlock;
 850         }
 851
 852         src_page = pmd_page(pmd);
 853         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
 854         get_page(src_page);
 855         page_dup_rmap(src_page, true);
 856         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 857         atomic_long_inc(&dst_mm->nr_ptes);
 858         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
 859
 860         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 861         pmd = pmd_mkold(pmd_wrprotect(pmd));
 862         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 863
 864         ret = 0;
 865 out_unlock:
 866         spin_unlock(src_ptl);
 867         spin_unlock(dst_ptl);
 868 out:
 869         return ret;
 870 }
 871
 872 void huge_pmd_set_accessed(struct fault_env *fe, pmd_t orig_pmd)
 873 {
 874         pmd_t entry;
 875         unsigned long haddr;
 876         bool write = fe->flags & FAULT_FLAG_WRITE;
 877
 878         fe->ptl = pmd_lock(fe->vma->vm_mm, fe->pmd);
 879         if (unlikely(!pmd_same(*fe->pmd, orig_pmd)))
 880                 goto unlock;
 881
 882         entry = pmd_mkyoung(orig_pmd);
 883         if (write)
 884                 entry = pmd_mkdirty(entry);
 885         haddr = fe->address & HPAGE_PMD_MASK;
 886         if (pmdp_set_access_flags(fe->vma, haddr, fe->pmd, entry, write))
 887                 update_mmu_cache_pmd(fe->vma, fe->address, fe->pmd);
 888
 889 unlock:
 890         spin_unlock(fe->ptl);
 891 }
 892
 893 static int do_huge_pmd_wp_page_fallback(struct fault_env *fe, pmd_t orig_pmd,
 894                 struct page *page)
 895 {
 896         struct vm_area_struct *vma = fe->vma;
 897         unsigned long haddr = fe->address & HPAGE_PMD_MASK;
 898         struct mem_cgroup *memcg;
 899         pgtable_t pgtable;
 900         pmd_t _pmd;
 901         int ret = 0, i;
 902         struct page **pages;
 903         unsigned long mmun_start;       /* For mmu_notifiers */
 904         unsigned long mmun_end;         /* For mmu_notifiers */
 905
 906         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
 907                         GFP_KERNEL);
 908         if (unlikely(!pages)) {
 909                 ret |= VM_FAULT_OOM;
 910                 goto out;
 911         }
 912
 913         for (i = 0; i < HPAGE_PMD_NR; i++) {
 914                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
 915                                                __GFP_OTHER_NODE, vma,
 916                                                fe->address, page_to_nid(page));
 917                 if (unlikely(!pages[i] ||
 918                              mem_cgroup_try_charge(pages[i], vma->vm_mm,
 919                                      GFP_KERNEL, &memcg, false))) {
 920                         if (pages[i])
 921                                 put_page(pages[i]);
 922                         while (--i >= 0) {
 923                                 memcg = (void *)page_private(pages[i]);
 924                                 set_page_private(pages[i], 0);
 925                                 mem_cgroup_cancel_charge(pages[i], memcg,
 926                                                 false);
 927                                 put_page(pages[i]);
 928                         }
 929                         kfree(pages);
 930                         ret |= VM_FAULT_OOM;
 931                         goto out;
 932                 }
 933                 set_page_private(pages[i], (unsigned long)memcg);
 934         }
 935
 936         for (i = 0; i < HPAGE_PMD_NR; i++) {
 937                 copy_user_highpage(pages[i], page + i,
 938                                    haddr + PAGE_SIZE * i, vma);
 939                 __SetPageUptodate(pages[i]);
 940                 cond_resched();
 941         }
 942
 943         mmun_start = haddr;
 944         mmun_end   = haddr + HPAGE_PMD_SIZE;
 945         mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
 946
 947         fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
 948         if (unlikely(!pmd_same(*fe->pmd, orig_pmd)))
 949                 goto out_free_pages;
 950         VM_BUG_ON_PAGE(!PageHead(page), page);
 951
 952         pmdp_huge_clear_flush_notify(vma, haddr, fe->pmd);
 953         /* leave pmd empty until pte is filled */
 954
 955         pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, fe->pmd);
 956         pmd_populate(vma->vm_mm, &_pmd, pgtable);
 957
 958         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
 959                 pte_t entry;
 960                 entry = mk_pte(pages[i], vma->vm_page_prot);
 961                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
 962                 memcg = (void *)page_private(pages[i]);
 963                 set_page_private(pages[i], 0);
 964                 page_add_new_anon_rmap(pages[i], fe->vma, haddr, false);
 965                 mem_cgroup_commit_charge(pages[i], memcg, false, false);
 966                 lru_cache_add_active_or_unevictable(pages[i], vma);
 967                 fe->pte = pte_offset_map(&_pmd, haddr);
 968                 VM_BUG_ON(!pte_none(*fe->pte));
 969                 set_pte_at(vma->vm_mm, haddr, fe->pte, entry);
 970                 pte_unmap(fe->pte);
 971         }
 972         kfree(pages);
 973
 974         smp_wmb(); /* make pte visible before pmd */
 975         pmd_populate(vma->vm_mm, fe->pmd, pgtable);
 976         page_remove_rmap(page, true);
 977         spin_unlock(fe->ptl);
 978
 979         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
 980
 981         ret |= VM_FAULT_WRITE;
 982         put_page(page);
 983
 984 out:
 985         return ret;
 986
 987 out_free_pages:
 988         spin_unlock(fe->ptl);
 989         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
 990         for (i = 0; i < HPAGE_PMD_NR; i++) {
 991                 memcg = (void *)page_private(pages[i]);
 992                 set_page_private(pages[i], 0);
 993                 mem_cgroup_cancel_charge(pages[i], memcg, false);
 994                 put_page(pages[i]);
 995         }
 996         kfree(pages);
 997         goto out;
 998 }
 999
1000 int do_huge_pmd_wp_page(struct fault_env *fe, pmd_t orig_pmd)
1001 {
1002         struct vm_area_struct *vma = fe->vma;
1003         struct page *page = NULL, *new_page;
1004         struct mem_cgroup *memcg;
1005         unsigned long haddr = fe->address & HPAGE_PMD_MASK;
1006         unsigned long mmun_start;       /* For mmu_notifiers */
1007         unsigned long mmun_end;         /* For mmu_notifiers */
1008         gfp_t huge_gfp;                 /* for allocation and charge */
1009         int ret = 0;
1010
1011         fe->ptl = pmd_lockptr(vma->vm_mm, fe->pmd);
1012         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1013         if (is_huge_zero_pmd(orig_pmd))
1014                 goto alloc;
1015         spin_lock(fe->ptl);
1016         if (unlikely(!pmd_same(*fe->pmd, orig_pmd)))
1017                 goto out_unlock;
1018
1019         page = pmd_page(orig_pmd);
1020         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1021         /*
1022          * We can only reuse the page if nobody else maps the huge page or it's
1023          * part.
1024          */
1025         if (page_trans_huge_mapcount(page, NULL) == 1) {
1026                 pmd_t entry;
1027                 entry = pmd_mkyoung(orig_pmd);
1028                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1029                 if (pmdp_set_access_flags(vma, haddr, fe->pmd, entry,  1))
1030                         update_mmu_cache_pmd(vma, fe->address, fe->pmd);
1031                 ret |= VM_FAULT_WRITE;
1032                 goto out_unlock;
1033         }
1034         get_page(page);
1035         spin_unlock(fe->ptl);
1036 alloc:
1037         if (transparent_hugepage_enabled(vma) &&
1038             !transparent_hugepage_debug_cow()) {
1039                 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1040                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1041         } else
1042                 new_page = NULL;
1043
1044         if (likely(new_page)) {
1045                 prep_transhuge_page(new_page);
1046         } else {
1047                 if (!page) {
1048                         split_huge_pmd(vma, fe->pmd, fe->address);
1049                         ret |= VM_FAULT_FALLBACK;
1050                 } else {
1051                         ret = do_huge_pmd_wp_page_fallback(fe, orig_pmd, page);
1052                         if (ret & VM_FAULT_OOM) {
1053                                 split_huge_pmd(vma, fe->pmd, fe->address);
1054                                 ret |= VM_FAULT_FALLBACK;
1055                         }
1056                         put_page(page);
1057                 }
1058                 count_vm_event(THP_FAULT_FALLBACK);
1059                 goto out;
1060         }
1061
1062         if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm,
1063                                         huge_gfp, &memcg, true))) {
1064                 put_page(new_page);
1065                 split_huge_pmd(vma, fe->pmd, fe->address);
1066                 if (page)
1067                         put_page(page);
1068                 ret |= VM_FAULT_FALLBACK;
1069                 count_vm_event(THP_FAULT_FALLBACK);
1070                 goto out;
1071         }
1072
1073         count_vm_event(THP_FAULT_ALLOC);
1074
1075         if (!page)
1076                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1077         else
1078                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1079         __SetPageUptodate(new_page);
1080
1081         mmun_start = haddr;
1082         mmun_end   = haddr + HPAGE_PMD_SIZE;
1083         mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
1084
1085         spin_lock(fe->ptl);
1086         if (page)
1087                 put_page(page);
1088         if (unlikely(!pmd_same(*fe->pmd, orig_pmd))) {
1089                 spin_unlock(fe->ptl);
1090                 mem_cgroup_cancel_charge(new_page, memcg, true);
1091                 put_page(new_page);
1092                 goto out_mn;
1093         } else {
1094                 pmd_t entry;
1095                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1096                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1097                 pmdp_huge_clear_flush_notify(vma, haddr, fe->pmd);
1098                 page_add_new_anon_rmap(new_page, vma, haddr, true);
1099                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1100                 lru_cache_add_active_or_unevictable(new_page, vma);
1101                 set_pmd_at(vma->vm_mm, haddr, fe->pmd, entry);
1102                 update_mmu_cache_pmd(vma, fe->address, fe->pmd);
1103                 if (!page) {
1104                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1105                 } else {
1106                         VM_BUG_ON_PAGE(!PageHead(page), page);
1107                         page_remove_rmap(page, true);
1108                         put_page(page);
1109                 }
1110                 ret |= VM_FAULT_WRITE;
1111         }
1112         spin_unlock(fe->ptl);
1113 out_mn:
1114         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1115 out:
1116         return ret;
1117 out_unlock:
1118         spin_unlock(fe->ptl);
1119         return ret;
1120 }
1121
1122 /*
1123  * FOLL_FORCE can write to even unwritable pmd's, but only
1124  * after we've gone through a COW cycle and they are dirty.
1125  */
1126 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1127 {
1128         return pmd_write(pmd) ||
1129                ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1130 }
1131
1132 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1133                                    unsigned long addr,
1134                                    pmd_t *pmd,
1135                                    unsigned int flags)
1136 {
1137         struct mm_struct *mm = vma->vm_mm;
1138         struct page *page = NULL;
1139
1140         assert_spin_locked(pmd_lockptr(mm, pmd));
1141
1142         if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1143                 goto out;
1144
1145         /* Avoid dumping huge zero page */
1146         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1147                 return ERR_PTR(-EFAULT);
1148
1149         /* Full NUMA hinting faults to serialise migration in fault paths */
1150         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1151                 goto out;
1152
1153         page = pmd_page(*pmd);
1154         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1155         if (flags & FOLL_TOUCH)
1156                 touch_pmd(vma, addr, pmd, flags);
1157         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1158                 /*
1159                  * We don't mlock() pte-mapped THPs. This way we can avoid
1160                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1161                  *
1162                  * For anon THP:
1163                  *
1164                  * In most cases the pmd is the only mapping of the page as we
1165                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1166                  * writable private mappings in populate_vma_page_range().
1167                  *
1168                  * The only scenario when we have the page shared here is if we
1169                  * mlocking read-only mapping shared over fork(). We skip
1170                  * mlocking such pages.
1171                  *
1172                  * For file THP:
1173                  *
1174                  * We can expect PageDoubleMap() to be stable under page lock:
1175                  * for file pages we set it in page_add_file_rmap(), which
1176                  * requires page to be locked.
1177                  */
1178
1179                 if (PageAnon(page) && compound_mapcount(page) != 1)
1180                         goto skip_mlock;
1181                 if (PageDoubleMap(page) || !page->mapping)
1182                         goto skip_mlock;
1183                 if (!trylock_page(page))
1184                         goto skip_mlock;
1185                 lru_add_drain();
1186                 if (page->mapping && !PageDoubleMap(page))
1187                         mlock_vma_page(page);
1188                 unlock_page(page);
1189         }
1190 skip_mlock:
1191         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1192         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1193         if (flags & FOLL_GET)
1194                 get_page(page);
1195
1196 out:
1197         return page;
1198 }
1199
1200 /* NUMA hinting page fault entry point for trans huge pmds */
1201 int do_huge_pmd_numa_page(struct fault_env *fe, pmd_t pmd)
1202 {
1203         struct vm_area_struct *vma = fe->vma;
1204         struct anon_vma *anon_vma = NULL;
1205         struct page *page;
1206         unsigned long haddr = fe->address & HPAGE_PMD_MASK;
1207         int page_nid = -1, this_nid = numa_node_id();
1208         int target_nid, last_cpupid = -1;
1209         bool page_locked;
1210         bool migrated = false;
1211         bool was_writable;
1212         int flags = 0;
1213
1214         fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
1215         if (unlikely(!pmd_same(pmd, *fe->pmd)))
1216                 goto out_unlock;
1217
1218         /*
1219          * If there are potential migrations, wait for completion and retry
1220          * without disrupting NUMA hinting information. Do not relock and
1221          * check_same as the page may no longer be mapped.
1222          */
1223         if (unlikely(pmd_trans_migrating(*fe->pmd))) {
1224                 page = pmd_page(*fe->pmd);
1225                 if (!get_page_unless_zero(page))
1226                         goto out_unlock;
1227                 spin_unlock(fe->ptl);
1228                 wait_on_page_locked(page);
1229                 put_page(page);
1230                 goto out;
1231         }
1232
1233         page = pmd_page(pmd);
1234         BUG_ON(is_huge_zero_page(page));
1235         page_nid = page_to_nid(page);
1236         last_cpupid = page_cpupid_last(page);
1237         count_vm_numa_event(NUMA_HINT_FAULTS);
1238         if (page_nid == this_nid) {
1239                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1240                 flags |= TNF_FAULT_LOCAL;
1241         }
1242
1243         /* See similar comment in do_numa_page for explanation */
1244         if (!pmd_write(pmd))
1245                 flags |= TNF_NO_GROUP;
1246
1247         /*
1248          * Acquire the page lock to serialise THP migrations but avoid dropping
1249          * page_table_lock if at all possible
1250          */
1251         page_locked = trylock_page(page);
1252         target_nid = mpol_misplaced(page, vma, haddr);
1253         if (target_nid == -1) {
1254                 /* If the page was locked, there are no parallel migrations */
1255                 if (page_locked)
1256                         goto clear_pmdnuma;
1257         }
1258
1259         /* Migration could have started since the pmd_trans_migrating check */
1260         if (!page_locked) {
1261                 if (!get_page_unless_zero(page))
1262                         goto out_unlock;
1263                 spin_unlock(fe->ptl);
1264                 wait_on_page_locked(page);
1265                 put_page(page);
1266                 page_nid = -1;
1267                 goto out;
1268         }
1269
1270         /*
1271          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1272          * to serialises splits
1273          */
1274         get_page(page);
1275         spin_unlock(fe->ptl);
1276         anon_vma = page_lock_anon_vma_read(page);
1277
1278         /* Confirm the PMD did not change while page_table_lock was released */
1279         spin_lock(fe->ptl);
1280         if (unlikely(!pmd_same(pmd, *fe->pmd))) {
1281                 unlock_page(page);
1282                 put_page(page);
1283                 page_nid = -1;
1284                 goto out_unlock;
1285         }
1286
1287         /* Bail if we fail to protect against THP splits for any reason */
1288         if (unlikely(!anon_vma)) {
1289                 put_page(page);
1290                 page_nid = -1;
1291                 goto clear_pmdnuma;
1292         }
1293
1294         /*
1295          * Migrate the THP to the requested node, returns with page unlocked
1296          * and access rights restored.
1297          */
1298         spin_unlock(fe->ptl);
1299         migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1300                                 fe->pmd, pmd, fe->address, page, target_nid);
1301         if (migrated) {
1302                 flags |= TNF_MIGRATED;
1303                 page_nid = target_nid;
1304         } else
1305                 flags |= TNF_MIGRATE_FAIL;
1306
1307         goto out;
1308 clear_pmdnuma:
1309         BUG_ON(!PageLocked(page));
1310         was_writable = pmd_write(pmd);
1311         pmd = pmd_modify(pmd, vma->vm_page_prot);
1312         pmd = pmd_mkyoung(pmd);
1313         if (was_writable)
1314                 pmd = pmd_mkwrite(pmd);
1315         set_pmd_at(vma->vm_mm, haddr, fe->pmd, pmd);
1316         update_mmu_cache_pmd(vma, fe->address, fe->pmd);
1317         unlock_page(page);
1318 out_unlock:
1319         spin_unlock(fe->ptl);
1320
1321 out:
1322         if (anon_vma)
1323                 page_unlock_anon_vma_read(anon_vma);
1324
1325         if (page_nid != -1)
1326                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, fe->flags);
1327
1328         return 0;
1329 }
1330
1331 /*
1332  * Return true if we do MADV_FREE successfully on entire pmd page.
1333  * Otherwise, return false.
1334  */
1335 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1336                 pmd_t *pmd, unsigned long addr, unsigned long next)
1337 {
1338         spinlock_t *ptl;
1339         pmd_t orig_pmd;
1340         struct page *page;
1341         struct mm_struct *mm = tlb->mm;
1342         bool ret = false;
1343
1344         ptl = pmd_trans_huge_lock(pmd, vma);
1345         if (!ptl)
1346                 goto out_unlocked;
1347
1348         orig_pmd = *pmd;
1349         if (is_huge_zero_pmd(orig_pmd))
1350                 goto out;
1351
1352         page = pmd_page(orig_pmd);
1353         /*
1354          * If other processes are mapping this page, we couldn't discard
1355          * the page unless they all do MADV_FREE so let's skip the page.
1356          */
1357         if (page_mapcount(page) != 1)
1358                 goto out;
1359
1360         if (!trylock_page(page))
1361                 goto out;
1362
1363         /*
1364          * If user want to discard part-pages of THP, split it so MADV_FREE
1365          * will deactivate only them.
1366          */
1367         if (next - addr != HPAGE_PMD_SIZE) {
1368                 get_page(page);
1369                 spin_unlock(ptl);
1370                 split_huge_page(page);
1371                 unlock_page(page);
1372                 put_page(page);
1373                 goto out_unlocked;
1374         }
1375
1376         if (PageDirty(page))
1377                 ClearPageDirty(page);
1378         unlock_page(page);
1379
1380         if (PageActive(page))
1381                 deactivate_page(page);
1382
1383         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1384                 pmdp_invalidate(vma, addr, pmd);
1385                 orig_pmd = pmd_mkold(orig_pmd);
1386                 orig_pmd = pmd_mkclean(orig_pmd);
1387
1388                 set_pmd_at(mm, addr, pmd, orig_pmd);
1389                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1390         }
1391         ret = true;
1392 out:
1393         spin_unlock(ptl);
1394 out_unlocked:
1395         return ret;
1396 }
1397
1398 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1399                  pmd_t *pmd, unsigned long addr)
1400 {
1401         pmd_t orig_pmd;
1402         spinlock_t *ptl;
1403
1404         ptl = __pmd_trans_huge_lock(pmd, vma);
1405         if (!ptl)
1406                 return 0;
1407         /*
1408          * For architectures like ppc64 we look at deposited pgtable
1409          * when calling pmdp_huge_get_and_clear. So do the
1410          * pgtable_trans_huge_withdraw after finishing pmdp related
1411          * operations.
1412          */
1413         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1414                         tlb->fullmm);
1415         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1416         if (vma_is_dax(vma)) {
1417                 spin_unlock(ptl);
1418                 if (is_huge_zero_pmd(orig_pmd))
1419                         tlb_remove_page(tlb, pmd_page(orig_pmd));
1420         } else if (is_huge_zero_pmd(orig_pmd)) {
1421                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1422                 atomic_long_dec(&tlb->mm->nr_ptes);
1423                 spin_unlock(ptl);
1424                 tlb_remove_page(tlb, pmd_page(orig_pmd));
1425         } else {
1426                 struct page *page = pmd_page(orig_pmd);
1427                 page_remove_rmap(page, true);
1428                 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1429                 VM_BUG_ON_PAGE(!PageHead(page), page);
1430                 if (PageAnon(page)) {
1431                         pgtable_t pgtable;
1432                         pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1433                         pte_free(tlb->mm, pgtable);
1434                         atomic_long_dec(&tlb->mm->nr_ptes);
1435                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1436                 } else {
1437                         add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1438                 }
1439                 spin_unlock(ptl);
1440                 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1441         }
1442         return 1;
1443 }
1444
1445 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1446                   unsigned long new_addr, unsigned long old_end,
1447                   pmd_t *old_pmd, pmd_t *new_pmd, bool *need_flush)
1448 {
1449         spinlock_t *old_ptl, *new_ptl;
1450         pmd_t pmd;
1451         struct mm_struct *mm = vma->vm_mm;
1452         bool force_flush = false;
1453
1454         if ((old_addr & ~HPAGE_PMD_MASK) ||
1455             (new_addr & ~HPAGE_PMD_MASK) ||
1456             old_end - old_addr < HPAGE_PMD_SIZE)
1457                 return false;
1458
1459         /*
1460          * The destination pmd shouldn't be established, free_pgtables()
1461          * should have release it.
1462          */
1463         if (WARN_ON(!pmd_none(*new_pmd))) {
1464                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1465                 return false;
1466         }
1467
1468         /*
1469          * We don't have to worry about the ordering of src and dst
1470          * ptlocks because exclusive mmap_sem prevents deadlock.
1471          */
1472         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1473         if (old_ptl) {
1474                 new_ptl = pmd_lockptr(mm, new_pmd);
1475                 if (new_ptl != old_ptl)
1476                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1477                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1478                 if (pmd_present(pmd) && pmd_dirty(pmd))
1479                         force_flush = true;
1480                 VM_BUG_ON(!pmd_none(*new_pmd));
1481
1482                 if (pmd_move_must_withdraw(new_ptl, old_ptl) &&
1483                                 vma_is_anonymous(vma)) {
1484                         pgtable_t pgtable;
1485                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1486                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1487                 }
1488                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1489                 if (new_ptl != old_ptl)
1490                         spin_unlock(new_ptl);
1491                 if (force_flush)
1492                         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1493                 else
1494                         *need_flush = true;
1495                 spin_unlock(old_ptl);
1496                 return true;
1497         }
1498         return false;
1499 }
1500
1501 /*
1502  * Returns
1503  *  - 0 if PMD could not be locked
1504  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1505  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1506  */
1507 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1508                 unsigned long addr, pgprot_t newprot, int prot_numa)
1509 {
1510         struct mm_struct *mm = vma->vm_mm;
1511         spinlock_t *ptl;
1512         pmd_t entry;
1513         bool preserve_write;
1514         int ret;
1515
1516         ptl = __pmd_trans_huge_lock(pmd, vma);
1517         if (!ptl)
1518                 return 0;
1519
1520         preserve_write = prot_numa && pmd_write(*pmd);
1521         ret = 1;
1522
1523         /*
1524          * Avoid trapping faults against the zero page. The read-only
1525          * data is likely to be read-cached on the local CPU and
1526          * local/remote hits to the zero page are not interesting.
1527          */
1528         if (prot_numa && is_huge_zero_pmd(*pmd))
1529                 goto unlock;
1530
1531         if (prot_numa && pmd_protnone(*pmd))
1532                 goto unlock;
1533
1534         /*
1535          * In case prot_numa, we are under down_read(mmap_sem). It's critical
1536          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1537          * which is also under down_read(mmap_sem):
1538          *
1539          *      CPU0:                           CPU1:
1540          *                              change_huge_pmd(prot_numa=1)
1541          *                               pmdp_huge_get_and_clear_notify()
1542          * madvise_dontneed()
1543          *  zap_pmd_range()
1544          *   pmd_trans_huge(*pmd) == 0 (without ptl)
1545          *   // skip the pmd
1546          *                               set_pmd_at();
1547          *                               // pmd is re-established
1548          *
1549          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1550          * which may break userspace.
1551          *
1552          * pmdp_invalidate() is required to make sure we don't miss
1553          * dirty/young flags set by hardware.
1554          */
1555         entry = *pmd;
1556         pmdp_invalidate(vma, addr, pmd);
1557
1558         /*
1559          * Recover dirty/young flags.  It relies on pmdp_invalidate to not
1560          * corrupt them.
1561          */
1562         if (pmd_dirty(*pmd))
1563                 entry = pmd_mkdirty(entry);
1564         if (pmd_young(*pmd))
1565                 entry = pmd_mkyoung(entry);
1566
1567         entry = pmd_modify(entry, newprot);
1568         if (preserve_write)
1569                 entry = pmd_mkwrite(entry);
1570         ret = HPAGE_PMD_NR;
1571         set_pmd_at(mm, addr, pmd, entry);
1572         BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1573 unlock:
1574         spin_unlock(ptl);
1575         return ret;
1576 }
1577
1578 /*
1579  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1580  *
1581  * Note that if it returns page table lock pointer, this routine returns without
1582  * unlocking page table lock. So callers must unlock it.
1583  */
1584 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1585 {
1586         spinlock_t *ptl;
1587         ptl = pmd_lock(vma->vm_mm, pmd);
1588         if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd)))
1589                 return ptl;
1590         spin_unlock(ptl);
1591         return NULL;
1592 }
1593
1594 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1595                 unsigned long haddr, pmd_t *pmd)
1596 {
1597         struct mm_struct *mm = vma->vm_mm;
1598         pgtable_t pgtable;
1599         pmd_t _pmd;
1600         int i;
1601
1602         /* leave pmd empty until pte is filled */
1603         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1604
1605         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1606         pmd_populate(mm, &_pmd, pgtable);
1607
1608         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1609                 pte_t *pte, entry;
1610                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
1611                 entry = pte_mkspecial(entry);
1612                 pte = pte_offset_map(&_pmd, haddr);
1613                 VM_BUG_ON(!pte_none(*pte));
1614                 set_pte_at(mm, haddr, pte, entry);
1615                 pte_unmap(pte);
1616         }
1617         smp_wmb(); /* make pte visible before pmd */
1618         pmd_populate(mm, pmd, pgtable);
1619 }
1620
1621 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
1622                 unsigned long haddr, bool freeze)
1623 {
1624         struct mm_struct *mm = vma->vm_mm;
1625         struct page *page;
1626         pgtable_t pgtable;
1627         pmd_t _pmd;
1628         bool young, write, dirty, soft_dirty;
1629         unsigned long addr;
1630         int i;
1631
1632         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
1633         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1634         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
1635         VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd));
1636
1637         count_vm_event(THP_SPLIT_PMD);
1638
1639         if (!vma_is_anonymous(vma)) {
1640                 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1641                 if (vma_is_dax(vma))
1642                         return;
1643                 page = pmd_page(_pmd);
1644                 if (!PageReferenced(page) && pmd_young(_pmd))
1645                         SetPageReferenced(page);
1646                 page_remove_rmap(page, true);
1647                 put_page(page);
1648                 add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1649                 return;
1650         } else if (is_huge_zero_pmd(*pmd)) {
1651                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
1652         }
1653
1654         page = pmd_page(*pmd);
1655         VM_BUG_ON_PAGE(!page_count(page), page);
1656         page_ref_add(page, HPAGE_PMD_NR - 1);
1657         write = pmd_write(*pmd);
1658         young = pmd_young(*pmd);
1659         dirty = pmd_dirty(*pmd);
1660         soft_dirty = pmd_soft_dirty(*pmd);
1661
1662         pmdp_huge_split_prepare(vma, haddr, pmd);
1663         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1664         pmd_populate(mm, &_pmd, pgtable);
1665
1666         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
1667                 pte_t entry, *pte;
1668                 /*
1669                  * Note that NUMA hinting access restrictions are not
1670                  * transferred to avoid any possibility of altering
1671                  * permissions across VMAs.
1672                  */
1673                 if (freeze) {
1674                         swp_entry_t swp_entry;
1675                         swp_entry = make_migration_entry(page + i, write);
1676                         entry = swp_entry_to_pte(swp_entry);
1677                         if (soft_dirty)
1678                                 entry = pte_swp_mksoft_dirty(entry);
1679                 } else {
1680                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
1681                         entry = maybe_mkwrite(entry, vma);
1682                         if (!write)
1683                                 entry = pte_wrprotect(entry);
1684                         if (!young)
1685                                 entry = pte_mkold(entry);
1686                         if (soft_dirty)
1687                                 entry = pte_mksoft_dirty(entry);
1688                 }
1689                 if (dirty)
1690                         SetPageDirty(page + i);
1691                 pte = pte_offset_map(&_pmd, addr);
1692                 BUG_ON(!pte_none(*pte));
1693                 set_pte_at(mm, addr, pte, entry);
1694                 atomic_inc(&page[i]._mapcount);
1695                 pte_unmap(pte);
1696         }
1697
1698         /*
1699          * Set PG_double_map before dropping compound_mapcount to avoid
1700          * false-negative page_mapped().
1701          */
1702         if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
1703                 for (i = 0; i < HPAGE_PMD_NR; i++)
1704                         atomic_inc(&page[i]._mapcount);
1705         }
1706
1707         if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
1708                 /* Last compound_mapcount is gone. */
1709                 __dec_node_page_state(page, NR_ANON_THPS);
1710                 if (TestClearPageDoubleMap(page)) {
1711                         /* No need in mapcount reference anymore */
1712                         for (i = 0; i < HPAGE_PMD_NR; i++)
1713                                 atomic_dec(&page[i]._mapcount);
1714                 }
1715         }
1716
1717         smp_wmb(); /* make pte visible before pmd */
1718         /*
1719          * Up to this point the pmd is present and huge and userland has the
1720          * whole access to the hugepage during the split (which happens in
1721          * place). If we overwrite the pmd with the not-huge version pointing
1722          * to the pte here (which of course we could if all CPUs were bug
1723          * free), userland could trigger a small page size TLB miss on the
1724          * small sized TLB while the hugepage TLB entry is still established in
1725          * the huge TLB. Some CPU doesn't like that.
1726          * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
1727          * 383 on page 93. Intel should be safe but is also warns that it's
1728          * only safe if the permission and cache attributes of the two entries
1729          * loaded in the two TLB is identical (which should be the case here).
1730          * But it is generally safer to never allow small and huge TLB entries
1731          * for the same virtual address to be loaded simultaneously. So instead
1732          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
1733          * current pmd notpresent (atomically because here the pmd_trans_huge
1734          * and pmd_trans_splitting must remain set at all times on the pmd
1735          * until the split is complete for this pmd), then we flush the SMP TLB
1736          * and finally we write the non-huge version of the pmd entry with
1737          * pmd_populate.
1738          */
1739         pmdp_invalidate(vma, haddr, pmd);
1740         pmd_populate(mm, pmd, pgtable);
1741
1742         if (freeze) {
1743                 for (i = 0; i < HPAGE_PMD_NR; i++) {
1744                         page_remove_rmap(page + i, false);
1745                         put_page(page + i);
1746                 }
1747         }
1748 }
1749
1750 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1751                 unsigned long address, bool freeze, struct page *page)
1752 {
1753         spinlock_t *ptl;
1754         struct mm_struct *mm = vma->vm_mm;
1755         unsigned long haddr = address & HPAGE_PMD_MASK;
1756
1757         mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
1758         ptl = pmd_lock(mm, pmd);
1759
1760         /*
1761          * If caller asks to setup a migration entries, we need a page to check
1762          * pmd against. Otherwise we can end up replacing wrong page.
1763          */
1764         VM_BUG_ON(freeze && !page);
1765         if (page && page != pmd_page(*pmd))
1766                 goto out;
1767
1768         if (pmd_trans_huge(*pmd)) {
1769                 page = pmd_page(*pmd);
1770                 if (PageMlocked(page))
1771                         clear_page_mlock(page);
1772         } else if (!pmd_devmap(*pmd))
1773                 goto out;
1774         __split_huge_pmd_locked(vma, pmd, haddr, freeze);
1775 out:
1776         spin_unlock(ptl);
1777         mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
1778 }
1779
1780 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
1781                 bool freeze, struct page *page)
1782 {
1783         pgd_t *pgd;
1784         pud_t *pud;
1785         pmd_t *pmd;
1786
1787         pgd = pgd_offset(vma->vm_mm, address);
1788         if (!pgd_present(*pgd))
1789                 return;
1790
1791         pud = pud_offset(pgd, address);
1792         if (!pud_present(*pud))
1793                 return;
1794
1795         pmd = pmd_offset(pud, address);
1796
1797         __split_huge_pmd(vma, pmd, address, freeze, page);
1798 }
1799
1800 void vma_adjust_trans_huge(struct vm_area_struct *vma,
1801                              unsigned long start,
1802                              unsigned long end,
1803                              long adjust_next)
1804 {
1805         /*
1806          * If the new start address isn't hpage aligned and it could
1807          * previously contain an hugepage: check if we need to split
1808          * an huge pmd.
1809          */
1810         if (start & ~HPAGE_PMD_MASK &&
1811             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
1812             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
1813                 split_huge_pmd_address(vma, start, false, NULL);
1814
1815         /*
1816          * If the new end address isn't hpage aligned and it could
1817          * previously contain an hugepage: check if we need to split
1818          * an huge pmd.
1819          */
1820         if (end & ~HPAGE_PMD_MASK &&
1821             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
1822             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
1823                 split_huge_pmd_address(vma, end, false, NULL);
1824
1825         /*
1826          * If we're also updating the vma->vm_next->vm_start, if the new
1827          * vm_next->vm_start isn't page aligned and it could previously
1828          * contain an hugepage: check if we need to split an huge pmd.
1829          */
1830         if (adjust_next > 0) {
1831                 struct vm_area_struct *next = vma->vm_next;
1832                 unsigned long nstart = next->vm_start;
1833                 nstart += adjust_next << PAGE_SHIFT;
1834                 if (nstart & ~HPAGE_PMD_MASK &&
1835                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
1836                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
1837                         split_huge_pmd_address(next, nstart, false, NULL);
1838         }
1839 }
1840
1841 static void freeze_page(struct page *page)
1842 {
1843         enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
1844                 TTU_RMAP_LOCKED;
1845         int i, ret;
1846
1847         VM_BUG_ON_PAGE(!PageHead(page), page);
1848
1849         if (PageAnon(page))
1850                 ttu_flags |= TTU_MIGRATION;
1851
1852         /* We only need TTU_SPLIT_HUGE_PMD once */
1853         ret = try_to_unmap(page, ttu_flags | TTU_SPLIT_HUGE_PMD);
1854         for (i = 1; !ret && i < HPAGE_PMD_NR; i++) {
1855                 /* Cut short if the page is unmapped */
1856                 if (page_count(page) == 1)
1857                         return;
1858
1859                 ret = try_to_unmap(page + i, ttu_flags);
1860         }
1861         VM_BUG_ON_PAGE(ret, page + i - 1);
1862 }
1863
1864 static void unfreeze_page(struct page *page)
1865 {
1866         int i;
1867
1868         for (i = 0; i < HPAGE_PMD_NR; i++)
1869                 remove_migration_ptes(page + i, page + i, true);
1870 }
1871
1872 static void __split_huge_page_tail(struct page *head, int tail,
1873                 struct lruvec *lruvec, struct list_head *list)
1874 {
1875         struct page *page_tail = head + tail;
1876
1877         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
1878         VM_BUG_ON_PAGE(page_ref_count(page_tail) != 0, page_tail);
1879
1880         /*
1881          * tail_page->_refcount is zero and not changing from under us. But
1882          * get_page_unless_zero() may be running from under us on the
1883          * tail_page. If we used atomic_set() below instead of atomic_inc() or
1884          * atomic_add(), we would then run atomic_set() concurrently with
1885          * get_page_unless_zero(), and atomic_set() is implemented in C not
1886          * using locked ops. spin_unlock on x86 sometime uses locked ops
1887          * because of PPro errata 66, 92, so unless somebody can guarantee
1888          * atomic_set() here would be safe on all archs (and not only on x86),
1889          * it's safer to use atomic_inc()/atomic_add().
1890          */
1891         if (PageAnon(head)) {
1892                 page_ref_inc(page_tail);
1893         } else {
1894                 /* Additional pin to radix tree */
1895                 page_ref_add(page_tail, 2);
1896         }
1897
1898         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1899         page_tail->flags |= (head->flags &
1900                         ((1L << PG_referenced) |
1901                          (1L << PG_swapbacked) |
1902                          (1L << PG_mlocked) |
1903                          (1L << PG_uptodate) |
1904                          (1L << PG_active) |
1905                          (1L << PG_locked) |
1906                          (1L << PG_unevictable) |
1907                          (1L << PG_dirty)));
1908
1909         /*
1910          * After clearing PageTail the gup refcount can be released.
1911          * Page flags also must be visible before we make the page non-compound.
1912          */
1913         smp_wmb();
1914
1915         clear_compound_head(page_tail);
1916
1917         if (page_is_young(head))
1918                 set_page_young(page_tail);
1919         if (page_is_idle(head))
1920                 set_page_idle(page_tail);
1921
1922         /* ->mapping in first tail page is compound_mapcount */
1923         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
1924                         page_tail);
1925         page_tail->mapping = head->mapping;
1926
1927         page_tail->index = head->index + tail;
1928         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
1929         lru_add_page_tail(head, page_tail, lruvec, list);
1930 }
1931
1932 static void __split_huge_page(struct page *page, struct list_head *list,
1933                 unsigned long flags)
1934 {
1935         struct page *head = compound_head(page);
1936         struct zone *zone = page_zone(head);
1937         struct lruvec *lruvec;
1938         pgoff_t end = -1;
1939         int i;
1940
1941         lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat);
1942
1943         /* complete memcg works before add pages to LRU */
1944         mem_cgroup_split_huge_fixup(head);
1945
1946         if (!PageAnon(page))
1947                 end = DIV_ROUND_UP(i_size_read(head->mapping->host), PAGE_SIZE);
1948
1949         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1950                 __split_huge_page_tail(head, i, lruvec, list);
1951                 /* Some pages can be beyond i_size: drop them from page cache */
1952                 if (head[i].index >= end) {
1953                         __ClearPageDirty(head + i);
1954                         __delete_from_page_cache(head + i, NULL);
1955                         if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
1956                                 shmem_uncharge(head->mapping->host, 1);
1957                         put_page(head + i);
1958                 }
1959         }
1960
1961         ClearPageCompound(head);
1962         /* See comment in __split_huge_page_tail() */
1963         if (PageAnon(head)) {
1964                 page_ref_inc(head);
1965         } else {
1966                 /* Additional pin to radix tree */
1967                 page_ref_add(head, 2);
1968                 spin_unlock(&head->mapping->tree_lock);
1969         }
1970
1971         spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
1972
1973         unfreeze_page(head);
1974
1975         for (i = 0; i < HPAGE_PMD_NR; i++) {
1976                 struct page *subpage = head + i;
1977                 if (subpage == page)
1978                         continue;
1979                 unlock_page(subpage);
1980
1981                 /*
1982                  * Subpages may be freed if there wasn't any mapping
1983                  * like if add_to_swap() is running on a lru page that
1984                  * had its mapping zapped. And freeing these pages
1985                  * requires taking the lru_lock so we do the put_page
1986                  * of the tail pages after the split is complete.
1987                  */
1988                 put_page(subpage);
1989         }
1990 }
1991
1992 int total_mapcount(struct page *page)
1993 {
1994         int i, compound, ret;
1995
1996         VM_BUG_ON_PAGE(PageTail(page), page);
1997
1998         if (likely(!PageCompound(page)))
1999                 return atomic_read(&page->_mapcount) + 1;
2000
2001         compound = compound_mapcount(page);
2002         if (PageHuge(page))
2003                 return compound;
2004         ret = compound;
2005         for (i = 0; i < HPAGE_PMD_NR; i++)
2006                 ret += atomic_read(&page[i]._mapcount) + 1;
2007         /* File pages has compound_mapcount included in _mapcount */
2008         if (!PageAnon(page))
2009                 return ret - compound * HPAGE_PMD_NR;
2010         if (PageDoubleMap(page))
2011                 ret -= HPAGE_PMD_NR;
2012         return ret;
2013 }
2014
2015 /*
2016  * This calculates accurately how many mappings a transparent hugepage
2017  * has (unlike page_mapcount() which isn't fully accurate). This full
2018  * accuracy is primarily needed to know if copy-on-write faults can
2019  * reuse the page and change the mapping to read-write instead of
2020  * copying them. At the same time this returns the total_mapcount too.
2021  *
2022  * The function returns the highest mapcount any one of the subpages
2023  * has. If the return value is one, even if different processes are
2024  * mapping different subpages of the transparent hugepage, they can
2025  * all reuse it, because each process is reusing a different subpage.
2026  *
2027  * The total_mapcount is instead counting all virtual mappings of the
2028  * subpages. If the total_mapcount is equal to "one", it tells the
2029  * caller all mappings belong to the same "mm" and in turn the
2030  * anon_vma of the transparent hugepage can become the vma->anon_vma
2031  * local one as no other process may be mapping any of the subpages.
2032  *
2033  * It would be more accurate to replace page_mapcount() with
2034  * page_trans_huge_mapcount(), however we only use
2035  * page_trans_huge_mapcount() in the copy-on-write faults where we
2036  * need full accuracy to avoid breaking page pinning, because
2037  * page_trans_huge_mapcount() is slower than page_mapcount().
2038  */
2039 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2040 {
2041         int i, ret, _total_mapcount, mapcount;
2042
2043         /* hugetlbfs shouldn't call it */
2044         VM_BUG_ON_PAGE(PageHuge(page), page);
2045
2046         if (likely(!PageTransCompound(page))) {
2047                 mapcount = atomic_read(&page->_mapcount) + 1;
2048                 if (total_mapcount)
2049                         *total_mapcount = mapcount;
2050                 return mapcount;
2051         }
2052
2053         page = compound_head(page);
2054
2055         _total_mapcount = ret = 0;
2056         for (i = 0; i < HPAGE_PMD_NR; i++) {
2057                 mapcount = atomic_read(&page[i]._mapcount) + 1;
2058                 ret = max(ret, mapcount);
2059                 _total_mapcount += mapcount;
2060         }
2061         if (PageDoubleMap(page)) {
2062                 ret -= 1;
2063                 _total_mapcount -= HPAGE_PMD_NR;
2064         }
2065         mapcount = compound_mapcount(page);
2066         ret += mapcount;
2067         _total_mapcount += mapcount;
2068         if (total_mapcount)
2069                 *total_mapcount = _total_mapcount;
2070         return ret;
2071 }
2072
2073 /*
2074  * This function splits huge page into normal pages. @page can point to any
2075  * subpage of huge page to split. Split doesn't change the position of @page.
2076  *
2077  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2078  * The huge page must be locked.
2079  *
2080  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2081  *
2082  * Both head page and tail pages will inherit mapping, flags, and so on from
2083  * the hugepage.
2084  *
2085  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2086  * they are not mapped.
2087  *
2088  * Returns 0 if the hugepage is split successfully.
2089  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2090  * us.
2091  */
2092 int split_huge_page_to_list(struct page *page, struct list_head *list)
2093 {
2094         struct page *head = compound_head(page);
2095         struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2096         struct anon_vma *anon_vma = NULL;
2097         struct address_space *mapping = NULL;
2098         int count, mapcount, extra_pins, ret;
2099         bool mlocked;
2100         unsigned long flags;
2101
2102         VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
2103         VM_BUG_ON_PAGE(!PageLocked(page), page);
2104         VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2105         VM_BUG_ON_PAGE(!PageCompound(page), page);
2106
2107         if (PageAnon(head)) {
2108                 /*
2109                  * The caller does not necessarily hold an mmap_sem that would
2110                  * prevent the anon_vma disappearing so we first we take a
2111                  * reference to it and then lock the anon_vma for write. This
2112                  * is similar to page_lock_anon_vma_read except the write lock
2113                  * is taken to serialise against parallel split or collapse
2114                  * operations.
2115                  */
2116                 anon_vma = page_get_anon_vma(head);
2117                 if (!anon_vma) {
2118                         ret = -EBUSY;
2119                         goto out;
2120                 }
2121                 extra_pins = 0;
2122                 mapping = NULL;
2123                 anon_vma_lock_write(anon_vma);
2124         } else {
2125                 mapping = head->mapping;
2126
2127                 /* Truncated ? */
2128                 if (!mapping) {
2129                         ret = -EBUSY;
2130                         goto out;
2131                 }
2132
2133                 /* Addidional pins from radix tree */
2134                 extra_pins = HPAGE_PMD_NR;
2135                 anon_vma = NULL;
2136                 i_mmap_lock_read(mapping);
2137         }
2138
2139         /*
2140          * Racy check if we can split the page, before freeze_page() will
2141          * split PMDs
2142          */
2143         if (total_mapcount(head) != page_count(head) - extra_pins - 1) {
2144                 ret = -EBUSY;
2145                 goto out_unlock;
2146         }
2147
2148         mlocked = PageMlocked(page);
2149         freeze_page(head);
2150         VM_BUG_ON_PAGE(compound_mapcount(head), head);
2151
2152         /* Make sure the page is not on per-CPU pagevec as it takes pin */
2153         if (mlocked)
2154                 lru_add_drain();
2155
2156         /* prevent PageLRU to go away from under us, and freeze lru stats */
2157         spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags);
2158
2159         if (mapping) {
2160                 void **pslot;
2161
2162                 spin_lock(&mapping->tree_lock);
2163                 pslot = radix_tree_lookup_slot(&mapping->page_tree,
2164                                 page_index(head));
2165                 /*
2166                  * Check if the head page is present in radix tree.
2167                  * We assume all tail are present too, if head is there.
2168                  */
2169                 if (radix_tree_deref_slot_protected(pslot,
2170                                         &mapping->tree_lock) != head)
2171                         goto fail;
2172         }
2173
2174         /* Prevent deferred_split_scan() touching ->_refcount */
2175         spin_lock(&pgdata->split_queue_lock);
2176         count = page_count(head);
2177         mapcount = total_mapcount(head);
2178         if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2179                 if (!list_empty(page_deferred_list(head))) {
2180                         pgdata->split_queue_len--;
2181                         list_del(page_deferred_list(head));
2182                 }
2183                 if (mapping)
2184                         __dec_node_page_state(page, NR_SHMEM_THPS);
2185                 spin_unlock(&pgdata->split_queue_lock);
2186                 __split_huge_page(page, list, flags);
2187                 ret = 0;
2188         } else {
2189                 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2190                         pr_alert("total_mapcount: %u, page_count(): %u\n",
2191                                         mapcount, count);
2192                         if (PageTail(page))
2193                                 dump_page(head, NULL);
2194                         dump_page(page, "total_mapcount(head) > 0");
2195                         BUG();
2196                 }
2197                 spin_unlock(&pgdata->split_queue_lock);
2198 fail:           if (mapping)
2199                         spin_unlock(&mapping->tree_lock);
2200                 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
2201                 unfreeze_page(head);
2202                 ret = -EBUSY;
2203         }
2204
2205 out_unlock:
2206         if (anon_vma) {
2207                 anon_vma_unlock_write(anon_vma);
2208                 put_anon_vma(anon_vma);
2209         }
2210         if (mapping)
2211                 i_mmap_unlock_read(mapping);
2212 out:
2213         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2214         return ret;
2215 }
2216
2217 void free_transhuge_page(struct page *page)
2218 {
2219         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2220         unsigned long flags;
2221
2222         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2223         if (!list_empty(page_deferred_list(page))) {
2224                 pgdata->split_queue_len--;
2225                 list_del(page_deferred_list(page));
2226         }
2227         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2228         free_compound_page(page);
2229 }
2230
2231 void deferred_split_huge_page(struct page *page)
2232 {
2233         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2234         unsigned long flags;
2235
2236         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2237
2238         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2239         if (list_empty(page_deferred_list(page))) {
2240                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2241                 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2242                 pgdata->split_queue_len++;
2243         }
2244         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2245 }
2246
2247 static unsigned long deferred_split_count(struct shrinker *shrink,
2248                 struct shrink_control *sc)
2249 {
2250         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2251         return ACCESS_ONCE(pgdata->split_queue_len);
2252 }
2253
2254 static unsigned long deferred_split_scan(struct shrinker *shrink,
2255                 struct shrink_control *sc)
2256 {
2257         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2258         unsigned long flags;
2259         LIST_HEAD(list), *pos, *next;
2260         struct page *page;
2261         int split = 0;
2262
2263         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2264         /* Take pin on all head pages to avoid freeing them under us */
2265         list_for_each_safe(pos, next, &pgdata->split_queue) {
2266                 page = list_entry((void *)pos, struct page, mapping);
2267                 page = compound_head(page);
2268                 if (get_page_unless_zero(page)) {
2269                         list_move(page_deferred_list(page), &list);
2270                 } else {
2271                         /* We lost race with put_compound_page() */
2272                         list_del_init(page_deferred_list(page));
2273                         pgdata->split_queue_len--;
2274                 }
2275                 if (!--sc->nr_to_scan)
2276                         break;
2277         }
2278         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2279
2280         list_for_each_safe(pos, next, &list) {
2281                 page = list_entry((void *)pos, struct page, mapping);
2282                 if (!trylock_page(page))
2283                         goto next;
2284                 /* split_huge_page() removes page from list on success */
2285                 if (!split_huge_page(page))
2286                         split++;
2287                 unlock_page(page);
2288 next:
2289                 put_page(page);
2290         }
2291
2292         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2293         list_splice_tail(&list, &pgdata->split_queue);
2294         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2295
2296         /*
2297          * Stop shrinker if we didn't split any page, but the queue is empty.
2298          * This can happen if pages were freed under us.
2299          */
2300         if (!split && list_empty(&pgdata->split_queue))
2301                 return SHRINK_STOP;
2302         return split;
2303 }
2304
2305 static struct shrinker deferred_split_shrinker = {
2306         .count_objects = deferred_split_count,
2307         .scan_objects = deferred_split_scan,
2308         .seeks = DEFAULT_SEEKS,
2309         .flags = SHRINKER_NUMA_AWARE,
2310 };
2311
2312 #ifdef CONFIG_DEBUG_FS
2313 static int split_huge_pages_set(void *data, u64 val)
2314 {
2315         struct zone *zone;
2316         struct page *page;
2317         unsigned long pfn, max_zone_pfn;
2318         unsigned long total = 0, split = 0;
2319
2320         if (val != 1)
2321                 return -EINVAL;
2322
2323         for_each_populated_zone(zone) {
2324                 max_zone_pfn = zone_end_pfn(zone);
2325                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2326                         if (!pfn_valid(pfn))
2327                                 continue;
2328
2329                         page = pfn_to_page(pfn);
2330                         if (!get_page_unless_zero(page))
2331                                 continue;
2332
2333                         if (zone != page_zone(page))
2334                                 goto next;
2335
2336                         if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2337                                 goto next;
2338
2339                         total++;
2340                         lock_page(page);
2341                         if (!split_huge_page(page))
2342                                 split++;
2343                         unlock_page(page);
2344 next:
2345                         put_page(page);
2346                 }
2347         }
2348
2349         pr_info("%lu of %lu THP split\n", split, total);
2350
2351         return 0;
2352 }
2353 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2354                 "%llu\n");
2355
2356 static int __init split_huge_pages_debugfs(void)
2357 {
2358         void *ret;
2359
2360         ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2361                         &split_huge_pages_fops);
2362         if (!ret)
2363                 pr_warn("Failed to create split_huge_pages in debugfs");
2364         return 0;
2365 }
2366 late_initcall(split_huge_pages_debugfs);
2367 #endif