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