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/sched/coredump.h>
  13 #include <linux/sched/numa_balancing.h>
  14 #include <linux/highmem.h>
  15 #include <linux/hugetlb.h>
  16 #include <linux/mmu_notifier.h>
  17 #include <linux/rmap.h>
  18 #include <linux/swap.h>
  19 #include <linux/shrinker.h>
  20 #include <linux/mm_inline.h>
  21 #include <linux/swapops.h>
  22 #include <linux/dax.h>
  23 #include <linux/khugepaged.h>
  24 #include <linux/freezer.h>
  25 #include <linux/pfn_t.h>
  26 #include <linux/mman.h>
  27 #include <linux/memremap.h>
  28 #include <linux/pagemap.h>
  29 #include <linux/debugfs.h>
  30 #include <linux/migrate.h>
  31 #include <linux/hashtable.h>
  32 #include <linux/userfaultfd_k.h>
  33 #include <linux/page_idle.h>
  34 #include <linux/shmem_fs.h>
  35 #include <linux/oom.h>
  36 #include <linux/page_owner.h>
  37
  38 #include <asm/tlb.h>
  39 #include <asm/pgalloc.h>
  40 #include "internal.h"
  41
  42 /*
  43  * By default transparent hugepage support is disabled in order that avoid
  44  * to risk increase the memory footprint of applications without a guaranteed
  45  * benefit. When transparent hugepage support is enabled, is for all mappings,
  46  * and khugepaged scans all mappings.
  47  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  48  * for all hugepage allocations.
  49  */
  50 unsigned long transparent_hugepage_flags __read_mostly =
  51 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  52         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  53 #endif
  54 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  55         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  56 #endif
  57         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
  58         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  59         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  60
  61 static struct shrinker deferred_split_shrinker;
  62
  63 static atomic_t huge_zero_refcount;
  64 struct page *huge_zero_page __read_mostly;
  65
  66 static struct page *get_huge_zero_page(void)
  67 {
  68         struct page *zero_page;
  69 retry:
  70         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
  71                 return READ_ONCE(huge_zero_page);
  72
  73         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
  74                         HPAGE_PMD_ORDER);
  75         if (!zero_page) {
  76                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
  77                 return NULL;
  78         }
  79         count_vm_event(THP_ZERO_PAGE_ALLOC);
  80         preempt_disable();
  81         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
  82                 preempt_enable();
  83                 __free_pages(zero_page, compound_order(zero_page));
  84                 goto retry;
  85         }
  86
  87         /* We take additional reference here. It will be put back by shrinker */
  88         atomic_set(&huge_zero_refcount, 2);
  89         preempt_enable();
  90         return READ_ONCE(huge_zero_page);
  91 }
  92
  93 static void put_huge_zero_page(void)
  94 {
  95         /*
  96          * Counter should never go to zero here. Only shrinker can put
  97          * last reference.
  98          */
  99         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 100 }
 101
 102 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
 103 {
 104         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 105                 return READ_ONCE(huge_zero_page);
 106
 107         if (!get_huge_zero_page())
 108                 return NULL;
 109
 110         if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 111                 put_huge_zero_page();
 112
 113         return READ_ONCE(huge_zero_page);
 114 }
 115
 116 void mm_put_huge_zero_page(struct mm_struct *mm)
 117 {
 118         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 119                 put_huge_zero_page();
 120 }
 121
 122 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 123                                         struct shrink_control *sc)
 124 {
 125         /* we can free zero page only if last reference remains */
 126         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 127 }
 128
 129 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 130                                        struct shrink_control *sc)
 131 {
 132         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 133                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 134                 BUG_ON(zero_page == NULL);
 135                 __free_pages(zero_page, compound_order(zero_page));
 136                 return HPAGE_PMD_NR;
 137         }
 138
 139         return 0;
 140 }
 141
 142 static struct shrinker huge_zero_page_shrinker = {
 143         .count_objects = shrink_huge_zero_page_count,
 144         .scan_objects = shrink_huge_zero_page_scan,
 145         .seeks = DEFAULT_SEEKS,
 146 };
 147
 148 #ifdef CONFIG_SYSFS
 149 static ssize_t enabled_show(struct kobject *kobj,
 150                             struct kobj_attribute *attr, char *buf)
 151 {
 152         if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
 153                 return sprintf(buf, "[always] madvise never\n");
 154         else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
 155                 return sprintf(buf, "always [madvise] never\n");
 156         else
 157                 return sprintf(buf, "always madvise [never]\n");
 158 }
 159
 160 static ssize_t enabled_store(struct kobject *kobj,
 161                              struct kobj_attribute *attr,
 162                              const char *buf, size_t count)
 163 {
 164         ssize_t ret = count;
 165
 166         if (sysfs_streq(buf, "always")) {
 167                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 168                 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 169         } else if (sysfs_streq(buf, "madvise")) {
 170                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 171                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 172         } else if (sysfs_streq(buf, "never")) {
 173                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 174                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 175         } else
 176                 ret = -EINVAL;
 177
 178         if (ret > 0) {
 179                 int err = start_stop_khugepaged();
 180                 if (err)
 181                         ret = err;
 182         }
 183         return ret;
 184 }
 185 static struct kobj_attribute enabled_attr =
 186         __ATTR(enabled, 0644, enabled_show, enabled_store);
 187
 188 ssize_t single_hugepage_flag_show(struct kobject *kobj,
 189                                 struct kobj_attribute *attr, char *buf,
 190                                 enum transparent_hugepage_flag flag)
 191 {
 192         return sprintf(buf, "%d\n",
 193                        !!test_bit(flag, &transparent_hugepage_flags));
 194 }
 195
 196 ssize_t single_hugepage_flag_store(struct kobject *kobj,
 197                                  struct kobj_attribute *attr,
 198                                  const char *buf, size_t count,
 199                                  enum transparent_hugepage_flag flag)
 200 {
 201         unsigned long value;
 202         int ret;
 203
 204         ret = kstrtoul(buf, 10, &value);
 205         if (ret < 0)
 206                 return ret;
 207         if (value > 1)
 208                 return -EINVAL;
 209
 210         if (value)
 211                 set_bit(flag, &transparent_hugepage_flags);
 212         else
 213                 clear_bit(flag, &transparent_hugepage_flags);
 214
 215         return count;
 216 }
 217
 218 static ssize_t defrag_show(struct kobject *kobj,
 219                            struct kobj_attribute *attr, char *buf)
 220 {
 221         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 222                 return sprintf(buf, "[always] defer defer+madvise madvise never\n");
 223         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 224                 return sprintf(buf, "always [defer] defer+madvise madvise never\n");
 225         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 226                 return sprintf(buf, "always defer [defer+madvise] madvise never\n");
 227         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 228                 return sprintf(buf, "always defer defer+madvise [madvise] never\n");
 229         return sprintf(buf, "always defer defer+madvise madvise [never]\n");
 230 }
 231
 232 static ssize_t defrag_store(struct kobject *kobj,
 233                             struct kobj_attribute *attr,
 234                             const char *buf, size_t count)
 235 {
 236         if (sysfs_streq(buf, "always")) {
 237                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 238                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 239                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 240                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 241         } else if (sysfs_streq(buf, "defer+madvise")) {
 242                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 243                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 244                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 245                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 246         } else if (sysfs_streq(buf, "defer")) {
 247                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 248                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 249                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 250                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 251         } else if (sysfs_streq(buf, "madvise")) {
 252                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 253                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 254                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 255                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 256         } else if (sysfs_streq(buf, "never")) {
 257                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 258                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 259                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 260                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 261         } else
 262                 return -EINVAL;
 263
 264         return count;
 265 }
 266 static struct kobj_attribute defrag_attr =
 267         __ATTR(defrag, 0644, defrag_show, defrag_store);
 268
 269 static ssize_t use_zero_page_show(struct kobject *kobj,
 270                 struct kobj_attribute *attr, char *buf)
 271 {
 272         return single_hugepage_flag_show(kobj, attr, buf,
 273                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 274 }
 275 static ssize_t use_zero_page_store(struct kobject *kobj,
 276                 struct kobj_attribute *attr, const char *buf, size_t count)
 277 {
 278         return single_hugepage_flag_store(kobj, attr, buf, count,
 279                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 280 }
 281 static struct kobj_attribute use_zero_page_attr =
 282         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 283
 284 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
 285                 struct kobj_attribute *attr, char *buf)
 286 {
 287         return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE);
 288 }
 289 static struct kobj_attribute hpage_pmd_size_attr =
 290         __ATTR_RO(hpage_pmd_size);
 291
 292 #ifdef CONFIG_DEBUG_VM
 293 static ssize_t debug_cow_show(struct kobject *kobj,
 294                                 struct kobj_attribute *attr, char *buf)
 295 {
 296         return single_hugepage_flag_show(kobj, attr, buf,
 297                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 298 }
 299 static ssize_t debug_cow_store(struct kobject *kobj,
 300                                struct kobj_attribute *attr,
 301                                const char *buf, size_t count)
 302 {
 303         return single_hugepage_flag_store(kobj, attr, buf, count,
 304                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 305 }
 306 static struct kobj_attribute debug_cow_attr =
 307         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 308 #endif /* CONFIG_DEBUG_VM */
 309
 310 static struct attribute *hugepage_attr[] = {
 311         &enabled_attr.attr,
 312         &defrag_attr.attr,
 313         &use_zero_page_attr.attr,
 314         &hpage_pmd_size_attr.attr,
 315 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
 316         &shmem_enabled_attr.attr,
 317 #endif
 318 #ifdef CONFIG_DEBUG_VM
 319         &debug_cow_attr.attr,
 320 #endif
 321         NULL,
 322 };
 323
 324 static const struct attribute_group hugepage_attr_group = {
 325         .attrs = hugepage_attr,
 326 };
 327
 328 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 329 {
 330         int err;
 331
 332         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 333         if (unlikely(!*hugepage_kobj)) {
 334                 pr_err("failed to create transparent hugepage kobject\n");
 335                 return -ENOMEM;
 336         }
 337
 338         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 339         if (err) {
 340                 pr_err("failed to register transparent hugepage group\n");
 341                 goto delete_obj;
 342         }
 343
 344         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 345         if (err) {
 346                 pr_err("failed to register transparent hugepage group\n");
 347                 goto remove_hp_group;
 348         }
 349
 350         return 0;
 351
 352 remove_hp_group:
 353         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 354 delete_obj:
 355         kobject_put(*hugepage_kobj);
 356         return err;
 357 }
 358
 359 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 360 {
 361         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 362         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 363         kobject_put(hugepage_kobj);
 364 }
 365 #else
 366 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 367 {
 368         return 0;
 369 }
 370
 371 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 372 {
 373 }
 374 #endif /* CONFIG_SYSFS */
 375
 376 static int __init hugepage_init(void)
 377 {
 378         int err;
 379         struct kobject *hugepage_kobj;
 380
 381         if (!has_transparent_hugepage()) {
 382                 transparent_hugepage_flags = 0;
 383                 return -EINVAL;
 384         }
 385
 386         /*
 387          * hugepages can't be allocated by the buddy allocator
 388          */
 389         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
 390         /*
 391          * we use page->mapping and page->index in second tail page
 392          * as list_head: assuming THP order >= 2
 393          */
 394         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 395
 396         err = hugepage_init_sysfs(&hugepage_kobj);
 397         if (err)
 398                 goto err_sysfs;
 399
 400         err = khugepaged_init();
 401         if (err)
 402                 goto err_slab;
 403
 404         err = register_shrinker(&huge_zero_page_shrinker);
 405         if (err)
 406                 goto err_hzp_shrinker;
 407         err = register_shrinker(&deferred_split_shrinker);
 408         if (err)
 409                 goto err_split_shrinker;
 410
 411         /*
 412          * By default disable transparent hugepages on smaller systems,
 413          * where the extra memory used could hurt more than TLB overhead
 414          * is likely to save.  The admin can still enable it through /sys.
 415          */
 416         if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
 417                 transparent_hugepage_flags = 0;
 418                 return 0;
 419         }
 420
 421         err = start_stop_khugepaged();
 422         if (err)
 423                 goto err_khugepaged;
 424
 425         return 0;
 426 err_khugepaged:
 427         unregister_shrinker(&deferred_split_shrinker);
 428 err_split_shrinker:
 429         unregister_shrinker(&huge_zero_page_shrinker);
 430 err_hzp_shrinker:
 431         khugepaged_destroy();
 432 err_slab:
 433         hugepage_exit_sysfs(hugepage_kobj);
 434 err_sysfs:
 435         return err;
 436 }
 437 subsys_initcall(hugepage_init);
 438
 439 static int __init setup_transparent_hugepage(char *str)
 440 {
 441         int ret = 0;
 442         if (!str)
 443                 goto out;
 444         if (!strcmp(str, "always")) {
 445                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 446                         &transparent_hugepage_flags);
 447                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 448                           &transparent_hugepage_flags);
 449                 ret = 1;
 450         } else if (!strcmp(str, "madvise")) {
 451                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 452                           &transparent_hugepage_flags);
 453                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 454                         &transparent_hugepage_flags);
 455                 ret = 1;
 456         } else if (!strcmp(str, "never")) {
 457                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 458                           &transparent_hugepage_flags);
 459                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 460                           &transparent_hugepage_flags);
 461                 ret = 1;
 462         }
 463 out:
 464         if (!ret)
 465                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 466         return ret;
 467 }
 468 __setup("transparent_hugepage=", setup_transparent_hugepage);
 469
 470 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 471 {
 472         if (likely(vma->vm_flags & VM_WRITE))
 473                 pmd = pmd_mkwrite(pmd);
 474         return pmd;
 475 }
 476
 477 static inline struct list_head *page_deferred_list(struct page *page)
 478 {
 479         /*
 480          * ->lru in the tail pages is occupied by compound_head.
 481          * Let's use ->mapping + ->index in the second tail page as list_head.
 482          */
 483         return (struct list_head *)&page[2].mapping;
 484 }
 485
 486 void prep_transhuge_page(struct page *page)
 487 {
 488         /*
 489          * we use page->mapping and page->indexlru in second tail page
 490          * as list_head: assuming THP order >= 2
 491          */
 492
 493         INIT_LIST_HEAD(page_deferred_list(page));
 494         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 495 }
 496
 497 static unsigned long __thp_get_unmapped_area(struct file *filp,
 498                 unsigned long addr, unsigned long len,
 499                 loff_t off, unsigned long flags, unsigned long size)
 500 {
 501         loff_t off_end = off + len;
 502         loff_t off_align = round_up(off, size);
 503         unsigned long len_pad, ret;
 504
 505         if (off_end <= off_align || (off_end - off_align) < size)
 506                 return 0;
 507
 508         len_pad = len + size;
 509         if (len_pad < len || (off + len_pad) < off)
 510                 return 0;
 511
 512         ret = current->mm->get_unmapped_area(filp, addr, len_pad,
 513                                               off >> PAGE_SHIFT, flags);
 514
 515         /*
 516          * The failure might be due to length padding. The caller will retry
 517          * without the padding.
 518          */
 519         if (IS_ERR_VALUE(ret))
 520                 return 0;
 521
 522         /*
 523          * Do not try to align to THP boundary if allocation at the address
 524          * hint succeeds.
 525          */
 526         if (ret == addr)
 527                 return addr;
 528
 529         ret += (off - ret) & (size - 1);
 530         return ret;
 531 }
 532
 533 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 534                 unsigned long len, unsigned long pgoff, unsigned long flags)
 535 {
 536         unsigned long ret;
 537         loff_t off = (loff_t)pgoff << PAGE_SHIFT;
 538
 539         if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
 540                 goto out;
 541
 542         ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE);
 543         if (ret)
 544                 return ret;
 545 out:
 546         return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
 547 }
 548 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
 549
 550 static int __do_huge_pmd_anonymous_page(struct vm_fault *vmf, struct page *page,
 551                 gfp_t gfp)
 552 {
 553         struct vm_area_struct *vma = vmf->vma;
 554         struct mem_cgroup *memcg;
 555         pgtable_t pgtable;
 556         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 557         int ret = 0;
 558
 559         VM_BUG_ON_PAGE(!PageCompound(page), page);
 560
 561         if (mem_cgroup_try_charge(page, vma->vm_mm, gfp | __GFP_NORETRY, &memcg,
 562                                   true)) {
 563                 put_page(page);
 564                 count_vm_event(THP_FAULT_FALLBACK);
 565                 return VM_FAULT_FALLBACK;
 566         }
 567
 568         pgtable = pte_alloc_one(vma->vm_mm, haddr);
 569         if (unlikely(!pgtable)) {
 570                 ret = VM_FAULT_OOM;
 571                 goto release;
 572         }
 573
 574         clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
 575         /*
 576          * The memory barrier inside __SetPageUptodate makes sure that
 577          * clear_huge_page writes become visible before the set_pmd_at()
 578          * write.
 579          */
 580         __SetPageUptodate(page);
 581
 582         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 583         if (unlikely(!pmd_none(*vmf->pmd))) {
 584                 goto unlock_release;
 585         } else {
 586                 pmd_t entry;
 587
 588                 ret = check_stable_address_space(vma->vm_mm);
 589                 if (ret)
 590                         goto unlock_release;
 591
 592                 /* Deliver the page fault to userland */
 593                 if (userfaultfd_missing(vma)) {
 594                         int ret;
 595
 596                         spin_unlock(vmf->ptl);
 597                         mem_cgroup_cancel_charge(page, memcg, true);
 598                         put_page(page);
 599                         pte_free(vma->vm_mm, pgtable);
 600                         ret = handle_userfault(vmf, VM_UFFD_MISSING);
 601                         VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 602                         return ret;
 603                 }
 604
 605                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 606                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 607                 page_add_new_anon_rmap(page, vma, haddr, true);
 608                 mem_cgroup_commit_charge(page, memcg, false, true);
 609                 lru_cache_add_active_or_unevictable(page, vma);
 610                 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
 611                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
 612                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 613                 atomic_long_inc(&vma->vm_mm->nr_ptes);
 614                 spin_unlock(vmf->ptl);
 615                 count_vm_event(THP_FAULT_ALLOC);
 616         }
 617
 618         return 0;
 619 unlock_release:
 620         spin_unlock(vmf->ptl);
 621 release:
 622         if (pgtable)
 623                 pte_free(vma->vm_mm, pgtable);
 624         mem_cgroup_cancel_charge(page, memcg, true);
 625         put_page(page);
 626         return ret;
 627
 628 }
 629
 630 /*
 631  * always: directly stall for all thp allocations
 632  * defer: wake kswapd and fail if not immediately available
 633  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
 634  *                fail if not immediately available
 635  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
 636  *          available
 637  * never: never stall for any thp allocation
 638  */
 639 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
 640 {
 641         const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
 642
 643         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 644                 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
 645         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 646                 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
 647         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 648                 return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM :
 649                                                              __GFP_KSWAPD_RECLAIM);
 650         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 651                 return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM :
 652                                                              0);
 653         return GFP_TRANSHUGE_LIGHT;
 654 }
 655
 656 /* Caller must hold page table lock. */
 657 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 658                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 659                 struct page *zero_page)
 660 {
 661         pmd_t entry;
 662         if (!pmd_none(*pmd))
 663                 return false;
 664         entry = mk_pmd(zero_page, vma->vm_page_prot);
 665         entry = pmd_mkhuge(entry);
 666         if (pgtable)
 667                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 668         set_pmd_at(mm, haddr, pmd, entry);
 669         atomic_long_inc(&mm->nr_ptes);
 670         return true;
 671 }
 672
 673 int do_huge_pmd_anonymous_page(struct vm_fault *vmf)
 674 {
 675         struct vm_area_struct *vma = vmf->vma;
 676         gfp_t gfp;
 677         struct page *page;
 678         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 679
 680         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 681                 return VM_FAULT_FALLBACK;
 682         if (unlikely(anon_vma_prepare(vma)))
 683                 return VM_FAULT_OOM;
 684         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 685                 return VM_FAULT_OOM;
 686         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
 687                         !mm_forbids_zeropage(vma->vm_mm) &&
 688                         transparent_hugepage_use_zero_page()) {
 689                 pgtable_t pgtable;
 690                 struct page *zero_page;
 691                 int ret;
 692                 pgtable = pte_alloc_one(vma->vm_mm, haddr);
 693                 if (unlikely(!pgtable))
 694                         return VM_FAULT_OOM;
 695                 zero_page = mm_get_huge_zero_page(vma->vm_mm);
 696                 if (unlikely(!zero_page)) {
 697                         pte_free(vma->vm_mm, pgtable);
 698                         count_vm_event(THP_FAULT_FALLBACK);
 699                         return VM_FAULT_FALLBACK;
 700                 }
 701                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 702                 ret = 0;
 703                 if (pmd_none(*vmf->pmd)) {
 704                         ret = check_stable_address_space(vma->vm_mm);
 705                         if (ret) {
 706                                 spin_unlock(vmf->ptl);
 707                                 pte_free(vma->vm_mm, pgtable);
 708                         } else if (userfaultfd_missing(vma)) {
 709                                 spin_unlock(vmf->ptl);
 710                                 pte_free(vma->vm_mm, pgtable);
 711                                 ret = handle_userfault(vmf, VM_UFFD_MISSING);
 712                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 713                         } else {
 714                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
 715                                                    haddr, vmf->pmd, zero_page);
 716                                 spin_unlock(vmf->ptl);
 717                         }
 718                 } else {
 719                         spin_unlock(vmf->ptl);
 720                         pte_free(vma->vm_mm, pgtable);
 721                 }
 722                 return ret;
 723         }
 724         gfp = alloc_hugepage_direct_gfpmask(vma);
 725         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 726         if (unlikely(!page)) {
 727                 count_vm_event(THP_FAULT_FALLBACK);
 728                 return VM_FAULT_FALLBACK;
 729         }
 730         prep_transhuge_page(page);
 731         return __do_huge_pmd_anonymous_page(vmf, page, gfp);
 732 }
 733
 734 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 735                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
 736                 pgtable_t pgtable)
 737 {
 738         struct mm_struct *mm = vma->vm_mm;
 739         pmd_t entry;
 740         spinlock_t *ptl;
 741
 742         ptl = pmd_lock(mm, pmd);
 743         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 744         if (pfn_t_devmap(pfn))
 745                 entry = pmd_mkdevmap(entry);
 746         if (write) {
 747                 entry = pmd_mkyoung(pmd_mkdirty(entry));
 748                 entry = maybe_pmd_mkwrite(entry, vma);
 749         }
 750
 751         if (pgtable) {
 752                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 753                 atomic_long_inc(&mm->nr_ptes);
 754         }
 755
 756         set_pmd_at(mm, addr, pmd, entry);
 757         update_mmu_cache_pmd(vma, addr, pmd);
 758         spin_unlock(ptl);
 759 }
 760
 761 int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 762                         pmd_t *pmd, pfn_t pfn, bool write)
 763 {
 764         pgprot_t pgprot = vma->vm_page_prot;
 765         pgtable_t pgtable = NULL;
 766         /*
 767          * If we had pmd_special, we could avoid all these restrictions,
 768          * but we need to be consistent with PTEs and architectures that
 769          * can't support a 'special' bit.
 770          */
 771         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
 772         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 773                                                 (VM_PFNMAP|VM_MIXEDMAP));
 774         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 775         BUG_ON(!pfn_t_devmap(pfn));
 776
 777         if (addr < vma->vm_start || addr >= vma->vm_end)
 778                 return VM_FAULT_SIGBUS;
 779
 780         if (arch_needs_pgtable_deposit()) {
 781                 pgtable = pte_alloc_one(vma->vm_mm, addr);
 782                 if (!pgtable)
 783                         return VM_FAULT_OOM;
 784         }
 785
 786         track_pfn_insert(vma, &pgprot, pfn);
 787
 788         insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write, pgtable);
 789         return VM_FAULT_NOPAGE;
 790 }
 791 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
 792
 793 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 794 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
 795 {
 796         if (likely(vma->vm_flags & VM_WRITE))
 797                 pud = pud_mkwrite(pud);
 798         return pud;
 799 }
 800
 801 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 802                 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
 803 {
 804         struct mm_struct *mm = vma->vm_mm;
 805         pud_t entry;
 806         spinlock_t *ptl;
 807
 808         ptl = pud_lock(mm, pud);
 809         entry = pud_mkhuge(pfn_t_pud(pfn, prot));
 810         if (pfn_t_devmap(pfn))
 811                 entry = pud_mkdevmap(entry);
 812         if (write) {
 813                 entry = pud_mkyoung(pud_mkdirty(entry));
 814                 entry = maybe_pud_mkwrite(entry, vma);
 815         }
 816         set_pud_at(mm, addr, pud, entry);
 817         update_mmu_cache_pud(vma, addr, pud);
 818         spin_unlock(ptl);
 819 }
 820
 821 int vmf_insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 822                         pud_t *pud, pfn_t pfn, bool write)
 823 {
 824         pgprot_t pgprot = vma->vm_page_prot;
 825         /*
 826          * If we had pud_special, we could avoid all these restrictions,
 827          * but we need to be consistent with PTEs and architectures that
 828          * can't support a 'special' bit.
 829          */
 830         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
 831         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 832                                                 (VM_PFNMAP|VM_MIXEDMAP));
 833         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 834         BUG_ON(!pfn_t_devmap(pfn));
 835
 836         if (addr < vma->vm_start || addr >= vma->vm_end)
 837                 return VM_FAULT_SIGBUS;
 838
 839         track_pfn_insert(vma, &pgprot, pfn);
 840
 841         insert_pfn_pud(vma, addr, pud, pfn, pgprot, write);
 842         return VM_FAULT_NOPAGE;
 843 }
 844 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
 845 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
 846
 847 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
 848                 pmd_t *pmd, int flags)
 849 {
 850         pmd_t _pmd;
 851
 852         _pmd = pmd_mkyoung(*pmd);
 853         if (flags & FOLL_WRITE)
 854                 _pmd = pmd_mkdirty(_pmd);
 855         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
 856                                 pmd, _pmd, flags & FOLL_WRITE))
 857                 update_mmu_cache_pmd(vma, addr, pmd);
 858 }
 859
 860 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 861                 pmd_t *pmd, int flags)
 862 {
 863         unsigned long pfn = pmd_pfn(*pmd);
 864         struct mm_struct *mm = vma->vm_mm;
 865         struct dev_pagemap *pgmap;
 866         struct page *page;
 867
 868         assert_spin_locked(pmd_lockptr(mm, pmd));
 869
 870         /*
 871          * When we COW a devmap PMD entry, we split it into PTEs, so we should
 872          * not be in this function with `flags & FOLL_COW` set.
 873          */
 874         WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
 875
 876         if (flags & FOLL_WRITE && !pmd_write(*pmd))
 877                 return NULL;
 878
 879         if (pmd_present(*pmd) && pmd_devmap(*pmd))
 880                 /* pass */;
 881         else
 882                 return NULL;
 883
 884         if (flags & FOLL_TOUCH)
 885                 touch_pmd(vma, addr, pmd, flags);
 886
 887         /*
 888          * device mapped pages can only be returned if the
 889          * caller will manage the page reference count.
 890          */
 891         if (!(flags & FOLL_GET))
 892                 return ERR_PTR(-EEXIST);
 893
 894         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
 895         pgmap = get_dev_pagemap(pfn, NULL);
 896         if (!pgmap)
 897                 return ERR_PTR(-EFAULT);
 898         page = pfn_to_page(pfn);
 899         get_page(page);
 900         put_dev_pagemap(pgmap);
 901
 902         return page;
 903 }
 904
 905 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 906                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 907                   struct vm_area_struct *vma)
 908 {
 909         spinlock_t *dst_ptl, *src_ptl;
 910         struct page *src_page;
 911         pmd_t pmd;
 912         pgtable_t pgtable = NULL;
 913         int ret = -ENOMEM;
 914
 915         /* Skip if can be re-fill on fault */
 916         if (!vma_is_anonymous(vma))
 917                 return 0;
 918
 919         pgtable = pte_alloc_one(dst_mm, addr);
 920         if (unlikely(!pgtable))
 921                 goto out;
 922
 923         dst_ptl = pmd_lock(dst_mm, dst_pmd);
 924         src_ptl = pmd_lockptr(src_mm, src_pmd);
 925         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 926
 927         ret = -EAGAIN;
 928         pmd = *src_pmd;
 929
 930 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
 931         if (unlikely(is_swap_pmd(pmd))) {
 932                 swp_entry_t entry = pmd_to_swp_entry(pmd);
 933
 934                 VM_BUG_ON(!is_pmd_migration_entry(pmd));
 935                 if (is_write_migration_entry(entry)) {
 936                         make_migration_entry_read(&entry);
 937                         pmd = swp_entry_to_pmd(entry);
 938                         if (pmd_swp_soft_dirty(*src_pmd))
 939                                 pmd = pmd_swp_mksoft_dirty(pmd);
 940                         set_pmd_at(src_mm, addr, src_pmd, pmd);
 941                 }
 942                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 943                 atomic_long_inc(&dst_mm->nr_ptes);
 944                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
 945                 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 946                 ret = 0;
 947                 goto out_unlock;
 948         }
 949 #endif
 950
 951         if (unlikely(!pmd_trans_huge(pmd))) {
 952                 pte_free(dst_mm, pgtable);
 953                 goto out_unlock;
 954         }
 955         /*
 956          * When page table lock is held, the huge zero pmd should not be
 957          * under splitting since we don't split the page itself, only pmd to
 958          * a page table.
 959          */
 960         if (is_huge_zero_pmd(pmd)) {
 961                 struct page *zero_page;
 962                 /*
 963                  * get_huge_zero_page() will never allocate a new page here,
 964                  * since we already have a zero page to copy. It just takes a
 965                  * reference.
 966                  */
 967                 zero_page = mm_get_huge_zero_page(dst_mm);
 968                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
 969                                 zero_page);
 970                 ret = 0;
 971                 goto out_unlock;
 972         }
 973
 974         src_page = pmd_page(pmd);
 975         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
 976         get_page(src_page);
 977         page_dup_rmap(src_page, true);
 978         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 979         atomic_long_inc(&dst_mm->nr_ptes);
 980         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
 981
 982         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 983         pmd = pmd_mkold(pmd_wrprotect(pmd));
 984         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 985
 986         ret = 0;
 987 out_unlock:
 988         spin_unlock(src_ptl);
 989         spin_unlock(dst_ptl);
 990 out:
 991         return ret;
 992 }
 993
 994 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 995 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
 996                 pud_t *pud, int flags)
 997 {
 998         pud_t _pud;
 999
1000         _pud = pud_mkyoung(*pud);
1001         if (flags & FOLL_WRITE)
1002                 _pud = pud_mkdirty(_pud);
1003         if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1004                                 pud, _pud, flags & FOLL_WRITE))
1005                 update_mmu_cache_pud(vma, addr, pud);
1006 }
1007
1008 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1009                 pud_t *pud, int flags)
1010 {
1011         unsigned long pfn = pud_pfn(*pud);
1012         struct mm_struct *mm = vma->vm_mm;
1013         struct dev_pagemap *pgmap;
1014         struct page *page;
1015
1016         assert_spin_locked(pud_lockptr(mm, pud));
1017
1018         if (flags & FOLL_WRITE && !pud_write(*pud))
1019                 return NULL;
1020
1021         if (pud_present(*pud) && pud_devmap(*pud))
1022                 /* pass */;
1023         else
1024                 return NULL;
1025
1026         if (flags & FOLL_TOUCH)
1027                 touch_pud(vma, addr, pud, flags);
1028
1029         /*
1030          * device mapped pages can only be returned if the
1031          * caller will manage the page reference count.
1032          */
1033         if (!(flags & FOLL_GET))
1034                 return ERR_PTR(-EEXIST);
1035
1036         pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1037         pgmap = get_dev_pagemap(pfn, NULL);
1038         if (!pgmap)
1039                 return ERR_PTR(-EFAULT);
1040         page = pfn_to_page(pfn);
1041         get_page(page);
1042         put_dev_pagemap(pgmap);
1043
1044         return page;
1045 }
1046
1047 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1048                   pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1049                   struct vm_area_struct *vma)
1050 {
1051         spinlock_t *dst_ptl, *src_ptl;
1052         pud_t pud;
1053         int ret;
1054
1055         dst_ptl = pud_lock(dst_mm, dst_pud);
1056         src_ptl = pud_lockptr(src_mm, src_pud);
1057         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1058
1059         ret = -EAGAIN;
1060         pud = *src_pud;
1061         if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1062                 goto out_unlock;
1063
1064         /*
1065          * When page table lock is held, the huge zero pud should not be
1066          * under splitting since we don't split the page itself, only pud to
1067          * a page table.
1068          */
1069         if (is_huge_zero_pud(pud)) {
1070                 /* No huge zero pud yet */
1071         }
1072
1073         pudp_set_wrprotect(src_mm, addr, src_pud);
1074         pud = pud_mkold(pud_wrprotect(pud));
1075         set_pud_at(dst_mm, addr, dst_pud, pud);
1076
1077         ret = 0;
1078 out_unlock:
1079         spin_unlock(src_ptl);
1080         spin_unlock(dst_ptl);
1081         return ret;
1082 }
1083
1084 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1085 {
1086         pud_t entry;
1087         unsigned long haddr;
1088         bool write = vmf->flags & FAULT_FLAG_WRITE;
1089
1090         vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1091         if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1092                 goto unlock;
1093
1094         entry = pud_mkyoung(orig_pud);
1095         if (write)
1096                 entry = pud_mkdirty(entry);
1097         haddr = vmf->address & HPAGE_PUD_MASK;
1098         if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1099                 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1100
1101 unlock:
1102         spin_unlock(vmf->ptl);
1103 }
1104 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1105
1106 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1107 {
1108         pmd_t entry;
1109         unsigned long haddr;
1110         bool write = vmf->flags & FAULT_FLAG_WRITE;
1111
1112         vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1113         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1114                 goto unlock;
1115
1116         entry = pmd_mkyoung(orig_pmd);
1117         if (write)
1118                 entry = pmd_mkdirty(entry);
1119         haddr = vmf->address & HPAGE_PMD_MASK;
1120         if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1121                 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1122
1123 unlock:
1124         spin_unlock(vmf->ptl);
1125 }
1126
1127 static int do_huge_pmd_wp_page_fallback(struct vm_fault *vmf, pmd_t orig_pmd,
1128                 struct page *page)
1129 {
1130         struct vm_area_struct *vma = vmf->vma;
1131         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1132         struct mem_cgroup *memcg;
1133         pgtable_t pgtable;
1134         pmd_t _pmd;
1135         int ret = 0, i;
1136         struct page **pages;
1137         unsigned long mmun_start;       /* For mmu_notifiers */
1138         unsigned long mmun_end;         /* For mmu_notifiers */
1139
1140         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1141                         GFP_KERNEL);
1142         if (unlikely(!pages)) {
1143                 ret |= VM_FAULT_OOM;
1144                 goto out;
1145         }
1146
1147         for (i = 0; i < HPAGE_PMD_NR; i++) {
1148                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
1149                                                vmf->address, page_to_nid(page));
1150                 if (unlikely(!pages[i] ||
1151                              mem_cgroup_try_charge(pages[i], vma->vm_mm,
1152                                      GFP_KERNEL, &memcg, false))) {
1153                         if (pages[i])
1154                                 put_page(pages[i]);
1155                         while (--i >= 0) {
1156                                 memcg = (void *)page_private(pages[i]);
1157                                 set_page_private(pages[i], 0);
1158                                 mem_cgroup_cancel_charge(pages[i], memcg,
1159                                                 false);
1160                                 put_page(pages[i]);
1161                         }
1162                         kfree(pages);
1163                         ret |= VM_FAULT_OOM;
1164                         goto out;
1165                 }
1166                 set_page_private(pages[i], (unsigned long)memcg);
1167         }
1168
1169         for (i = 0; i < HPAGE_PMD_NR; i++) {
1170                 copy_user_highpage(pages[i], page + i,
1171                                    haddr + PAGE_SIZE * i, vma);
1172                 __SetPageUptodate(pages[i]);
1173                 cond_resched();
1174         }
1175
1176         mmun_start = haddr;
1177         mmun_end   = haddr + HPAGE_PMD_SIZE;
1178         mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
1179
1180         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1181         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1182                 goto out_free_pages;
1183         VM_BUG_ON_PAGE(!PageHead(page), page);
1184
1185         pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1186         /* leave pmd empty until pte is filled */
1187
1188         pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
1189         pmd_populate(vma->vm_mm, &_pmd, pgtable);
1190
1191         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1192                 pte_t entry;
1193                 entry = mk_pte(pages[i], vma->vm_page_prot);
1194                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1195                 memcg = (void *)page_private(pages[i]);
1196                 set_page_private(pages[i], 0);
1197                 page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
1198                 mem_cgroup_commit_charge(pages[i], memcg, false, false);
1199                 lru_cache_add_active_or_unevictable(pages[i], vma);
1200                 vmf->pte = pte_offset_map(&_pmd, haddr);
1201                 VM_BUG_ON(!pte_none(*vmf->pte));
1202                 set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
1203                 pte_unmap(vmf->pte);
1204         }
1205         kfree(pages);
1206
1207         smp_wmb(); /* make pte visible before pmd */
1208         pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
1209         page_remove_rmap(page, true);
1210         spin_unlock(vmf->ptl);
1211
1212         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1213
1214         ret |= VM_FAULT_WRITE;
1215         put_page(page);
1216
1217 out:
1218         return ret;
1219
1220 out_free_pages:
1221         spin_unlock(vmf->ptl);
1222         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1223         for (i = 0; i < HPAGE_PMD_NR; i++) {
1224                 memcg = (void *)page_private(pages[i]);
1225                 set_page_private(pages[i], 0);
1226                 mem_cgroup_cancel_charge(pages[i], memcg, false);
1227                 put_page(pages[i]);
1228         }
1229         kfree(pages);
1230         goto out;
1231 }
1232
1233 int do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1234 {
1235         struct vm_area_struct *vma = vmf->vma;
1236         struct page *page = NULL, *new_page;
1237         struct mem_cgroup *memcg;
1238         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1239         unsigned long mmun_start;       /* For mmu_notifiers */
1240         unsigned long mmun_end;         /* For mmu_notifiers */
1241         gfp_t huge_gfp;                 /* for allocation and charge */
1242         int ret = 0;
1243
1244         vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1245         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1246         if (is_huge_zero_pmd(orig_pmd))
1247                 goto alloc;
1248         spin_lock(vmf->ptl);
1249         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1250                 goto out_unlock;
1251
1252         page = pmd_page(orig_pmd);
1253         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1254         /*
1255          * We can only reuse the page if nobody else maps the huge page or it's
1256          * part.
1257          */
1258         if (!trylock_page(page)) {
1259                 get_page(page);
1260                 spin_unlock(vmf->ptl);
1261                 lock_page(page);
1262                 spin_lock(vmf->ptl);
1263                 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1264                         unlock_page(page);
1265                         put_page(page);
1266                         goto out_unlock;
1267                 }
1268                 put_page(page);
1269         }
1270         if (reuse_swap_page(page, NULL)) {
1271                 pmd_t entry;
1272                 entry = pmd_mkyoung(orig_pmd);
1273                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1274                 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry,  1))
1275                         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1276                 ret |= VM_FAULT_WRITE;
1277                 unlock_page(page);
1278                 goto out_unlock;
1279         }
1280         unlock_page(page);
1281         get_page(page);
1282         spin_unlock(vmf->ptl);
1283 alloc:
1284         if (transparent_hugepage_enabled(vma) &&
1285             !transparent_hugepage_debug_cow()) {
1286                 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1287                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1288         } else
1289                 new_page = NULL;
1290
1291         if (likely(new_page)) {
1292                 prep_transhuge_page(new_page);
1293         } else {
1294                 if (!page) {
1295                         split_huge_pmd(vma, vmf->pmd, vmf->address);
1296                         ret |= VM_FAULT_FALLBACK;
1297                 } else {
1298                         ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
1299                         if (ret & VM_FAULT_OOM) {
1300                                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1301                                 ret |= VM_FAULT_FALLBACK;
1302                         }
1303                         put_page(page);
1304                 }
1305                 count_vm_event(THP_FAULT_FALLBACK);
1306                 goto out;
1307         }
1308
1309         if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm,
1310                                 huge_gfp | __GFP_NORETRY, &memcg, true))) {
1311                 put_page(new_page);
1312                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1313                 if (page)
1314                         put_page(page);
1315                 ret |= VM_FAULT_FALLBACK;
1316                 count_vm_event(THP_FAULT_FALLBACK);
1317                 goto out;
1318         }
1319
1320         count_vm_event(THP_FAULT_ALLOC);
1321
1322         if (!page)
1323                 clear_huge_page(new_page, vmf->address, HPAGE_PMD_NR);
1324         else
1325                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1326         __SetPageUptodate(new_page);
1327
1328         mmun_start = haddr;
1329         mmun_end   = haddr + HPAGE_PMD_SIZE;
1330         mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
1331
1332         spin_lock(vmf->ptl);
1333         if (page)
1334                 put_page(page);
1335         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1336                 spin_unlock(vmf->ptl);
1337                 mem_cgroup_cancel_charge(new_page, memcg, true);
1338                 put_page(new_page);
1339                 goto out_mn;
1340         } else {
1341                 pmd_t entry;
1342                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1343                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1344                 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1345                 page_add_new_anon_rmap(new_page, vma, haddr, true);
1346                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1347                 lru_cache_add_active_or_unevictable(new_page, vma);
1348                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1349                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1350                 if (!page) {
1351                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1352                 } else {
1353                         VM_BUG_ON_PAGE(!PageHead(page), page);
1354                         page_remove_rmap(page, true);
1355                         put_page(page);
1356                 }
1357                 ret |= VM_FAULT_WRITE;
1358         }
1359         spin_unlock(vmf->ptl);
1360 out_mn:
1361         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1362 out:
1363         return ret;
1364 out_unlock:
1365         spin_unlock(vmf->ptl);
1366         return ret;
1367 }
1368
1369 /*
1370  * FOLL_FORCE or a forced COW break can write even to unwritable pmd's,
1371  * but only after we've gone through a COW cycle and they are dirty.
1372  */
1373 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1374 {
1375         return pmd_write(pmd) || ((flags & FOLL_COW) && pmd_dirty(pmd));
1376 }
1377
1378 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1379                                    unsigned long addr,
1380                                    pmd_t *pmd,
1381                                    unsigned int flags)
1382 {
1383         struct mm_struct *mm = vma->vm_mm;
1384         struct page *page = NULL;
1385
1386         assert_spin_locked(pmd_lockptr(mm, pmd));
1387
1388         if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1389                 goto out;
1390
1391         /* Avoid dumping huge zero page */
1392         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1393                 return ERR_PTR(-EFAULT);
1394
1395         /* Full NUMA hinting faults to serialise migration in fault paths */
1396         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1397                 goto out;
1398
1399         page = pmd_page(*pmd);
1400         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1401         if (flags & FOLL_TOUCH)
1402                 touch_pmd(vma, addr, pmd, flags);
1403         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1404                 /*
1405                  * We don't mlock() pte-mapped THPs. This way we can avoid
1406                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1407                  *
1408                  * For anon THP:
1409                  *
1410                  * In most cases the pmd is the only mapping of the page as we
1411                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1412                  * writable private mappings in populate_vma_page_range().
1413                  *
1414                  * The only scenario when we have the page shared here is if we
1415                  * mlocking read-only mapping shared over fork(). We skip
1416                  * mlocking such pages.
1417                  *
1418                  * For file THP:
1419                  *
1420                  * We can expect PageDoubleMap() to be stable under page lock:
1421                  * for file pages we set it in page_add_file_rmap(), which
1422                  * requires page to be locked.
1423                  */
1424
1425                 if (PageAnon(page) && compound_mapcount(page) != 1)
1426                         goto skip_mlock;
1427                 if (PageDoubleMap(page) || !page->mapping)
1428                         goto skip_mlock;
1429                 if (!trylock_page(page))
1430                         goto skip_mlock;
1431                 lru_add_drain();
1432                 if (page->mapping && !PageDoubleMap(page))
1433                         mlock_vma_page(page);
1434                 unlock_page(page);
1435         }
1436 skip_mlock:
1437         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1438         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1439         if (flags & FOLL_GET)
1440                 get_page(page);
1441
1442 out:
1443         return page;
1444 }
1445
1446 /* NUMA hinting page fault entry point for trans huge pmds */
1447 int do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1448 {
1449         struct vm_area_struct *vma = vmf->vma;
1450         struct anon_vma *anon_vma = NULL;
1451         struct page *page;
1452         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1453         int page_nid = -1, this_nid = numa_node_id();
1454         int target_nid, last_cpupid = -1;
1455         bool page_locked;
1456         bool migrated = false;
1457         bool was_writable;
1458         int flags = 0;
1459
1460         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1461         if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1462                 goto out_unlock;
1463
1464         /*
1465          * If there are potential migrations, wait for completion and retry
1466          * without disrupting NUMA hinting information. Do not relock and
1467          * check_same as the page may no longer be mapped.
1468          */
1469         if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1470                 page = pmd_page(*vmf->pmd);
1471                 if (!get_page_unless_zero(page))
1472                         goto out_unlock;
1473                 spin_unlock(vmf->ptl);
1474                 wait_on_page_locked(page);
1475                 put_page(page);
1476                 goto out;
1477         }
1478
1479         page = pmd_page(pmd);
1480         BUG_ON(is_huge_zero_page(page));
1481         page_nid = page_to_nid(page);
1482         last_cpupid = page_cpupid_last(page);
1483         count_vm_numa_event(NUMA_HINT_FAULTS);
1484         if (page_nid == this_nid) {
1485                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1486                 flags |= TNF_FAULT_LOCAL;
1487         }
1488
1489         /* See similar comment in do_numa_page for explanation */
1490         if (!pmd_savedwrite(pmd))
1491                 flags |= TNF_NO_GROUP;
1492
1493         /*
1494          * Acquire the page lock to serialise THP migrations but avoid dropping
1495          * page_table_lock if at all possible
1496          */
1497         page_locked = trylock_page(page);
1498         target_nid = mpol_misplaced(page, vma, haddr);
1499         if (target_nid == -1) {
1500                 /* If the page was locked, there are no parallel migrations */
1501                 if (page_locked)
1502                         goto clear_pmdnuma;
1503         }
1504
1505         /* Migration could have started since the pmd_trans_migrating check */
1506         if (!page_locked) {
1507                 page_nid = -1;
1508                 if (!get_page_unless_zero(page))
1509                         goto out_unlock;
1510                 spin_unlock(vmf->ptl);
1511                 wait_on_page_locked(page);
1512                 put_page(page);
1513                 goto out;
1514         }
1515
1516         /*
1517          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1518          * to serialises splits
1519          */
1520         get_page(page);
1521         spin_unlock(vmf->ptl);
1522         anon_vma = page_lock_anon_vma_read(page);
1523
1524         /* Confirm the PMD did not change while page_table_lock was released */
1525         spin_lock(vmf->ptl);
1526         if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1527                 unlock_page(page);
1528                 put_page(page);
1529                 page_nid = -1;
1530                 goto out_unlock;
1531         }
1532
1533         /* Bail if we fail to protect against THP splits for any reason */
1534         if (unlikely(!anon_vma)) {
1535                 put_page(page);
1536                 page_nid = -1;
1537                 goto clear_pmdnuma;
1538         }
1539
1540         /*
1541          * Since we took the NUMA fault, we must have observed the !accessible
1542          * bit. Make sure all other CPUs agree with that, to avoid them
1543          * modifying the page we're about to migrate.
1544          *
1545          * Must be done under PTL such that we'll observe the relevant
1546          * inc_tlb_flush_pending().
1547          *
1548          * We are not sure a pending tlb flush here is for a huge page
1549          * mapping or not. Hence use the tlb range variant
1550          */
1551         if (mm_tlb_flush_pending(vma->vm_mm))
1552                 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1553
1554         /*
1555          * Migrate the THP to the requested node, returns with page unlocked
1556          * and access rights restored.
1557          */
1558         spin_unlock(vmf->ptl);
1559
1560         migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1561                                 vmf->pmd, pmd, vmf->address, page, target_nid);
1562         if (migrated) {
1563                 flags |= TNF_MIGRATED;
1564                 page_nid = target_nid;
1565         } else
1566                 flags |= TNF_MIGRATE_FAIL;
1567
1568         goto out;
1569 clear_pmdnuma:
1570         BUG_ON(!PageLocked(page));
1571         was_writable = pmd_savedwrite(pmd);
1572         pmd = pmd_modify(pmd, vma->vm_page_prot);
1573         pmd = pmd_mkyoung(pmd);
1574         if (was_writable)
1575                 pmd = pmd_mkwrite(pmd);
1576         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1577         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1578         unlock_page(page);
1579 out_unlock:
1580         spin_unlock(vmf->ptl);
1581
1582 out:
1583         if (anon_vma)
1584                 page_unlock_anon_vma_read(anon_vma);
1585
1586         if (page_nid != -1)
1587                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1588                                 flags);
1589
1590         return 0;
1591 }
1592
1593 /*
1594  * Return true if we do MADV_FREE successfully on entire pmd page.
1595  * Otherwise, return false.
1596  */
1597 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1598                 pmd_t *pmd, unsigned long addr, unsigned long next)
1599 {
1600         spinlock_t *ptl;
1601         pmd_t orig_pmd;
1602         struct page *page;
1603         struct mm_struct *mm = tlb->mm;
1604         bool ret = false;
1605
1606         tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1607
1608         ptl = pmd_trans_huge_lock(pmd, vma);
1609         if (!ptl)
1610                 goto out_unlocked;
1611
1612         orig_pmd = *pmd;
1613         if (is_huge_zero_pmd(orig_pmd))
1614                 goto out;
1615
1616         if (unlikely(!pmd_present(orig_pmd))) {
1617                 VM_BUG_ON(thp_migration_supported() &&
1618                                   !is_pmd_migration_entry(orig_pmd));
1619                 goto out;
1620         }
1621
1622         page = pmd_page(orig_pmd);
1623         /*
1624          * If other processes are mapping this page, we couldn't discard
1625          * the page unless they all do MADV_FREE so let's skip the page.
1626          */
1627         if (total_mapcount(page) != 1)
1628                 goto out;
1629
1630         if (!trylock_page(page))
1631                 goto out;
1632
1633         /*
1634          * If user want to discard part-pages of THP, split it so MADV_FREE
1635          * will deactivate only them.
1636          */
1637         if (next - addr != HPAGE_PMD_SIZE) {
1638                 get_page(page);
1639                 spin_unlock(ptl);
1640                 split_huge_page(page);
1641                 unlock_page(page);
1642                 put_page(page);
1643                 goto out_unlocked;
1644         }
1645
1646         if (PageDirty(page))
1647                 ClearPageDirty(page);
1648         unlock_page(page);
1649
1650         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1651                 pmdp_invalidate(vma, addr, pmd);
1652                 orig_pmd = pmd_mkold(orig_pmd);
1653                 orig_pmd = pmd_mkclean(orig_pmd);
1654
1655                 set_pmd_at(mm, addr, pmd, orig_pmd);
1656                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1657         }
1658
1659         mark_page_lazyfree(page);
1660         ret = true;
1661 out:
1662         spin_unlock(ptl);
1663 out_unlocked:
1664         return ret;
1665 }
1666
1667 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1668 {
1669         pgtable_t pgtable;
1670
1671         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1672         pte_free(mm, pgtable);
1673         atomic_long_dec(&mm->nr_ptes);
1674 }
1675
1676 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1677                  pmd_t *pmd, unsigned long addr)
1678 {
1679         pmd_t orig_pmd;
1680         spinlock_t *ptl;
1681
1682         tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1683
1684         ptl = __pmd_trans_huge_lock(pmd, vma);
1685         if (!ptl)
1686                 return 0;
1687         /*
1688          * For architectures like ppc64 we look at deposited pgtable
1689          * when calling pmdp_huge_get_and_clear. So do the
1690          * pgtable_trans_huge_withdraw after finishing pmdp related
1691          * operations.
1692          */
1693         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1694                         tlb->fullmm);
1695         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1696         if (vma_is_dax(vma)) {
1697                 if (arch_needs_pgtable_deposit())
1698                         zap_deposited_table(tlb->mm, pmd);
1699                 spin_unlock(ptl);
1700                 if (is_huge_zero_pmd(orig_pmd))
1701                         tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1702         } else if (is_huge_zero_pmd(orig_pmd)) {
1703                 zap_deposited_table(tlb->mm, pmd);
1704                 spin_unlock(ptl);
1705                 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1706         } else {
1707                 struct page *page = NULL;
1708                 int flush_needed = 1;
1709
1710                 if (pmd_present(orig_pmd)) {
1711                         page = pmd_page(orig_pmd);
1712                         page_remove_rmap(page, true);
1713                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1714                         VM_BUG_ON_PAGE(!PageHead(page), page);
1715                 } else if (thp_migration_supported()) {
1716                         swp_entry_t entry;
1717
1718                         VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1719                         entry = pmd_to_swp_entry(orig_pmd);
1720                         page = pfn_to_page(swp_offset(entry));
1721                         flush_needed = 0;
1722                 } else
1723                         WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1724
1725                 if (PageAnon(page)) {
1726                         zap_deposited_table(tlb->mm, pmd);
1727                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1728                 } else {
1729                         if (arch_needs_pgtable_deposit())
1730                                 zap_deposited_table(tlb->mm, pmd);
1731                         add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1732                 }
1733
1734                 spin_unlock(ptl);
1735                 if (flush_needed)
1736                         tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1737         }
1738         return 1;
1739 }
1740
1741 #ifndef pmd_move_must_withdraw
1742 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1743                                          spinlock_t *old_pmd_ptl,
1744                                          struct vm_area_struct *vma)
1745 {
1746         /*
1747          * With split pmd lock we also need to move preallocated
1748          * PTE page table if new_pmd is on different PMD page table.
1749          *
1750          * We also don't deposit and withdraw tables for file pages.
1751          */
1752         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1753 }
1754 #endif
1755
1756 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1757 {
1758 #ifdef CONFIG_MEM_SOFT_DIRTY
1759         if (unlikely(is_pmd_migration_entry(pmd)))
1760                 pmd = pmd_swp_mksoft_dirty(pmd);
1761         else if (pmd_present(pmd))
1762                 pmd = pmd_mksoft_dirty(pmd);
1763 #endif
1764         return pmd;
1765 }
1766
1767 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1768                   unsigned long new_addr, unsigned long old_end,
1769                   pmd_t *old_pmd, pmd_t *new_pmd)
1770 {
1771         spinlock_t *old_ptl, *new_ptl;
1772         pmd_t pmd;
1773         struct mm_struct *mm = vma->vm_mm;
1774         bool force_flush = false;
1775
1776         if ((old_addr & ~HPAGE_PMD_MASK) ||
1777             (new_addr & ~HPAGE_PMD_MASK) ||
1778             old_end - old_addr < HPAGE_PMD_SIZE)
1779                 return false;
1780
1781         /*
1782          * The destination pmd shouldn't be established, free_pgtables()
1783          * should have release it.
1784          */
1785         if (WARN_ON(!pmd_none(*new_pmd))) {
1786                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1787                 return false;
1788         }
1789
1790         /*
1791          * We don't have to worry about the ordering of src and dst
1792          * ptlocks because exclusive mmap_sem prevents deadlock.
1793          */
1794         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1795         if (old_ptl) {
1796                 new_ptl = pmd_lockptr(mm, new_pmd);
1797                 if (new_ptl != old_ptl)
1798                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1799                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1800                 if (pmd_present(pmd))
1801                         force_flush = true;
1802                 VM_BUG_ON(!pmd_none(*new_pmd));
1803
1804                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1805                         pgtable_t pgtable;
1806                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1807                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1808                 }
1809                 pmd = move_soft_dirty_pmd(pmd);
1810                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1811                 if (force_flush)
1812                         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1813                 if (new_ptl != old_ptl)
1814                         spin_unlock(new_ptl);
1815                 spin_unlock(old_ptl);
1816                 return true;
1817         }
1818         return false;
1819 }
1820
1821 /*
1822  * Returns
1823  *  - 0 if PMD could not be locked
1824  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1825  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1826  */
1827 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1828                 unsigned long addr, pgprot_t newprot, int prot_numa)
1829 {
1830         struct mm_struct *mm = vma->vm_mm;
1831         spinlock_t *ptl;
1832         pmd_t entry;
1833         bool preserve_write;
1834         int ret;
1835
1836         ptl = __pmd_trans_huge_lock(pmd, vma);
1837         if (!ptl)
1838                 return 0;
1839
1840         preserve_write = prot_numa && pmd_write(*pmd);
1841         ret = 1;
1842
1843 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1844         if (is_swap_pmd(*pmd)) {
1845                 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1846
1847                 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1848                 if (is_write_migration_entry(entry)) {
1849                         pmd_t newpmd;
1850                         /*
1851                          * A protection check is difficult so
1852                          * just be safe and disable write
1853                          */
1854                         make_migration_entry_read(&entry);
1855                         newpmd = swp_entry_to_pmd(entry);
1856                         if (pmd_swp_soft_dirty(*pmd))
1857                                 newpmd = pmd_swp_mksoft_dirty(newpmd);
1858                         set_pmd_at(mm, addr, pmd, newpmd);
1859                 }
1860                 goto unlock;
1861         }
1862 #endif
1863
1864         /*
1865          * Avoid trapping faults against the zero page. The read-only
1866          * data is likely to be read-cached on the local CPU and
1867          * local/remote hits to the zero page are not interesting.
1868          */
1869         if (prot_numa && is_huge_zero_pmd(*pmd))
1870                 goto unlock;
1871
1872         if (prot_numa && pmd_protnone(*pmd))
1873                 goto unlock;
1874
1875         /*
1876          * In case prot_numa, we are under down_read(mmap_sem). It's critical
1877          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1878          * which is also under down_read(mmap_sem):
1879          *
1880          *      CPU0:                           CPU1:
1881          *                              change_huge_pmd(prot_numa=1)
1882          *                               pmdp_huge_get_and_clear_notify()
1883          * madvise_dontneed()
1884          *  zap_pmd_range()
1885          *   pmd_trans_huge(*pmd) == 0 (without ptl)
1886          *   // skip the pmd
1887          *                               set_pmd_at();
1888          *                               // pmd is re-established
1889          *
1890          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1891          * which may break userspace.
1892          *
1893          * pmdp_invalidate() is required to make sure we don't miss
1894          * dirty/young flags set by hardware.
1895          */
1896         entry = *pmd;
1897         pmdp_invalidate(vma, addr, pmd);
1898
1899         /*
1900          * Recover dirty/young flags.  It relies on pmdp_invalidate to not
1901          * corrupt them.
1902          */
1903         if (pmd_dirty(*pmd))
1904                 entry = pmd_mkdirty(entry);
1905         if (pmd_young(*pmd))
1906                 entry = pmd_mkyoung(entry);
1907
1908         entry = pmd_modify(entry, newprot);
1909         if (preserve_write)
1910                 entry = pmd_mk_savedwrite(entry);
1911         ret = HPAGE_PMD_NR;
1912         set_pmd_at(mm, addr, pmd, entry);
1913         BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1914 unlock:
1915         spin_unlock(ptl);
1916         return ret;
1917 }
1918
1919 /*
1920  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1921  *
1922  * Note that if it returns page table lock pointer, this routine returns without
1923  * unlocking page table lock. So callers must unlock it.
1924  */
1925 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1926 {
1927         spinlock_t *ptl;
1928         ptl = pmd_lock(vma->vm_mm, pmd);
1929         if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1930                         pmd_devmap(*pmd)))
1931                 return ptl;
1932         spin_unlock(ptl);
1933         return NULL;
1934 }
1935
1936 /*
1937  * Returns true if a given pud maps a thp, false otherwise.
1938  *
1939  * Note that if it returns true, this routine returns without unlocking page
1940  * table lock. So callers must unlock it.
1941  */
1942 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1943 {
1944         spinlock_t *ptl;
1945
1946         ptl = pud_lock(vma->vm_mm, pud);
1947         if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1948                 return ptl;
1949         spin_unlock(ptl);
1950         return NULL;
1951 }
1952
1953 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1954 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1955                  pud_t *pud, unsigned long addr)
1956 {
1957         pud_t orig_pud;
1958         spinlock_t *ptl;
1959
1960         ptl = __pud_trans_huge_lock(pud, vma);
1961         if (!ptl)
1962                 return 0;
1963         /*
1964          * For architectures like ppc64 we look at deposited pgtable
1965          * when calling pudp_huge_get_and_clear. So do the
1966          * pgtable_trans_huge_withdraw after finishing pudp related
1967          * operations.
1968          */
1969         orig_pud = pudp_huge_get_and_clear_full(tlb->mm, addr, pud,
1970                         tlb->fullmm);
1971         tlb_remove_pud_tlb_entry(tlb, pud, addr);
1972         if (vma_is_dax(vma)) {
1973                 spin_unlock(ptl);
1974                 /* No zero page support yet */
1975         } else {
1976                 /* No support for anonymous PUD pages yet */
1977                 BUG();
1978         }
1979         return 1;
1980 }
1981
1982 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1983                 unsigned long haddr)
1984 {
1985         VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1986         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1987         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
1988         VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
1989
1990         count_vm_event(THP_SPLIT_PUD);
1991
1992         pudp_huge_clear_flush_notify(vma, haddr, pud);
1993 }
1994
1995 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
1996                 unsigned long address)
1997 {
1998         spinlock_t *ptl;
1999         struct mm_struct *mm = vma->vm_mm;
2000         unsigned long haddr = address & HPAGE_PUD_MASK;
2001
2002         mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PUD_SIZE);
2003         ptl = pud_lock(mm, pud);
2004         if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2005                 goto out;
2006         __split_huge_pud_locked(vma, pud, haddr);
2007
2008 out:
2009         spin_unlock(ptl);
2010         mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PUD_SIZE);
2011 }
2012 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2013
2014 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2015                 unsigned long haddr, pmd_t *pmd)
2016 {
2017         struct mm_struct *mm = vma->vm_mm;
2018         pgtable_t pgtable;
2019         pmd_t _pmd;
2020         int i;
2021
2022         /* leave pmd empty until pte is filled */
2023         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2024
2025         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2026         pmd_populate(mm, &_pmd, pgtable);
2027
2028         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2029                 pte_t *pte, entry;
2030                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2031                 entry = pte_mkspecial(entry);
2032                 pte = pte_offset_map(&_pmd, haddr);
2033                 VM_BUG_ON(!pte_none(*pte));
2034                 set_pte_at(mm, haddr, pte, entry);
2035                 pte_unmap(pte);
2036         }
2037         smp_wmb(); /* make pte visible before pmd */
2038         pmd_populate(mm, pmd, pgtable);
2039 }
2040
2041 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2042                 unsigned long haddr, bool freeze)
2043 {
2044         struct mm_struct *mm = vma->vm_mm;
2045         struct page *page;
2046         pgtable_t pgtable;
2047         pmd_t _pmd;
2048         bool young, write, dirty, soft_dirty, pmd_migration = false;
2049         unsigned long addr;
2050         int i;
2051
2052         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2053         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2054         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2055         VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2056                                 && !pmd_devmap(*pmd));
2057
2058         count_vm_event(THP_SPLIT_PMD);
2059
2060         if (!vma_is_anonymous(vma)) {
2061                 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2062                 /*
2063                  * We are going to unmap this huge page. So
2064                  * just go ahead and zap it
2065                  */
2066                 if (arch_needs_pgtable_deposit())
2067                         zap_deposited_table(mm, pmd);
2068                 if (vma_is_dax(vma))
2069                         return;
2070                 page = pmd_page(_pmd);
2071                 if (!PageDirty(page) && pmd_dirty(_pmd))
2072                         set_page_dirty(page);
2073                 if (!PageReferenced(page) && pmd_young(_pmd))
2074                         SetPageReferenced(page);
2075                 page_remove_rmap(page, true);
2076                 put_page(page);
2077                 add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR);
2078                 return;
2079         } else if (pmd_trans_huge(*pmd) && is_huge_zero_pmd(*pmd)) {
2080                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2081         }
2082
2083 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2084         pmd_migration = is_pmd_migration_entry(*pmd);
2085         if (pmd_migration) {
2086                 swp_entry_t entry;
2087
2088                 entry = pmd_to_swp_entry(*pmd);
2089                 page = pfn_to_page(swp_offset(entry));
2090         } else
2091 #endif
2092                 page = pmd_page(*pmd);
2093         VM_BUG_ON_PAGE(!page_count(page), page);
2094         page_ref_add(page, HPAGE_PMD_NR - 1);
2095         write = pmd_write(*pmd);
2096         young = pmd_young(*pmd);
2097         dirty = pmd_dirty(*pmd);
2098         soft_dirty = pmd_soft_dirty(*pmd);
2099
2100         pmdp_huge_split_prepare(vma, haddr, pmd);
2101         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2102         pmd_populate(mm, &_pmd, pgtable);
2103
2104         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2105                 pte_t entry, *pte;
2106                 /*
2107                  * Note that NUMA hinting access restrictions are not
2108                  * transferred to avoid any possibility of altering
2109                  * permissions across VMAs.
2110                  */
2111                 if (freeze || pmd_migration) {
2112                         swp_entry_t swp_entry;
2113                         swp_entry = make_migration_entry(page + i, write);
2114                         entry = swp_entry_to_pte(swp_entry);
2115                         if (soft_dirty)
2116                                 entry = pte_swp_mksoft_dirty(entry);
2117                 } else {
2118                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2119                         entry = maybe_mkwrite(entry, vma);
2120                         if (!write)
2121                                 entry = pte_wrprotect(entry);
2122                         if (!young)
2123                                 entry = pte_mkold(entry);
2124                         if (soft_dirty)
2125                                 entry = pte_mksoft_dirty(entry);
2126                 }
2127                 if (dirty)
2128                         SetPageDirty(page + i);
2129                 pte = pte_offset_map(&_pmd, addr);
2130                 BUG_ON(!pte_none(*pte));
2131                 set_pte_at(mm, addr, pte, entry);
2132                 if (!pmd_migration)
2133                         atomic_inc(&page[i]._mapcount);
2134                 pte_unmap(pte);
2135         }
2136
2137         if (!pmd_migration) {
2138                 /*
2139                  * Set PG_double_map before dropping compound_mapcount to avoid
2140                  * false-negative page_mapped().
2141                  */
2142                 if (compound_mapcount(page) > 1 &&
2143                     !TestSetPageDoubleMap(page)) {
2144                         for (i = 0; i < HPAGE_PMD_NR; i++)
2145                                 atomic_inc(&page[i]._mapcount);
2146                 }
2147
2148                 lock_page_memcg(page);
2149                 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2150                         /* Last compound_mapcount is gone. */
2151                         __dec_lruvec_page_state(page, NR_ANON_THPS);
2152                         if (TestClearPageDoubleMap(page)) {
2153                                 /* No need in mapcount reference anymore */
2154                                 for (i = 0; i < HPAGE_PMD_NR; i++)
2155                                         atomic_dec(&page[i]._mapcount);
2156                         }
2157                 }
2158                 unlock_page_memcg(page);
2159         }
2160
2161         smp_wmb(); /* make pte visible before pmd */
2162         /*
2163          * Up to this point the pmd is present and huge and userland has the
2164          * whole access to the hugepage during the split (which happens in
2165          * place). If we overwrite the pmd with the not-huge version pointing
2166          * to the pte here (which of course we could if all CPUs were bug
2167          * free), userland could trigger a small page size TLB miss on the
2168          * small sized TLB while the hugepage TLB entry is still established in
2169          * the huge TLB. Some CPU doesn't like that.
2170          * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2171          * 383 on page 93. Intel should be safe but is also warns that it's
2172          * only safe if the permission and cache attributes of the two entries
2173          * loaded in the two TLB is identical (which should be the case here).
2174          * But it is generally safer to never allow small and huge TLB entries
2175          * for the same virtual address to be loaded simultaneously. So instead
2176          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2177          * current pmd notpresent (atomically because here the pmd_trans_huge
2178          * and pmd_trans_splitting must remain set at all times on the pmd
2179          * until the split is complete for this pmd), then we flush the SMP TLB
2180          * and finally we write the non-huge version of the pmd entry with
2181          * pmd_populate.
2182          */
2183         pmdp_invalidate(vma, haddr, pmd);
2184         pmd_populate(mm, pmd, pgtable);
2185
2186         if (freeze) {
2187                 for (i = 0; i < HPAGE_PMD_NR; i++) {
2188                         page_remove_rmap(page + i, false);
2189                         put_page(page + i);
2190                 }
2191         }
2192 }
2193
2194 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2195                 unsigned long address, bool freeze, struct page *page)
2196 {
2197         spinlock_t *ptl;
2198         struct mm_struct *mm = vma->vm_mm;
2199         unsigned long haddr = address & HPAGE_PMD_MASK;
2200         bool do_unlock_page = false;
2201         pmd_t _pmd;
2202
2203         mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
2204         ptl = pmd_lock(mm, pmd);
2205
2206         /*
2207          * If caller asks to setup a migration entries, we need a page to check
2208          * pmd against. Otherwise we can end up replacing wrong page.
2209          */
2210         VM_BUG_ON(freeze && !page);
2211         if (page) {
2212                 VM_WARN_ON_ONCE(!PageLocked(page));
2213                 if (page != pmd_page(*pmd))
2214                         goto out;
2215         }
2216
2217 repeat:
2218         if (pmd_trans_huge(*pmd)) {
2219                 if (!page) {
2220                         page = pmd_page(*pmd);
2221                         /*
2222                          * An anonymous page must be locked, to ensure that a
2223                          * concurrent reuse_swap_page() sees stable mapcount;
2224                          * but reuse_swap_page() is not used on shmem or file,
2225                          * and page lock must not be taken when zap_pmd_range()
2226                          * calls __split_huge_pmd() while i_mmap_lock is held.
2227                          */
2228                         if (PageAnon(page)) {
2229                                 if (unlikely(!trylock_page(page))) {
2230                                         get_page(page);
2231                                         _pmd = *pmd;
2232                                         spin_unlock(ptl);
2233                                         lock_page(page);
2234                                         spin_lock(ptl);
2235                                         if (unlikely(!pmd_same(*pmd, _pmd))) {
2236                                                 unlock_page(page);
2237                                                 put_page(page);
2238                                                 page = NULL;
2239                                                 goto repeat;
2240                                         }
2241                                         put_page(page);
2242                                 }
2243                                 do_unlock_page = true;
2244                         }
2245                 }
2246                 if (PageMlocked(page))
2247                         clear_page_mlock(page);
2248         } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2249                 goto out;
2250         __split_huge_pmd_locked(vma, pmd, haddr, freeze);
2251 out:
2252         spin_unlock(ptl);
2253         if (do_unlock_page)
2254                 unlock_page(page);
2255         mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
2256 }
2257
2258 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2259                 bool freeze, struct page *page)
2260 {
2261         pgd_t *pgd;
2262         p4d_t *p4d;
2263         pud_t *pud;
2264         pmd_t *pmd;
2265
2266         pgd = pgd_offset(vma->vm_mm, address);
2267         if (!pgd_present(*pgd))
2268                 return;
2269
2270         p4d = p4d_offset(pgd, address);
2271         if (!p4d_present(*p4d))
2272                 return;
2273
2274         pud = pud_offset(p4d, address);
2275         if (!pud_present(*pud))
2276                 return;
2277
2278         pmd = pmd_offset(pud, address);
2279
2280         __split_huge_pmd(vma, pmd, address, freeze, page);
2281 }
2282
2283 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2284                              unsigned long start,
2285                              unsigned long end,
2286                              long adjust_next)
2287 {
2288         /*
2289          * If the new start address isn't hpage aligned and it could
2290          * previously contain an hugepage: check if we need to split
2291          * an huge pmd.
2292          */
2293         if (start & ~HPAGE_PMD_MASK &&
2294             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2295             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2296                 split_huge_pmd_address(vma, start, false, NULL);
2297
2298         /*
2299          * If the new end address isn't hpage aligned and it could
2300          * previously contain an hugepage: check if we need to split
2301          * an huge pmd.
2302          */
2303         if (end & ~HPAGE_PMD_MASK &&
2304             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2305             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2306                 split_huge_pmd_address(vma, end, false, NULL);
2307
2308         /*
2309          * If we're also updating the vma->vm_next->vm_start, if the new
2310          * vm_next->vm_start isn't page aligned and it could previously
2311          * contain an hugepage: check if we need to split an huge pmd.
2312          */
2313         if (adjust_next > 0) {
2314                 struct vm_area_struct *next = vma->vm_next;
2315                 unsigned long nstart = next->vm_start;
2316                 nstart += adjust_next << PAGE_SHIFT;
2317                 if (nstart & ~HPAGE_PMD_MASK &&
2318                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2319                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2320                         split_huge_pmd_address(next, nstart, false, NULL);
2321         }
2322 }
2323
2324 static void unmap_page(struct page *page)
2325 {
2326         enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
2327                 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD | TTU_SYNC;
2328
2329         VM_BUG_ON_PAGE(!PageHead(page), page);
2330
2331         if (PageAnon(page))
2332                 ttu_flags |= TTU_SPLIT_FREEZE;
2333
2334         try_to_unmap(page, ttu_flags);
2335
2336         VM_WARN_ON_ONCE_PAGE(page_mapped(page), page);
2337 }
2338
2339 static void remap_page(struct page *page)
2340 {
2341         int i;
2342         if (PageTransHuge(page)) {
2343                 remove_migration_ptes(page, page, true);
2344         } else {
2345                 for (i = 0; i < HPAGE_PMD_NR; i++)
2346                         remove_migration_ptes(page + i, page + i, true);
2347         }
2348 }
2349
2350 static void __split_huge_page_tail(struct page *head, int tail,
2351                 struct lruvec *lruvec, struct list_head *list)
2352 {
2353         struct page *page_tail = head + tail;
2354
2355         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2356
2357         /*
2358          * Clone page flags before unfreezing refcount.
2359          *
2360          * After successful get_page_unless_zero() might follow flags change,
2361          * for exmaple lock_page() which set PG_waiters.
2362          */
2363         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2364         page_tail->flags |= (head->flags &
2365                         ((1L << PG_referenced) |
2366                          (1L << PG_swapbacked) |
2367                          (1L << PG_swapcache) |
2368                          (1L << PG_mlocked) |
2369                          (1L << PG_uptodate) |
2370                          (1L << PG_active) |
2371                          (1L << PG_locked) |
2372                          (1L << PG_unevictable) |
2373                          (1L << PG_dirty)));
2374
2375         /* ->mapping in first tail page is compound_mapcount */
2376         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2377                         page_tail);
2378         page_tail->mapping = head->mapping;
2379         page_tail->index = head->index + tail;
2380
2381         /* Page flags must be visible before we make the page non-compound. */
2382         smp_wmb();
2383
2384         /*
2385          * Clear PageTail before unfreezing page refcount.
2386          *
2387          * After successful get_page_unless_zero() might follow put_page()
2388          * which needs correct compound_head().
2389          */
2390         clear_compound_head(page_tail);
2391
2392         /* Finally unfreeze refcount. Additional reference from page cache. */
2393         page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2394                                           PageSwapCache(head)));
2395
2396         if (page_is_young(head))
2397                 set_page_young(page_tail);
2398         if (page_is_idle(head))
2399                 set_page_idle(page_tail);
2400
2401         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2402         lru_add_page_tail(head, page_tail, lruvec, list);
2403 }
2404
2405 static void __split_huge_page(struct page *page, struct list_head *list,
2406                 pgoff_t end, unsigned long flags)
2407 {
2408         struct page *head = compound_head(page);
2409         struct zone *zone = page_zone(head);
2410         struct lruvec *lruvec;
2411         int i;
2412
2413         lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat);
2414
2415         /* complete memcg works before add pages to LRU */
2416         mem_cgroup_split_huge_fixup(head);
2417
2418         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
2419                 __split_huge_page_tail(head, i, lruvec, list);
2420                 /* Some pages can be beyond i_size: drop them from page cache */
2421                 if (head[i].index >= end) {
2422                         ClearPageDirty(head + i);
2423                         __delete_from_page_cache(head + i, NULL);
2424                         if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2425                                 shmem_uncharge(head->mapping->host, 1);
2426                         put_page(head + i);
2427                 }
2428         }
2429
2430         ClearPageCompound(head);
2431
2432         split_page_owner(head, HPAGE_PMD_ORDER);
2433
2434         /* See comment in __split_huge_page_tail() */
2435         if (PageAnon(head)) {
2436                 /* Additional pin to radix tree of swap cache */
2437                 if (PageSwapCache(head))
2438                         page_ref_add(head, 2);
2439                 else
2440                         page_ref_inc(head);
2441         } else {
2442                 /* Additional pin to radix tree */
2443                 page_ref_add(head, 2);
2444                 spin_unlock(&head->mapping->tree_lock);
2445         }
2446
2447         spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
2448
2449         remap_page(head);
2450
2451         for (i = 0; i < HPAGE_PMD_NR; i++) {
2452                 struct page *subpage = head + i;
2453                 if (subpage == page)
2454                         continue;
2455                 unlock_page(subpage);
2456
2457                 /*
2458                  * Subpages may be freed if there wasn't any mapping
2459                  * like if add_to_swap() is running on a lru page that
2460                  * had its mapping zapped. And freeing these pages
2461                  * requires taking the lru_lock so we do the put_page
2462                  * of the tail pages after the split is complete.
2463                  */
2464                 put_page(subpage);
2465         }
2466 }
2467
2468 int total_mapcount(struct page *page)
2469 {
2470         int i, compound, ret;
2471
2472         VM_BUG_ON_PAGE(PageTail(page), page);
2473
2474         if (likely(!PageCompound(page)))
2475                 return atomic_read(&page->_mapcount) + 1;
2476
2477         compound = compound_mapcount(page);
2478         if (PageHuge(page))
2479                 return compound;
2480         ret = compound;
2481         for (i = 0; i < HPAGE_PMD_NR; i++)
2482                 ret += atomic_read(&page[i]._mapcount) + 1;
2483         /* File pages has compound_mapcount included in _mapcount */
2484         if (!PageAnon(page))
2485                 return ret - compound * HPAGE_PMD_NR;
2486         if (PageDoubleMap(page))
2487                 ret -= HPAGE_PMD_NR;
2488         return ret;
2489 }
2490
2491 /*
2492  * This calculates accurately how many mappings a transparent hugepage
2493  * has (unlike page_mapcount() which isn't fully accurate). This full
2494  * accuracy is primarily needed to know if copy-on-write faults can
2495  * reuse the page and change the mapping to read-write instead of
2496  * copying them. At the same time this returns the total_mapcount too.
2497  *
2498  * The function returns the highest mapcount any one of the subpages
2499  * has. If the return value is one, even if different processes are
2500  * mapping different subpages of the transparent hugepage, they can
2501  * all reuse it, because each process is reusing a different subpage.
2502  *
2503  * The total_mapcount is instead counting all virtual mappings of the
2504  * subpages. If the total_mapcount is equal to "one", it tells the
2505  * caller all mappings belong to the same "mm" and in turn the
2506  * anon_vma of the transparent hugepage can become the vma->anon_vma
2507  * local one as no other process may be mapping any of the subpages.
2508  *
2509  * It would be more accurate to replace page_mapcount() with
2510  * page_trans_huge_mapcount(), however we only use
2511  * page_trans_huge_mapcount() in the copy-on-write faults where we
2512  * need full accuracy to avoid breaking page pinning, because
2513  * page_trans_huge_mapcount() is slower than page_mapcount().
2514  */
2515 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2516 {
2517         int i, ret, _total_mapcount, mapcount;
2518
2519         /* hugetlbfs shouldn't call it */
2520         VM_BUG_ON_PAGE(PageHuge(page), page);
2521
2522         if (likely(!PageTransCompound(page))) {
2523                 mapcount = atomic_read(&page->_mapcount) + 1;
2524                 if (total_mapcount)
2525                         *total_mapcount = mapcount;
2526                 return mapcount;
2527         }
2528
2529         page = compound_head(page);
2530
2531         _total_mapcount = ret = 0;
2532         for (i = 0; i < HPAGE_PMD_NR; i++) {
2533                 mapcount = atomic_read(&page[i]._mapcount) + 1;
2534                 ret = max(ret, mapcount);
2535                 _total_mapcount += mapcount;
2536         }
2537         if (PageDoubleMap(page)) {
2538                 ret -= 1;
2539                 _total_mapcount -= HPAGE_PMD_NR;
2540         }
2541         mapcount = compound_mapcount(page);
2542         ret += mapcount;
2543         _total_mapcount += mapcount;
2544         if (total_mapcount)
2545                 *total_mapcount = _total_mapcount;
2546         return ret;
2547 }
2548
2549 /* Racy check whether the huge page can be split */
2550 bool can_split_huge_page(struct page *page, int *pextra_pins)
2551 {
2552         int extra_pins;
2553
2554         /* Additional pins from radix tree */
2555         if (PageAnon(page))
2556                 extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0;
2557         else
2558                 extra_pins = HPAGE_PMD_NR;
2559         if (pextra_pins)
2560                 *pextra_pins = extra_pins;
2561         return total_mapcount(page) == page_count(page) - extra_pins - 1;
2562 }
2563
2564 /*
2565  * This function splits huge page into normal pages. @page can point to any
2566  * subpage of huge page to split. Split doesn't change the position of @page.
2567  *
2568  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2569  * The huge page must be locked.
2570  *
2571  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2572  *
2573  * Both head page and tail pages will inherit mapping, flags, and so on from
2574  * the hugepage.
2575  *
2576  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2577  * they are not mapped.
2578  *
2579  * Returns 0 if the hugepage is split successfully.
2580  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2581  * us.
2582  */
2583 int split_huge_page_to_list(struct page *page, struct list_head *list)
2584 {
2585         struct page *head = compound_head(page);
2586         struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2587         struct anon_vma *anon_vma = NULL;
2588         struct address_space *mapping = NULL;
2589         int extra_pins, ret;
2590         bool mlocked;
2591         unsigned long flags;
2592         pgoff_t end;
2593
2594         VM_BUG_ON_PAGE(is_huge_zero_page(head), head);
2595         VM_BUG_ON_PAGE(!PageLocked(page), page);
2596         VM_BUG_ON_PAGE(!PageCompound(page), page);
2597
2598         if (PageWriteback(page))
2599                 return -EBUSY;
2600
2601         if (PageAnon(head)) {
2602                 /*
2603                  * The caller does not necessarily hold an mmap_sem that would
2604                  * prevent the anon_vma disappearing so we first we take a
2605                  * reference to it and then lock the anon_vma for write. This
2606                  * is similar to page_lock_anon_vma_read except the write lock
2607                  * is taken to serialise against parallel split or collapse
2608                  * operations.
2609                  */
2610                 anon_vma = page_get_anon_vma(head);
2611                 if (!anon_vma) {
2612                         ret = -EBUSY;
2613                         goto out;
2614                 }
2615                 end = -1;
2616                 mapping = NULL;
2617                 anon_vma_lock_write(anon_vma);
2618         } else {
2619                 mapping = head->mapping;
2620
2621                 /* Truncated ? */
2622                 if (!mapping) {
2623                         ret = -EBUSY;
2624                         goto out;
2625                 }
2626
2627                 anon_vma = NULL;
2628                 i_mmap_lock_read(mapping);
2629
2630                 /*
2631                  *__split_huge_page() may need to trim off pages beyond EOF:
2632                  * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2633                  * which cannot be nested inside the page tree lock. So note
2634                  * end now: i_size itself may be changed at any moment, but
2635                  * head page lock is good enough to serialize the trimming.
2636                  */
2637                 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2638         }
2639
2640         /*
2641          * Racy check if we can split the page, before unmap_page() will
2642          * split PMDs
2643          */
2644         if (!can_split_huge_page(head, &extra_pins)) {
2645                 ret = -EBUSY;
2646                 goto out_unlock;
2647         }
2648
2649         mlocked = PageMlocked(page);
2650         unmap_page(head);
2651
2652         /* Make sure the page is not on per-CPU pagevec as it takes pin */
2653         if (mlocked)
2654                 lru_add_drain();
2655
2656         /* prevent PageLRU to go away from under us, and freeze lru stats */
2657         spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags);
2658
2659         if (mapping) {
2660                 void **pslot;
2661
2662                 spin_lock(&mapping->tree_lock);
2663                 pslot = radix_tree_lookup_slot(&mapping->page_tree,
2664                                 page_index(head));
2665                 /*
2666                  * Check if the head page is present in radix tree.
2667                  * We assume all tail are present too, if head is there.
2668                  */
2669                 if (radix_tree_deref_slot_protected(pslot,
2670                                         &mapping->tree_lock) != head)
2671                         goto fail;
2672         }
2673
2674         /* Prevent deferred_split_scan() touching ->_refcount */
2675         spin_lock(&pgdata->split_queue_lock);
2676         if (page_ref_freeze(head, 1 + extra_pins)) {
2677                 if (!list_empty(page_deferred_list(head))) {
2678                         pgdata->split_queue_len--;
2679                         list_del(page_deferred_list(head));
2680                 }
2681                 if (mapping)
2682                         __dec_node_page_state(page, NR_SHMEM_THPS);
2683                 spin_unlock(&pgdata->split_queue_lock);
2684                 __split_huge_page(page, list, end, flags);
2685                 if (PageSwapCache(head)) {
2686                         swp_entry_t entry = { .val = page_private(head) };
2687
2688                         ret = split_swap_cluster(entry);
2689                 } else
2690                         ret = 0;
2691         } else {
2692                 spin_unlock(&pgdata->split_queue_lock);
2693 fail:
2694                 if (mapping)
2695                         spin_unlock(&mapping->tree_lock);
2696                 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
2697                 remap_page(head);
2698                 ret = -EBUSY;
2699         }
2700
2701 out_unlock:
2702         if (anon_vma) {
2703                 anon_vma_unlock_write(anon_vma);
2704                 put_anon_vma(anon_vma);
2705         }
2706         if (mapping)
2707                 i_mmap_unlock_read(mapping);
2708 out:
2709         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2710         return ret;
2711 }
2712
2713 void free_transhuge_page(struct page *page)
2714 {
2715         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2716         unsigned long flags;
2717
2718         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2719         if (!list_empty(page_deferred_list(page))) {
2720                 pgdata->split_queue_len--;
2721                 list_del(page_deferred_list(page));
2722         }
2723         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2724         free_compound_page(page);
2725 }
2726
2727 void deferred_split_huge_page(struct page *page)
2728 {
2729         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2730         unsigned long flags;
2731
2732         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2733
2734         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2735         if (list_empty(page_deferred_list(page))) {
2736                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2737                 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2738                 pgdata->split_queue_len++;
2739         }
2740         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2741 }
2742
2743 static unsigned long deferred_split_count(struct shrinker *shrink,
2744                 struct shrink_control *sc)
2745 {
2746         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2747         return ACCESS_ONCE(pgdata->split_queue_len);
2748 }
2749
2750 static unsigned long deferred_split_scan(struct shrinker *shrink,
2751                 struct shrink_control *sc)
2752 {
2753         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2754         unsigned long flags;
2755         LIST_HEAD(list), *pos, *next;
2756         struct page *page;
2757         int split = 0;
2758
2759         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2760         /* Take pin on all head pages to avoid freeing them under us */
2761         list_for_each_safe(pos, next, &pgdata->split_queue) {
2762                 page = list_entry((void *)pos, struct page, mapping);
2763                 page = compound_head(page);
2764                 if (get_page_unless_zero(page)) {
2765                         list_move(page_deferred_list(page), &list);
2766                 } else {
2767                         /* We lost race with put_compound_page() */
2768                         list_del_init(page_deferred_list(page));
2769                         pgdata->split_queue_len--;
2770                 }
2771                 if (!--sc->nr_to_scan)
2772                         break;
2773         }
2774         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2775
2776         list_for_each_safe(pos, next, &list) {
2777                 page = list_entry((void *)pos, struct page, mapping);
2778                 if (!trylock_page(page))
2779                         goto next;
2780                 /* split_huge_page() removes page from list on success */
2781                 if (!split_huge_page(page))
2782                         split++;
2783                 unlock_page(page);
2784 next:
2785                 put_page(page);
2786         }
2787
2788         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2789         list_splice_tail(&list, &pgdata->split_queue);
2790         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2791
2792         /*
2793          * Stop shrinker if we didn't split any page, but the queue is empty.
2794          * This can happen if pages were freed under us.
2795          */
2796         if (!split && list_empty(&pgdata->split_queue))
2797                 return SHRINK_STOP;
2798         return split;
2799 }
2800
2801 static struct shrinker deferred_split_shrinker = {
2802         .count_objects = deferred_split_count,
2803         .scan_objects = deferred_split_scan,
2804         .seeks = DEFAULT_SEEKS,
2805         .flags = SHRINKER_NUMA_AWARE,
2806 };
2807
2808 #ifdef CONFIG_DEBUG_FS
2809 static int split_huge_pages_set(void *data, u64 val)
2810 {
2811         struct zone *zone;
2812         struct page *page;
2813         unsigned long pfn, max_zone_pfn;
2814         unsigned long total = 0, split = 0;
2815
2816         if (val != 1)
2817                 return -EINVAL;
2818
2819         for_each_populated_zone(zone) {
2820                 max_zone_pfn = zone_end_pfn(zone);
2821                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2822                         if (!pfn_valid(pfn))
2823                                 continue;
2824
2825                         page = pfn_to_page(pfn);
2826                         if (!get_page_unless_zero(page))
2827                                 continue;
2828
2829                         if (zone != page_zone(page))
2830                                 goto next;
2831
2832                         if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2833                                 goto next;
2834
2835                         total++;
2836                         lock_page(page);
2837                         if (!split_huge_page(page))
2838                                 split++;
2839                         unlock_page(page);
2840 next:
2841                         put_page(page);
2842                 }
2843         }
2844
2845         pr_info("%lu of %lu THP split\n", split, total);
2846
2847         return 0;
2848 }
2849 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2850                 "%llu\n");
2851
2852 static int __init split_huge_pages_debugfs(void)
2853 {
2854         void *ret;
2855
2856         ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2857                         &split_huge_pages_fops);
2858         if (!ret)
2859                 pr_warn("Failed to create split_huge_pages in debugfs");
2860         return 0;
2861 }
2862 late_initcall(split_huge_pages_debugfs);
2863 #endif
2864
2865 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2866 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
2867                 struct page *page)
2868 {
2869         struct vm_area_struct *vma = pvmw->vma;
2870         struct mm_struct *mm = vma->vm_mm;
2871         unsigned long address = pvmw->address;
2872         pmd_t pmdval;
2873         swp_entry_t entry;
2874         pmd_t pmdswp;
2875
2876         if (!(pvmw->pmd && !pvmw->pte))
2877                 return;
2878
2879         flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
2880         pmdval = *pvmw->pmd;
2881         pmdp_invalidate(vma, address, pvmw->pmd);
2882         if (pmd_dirty(pmdval))
2883                 set_page_dirty(page);
2884         entry = make_migration_entry(page, pmd_write(pmdval));
2885         pmdswp = swp_entry_to_pmd(entry);
2886         if (pmd_soft_dirty(pmdval))
2887                 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
2888         set_pmd_at(mm, address, pvmw->pmd, pmdswp);
2889         page_remove_rmap(page, true);
2890         put_page(page);
2891 }
2892
2893 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
2894 {
2895         struct vm_area_struct *vma = pvmw->vma;
2896         struct mm_struct *mm = vma->vm_mm;
2897         unsigned long address = pvmw->address;
2898         unsigned long mmun_start = address & HPAGE_PMD_MASK;
2899         pmd_t pmde;
2900         swp_entry_t entry;
2901
2902         if (!(pvmw->pmd && !pvmw->pte))
2903                 return;
2904
2905         entry = pmd_to_swp_entry(*pvmw->pmd);
2906         get_page(new);
2907         pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
2908         if (pmd_swp_soft_dirty(*pvmw->pmd))
2909                 pmde = pmd_mksoft_dirty(pmde);
2910         if (is_write_migration_entry(entry))
2911                 pmde = maybe_pmd_mkwrite(pmde, vma);
2912
2913         flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
2914         page_add_anon_rmap(new, vma, mmun_start, true);
2915         set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
2916         if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new))
2917                 mlock_vma_page(new);
2918         update_mmu_cache_pmd(vma, address, pvmw->pmd);
2919 }
2920 #endif