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
  37 #include <asm/tlb.h>
  38 #include <asm/pgalloc.h>
  39 #include "internal.h"
  40
  41 /*
  42  * By default transparent hugepage support is disabled in order that avoid
  43  * to risk increase the memory footprint of applications without a guaranteed
  44  * benefit. When transparent hugepage support is enabled, is for all mappings,
  45  * and khugepaged scans all mappings.
  46  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  47  * for all hugepage allocations.
  48  */
  49 unsigned long transparent_hugepage_flags __read_mostly =
  50 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  51         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  52 #endif
  53 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  54         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  55 #endif
  56         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
  57         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  58         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  59
  60 static struct shrinker deferred_split_shrinker;
  61
  62 static atomic_t huge_zero_refcount;
  63 struct page *huge_zero_page __read_mostly;
  64
  65 static struct page *get_huge_zero_page(void)
  66 {
  67         struct page *zero_page;
  68 retry:
  69         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
  70                 return READ_ONCE(huge_zero_page);
  71
  72         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
  73                         HPAGE_PMD_ORDER);
  74         if (!zero_page) {
  75                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
  76                 return NULL;
  77         }
  78         count_vm_event(THP_ZERO_PAGE_ALLOC);
  79         preempt_disable();
  80         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
  81                 preempt_enable();
  82                 __free_pages(zero_page, compound_order(zero_page));
  83                 goto retry;
  84         }
  85
  86         /* We take additional reference here. It will be put back by shrinker */
  87         atomic_set(&huge_zero_refcount, 2);
  88         preempt_enable();
  89         return READ_ONCE(huge_zero_page);
  90 }
  91
  92 static void put_huge_zero_page(void)
  93 {
  94         /*
  95          * Counter should never go to zero here. Only shrinker can put
  96          * last reference.
  97          */
  98         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
  99 }
 100
 101 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
 102 {
 103         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 104                 return READ_ONCE(huge_zero_page);
 105
 106         if (!get_huge_zero_page())
 107                 return NULL;
 108
 109         if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 110                 put_huge_zero_page();
 111
 112         return READ_ONCE(huge_zero_page);
 113 }
 114
 115 void mm_put_huge_zero_page(struct mm_struct *mm)
 116 {
 117         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 118                 put_huge_zero_page();
 119 }
 120
 121 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 122                                         struct shrink_control *sc)
 123 {
 124         /* we can free zero page only if last reference remains */
 125         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 126 }
 127
 128 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 129                                        struct shrink_control *sc)
 130 {
 131         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 132                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 133                 BUG_ON(zero_page == NULL);
 134                 __free_pages(zero_page, compound_order(zero_page));
 135                 return HPAGE_PMD_NR;
 136         }
 137
 138         return 0;
 139 }
 140
 141 static struct shrinker huge_zero_page_shrinker = {
 142         .count_objects = shrink_huge_zero_page_count,
 143         .scan_objects = shrink_huge_zero_page_scan,
 144         .seeks = DEFAULT_SEEKS,
 145 };
 146
 147 #ifdef CONFIG_SYSFS
 148 static ssize_t enabled_show(struct kobject *kobj,
 149                             struct kobj_attribute *attr, char *buf)
 150 {
 151         if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
 152                 return sprintf(buf, "[always] madvise never\n");
 153         else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
 154                 return sprintf(buf, "always [madvise] never\n");
 155         else
 156                 return sprintf(buf, "always madvise [never]\n");
 157 }
 158
 159 static ssize_t enabled_store(struct kobject *kobj,
 160                              struct kobj_attribute *attr,
 161                              const char *buf, size_t count)
 162 {
 163         ssize_t ret = count;
 164
 165         if (!memcmp("always", buf,
 166                     min(sizeof("always")-1, count))) {
 167                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 168                 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 169         } else if (!memcmp("madvise", buf,
 170                            min(sizeof("madvise")-1, count))) {
 171                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 172                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 173         } else if (!memcmp("never", buf,
 174                            min(sizeof("never")-1, count))) {
 175                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 176                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 177         } else
 178                 ret = -EINVAL;
 179
 180         if (ret > 0) {
 181                 int err = start_stop_khugepaged();
 182                 if (err)
 183                         ret = err;
 184         }
 185         return ret;
 186 }
 187 static struct kobj_attribute enabled_attr =
 188         __ATTR(enabled, 0644, enabled_show, enabled_store);
 189
 190 ssize_t single_hugepage_flag_show(struct kobject *kobj,
 191                                 struct kobj_attribute *attr, char *buf,
 192                                 enum transparent_hugepage_flag flag)
 193 {
 194         return sprintf(buf, "%d\n",
 195                        !!test_bit(flag, &transparent_hugepage_flags));
 196 }
 197
 198 ssize_t single_hugepage_flag_store(struct kobject *kobj,
 199                                  struct kobj_attribute *attr,
 200                                  const char *buf, size_t count,
 201                                  enum transparent_hugepage_flag flag)
 202 {
 203         unsigned long value;
 204         int ret;
 205
 206         ret = kstrtoul(buf, 10, &value);
 207         if (ret < 0)
 208                 return ret;
 209         if (value > 1)
 210                 return -EINVAL;
 211
 212         if (value)
 213                 set_bit(flag, &transparent_hugepage_flags);
 214         else
 215                 clear_bit(flag, &transparent_hugepage_flags);
 216
 217         return count;
 218 }
 219
 220 static ssize_t defrag_show(struct kobject *kobj,
 221                            struct kobj_attribute *attr, char *buf)
 222 {
 223         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 224                 return sprintf(buf, "[always] defer defer+madvise madvise never\n");
 225         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 226                 return sprintf(buf, "always [defer] defer+madvise madvise never\n");
 227         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 228                 return sprintf(buf, "always defer [defer+madvise] madvise never\n");
 229         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 230                 return sprintf(buf, "always defer defer+madvise [madvise] never\n");
 231         return sprintf(buf, "always defer defer+madvise madvise [never]\n");
 232 }
 233
 234 static ssize_t defrag_store(struct kobject *kobj,
 235                             struct kobj_attribute *attr,
 236                             const char *buf, size_t count)
 237 {
 238         if (!memcmp("always", buf,
 239                     min(sizeof("always")-1, count))) {
 240                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 241                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 242                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 243                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 244         } else if (!memcmp("defer+madvise", buf,
 245                     min(sizeof("defer+madvise")-1, count))) {
 246                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 247                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 248                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 249                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 250         } else if (!memcmp("defer", buf,
 251                     min(sizeof("defer")-1, count))) {
 252                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 253                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 254                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 255                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 256         } else if (!memcmp("madvise", buf,
 257                            min(sizeof("madvise")-1, count))) {
 258                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 259                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 260                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 261                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 262         } else if (!memcmp("never", buf,
 263                            min(sizeof("never")-1, count))) {
 264                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 265                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 266                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 267                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 268         } else
 269                 return -EINVAL;
 270
 271         return count;
 272 }
 273 static struct kobj_attribute defrag_attr =
 274         __ATTR(defrag, 0644, defrag_show, defrag_store);
 275
 276 static ssize_t use_zero_page_show(struct kobject *kobj,
 277                 struct kobj_attribute *attr, char *buf)
 278 {
 279         return single_hugepage_flag_show(kobj, attr, buf,
 280                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 281 }
 282 static ssize_t use_zero_page_store(struct kobject *kobj,
 283                 struct kobj_attribute *attr, const char *buf, size_t count)
 284 {
 285         return single_hugepage_flag_store(kobj, attr, buf, count,
 286                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 287 }
 288 static struct kobj_attribute use_zero_page_attr =
 289         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 290
 291 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
 292                 struct kobj_attribute *attr, char *buf)
 293 {
 294         return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE);
 295 }
 296 static struct kobj_attribute hpage_pmd_size_attr =
 297         __ATTR_RO(hpage_pmd_size);
 298
 299 #ifdef CONFIG_DEBUG_VM
 300 static ssize_t debug_cow_show(struct kobject *kobj,
 301                                 struct kobj_attribute *attr, char *buf)
 302 {
 303         return single_hugepage_flag_show(kobj, attr, buf,
 304                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 305 }
 306 static ssize_t debug_cow_store(struct kobject *kobj,
 307                                struct kobj_attribute *attr,
 308                                const char *buf, size_t count)
 309 {
 310         return single_hugepage_flag_store(kobj, attr, buf, count,
 311                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 312 }
 313 static struct kobj_attribute debug_cow_attr =
 314         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 315 #endif /* CONFIG_DEBUG_VM */
 316
 317 static struct attribute *hugepage_attr[] = {
 318         &enabled_attr.attr,
 319         &defrag_attr.attr,
 320         &use_zero_page_attr.attr,
 321         &hpage_pmd_size_attr.attr,
 322 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
 323         &shmem_enabled_attr.attr,
 324 #endif
 325 #ifdef CONFIG_DEBUG_VM
 326         &debug_cow_attr.attr,
 327 #endif
 328         NULL,
 329 };
 330
 331 static const struct attribute_group hugepage_attr_group = {
 332         .attrs = hugepage_attr,
 333 };
 334
 335 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 336 {
 337         int err;
 338
 339         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 340         if (unlikely(!*hugepage_kobj)) {
 341                 pr_err("failed to create transparent hugepage kobject\n");
 342                 return -ENOMEM;
 343         }
 344
 345         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 346         if (err) {
 347                 pr_err("failed to register transparent hugepage group\n");
 348                 goto delete_obj;
 349         }
 350
 351         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 352         if (err) {
 353                 pr_err("failed to register transparent hugepage group\n");
 354                 goto remove_hp_group;
 355         }
 356
 357         return 0;
 358
 359 remove_hp_group:
 360         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 361 delete_obj:
 362         kobject_put(*hugepage_kobj);
 363         return err;
 364 }
 365
 366 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 367 {
 368         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 369         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 370         kobject_put(hugepage_kobj);
 371 }
 372 #else
 373 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 374 {
 375         return 0;
 376 }
 377
 378 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 379 {
 380 }
 381 #endif /* CONFIG_SYSFS */
 382
 383 static int __init hugepage_init(void)
 384 {
 385         int err;
 386         struct kobject *hugepage_kobj;
 387
 388         if (!has_transparent_hugepage()) {
 389                 transparent_hugepage_flags = 0;
 390                 return -EINVAL;
 391         }
 392
 393         /*
 394          * hugepages can't be allocated by the buddy allocator
 395          */
 396         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
 397         /*
 398          * we use page->mapping and page->index in second tail page
 399          * as list_head: assuming THP order >= 2
 400          */
 401         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 402
 403         err = hugepage_init_sysfs(&hugepage_kobj);
 404         if (err)
 405                 goto err_sysfs;
 406
 407         err = khugepaged_init();
 408         if (err)
 409                 goto err_slab;
 410
 411         err = register_shrinker(&huge_zero_page_shrinker);
 412         if (err)
 413                 goto err_hzp_shrinker;
 414         err = register_shrinker(&deferred_split_shrinker);
 415         if (err)
 416                 goto err_split_shrinker;
 417
 418         /*
 419          * By default disable transparent hugepages on smaller systems,
 420          * where the extra memory used could hurt more than TLB overhead
 421          * is likely to save.  The admin can still enable it through /sys.
 422          */
 423         if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
 424                 transparent_hugepage_flags = 0;
 425                 return 0;
 426         }
 427
 428         err = start_stop_khugepaged();
 429         if (err)
 430                 goto err_khugepaged;
 431
 432         return 0;
 433 err_khugepaged:
 434         unregister_shrinker(&deferred_split_shrinker);
 435 err_split_shrinker:
 436         unregister_shrinker(&huge_zero_page_shrinker);
 437 err_hzp_shrinker:
 438         khugepaged_destroy();
 439 err_slab:
 440         hugepage_exit_sysfs(hugepage_kobj);
 441 err_sysfs:
 442         return err;
 443 }
 444 subsys_initcall(hugepage_init);
 445
 446 static int __init setup_transparent_hugepage(char *str)
 447 {
 448         int ret = 0;
 449         if (!str)
 450                 goto out;
 451         if (!strcmp(str, "always")) {
 452                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 453                         &transparent_hugepage_flags);
 454                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 455                           &transparent_hugepage_flags);
 456                 ret = 1;
 457         } else if (!strcmp(str, "madvise")) {
 458                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 459                           &transparent_hugepage_flags);
 460                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 461                         &transparent_hugepage_flags);
 462                 ret = 1;
 463         } else if (!strcmp(str, "never")) {
 464                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 465                           &transparent_hugepage_flags);
 466                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 467                           &transparent_hugepage_flags);
 468                 ret = 1;
 469         }
 470 out:
 471         if (!ret)
 472                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 473         return ret;
 474 }
 475 __setup("transparent_hugepage=", setup_transparent_hugepage);
 476
 477 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 478 {
 479         if (likely(vma->vm_flags & VM_WRITE))
 480                 pmd = pmd_mkwrite(pmd);
 481         return pmd;
 482 }
 483
 484 static inline struct list_head *page_deferred_list(struct page *page)
 485 {
 486         /*
 487          * ->lru in the tail pages is occupied by compound_head.
 488          * Let's use ->mapping + ->index in the second tail page as list_head.
 489          */
 490         return (struct list_head *)&page[2].mapping;
 491 }
 492
 493 void prep_transhuge_page(struct page *page)
 494 {
 495         /*
 496          * we use page->mapping and page->indexlru in second tail page
 497          * as list_head: assuming THP order >= 2
 498          */
 499
 500         INIT_LIST_HEAD(page_deferred_list(page));
 501         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 502 }
 503
 504 unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len,
 505                 loff_t off, unsigned long flags, unsigned long size)
 506 {
 507         unsigned long addr;
 508         loff_t off_end = off + len;
 509         loff_t off_align = round_up(off, size);
 510         unsigned long len_pad;
 511
 512         if (off_end <= off_align || (off_end - off_align) < size)
 513                 return 0;
 514
 515         len_pad = len + size;
 516         if (len_pad < len || (off + len_pad) < off)
 517                 return 0;
 518
 519         addr = current->mm->get_unmapped_area(filp, 0, len_pad,
 520                                               off >> PAGE_SHIFT, flags);
 521         if (IS_ERR_VALUE(addr))
 522                 return 0;
 523
 524         addr += (off - addr) & (size - 1);
 525         return addr;
 526 }
 527
 528 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 529                 unsigned long len, unsigned long pgoff, unsigned long flags)
 530 {
 531         loff_t off = (loff_t)pgoff << PAGE_SHIFT;
 532
 533         if (addr)
 534                 goto out;
 535         if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
 536                 goto out;
 537
 538         addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE);
 539         if (addr)
 540                 return addr;
 541
 542  out:
 543         return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
 544 }
 545 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
 546
 547 static int __do_huge_pmd_anonymous_page(struct vm_fault *vmf, struct page *page,
 548                 gfp_t gfp)
 549 {
 550         struct vm_area_struct *vma = vmf->vma;
 551         struct mem_cgroup *memcg;
 552         pgtable_t pgtable;
 553         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 554         int ret = 0;
 555
 556         VM_BUG_ON_PAGE(!PageCompound(page), page);
 557
 558         if (mem_cgroup_try_charge(page, vma->vm_mm, gfp | __GFP_NORETRY, &memcg,
 559                                   true)) {
 560                 put_page(page);
 561                 count_vm_event(THP_FAULT_FALLBACK);
 562                 return VM_FAULT_FALLBACK;
 563         }
 564
 565         pgtable = pte_alloc_one(vma->vm_mm, haddr);
 566         if (unlikely(!pgtable)) {
 567                 ret = VM_FAULT_OOM;
 568                 goto release;
 569         }
 570
 571         clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
 572         /*
 573          * The memory barrier inside __SetPageUptodate makes sure that
 574          * clear_huge_page writes become visible before the set_pmd_at()
 575          * write.
 576          */
 577         __SetPageUptodate(page);
 578
 579         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 580         if (unlikely(!pmd_none(*vmf->pmd))) {
 581                 goto unlock_release;
 582         } else {
 583                 pmd_t entry;
 584
 585                 ret = check_stable_address_space(vma->vm_mm);
 586                 if (ret)
 587                         goto unlock_release;
 588
 589                 /* Deliver the page fault to userland */
 590                 if (userfaultfd_missing(vma)) {
 591                         int ret;
 592
 593                         spin_unlock(vmf->ptl);
 594                         mem_cgroup_cancel_charge(page, memcg, true);
 595                         put_page(page);
 596                         pte_free(vma->vm_mm, pgtable);
 597                         ret = handle_userfault(vmf, VM_UFFD_MISSING);
 598                         VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 599                         return ret;
 600                 }
 601
 602                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 603                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 604                 page_add_new_anon_rmap(page, vma, haddr, true);
 605                 mem_cgroup_commit_charge(page, memcg, false, true);
 606                 lru_cache_add_active_or_unevictable(page, vma);
 607                 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
 608                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
 609                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 610                 atomic_long_inc(&vma->vm_mm->nr_ptes);
 611                 spin_unlock(vmf->ptl);
 612                 count_vm_event(THP_FAULT_ALLOC);
 613         }
 614
 615         return 0;
 616 unlock_release:
 617         spin_unlock(vmf->ptl);
 618 release:
 619         if (pgtable)
 620                 pte_free(vma->vm_mm, pgtable);
 621         mem_cgroup_cancel_charge(page, memcg, true);
 622         put_page(page);
 623         return ret;
 624
 625 }
 626
 627 /*
 628  * always: directly stall for all thp allocations
 629  * defer: wake kswapd and fail if not immediately available
 630  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
 631  *                fail if not immediately available
 632  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
 633  *          available
 634  * never: never stall for any thp allocation
 635  */
 636 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
 637 {
 638         const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
 639
 640         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 641                 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
 642         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 643                 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
 644         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 645                 return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM :
 646                                                              __GFP_KSWAPD_RECLAIM);
 647         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 648                 return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM :
 649                                                              0);
 650         return GFP_TRANSHUGE_LIGHT;
 651 }
 652
 653 /* Caller must hold page table lock. */
 654 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 655                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 656                 struct page *zero_page)
 657 {
 658         pmd_t entry;
 659         if (!pmd_none(*pmd))
 660                 return false;
 661         entry = mk_pmd(zero_page, vma->vm_page_prot);
 662         entry = pmd_mkhuge(entry);
 663         if (pgtable)
 664                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 665         set_pmd_at(mm, haddr, pmd, entry);
 666         atomic_long_inc(&mm->nr_ptes);
 667         return true;
 668 }
 669
 670 int do_huge_pmd_anonymous_page(struct vm_fault *vmf)
 671 {
 672         struct vm_area_struct *vma = vmf->vma;
 673         gfp_t gfp;
 674         struct page *page;
 675         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 676
 677         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 678                 return VM_FAULT_FALLBACK;
 679         if (unlikely(anon_vma_prepare(vma)))
 680                 return VM_FAULT_OOM;
 681         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 682                 return VM_FAULT_OOM;
 683         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
 684                         !mm_forbids_zeropage(vma->vm_mm) &&
 685                         transparent_hugepage_use_zero_page()) {
 686                 pgtable_t pgtable;
 687                 struct page *zero_page;
 688                 bool set;
 689                 int ret;
 690                 pgtable = pte_alloc_one(vma->vm_mm, haddr);
 691                 if (unlikely(!pgtable))
 692                         return VM_FAULT_OOM;
 693                 zero_page = mm_get_huge_zero_page(vma->vm_mm);
 694                 if (unlikely(!zero_page)) {
 695                         pte_free(vma->vm_mm, pgtable);
 696                         count_vm_event(THP_FAULT_FALLBACK);
 697                         return VM_FAULT_FALLBACK;
 698                 }
 699                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 700                 ret = 0;
 701                 set = false;
 702                 if (pmd_none(*vmf->pmd)) {
 703                         ret = check_stable_address_space(vma->vm_mm);
 704                         if (ret) {
 705                                 spin_unlock(vmf->ptl);
 706                         } else if (userfaultfd_missing(vma)) {
 707                                 spin_unlock(vmf->ptl);
 708                                 ret = handle_userfault(vmf, VM_UFFD_MISSING);
 709                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 710                         } else {
 711                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
 712                                                    haddr, vmf->pmd, zero_page);
 713                                 spin_unlock(vmf->ptl);
 714                                 set = true;
 715                         }
 716                 } else
 717                         spin_unlock(vmf->ptl);
 718                 if (!set)
 719                         pte_free(vma->vm_mm, pgtable);
 720                 return ret;
 721         }
 722         gfp = alloc_hugepage_direct_gfpmask(vma);
 723         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 724         if (unlikely(!page)) {
 725                 count_vm_event(THP_FAULT_FALLBACK);
 726                 return VM_FAULT_FALLBACK;
 727         }
 728         prep_transhuge_page(page);
 729         return __do_huge_pmd_anonymous_page(vmf, page, gfp);
 730 }
 731
 732 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 733                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
 734                 pgtable_t pgtable)
 735 {
 736         struct mm_struct *mm = vma->vm_mm;
 737         pmd_t entry;
 738         spinlock_t *ptl;
 739
 740         ptl = pmd_lock(mm, pmd);
 741         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 742         if (pfn_t_devmap(pfn))
 743                 entry = pmd_mkdevmap(entry);
 744         if (write) {
 745                 entry = pmd_mkyoung(pmd_mkdirty(entry));
 746                 entry = maybe_pmd_mkwrite(entry, vma);
 747         }
 748
 749         if (pgtable) {
 750                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 751                 atomic_long_inc(&mm->nr_ptes);
 752         }
 753
 754         set_pmd_at(mm, addr, pmd, entry);
 755         update_mmu_cache_pmd(vma, addr, pmd);
 756         spin_unlock(ptl);
 757 }
 758
 759 int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 760                         pmd_t *pmd, pfn_t pfn, bool write)
 761 {
 762         pgprot_t pgprot = vma->vm_page_prot;
 763         pgtable_t pgtable = NULL;
 764         /*
 765          * If we had pmd_special, we could avoid all these restrictions,
 766          * but we need to be consistent with PTEs and architectures that
 767          * can't support a 'special' bit.
 768          */
 769         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
 770         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 771                                                 (VM_PFNMAP|VM_MIXEDMAP));
 772         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 773         BUG_ON(!pfn_t_devmap(pfn));
 774
 775         if (addr < vma->vm_start || addr >= vma->vm_end)
 776                 return VM_FAULT_SIGBUS;
 777
 778         if (arch_needs_pgtable_deposit()) {
 779                 pgtable = pte_alloc_one(vma->vm_mm, addr);
 780                 if (!pgtable)
 781                         return VM_FAULT_OOM;
 782         }
 783
 784         track_pfn_insert(vma, &pgprot, pfn);
 785
 786         insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write, pgtable);
 787         return VM_FAULT_NOPAGE;
 788 }
 789 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
 790
 791 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 792 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
 793 {
 794         if (likely(vma->vm_flags & VM_WRITE))
 795                 pud = pud_mkwrite(pud);
 796         return pud;
 797 }
 798
 799 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 800                 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
 801 {
 802         struct mm_struct *mm = vma->vm_mm;
 803         pud_t entry;
 804         spinlock_t *ptl;
 805
 806         ptl = pud_lock(mm, pud);
 807         entry = pud_mkhuge(pfn_t_pud(pfn, prot));
 808         if (pfn_t_devmap(pfn))
 809                 entry = pud_mkdevmap(entry);
 810         if (write) {
 811                 entry = pud_mkyoung(pud_mkdirty(entry));
 812                 entry = maybe_pud_mkwrite(entry, vma);
 813         }
 814         set_pud_at(mm, addr, pud, entry);
 815         update_mmu_cache_pud(vma, addr, pud);
 816         spin_unlock(ptl);
 817 }
 818
 819 int vmf_insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 820                         pud_t *pud, pfn_t pfn, bool write)
 821 {
 822         pgprot_t pgprot = vma->vm_page_prot;
 823         /*
 824          * If we had pud_special, we could avoid all these restrictions,
 825          * but we need to be consistent with PTEs and architectures that
 826          * can't support a 'special' bit.
 827          */
 828         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
 829         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 830                                                 (VM_PFNMAP|VM_MIXEDMAP));
 831         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 832         BUG_ON(!pfn_t_devmap(pfn));
 833
 834         if (addr < vma->vm_start || addr >= vma->vm_end)
 835                 return VM_FAULT_SIGBUS;
 836
 837         track_pfn_insert(vma, &pgprot, pfn);
 838
 839         insert_pfn_pud(vma, addr, pud, pfn, pgprot, write);
 840         return VM_FAULT_NOPAGE;
 841 }
 842 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
 843 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
 844
 845 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
 846                 pmd_t *pmd, int flags)
 847 {
 848         pmd_t _pmd;
 849
 850         _pmd = pmd_mkyoung(*pmd);
 851         if (flags & FOLL_WRITE)
 852                 _pmd = pmd_mkdirty(_pmd);
 853         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
 854                                 pmd, _pmd, flags & FOLL_WRITE))
 855                 update_mmu_cache_pmd(vma, addr, pmd);
 856 }
 857
 858 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 859                 pmd_t *pmd, int flags)
 860 {
 861         unsigned long pfn = pmd_pfn(*pmd);
 862         struct mm_struct *mm = vma->vm_mm;
 863         struct dev_pagemap *pgmap;
 864         struct page *page;
 865
 866         assert_spin_locked(pmd_lockptr(mm, pmd));
 867
 868         /*
 869          * When we COW a devmap PMD entry, we split it into PTEs, so we should
 870          * not be in this function with `flags & FOLL_COW` set.
 871          */
 872         WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
 873
 874         if (flags & FOLL_WRITE && !pmd_write(*pmd))
 875                 return NULL;
 876
 877         if (pmd_present(*pmd) && pmd_devmap(*pmd))
 878                 /* pass */;
 879         else
 880                 return NULL;
 881
 882         if (flags & FOLL_TOUCH)
 883                 touch_pmd(vma, addr, pmd, flags);
 884
 885         /*
 886          * device mapped pages can only be returned if the
 887          * caller will manage the page reference count.
 888          */
 889         if (!(flags & FOLL_GET))
 890                 return ERR_PTR(-EEXIST);
 891
 892         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
 893         pgmap = get_dev_pagemap(pfn, NULL);
 894         if (!pgmap)
 895                 return ERR_PTR(-EFAULT);
 896         page = pfn_to_page(pfn);
 897         get_page(page);
 898         put_dev_pagemap(pgmap);
 899
 900         return page;
 901 }
 902
 903 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 904                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 905                   struct vm_area_struct *vma)
 906 {
 907         spinlock_t *dst_ptl, *src_ptl;
 908         struct page *src_page;
 909         pmd_t pmd;
 910         pgtable_t pgtable = NULL;
 911         int ret = -ENOMEM;
 912
 913         /* Skip if can be re-fill on fault */
 914         if (!vma_is_anonymous(vma))
 915                 return 0;
 916
 917         pgtable = pte_alloc_one(dst_mm, addr);
 918         if (unlikely(!pgtable))
 919                 goto out;
 920
 921         dst_ptl = pmd_lock(dst_mm, dst_pmd);
 922         src_ptl = pmd_lockptr(src_mm, src_pmd);
 923         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 924
 925         ret = -EAGAIN;
 926         pmd = *src_pmd;
 927
 928 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
 929         if (unlikely(is_swap_pmd(pmd))) {
 930                 swp_entry_t entry = pmd_to_swp_entry(pmd);
 931
 932                 VM_BUG_ON(!is_pmd_migration_entry(pmd));
 933                 if (is_write_migration_entry(entry)) {
 934                         make_migration_entry_read(&entry);
 935                         pmd = swp_entry_to_pmd(entry);
 936                         if (pmd_swp_soft_dirty(*src_pmd))
 937                                 pmd = pmd_swp_mksoft_dirty(pmd);
 938                         set_pmd_at(src_mm, addr, src_pmd, pmd);
 939                 }
 940                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 941                 atomic_long_inc(&dst_mm->nr_ptes);
 942                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
 943                 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 944                 ret = 0;
 945                 goto out_unlock;
 946         }
 947 #endif
 948
 949         if (unlikely(!pmd_trans_huge(pmd))) {
 950                 pte_free(dst_mm, pgtable);
 951                 goto out_unlock;
 952         }
 953         /*
 954          * When page table lock is held, the huge zero pmd should not be
 955          * under splitting since we don't split the page itself, only pmd to
 956          * a page table.
 957          */
 958         if (is_huge_zero_pmd(pmd)) {
 959                 struct page *zero_page;
 960                 /*
 961                  * get_huge_zero_page() will never allocate a new page here,
 962                  * since we already have a zero page to copy. It just takes a
 963                  * reference.
 964                  */
 965                 zero_page = mm_get_huge_zero_page(dst_mm);
 966                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
 967                                 zero_page);
 968                 ret = 0;
 969                 goto out_unlock;
 970         }
 971
 972         src_page = pmd_page(pmd);
 973         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
 974         get_page(src_page);
 975         page_dup_rmap(src_page, true);
 976         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 977         atomic_long_inc(&dst_mm->nr_ptes);
 978         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
 979
 980         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 981         pmd = pmd_mkold(pmd_wrprotect(pmd));
 982         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 983
 984         ret = 0;
 985 out_unlock:
 986         spin_unlock(src_ptl);
 987         spin_unlock(dst_ptl);
 988 out:
 989         return ret;
 990 }
 991
 992 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 993 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
 994                 pud_t *pud, int flags)
 995 {
 996         pud_t _pud;
 997
 998         _pud = pud_mkyoung(*pud);
 999         if (flags & FOLL_WRITE)
1000                 _pud = pud_mkdirty(_pud);
1001         if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1002                                 pud, _pud, flags & FOLL_WRITE))
1003                 update_mmu_cache_pud(vma, addr, pud);
1004 }
1005
1006 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1007                 pud_t *pud, int flags)
1008 {
1009         unsigned long pfn = pud_pfn(*pud);
1010         struct mm_struct *mm = vma->vm_mm;
1011         struct dev_pagemap *pgmap;
1012         struct page *page;
1013
1014         assert_spin_locked(pud_lockptr(mm, pud));
1015
1016         if (flags & FOLL_WRITE && !pud_write(*pud))
1017                 return NULL;
1018
1019         if (pud_present(*pud) && pud_devmap(*pud))
1020                 /* pass */;
1021         else
1022                 return NULL;
1023
1024         if (flags & FOLL_TOUCH)
1025                 touch_pud(vma, addr, pud, flags);
1026
1027         /*
1028          * device mapped pages can only be returned if the
1029          * caller will manage the page reference count.
1030          */
1031         if (!(flags & FOLL_GET))
1032                 return ERR_PTR(-EEXIST);
1033
1034         pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1035         pgmap = get_dev_pagemap(pfn, NULL);
1036         if (!pgmap)
1037                 return ERR_PTR(-EFAULT);
1038         page = pfn_to_page(pfn);
1039         get_page(page);
1040         put_dev_pagemap(pgmap);
1041
1042         return page;
1043 }
1044
1045 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1046                   pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1047                   struct vm_area_struct *vma)
1048 {
1049         spinlock_t *dst_ptl, *src_ptl;
1050         pud_t pud;
1051         int ret;
1052
1053         dst_ptl = pud_lock(dst_mm, dst_pud);
1054         src_ptl = pud_lockptr(src_mm, src_pud);
1055         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1056
1057         ret = -EAGAIN;
1058         pud = *src_pud;
1059         if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1060                 goto out_unlock;
1061
1062         /*
1063          * When page table lock is held, the huge zero pud should not be
1064          * under splitting since we don't split the page itself, only pud to
1065          * a page table.
1066          */
1067         if (is_huge_zero_pud(pud)) {
1068                 /* No huge zero pud yet */
1069         }
1070
1071         pudp_set_wrprotect(src_mm, addr, src_pud);
1072         pud = pud_mkold(pud_wrprotect(pud));
1073         set_pud_at(dst_mm, addr, dst_pud, pud);
1074
1075         ret = 0;
1076 out_unlock:
1077         spin_unlock(src_ptl);
1078         spin_unlock(dst_ptl);
1079         return ret;
1080 }
1081
1082 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1083 {
1084         pud_t entry;
1085         unsigned long haddr;
1086         bool write = vmf->flags & FAULT_FLAG_WRITE;
1087
1088         vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1089         if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1090                 goto unlock;
1091
1092         entry = pud_mkyoung(orig_pud);
1093         if (write)
1094                 entry = pud_mkdirty(entry);
1095         haddr = vmf->address & HPAGE_PUD_MASK;
1096         if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1097                 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1098
1099 unlock:
1100         spin_unlock(vmf->ptl);
1101 }
1102 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1103
1104 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1105 {
1106         pmd_t entry;
1107         unsigned long haddr;
1108         bool write = vmf->flags & FAULT_FLAG_WRITE;
1109
1110         vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1111         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1112                 goto unlock;
1113
1114         entry = pmd_mkyoung(orig_pmd);
1115         if (write)
1116                 entry = pmd_mkdirty(entry);
1117         haddr = vmf->address & HPAGE_PMD_MASK;
1118         if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1119                 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1120
1121 unlock:
1122         spin_unlock(vmf->ptl);
1123 }
1124
1125 static int do_huge_pmd_wp_page_fallback(struct vm_fault *vmf, pmd_t orig_pmd,
1126                 struct page *page)
1127 {
1128         struct vm_area_struct *vma = vmf->vma;
1129         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1130         struct mem_cgroup *memcg;
1131         pgtable_t pgtable;
1132         pmd_t _pmd;
1133         int ret = 0, i;
1134         struct page **pages;
1135         unsigned long mmun_start;       /* For mmu_notifiers */
1136         unsigned long mmun_end;         /* For mmu_notifiers */
1137
1138         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1139                         GFP_KERNEL);
1140         if (unlikely(!pages)) {
1141                 ret |= VM_FAULT_OOM;
1142                 goto out;
1143         }
1144
1145         for (i = 0; i < HPAGE_PMD_NR; i++) {
1146                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
1147                                                vmf->address, page_to_nid(page));
1148                 if (unlikely(!pages[i] ||
1149                              mem_cgroup_try_charge(pages[i], vma->vm_mm,
1150                                      GFP_KERNEL, &memcg, false))) {
1151                         if (pages[i])
1152                                 put_page(pages[i]);
1153                         while (--i >= 0) {
1154                                 memcg = (void *)page_private(pages[i]);
1155                                 set_page_private(pages[i], 0);
1156                                 mem_cgroup_cancel_charge(pages[i], memcg,
1157                                                 false);
1158                                 put_page(pages[i]);
1159                         }
1160                         kfree(pages);
1161                         ret |= VM_FAULT_OOM;
1162                         goto out;
1163                 }
1164                 set_page_private(pages[i], (unsigned long)memcg);
1165         }
1166
1167         for (i = 0; i < HPAGE_PMD_NR; i++) {
1168                 copy_user_highpage(pages[i], page + i,
1169                                    haddr + PAGE_SIZE * i, vma);
1170                 __SetPageUptodate(pages[i]);
1171                 cond_resched();
1172         }
1173
1174         mmun_start = haddr;
1175         mmun_end   = haddr + HPAGE_PMD_SIZE;
1176         mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
1177
1178         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1179         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1180                 goto out_free_pages;
1181         VM_BUG_ON_PAGE(!PageHead(page), page);
1182
1183         pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1184         /* leave pmd empty until pte is filled */
1185
1186         pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
1187         pmd_populate(vma->vm_mm, &_pmd, pgtable);
1188
1189         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1190                 pte_t entry;
1191                 entry = mk_pte(pages[i], vma->vm_page_prot);
1192                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1193                 memcg = (void *)page_private(pages[i]);
1194                 set_page_private(pages[i], 0);
1195                 page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
1196                 mem_cgroup_commit_charge(pages[i], memcg, false, false);
1197                 lru_cache_add_active_or_unevictable(pages[i], vma);
1198                 vmf->pte = pte_offset_map(&_pmd, haddr);
1199                 VM_BUG_ON(!pte_none(*vmf->pte));
1200                 set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
1201                 pte_unmap(vmf->pte);
1202         }
1203         kfree(pages);
1204
1205         smp_wmb(); /* make pte visible before pmd */
1206         pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
1207         page_remove_rmap(page, true);
1208         spin_unlock(vmf->ptl);
1209
1210         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1211
1212         ret |= VM_FAULT_WRITE;
1213         put_page(page);
1214
1215 out:
1216         return ret;
1217
1218 out_free_pages:
1219         spin_unlock(vmf->ptl);
1220         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1221         for (i = 0; i < HPAGE_PMD_NR; i++) {
1222                 memcg = (void *)page_private(pages[i]);
1223                 set_page_private(pages[i], 0);
1224                 mem_cgroup_cancel_charge(pages[i], memcg, false);
1225                 put_page(pages[i]);
1226         }
1227         kfree(pages);
1228         goto out;
1229 }
1230
1231 int do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1232 {
1233         struct vm_area_struct *vma = vmf->vma;
1234         struct page *page = NULL, *new_page;
1235         struct mem_cgroup *memcg;
1236         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1237         unsigned long mmun_start;       /* For mmu_notifiers */
1238         unsigned long mmun_end;         /* For mmu_notifiers */
1239         gfp_t huge_gfp;                 /* for allocation and charge */
1240         int ret = 0;
1241
1242         vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1243         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1244         if (is_huge_zero_pmd(orig_pmd))
1245                 goto alloc;
1246         spin_lock(vmf->ptl);
1247         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1248                 goto out_unlock;
1249
1250         page = pmd_page(orig_pmd);
1251         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1252         /*
1253          * We can only reuse the page if nobody else maps the huge page or it's
1254          * part.
1255          */
1256         if (!trylock_page(page)) {
1257                 get_page(page);
1258                 spin_unlock(vmf->ptl);
1259                 lock_page(page);
1260                 spin_lock(vmf->ptl);
1261                 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1262                         unlock_page(page);
1263                         put_page(page);
1264                         goto out_unlock;
1265                 }
1266                 put_page(page);
1267         }
1268         if (reuse_swap_page(page, NULL)) {
1269                 pmd_t entry;
1270                 entry = pmd_mkyoung(orig_pmd);
1271                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1272                 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry,  1))
1273                         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1274                 ret |= VM_FAULT_WRITE;
1275                 unlock_page(page);
1276                 goto out_unlock;
1277         }
1278         unlock_page(page);
1279         get_page(page);
1280         spin_unlock(vmf->ptl);
1281 alloc:
1282         if (transparent_hugepage_enabled(vma) &&
1283             !transparent_hugepage_debug_cow()) {
1284                 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1285                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1286         } else
1287                 new_page = NULL;
1288
1289         if (likely(new_page)) {
1290                 prep_transhuge_page(new_page);
1291         } else {
1292                 if (!page) {
1293                         split_huge_pmd(vma, vmf->pmd, vmf->address);
1294                         ret |= VM_FAULT_FALLBACK;
1295                 } else {
1296                         ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
1297                         if (ret & VM_FAULT_OOM) {
1298                                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1299                                 ret |= VM_FAULT_FALLBACK;
1300                         }
1301                         put_page(page);
1302                 }
1303                 count_vm_event(THP_FAULT_FALLBACK);
1304                 goto out;
1305         }
1306
1307         if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm,
1308                                 huge_gfp | __GFP_NORETRY, &memcg, true))) {
1309                 put_page(new_page);
1310                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1311                 if (page)
1312                         put_page(page);
1313                 ret |= VM_FAULT_FALLBACK;
1314                 count_vm_event(THP_FAULT_FALLBACK);
1315                 goto out;
1316         }
1317
1318         count_vm_event(THP_FAULT_ALLOC);
1319
1320         if (!page)
1321                 clear_huge_page(new_page, vmf->address, HPAGE_PMD_NR);
1322         else
1323                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1324         __SetPageUptodate(new_page);
1325
1326         mmun_start = haddr;
1327         mmun_end   = haddr + HPAGE_PMD_SIZE;
1328         mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
1329
1330         spin_lock(vmf->ptl);
1331         if (page)
1332                 put_page(page);
1333         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1334                 spin_unlock(vmf->ptl);
1335                 mem_cgroup_cancel_charge(new_page, memcg, true);
1336                 put_page(new_page);
1337                 goto out_mn;
1338         } else {
1339                 pmd_t entry;
1340                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1341                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1342                 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1343                 page_add_new_anon_rmap(new_page, vma, haddr, true);
1344                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1345                 lru_cache_add_active_or_unevictable(new_page, vma);
1346                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1347                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1348                 if (!page) {
1349                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1350                 } else {
1351                         VM_BUG_ON_PAGE(!PageHead(page), page);
1352                         page_remove_rmap(page, true);
1353                         put_page(page);
1354                 }
1355                 ret |= VM_FAULT_WRITE;
1356         }
1357         spin_unlock(vmf->ptl);
1358 out_mn:
1359         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1360 out:
1361         return ret;
1362 out_unlock:
1363         spin_unlock(vmf->ptl);
1364         return ret;
1365 }
1366
1367 /*
1368  * FOLL_FORCE can write to even unwritable pmd's, but only
1369  * after we've gone through a COW cycle and they are dirty.
1370  */
1371 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1372 {
1373         return pmd_write(pmd) ||
1374                ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1375 }
1376
1377 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1378                                    unsigned long addr,
1379                                    pmd_t *pmd,
1380                                    unsigned int flags)
1381 {
1382         struct mm_struct *mm = vma->vm_mm;
1383         struct page *page = NULL;
1384
1385         assert_spin_locked(pmd_lockptr(mm, pmd));
1386
1387         if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1388                 goto out;
1389
1390         /* Avoid dumping huge zero page */
1391         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1392                 return ERR_PTR(-EFAULT);
1393
1394         /* Full NUMA hinting faults to serialise migration in fault paths */
1395         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1396                 goto out;
1397
1398         page = pmd_page(*pmd);
1399         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1400         if (flags & FOLL_TOUCH)
1401                 touch_pmd(vma, addr, pmd, flags);
1402         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1403                 /*
1404                  * We don't mlock() pte-mapped THPs. This way we can avoid
1405                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1406                  *
1407                  * For anon THP:
1408                  *
1409                  * In most cases the pmd is the only mapping of the page as we
1410                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1411                  * writable private mappings in populate_vma_page_range().
1412                  *
1413                  * The only scenario when we have the page shared here is if we
1414                  * mlocking read-only mapping shared over fork(). We skip
1415                  * mlocking such pages.
1416                  *
1417                  * For file THP:
1418                  *
1419                  * We can expect PageDoubleMap() to be stable under page lock:
1420                  * for file pages we set it in page_add_file_rmap(), which
1421                  * requires page to be locked.
1422                  */
1423
1424                 if (PageAnon(page) && compound_mapcount(page) != 1)
1425                         goto skip_mlock;
1426                 if (PageDoubleMap(page) || !page->mapping)
1427                         goto skip_mlock;
1428                 if (!trylock_page(page))
1429                         goto skip_mlock;
1430                 lru_add_drain();
1431                 if (page->mapping && !PageDoubleMap(page))
1432                         mlock_vma_page(page);
1433                 unlock_page(page);
1434         }
1435 skip_mlock:
1436         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1437         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1438         if (flags & FOLL_GET)
1439                 get_page(page);
1440
1441 out:
1442         return page;
1443 }
1444
1445 /* NUMA hinting page fault entry point for trans huge pmds */
1446 int do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1447 {
1448         struct vm_area_struct *vma = vmf->vma;
1449         struct anon_vma *anon_vma = NULL;
1450         struct page *page;
1451         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1452         int page_nid = -1, this_nid = numa_node_id();
1453         int target_nid, last_cpupid = -1;
1454         bool page_locked;
1455         bool migrated = false;
1456         bool was_writable;
1457         int flags = 0;
1458
1459         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1460         if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1461                 goto out_unlock;
1462
1463         /*
1464          * If there are potential migrations, wait for completion and retry
1465          * without disrupting NUMA hinting information. Do not relock and
1466          * check_same as the page may no longer be mapped.
1467          */
1468         if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1469                 page = pmd_page(*vmf->pmd);
1470                 if (!get_page_unless_zero(page))
1471                         goto out_unlock;
1472                 spin_unlock(vmf->ptl);
1473                 wait_on_page_locked(page);
1474                 put_page(page);
1475                 goto out;
1476         }
1477
1478         page = pmd_page(pmd);
1479         BUG_ON(is_huge_zero_page(page));
1480         page_nid = page_to_nid(page);
1481         last_cpupid = page_cpupid_last(page);
1482         count_vm_numa_event(NUMA_HINT_FAULTS);
1483         if (page_nid == this_nid) {
1484                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1485                 flags |= TNF_FAULT_LOCAL;
1486         }
1487
1488         /* See similar comment in do_numa_page for explanation */
1489         if (!pmd_savedwrite(pmd))
1490                 flags |= TNF_NO_GROUP;
1491
1492         /*
1493          * Acquire the page lock to serialise THP migrations but avoid dropping
1494          * page_table_lock if at all possible
1495          */
1496         page_locked = trylock_page(page);
1497         target_nid = mpol_misplaced(page, vma, haddr);
1498         if (target_nid == -1) {
1499                 /* If the page was locked, there are no parallel migrations */
1500                 if (page_locked)
1501                         goto clear_pmdnuma;
1502         }
1503
1504         /* Migration could have started since the pmd_trans_migrating check */
1505         if (!page_locked) {
1506                 page_nid = -1;
1507                 if (!get_page_unless_zero(page))
1508                         goto out_unlock;
1509                 spin_unlock(vmf->ptl);
1510                 wait_on_page_locked(page);
1511                 put_page(page);
1512                 goto out;
1513         }
1514
1515         /*
1516          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1517          * to serialises splits
1518          */
1519         get_page(page);
1520         spin_unlock(vmf->ptl);
1521         anon_vma = page_lock_anon_vma_read(page);
1522
1523         /* Confirm the PMD did not change while page_table_lock was released */
1524         spin_lock(vmf->ptl);
1525         if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1526                 unlock_page(page);
1527                 put_page(page);
1528                 page_nid = -1;
1529                 goto out_unlock;
1530         }
1531
1532         /* Bail if we fail to protect against THP splits for any reason */
1533         if (unlikely(!anon_vma)) {
1534                 put_page(page);
1535                 page_nid = -1;
1536                 goto clear_pmdnuma;
1537         }
1538
1539         /*
1540          * Since we took the NUMA fault, we must have observed the !accessible
1541          * bit. Make sure all other CPUs agree with that, to avoid them
1542          * modifying the page we're about to migrate.
1543          *
1544          * Must be done under PTL such that we'll observe the relevant
1545          * inc_tlb_flush_pending().
1546          *
1547          * We are not sure a pending tlb flush here is for a huge page
1548          * mapping or not. Hence use the tlb range variant
1549          */
1550         if (mm_tlb_flush_pending(vma->vm_mm))
1551                 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1552
1553         /*
1554          * Migrate the THP to the requested node, returns with page unlocked
1555          * and access rights restored.
1556          */
1557         spin_unlock(vmf->ptl);
1558
1559         migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1560                                 vmf->pmd, pmd, vmf->address, page, target_nid);
1561         if (migrated) {
1562                 flags |= TNF_MIGRATED;
1563                 page_nid = target_nid;
1564         } else
1565                 flags |= TNF_MIGRATE_FAIL;
1566
1567         goto out;
1568 clear_pmdnuma:
1569         BUG_ON(!PageLocked(page));
1570         was_writable = pmd_savedwrite(pmd);
1571         pmd = pmd_modify(pmd, vma->vm_page_prot);
1572         pmd = pmd_mkyoung(pmd);
1573         if (was_writable)
1574                 pmd = pmd_mkwrite(pmd);
1575         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1576         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1577         unlock_page(page);
1578 out_unlock:
1579         spin_unlock(vmf->ptl);
1580
1581 out:
1582         if (anon_vma)
1583                 page_unlock_anon_vma_read(anon_vma);
1584
1585         if (page_nid != -1)
1586                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1587                                 flags);
1588
1589         return 0;
1590 }
1591
1592 /*
1593  * Return true if we do MADV_FREE successfully on entire pmd page.
1594  * Otherwise, return false.
1595  */
1596 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1597                 pmd_t *pmd, unsigned long addr, unsigned long next)
1598 {
1599         spinlock_t *ptl;
1600         pmd_t orig_pmd;
1601         struct page *page;
1602         struct mm_struct *mm = tlb->mm;
1603         bool ret = false;
1604
1605         tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1606
1607         ptl = pmd_trans_huge_lock(pmd, vma);
1608         if (!ptl)
1609                 goto out_unlocked;
1610
1611         orig_pmd = *pmd;
1612         if (is_huge_zero_pmd(orig_pmd))
1613                 goto out;
1614
1615         if (unlikely(!pmd_present(orig_pmd))) {
1616                 VM_BUG_ON(thp_migration_supported() &&
1617                                   !is_pmd_migration_entry(orig_pmd));
1618                 goto out;
1619         }
1620
1621         page = pmd_page(orig_pmd);
1622         /*
1623          * If other processes are mapping this page, we couldn't discard
1624          * the page unless they all do MADV_FREE so let's skip the page.
1625          */
1626         if (page_mapcount(page) != 1)
1627                 goto out;
1628
1629         if (!trylock_page(page))
1630                 goto out;
1631
1632         /*
1633          * If user want to discard part-pages of THP, split it so MADV_FREE
1634          * will deactivate only them.
1635          */
1636         if (next - addr != HPAGE_PMD_SIZE) {
1637                 get_page(page);
1638                 spin_unlock(ptl);
1639                 split_huge_page(page);
1640                 unlock_page(page);
1641                 put_page(page);
1642                 goto out_unlocked;
1643         }
1644
1645         if (PageDirty(page))
1646                 ClearPageDirty(page);
1647         unlock_page(page);
1648
1649         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1650                 pmdp_invalidate(vma, addr, pmd);
1651                 orig_pmd = pmd_mkold(orig_pmd);
1652                 orig_pmd = pmd_mkclean(orig_pmd);
1653
1654                 set_pmd_at(mm, addr, pmd, orig_pmd);
1655                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1656         }
1657
1658         mark_page_lazyfree(page);
1659         ret = true;
1660 out:
1661         spin_unlock(ptl);
1662 out_unlocked:
1663         return ret;
1664 }
1665
1666 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1667 {
1668         pgtable_t pgtable;
1669
1670         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1671         pte_free(mm, pgtable);
1672         atomic_long_dec(&mm->nr_ptes);
1673 }
1674
1675 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1676                  pmd_t *pmd, unsigned long addr)
1677 {
1678         pmd_t orig_pmd;
1679         spinlock_t *ptl;
1680
1681         tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1682
1683         ptl = __pmd_trans_huge_lock(pmd, vma);
1684         if (!ptl)
1685                 return 0;
1686         /*
1687          * For architectures like ppc64 we look at deposited pgtable
1688          * when calling pmdp_huge_get_and_clear. So do the
1689          * pgtable_trans_huge_withdraw after finishing pmdp related
1690          * operations.
1691          */
1692         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1693                         tlb->fullmm);
1694         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1695         if (vma_is_dax(vma)) {
1696                 if (arch_needs_pgtable_deposit())
1697                         zap_deposited_table(tlb->mm, pmd);
1698                 spin_unlock(ptl);
1699                 if (is_huge_zero_pmd(orig_pmd))
1700                         tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1701         } else if (is_huge_zero_pmd(orig_pmd)) {
1702                 zap_deposited_table(tlb->mm, pmd);
1703                 spin_unlock(ptl);
1704                 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1705         } else {
1706                 struct page *page = NULL;
1707                 int flush_needed = 1;
1708
1709                 if (pmd_present(orig_pmd)) {
1710                         page = pmd_page(orig_pmd);
1711                         page_remove_rmap(page, true);
1712                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1713                         VM_BUG_ON_PAGE(!PageHead(page), page);
1714                 } else if (thp_migration_supported()) {
1715                         swp_entry_t entry;
1716
1717                         VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1718                         entry = pmd_to_swp_entry(orig_pmd);
1719                         page = pfn_to_page(swp_offset(entry));
1720                         flush_needed = 0;
1721                 } else
1722                         WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1723
1724                 if (PageAnon(page)) {
1725                         zap_deposited_table(tlb->mm, pmd);
1726                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1727                 } else {
1728                         if (arch_needs_pgtable_deposit())
1729                                 zap_deposited_table(tlb->mm, pmd);
1730                         add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1731                 }
1732
1733                 spin_unlock(ptl);
1734                 if (flush_needed)
1735                         tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1736         }
1737         return 1;
1738 }
1739
1740 #ifndef pmd_move_must_withdraw
1741 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1742                                          spinlock_t *old_pmd_ptl,
1743                                          struct vm_area_struct *vma)
1744 {
1745         /*
1746          * With split pmd lock we also need to move preallocated
1747          * PTE page table if new_pmd is on different PMD page table.
1748          *
1749          * We also don't deposit and withdraw tables for file pages.
1750          */
1751         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1752 }
1753 #endif
1754
1755 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1756 {
1757 #ifdef CONFIG_MEM_SOFT_DIRTY
1758         if (unlikely(is_pmd_migration_entry(pmd)))
1759                 pmd = pmd_swp_mksoft_dirty(pmd);
1760         else if (pmd_present(pmd))
1761                 pmd = pmd_mksoft_dirty(pmd);
1762 #endif
1763         return pmd;
1764 }
1765
1766 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1767                   unsigned long new_addr, unsigned long old_end,
1768                   pmd_t *old_pmd, pmd_t *new_pmd, bool *need_flush)
1769 {
1770         spinlock_t *old_ptl, *new_ptl;
1771         pmd_t pmd;
1772         struct mm_struct *mm = vma->vm_mm;
1773         bool force_flush = false;
1774
1775         if ((old_addr & ~HPAGE_PMD_MASK) ||
1776             (new_addr & ~HPAGE_PMD_MASK) ||
1777             old_end - old_addr < HPAGE_PMD_SIZE)
1778                 return false;
1779
1780         /*
1781          * The destination pmd shouldn't be established, free_pgtables()
1782          * should have release it.
1783          */
1784         if (WARN_ON(!pmd_none(*new_pmd))) {
1785                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1786                 return false;
1787         }
1788
1789         /*
1790          * We don't have to worry about the ordering of src and dst
1791          * ptlocks because exclusive mmap_sem prevents deadlock.
1792          */
1793         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1794         if (old_ptl) {
1795                 new_ptl = pmd_lockptr(mm, new_pmd);
1796                 if (new_ptl != old_ptl)
1797                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1798                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1799                 if (pmd_present(pmd) && pmd_dirty(pmd))
1800                         force_flush = true;
1801                 VM_BUG_ON(!pmd_none(*new_pmd));
1802
1803                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1804                         pgtable_t pgtable;
1805                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1806                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1807                 }
1808                 pmd = move_soft_dirty_pmd(pmd);
1809                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1810                 if (new_ptl != old_ptl)
1811                         spin_unlock(new_ptl);
1812                 if (force_flush)
1813                         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1814                 else
1815                         *need_flush = true;
1816                 spin_unlock(old_ptl);
1817                 return true;
1818         }
1819         return false;
1820 }
1821
1822 /*
1823  * Returns
1824  *  - 0 if PMD could not be locked
1825  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1826  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1827  */
1828 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1829                 unsigned long addr, pgprot_t newprot, int prot_numa)
1830 {
1831         struct mm_struct *mm = vma->vm_mm;
1832         spinlock_t *ptl;
1833         pmd_t entry;
1834         bool preserve_write;
1835         int ret;
1836
1837         ptl = __pmd_trans_huge_lock(pmd, vma);
1838         if (!ptl)
1839                 return 0;
1840
1841         preserve_write = prot_numa && pmd_write(*pmd);
1842         ret = 1;
1843
1844 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1845         if (is_swap_pmd(*pmd)) {
1846                 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1847
1848                 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1849                 if (is_write_migration_entry(entry)) {
1850                         pmd_t newpmd;
1851                         /*
1852                          * A protection check is difficult so
1853                          * just be safe and disable write
1854                          */
1855                         make_migration_entry_read(&entry);
1856                         newpmd = swp_entry_to_pmd(entry);
1857                         if (pmd_swp_soft_dirty(*pmd))
1858                                 newpmd = pmd_swp_mksoft_dirty(newpmd);
1859                         set_pmd_at(mm, addr, pmd, newpmd);
1860                 }
1861                 goto unlock;
1862         }
1863 #endif
1864
1865         /*
1866          * Avoid trapping faults against the zero page. The read-only
1867          * data is likely to be read-cached on the local CPU and
1868          * local/remote hits to the zero page are not interesting.
1869          */
1870         if (prot_numa && is_huge_zero_pmd(*pmd))
1871                 goto unlock;
1872
1873         if (prot_numa && pmd_protnone(*pmd))
1874                 goto unlock;
1875
1876         /*
1877          * In case prot_numa, we are under down_read(mmap_sem). It's critical
1878          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1879          * which is also under down_read(mmap_sem):
1880          *
1881          *      CPU0:                           CPU1:
1882          *                              change_huge_pmd(prot_numa=1)
1883          *                               pmdp_huge_get_and_clear_notify()
1884          * madvise_dontneed()
1885          *  zap_pmd_range()
1886          *   pmd_trans_huge(*pmd) == 0 (without ptl)
1887          *   // skip the pmd
1888          *                               set_pmd_at();
1889          *                               // pmd is re-established
1890          *
1891          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1892          * which may break userspace.
1893          *
1894          * pmdp_invalidate() is required to make sure we don't miss
1895          * dirty/young flags set by hardware.
1896          */
1897         entry = *pmd;
1898         pmdp_invalidate(vma, addr, pmd);
1899
1900         /*
1901          * Recover dirty/young flags.  It relies on pmdp_invalidate to not
1902          * corrupt them.
1903          */
1904         if (pmd_dirty(*pmd))
1905                 entry = pmd_mkdirty(entry);
1906         if (pmd_young(*pmd))
1907                 entry = pmd_mkyoung(entry);
1908
1909         entry = pmd_modify(entry, newprot);
1910         if (preserve_write)
1911                 entry = pmd_mk_savedwrite(entry);
1912         ret = HPAGE_PMD_NR;
1913         set_pmd_at(mm, addr, pmd, entry);
1914         BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1915 unlock:
1916         spin_unlock(ptl);
1917         return ret;
1918 }
1919
1920 /*
1921  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1922  *
1923  * Note that if it returns page table lock pointer, this routine returns without
1924  * unlocking page table lock. So callers must unlock it.
1925  */
1926 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1927 {
1928         spinlock_t *ptl;
1929         ptl = pmd_lock(vma->vm_mm, pmd);
1930         if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1931                         pmd_devmap(*pmd)))
1932                 return ptl;
1933         spin_unlock(ptl);
1934         return NULL;
1935 }
1936
1937 /*
1938  * Returns true if a given pud maps a thp, false otherwise.
1939  *
1940  * Note that if it returns true, this routine returns without unlocking page
1941  * table lock. So callers must unlock it.
1942  */
1943 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1944 {
1945         spinlock_t *ptl;
1946
1947         ptl = pud_lock(vma->vm_mm, pud);
1948         if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1949                 return ptl;
1950         spin_unlock(ptl);
1951         return NULL;
1952 }
1953
1954 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1955 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1956                  pud_t *pud, unsigned long addr)
1957 {
1958         pud_t orig_pud;
1959         spinlock_t *ptl;
1960
1961         ptl = __pud_trans_huge_lock(pud, vma);
1962         if (!ptl)
1963                 return 0;
1964         /*
1965          * For architectures like ppc64 we look at deposited pgtable
1966          * when calling pudp_huge_get_and_clear. So do the
1967          * pgtable_trans_huge_withdraw after finishing pudp related
1968          * operations.
1969          */
1970         orig_pud = pudp_huge_get_and_clear_full(tlb->mm, addr, pud,
1971                         tlb->fullmm);
1972         tlb_remove_pud_tlb_entry(tlb, pud, addr);
1973         if (vma_is_dax(vma)) {
1974                 spin_unlock(ptl);
1975                 /* No zero page support yet */
1976         } else {
1977                 /* No support for anonymous PUD pages yet */
1978                 BUG();
1979         }
1980         return 1;
1981 }
1982
1983 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1984                 unsigned long haddr)
1985 {
1986         VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1987         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1988         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
1989         VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
1990
1991         count_vm_event(THP_SPLIT_PUD);
1992
1993         pudp_huge_clear_flush_notify(vma, haddr, pud);
1994 }
1995
1996 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
1997                 unsigned long address)
1998 {
1999         spinlock_t *ptl;
2000         struct mm_struct *mm = vma->vm_mm;
2001         unsigned long haddr = address & HPAGE_PUD_MASK;
2002
2003         mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PUD_SIZE);
2004         ptl = pud_lock(mm, pud);
2005         if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2006                 goto out;
2007         __split_huge_pud_locked(vma, pud, haddr);
2008
2009 out:
2010         spin_unlock(ptl);
2011         mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PUD_SIZE);
2012 }
2013 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2014
2015 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2016                 unsigned long haddr, pmd_t *pmd)
2017 {
2018         struct mm_struct *mm = vma->vm_mm;
2019         pgtable_t pgtable;
2020         pmd_t _pmd;
2021         int i;
2022
2023         /* leave pmd empty until pte is filled */
2024         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2025
2026         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2027         pmd_populate(mm, &_pmd, pgtable);
2028
2029         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2030                 pte_t *pte, entry;
2031                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2032                 entry = pte_mkspecial(entry);
2033                 pte = pte_offset_map(&_pmd, haddr);
2034                 VM_BUG_ON(!pte_none(*pte));
2035                 set_pte_at(mm, haddr, pte, entry);
2036                 pte_unmap(pte);
2037         }
2038         smp_wmb(); /* make pte visible before pmd */
2039         pmd_populate(mm, pmd, pgtable);
2040 }
2041
2042 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2043                 unsigned long haddr, bool freeze)
2044 {
2045         struct mm_struct *mm = vma->vm_mm;
2046         struct page *page;
2047         pgtable_t pgtable;
2048         pmd_t _pmd;
2049         bool young, write, dirty, soft_dirty, pmd_migration = false;
2050         unsigned long addr;
2051         int i;
2052
2053         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2054         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2055         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2056         VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2057                                 && !pmd_devmap(*pmd));
2058
2059         count_vm_event(THP_SPLIT_PMD);
2060
2061         if (!vma_is_anonymous(vma)) {
2062                 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2063                 /*
2064                  * We are going to unmap this huge page. So
2065                  * just go ahead and zap it
2066                  */
2067                 if (arch_needs_pgtable_deposit())
2068                         zap_deposited_table(mm, pmd);
2069                 if (vma_is_dax(vma))
2070                         return;
2071                 page = pmd_page(_pmd);
2072                 if (!PageDirty(page) && pmd_dirty(_pmd))
2073                         set_page_dirty(page);
2074                 if (!PageReferenced(page) && pmd_young(_pmd))
2075                         SetPageReferenced(page);
2076                 page_remove_rmap(page, true);
2077                 put_page(page);
2078                 add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR);
2079                 return;
2080         } else if (is_huge_zero_pmd(*pmd)) {
2081                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2082         }
2083
2084 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2085         pmd_migration = is_pmd_migration_entry(*pmd);
2086         if (pmd_migration) {
2087                 swp_entry_t entry;
2088
2089                 entry = pmd_to_swp_entry(*pmd);
2090                 page = pfn_to_page(swp_offset(entry));
2091         } else
2092 #endif
2093                 page = pmd_page(*pmd);
2094         VM_BUG_ON_PAGE(!page_count(page), page);
2095         page_ref_add(page, HPAGE_PMD_NR - 1);
2096         write = pmd_write(*pmd);
2097         young = pmd_young(*pmd);
2098         dirty = pmd_dirty(*pmd);
2099         soft_dirty = pmd_soft_dirty(*pmd);
2100
2101         pmdp_huge_split_prepare(vma, haddr, pmd);
2102         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2103         pmd_populate(mm, &_pmd, pgtable);
2104
2105         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2106                 pte_t entry, *pte;
2107                 /*
2108                  * Note that NUMA hinting access restrictions are not
2109                  * transferred to avoid any possibility of altering
2110                  * permissions across VMAs.
2111                  */
2112                 if (freeze || pmd_migration) {
2113                         swp_entry_t swp_entry;
2114                         swp_entry = make_migration_entry(page + i, write);
2115                         entry = swp_entry_to_pte(swp_entry);
2116                         if (soft_dirty)
2117                                 entry = pte_swp_mksoft_dirty(entry);
2118                 } else {
2119                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2120                         entry = maybe_mkwrite(entry, vma);
2121                         if (!write)
2122                                 entry = pte_wrprotect(entry);
2123                         if (!young)
2124                                 entry = pte_mkold(entry);
2125                         if (soft_dirty)
2126                                 entry = pte_mksoft_dirty(entry);
2127                 }
2128                 if (dirty)
2129                         SetPageDirty(page + i);
2130                 pte = pte_offset_map(&_pmd, addr);
2131                 BUG_ON(!pte_none(*pte));
2132                 set_pte_at(mm, addr, pte, entry);
2133                 atomic_inc(&page[i]._mapcount);
2134                 pte_unmap(pte);
2135         }
2136
2137         /*
2138          * Set PG_double_map before dropping compound_mapcount to avoid
2139          * false-negative page_mapped().
2140          */
2141         if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2142                 for (i = 0; i < HPAGE_PMD_NR; i++)
2143                         atomic_inc(&page[i]._mapcount);
2144         }
2145
2146         if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2147                 /* Last compound_mapcount is gone. */
2148                 __dec_node_page_state(page, NR_ANON_THPS);
2149                 if (TestClearPageDoubleMap(page)) {
2150                         /* No need in mapcount reference anymore */
2151                         for (i = 0; i < HPAGE_PMD_NR; i++)
2152                                 atomic_dec(&page[i]._mapcount);
2153                 }
2154         }
2155
2156         smp_wmb(); /* make pte visible before pmd */
2157         /*
2158          * Up to this point the pmd is present and huge and userland has the
2159          * whole access to the hugepage during the split (which happens in
2160          * place). If we overwrite the pmd with the not-huge version pointing
2161          * to the pte here (which of course we could if all CPUs were bug
2162          * free), userland could trigger a small page size TLB miss on the
2163          * small sized TLB while the hugepage TLB entry is still established in
2164          * the huge TLB. Some CPU doesn't like that.
2165          * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2166          * 383 on page 93. Intel should be safe but is also warns that it's
2167          * only safe if the permission and cache attributes of the two entries
2168          * loaded in the two TLB is identical (which should be the case here).
2169          * But it is generally safer to never allow small and huge TLB entries
2170          * for the same virtual address to be loaded simultaneously. So instead
2171          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2172          * current pmd notpresent (atomically because here the pmd_trans_huge
2173          * and pmd_trans_splitting must remain set at all times on the pmd
2174          * until the split is complete for this pmd), then we flush the SMP TLB
2175          * and finally we write the non-huge version of the pmd entry with
2176          * pmd_populate.
2177          */
2178         pmdp_invalidate(vma, haddr, pmd);
2179         pmd_populate(mm, pmd, pgtable);
2180
2181         if (freeze) {
2182                 for (i = 0; i < HPAGE_PMD_NR; i++) {
2183                         page_remove_rmap(page + i, false);
2184                         put_page(page + i);
2185                 }
2186         }
2187 }
2188
2189 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2190                 unsigned long address, bool freeze, struct page *page)
2191 {
2192         spinlock_t *ptl;
2193         struct mm_struct *mm = vma->vm_mm;
2194         unsigned long haddr = address & HPAGE_PMD_MASK;
2195
2196         mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
2197         ptl = pmd_lock(mm, pmd);
2198
2199         /*
2200          * If caller asks to setup a migration entries, we need a page to check
2201          * pmd against. Otherwise we can end up replacing wrong page.
2202          */
2203         VM_BUG_ON(freeze && !page);
2204         if (page && page != pmd_page(*pmd))
2205                 goto out;
2206
2207         if (pmd_trans_huge(*pmd)) {
2208                 page = pmd_page(*pmd);
2209                 if (PageMlocked(page))
2210                         clear_page_mlock(page);
2211         } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2212                 goto out;
2213         __split_huge_pmd_locked(vma, pmd, haddr, freeze);
2214 out:
2215         spin_unlock(ptl);
2216         mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
2217 }
2218
2219 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2220                 bool freeze, struct page *page)
2221 {
2222         pgd_t *pgd;
2223         p4d_t *p4d;
2224         pud_t *pud;
2225         pmd_t *pmd;
2226
2227         pgd = pgd_offset(vma->vm_mm, address);
2228         if (!pgd_present(*pgd))
2229                 return;
2230
2231         p4d = p4d_offset(pgd, address);
2232         if (!p4d_present(*p4d))
2233                 return;
2234
2235         pud = pud_offset(p4d, address);
2236         if (!pud_present(*pud))
2237                 return;
2238
2239         pmd = pmd_offset(pud, address);
2240
2241         __split_huge_pmd(vma, pmd, address, freeze, page);
2242 }
2243
2244 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2245                              unsigned long start,
2246                              unsigned long end,
2247                              long adjust_next)
2248 {
2249         /*
2250          * If the new start address isn't hpage aligned and it could
2251          * previously contain an hugepage: check if we need to split
2252          * an huge pmd.
2253          */
2254         if (start & ~HPAGE_PMD_MASK &&
2255             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2256             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2257                 split_huge_pmd_address(vma, start, false, NULL);
2258
2259         /*
2260          * If the new end address isn't hpage aligned and it could
2261          * previously contain an hugepage: check if we need to split
2262          * an huge pmd.
2263          */
2264         if (end & ~HPAGE_PMD_MASK &&
2265             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2266             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2267                 split_huge_pmd_address(vma, end, false, NULL);
2268
2269         /*
2270          * If we're also updating the vma->vm_next->vm_start, if the new
2271          * vm_next->vm_start isn't page aligned and it could previously
2272          * contain an hugepage: check if we need to split an huge pmd.
2273          */
2274         if (adjust_next > 0) {
2275                 struct vm_area_struct *next = vma->vm_next;
2276                 unsigned long nstart = next->vm_start;
2277                 nstart += adjust_next << PAGE_SHIFT;
2278                 if (nstart & ~HPAGE_PMD_MASK &&
2279                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2280                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2281                         split_huge_pmd_address(next, nstart, false, NULL);
2282         }
2283 }
2284
2285 static void freeze_page(struct page *page)
2286 {
2287         enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
2288                 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2289         bool unmap_success;
2290
2291         VM_BUG_ON_PAGE(!PageHead(page), page);
2292
2293         if (PageAnon(page))
2294                 ttu_flags |= TTU_SPLIT_FREEZE;
2295
2296         unmap_success = try_to_unmap(page, ttu_flags);
2297         VM_BUG_ON_PAGE(!unmap_success, page);
2298 }
2299
2300 static void unfreeze_page(struct page *page)
2301 {
2302         int i;
2303         if (PageTransHuge(page)) {
2304                 remove_migration_ptes(page, page, true);
2305         } else {
2306                 for (i = 0; i < HPAGE_PMD_NR; i++)
2307                         remove_migration_ptes(page + i, page + i, true);
2308         }
2309 }
2310
2311 static void __split_huge_page_tail(struct page *head, int tail,
2312                 struct lruvec *lruvec, struct list_head *list)
2313 {
2314         struct page *page_tail = head + tail;
2315
2316         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2317         VM_BUG_ON_PAGE(page_ref_count(page_tail) != 0, page_tail);
2318
2319         /*
2320          * tail_page->_refcount is zero and not changing from under us. But
2321          * get_page_unless_zero() may be running from under us on the
2322          * tail_page. If we used atomic_set() below instead of atomic_inc() or
2323          * atomic_add(), we would then run atomic_set() concurrently with
2324          * get_page_unless_zero(), and atomic_set() is implemented in C not
2325          * using locked ops. spin_unlock on x86 sometime uses locked ops
2326          * because of PPro errata 66, 92, so unless somebody can guarantee
2327          * atomic_set() here would be safe on all archs (and not only on x86),
2328          * it's safer to use atomic_inc()/atomic_add().
2329          */
2330         if (PageAnon(head) && !PageSwapCache(head)) {
2331                 page_ref_inc(page_tail);
2332         } else {
2333                 /* Additional pin to radix tree */
2334                 page_ref_add(page_tail, 2);
2335         }
2336
2337         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2338         page_tail->flags |= (head->flags &
2339                         ((1L << PG_referenced) |
2340                          (1L << PG_swapbacked) |
2341                          (1L << PG_swapcache) |
2342                          (1L << PG_mlocked) |
2343                          (1L << PG_uptodate) |
2344                          (1L << PG_active) |
2345                          (1L << PG_locked) |
2346                          (1L << PG_unevictable) |
2347                          (1L << PG_dirty)));
2348
2349         /*
2350          * After clearing PageTail the gup refcount can be released.
2351          * Page flags also must be visible before we make the page non-compound.
2352          */
2353         smp_wmb();
2354
2355         clear_compound_head(page_tail);
2356
2357         if (page_is_young(head))
2358                 set_page_young(page_tail);
2359         if (page_is_idle(head))
2360                 set_page_idle(page_tail);
2361
2362         /* ->mapping in first tail page is compound_mapcount */
2363         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2364                         page_tail);
2365         page_tail->mapping = head->mapping;
2366
2367         page_tail->index = head->index + tail;
2368         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2369         lru_add_page_tail(head, page_tail, lruvec, list);
2370 }
2371
2372 static void __split_huge_page(struct page *page, struct list_head *list,
2373                 unsigned long flags)
2374 {
2375         struct page *head = compound_head(page);
2376         struct zone *zone = page_zone(head);
2377         struct lruvec *lruvec;
2378         pgoff_t end = -1;
2379         int i;
2380
2381         lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat);
2382
2383         /* complete memcg works before add pages to LRU */
2384         mem_cgroup_split_huge_fixup(head);
2385
2386         if (!PageAnon(page))
2387                 end = DIV_ROUND_UP(i_size_read(head->mapping->host), PAGE_SIZE);
2388
2389         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
2390                 __split_huge_page_tail(head, i, lruvec, list);
2391                 /* Some pages can be beyond i_size: drop them from page cache */
2392                 if (head[i].index >= end) {
2393                         ClearPageDirty(head + i);
2394                         __delete_from_page_cache(head + i, NULL);
2395                         if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2396                                 shmem_uncharge(head->mapping->host, 1);
2397                         put_page(head + i);
2398                 }
2399         }
2400
2401         ClearPageCompound(head);
2402         /* See comment in __split_huge_page_tail() */
2403         if (PageAnon(head)) {
2404                 /* Additional pin to radix tree of swap cache */
2405                 if (PageSwapCache(head))
2406                         page_ref_add(head, 2);
2407                 else
2408                         page_ref_inc(head);
2409         } else {
2410                 /* Additional pin to radix tree */
2411                 page_ref_add(head, 2);
2412                 spin_unlock(&head->mapping->tree_lock);
2413         }
2414
2415         spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
2416
2417         unfreeze_page(head);
2418
2419         for (i = 0; i < HPAGE_PMD_NR; i++) {
2420                 struct page *subpage = head + i;
2421                 if (subpage == page)
2422                         continue;
2423                 unlock_page(subpage);
2424
2425                 /*
2426                  * Subpages may be freed if there wasn't any mapping
2427                  * like if add_to_swap() is running on a lru page that
2428                  * had its mapping zapped. And freeing these pages
2429                  * requires taking the lru_lock so we do the put_page
2430                  * of the tail pages after the split is complete.
2431                  */
2432                 put_page(subpage);
2433         }
2434 }
2435
2436 int total_mapcount(struct page *page)
2437 {
2438         int i, compound, ret;
2439
2440         VM_BUG_ON_PAGE(PageTail(page), page);
2441
2442         if (likely(!PageCompound(page)))
2443                 return atomic_read(&page->_mapcount) + 1;
2444
2445         compound = compound_mapcount(page);
2446         if (PageHuge(page))
2447                 return compound;
2448         ret = compound;
2449         for (i = 0; i < HPAGE_PMD_NR; i++)
2450                 ret += atomic_read(&page[i]._mapcount) + 1;
2451         /* File pages has compound_mapcount included in _mapcount */
2452         if (!PageAnon(page))
2453                 return ret - compound * HPAGE_PMD_NR;
2454         if (PageDoubleMap(page))
2455                 ret -= HPAGE_PMD_NR;
2456         return ret;
2457 }
2458
2459 /*
2460  * This calculates accurately how many mappings a transparent hugepage
2461  * has (unlike page_mapcount() which isn't fully accurate). This full
2462  * accuracy is primarily needed to know if copy-on-write faults can
2463  * reuse the page and change the mapping to read-write instead of
2464  * copying them. At the same time this returns the total_mapcount too.
2465  *
2466  * The function returns the highest mapcount any one of the subpages
2467  * has. If the return value is one, even if different processes are
2468  * mapping different subpages of the transparent hugepage, they can
2469  * all reuse it, because each process is reusing a different subpage.
2470  *
2471  * The total_mapcount is instead counting all virtual mappings of the
2472  * subpages. If the total_mapcount is equal to "one", it tells the
2473  * caller all mappings belong to the same "mm" and in turn the
2474  * anon_vma of the transparent hugepage can become the vma->anon_vma
2475  * local one as no other process may be mapping any of the subpages.
2476  *
2477  * It would be more accurate to replace page_mapcount() with
2478  * page_trans_huge_mapcount(), however we only use
2479  * page_trans_huge_mapcount() in the copy-on-write faults where we
2480  * need full accuracy to avoid breaking page pinning, because
2481  * page_trans_huge_mapcount() is slower than page_mapcount().
2482  */
2483 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2484 {
2485         int i, ret, _total_mapcount, mapcount;
2486
2487         /* hugetlbfs shouldn't call it */
2488         VM_BUG_ON_PAGE(PageHuge(page), page);
2489
2490         if (likely(!PageTransCompound(page))) {
2491                 mapcount = atomic_read(&page->_mapcount) + 1;
2492                 if (total_mapcount)
2493                         *total_mapcount = mapcount;
2494                 return mapcount;
2495         }
2496
2497         page = compound_head(page);
2498
2499         _total_mapcount = ret = 0;
2500         for (i = 0; i < HPAGE_PMD_NR; i++) {
2501                 mapcount = atomic_read(&page[i]._mapcount) + 1;
2502                 ret = max(ret, mapcount);
2503                 _total_mapcount += mapcount;
2504         }
2505         if (PageDoubleMap(page)) {
2506                 ret -= 1;
2507                 _total_mapcount -= HPAGE_PMD_NR;
2508         }
2509         mapcount = compound_mapcount(page);
2510         ret += mapcount;
2511         _total_mapcount += mapcount;
2512         if (total_mapcount)
2513                 *total_mapcount = _total_mapcount;
2514         return ret;
2515 }
2516
2517 /* Racy check whether the huge page can be split */
2518 bool can_split_huge_page(struct page *page, int *pextra_pins)
2519 {
2520         int extra_pins;
2521
2522         /* Additional pins from radix tree */
2523         if (PageAnon(page))
2524                 extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0;
2525         else
2526                 extra_pins = HPAGE_PMD_NR;
2527         if (pextra_pins)
2528                 *pextra_pins = extra_pins;
2529         return total_mapcount(page) == page_count(page) - extra_pins - 1;
2530 }
2531
2532 /*
2533  * This function splits huge page into normal pages. @page can point to any
2534  * subpage of huge page to split. Split doesn't change the position of @page.
2535  *
2536  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2537  * The huge page must be locked.
2538  *
2539  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2540  *
2541  * Both head page and tail pages will inherit mapping, flags, and so on from
2542  * the hugepage.
2543  *
2544  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2545  * they are not mapped.
2546  *
2547  * Returns 0 if the hugepage is split successfully.
2548  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2549  * us.
2550  */
2551 int split_huge_page_to_list(struct page *page, struct list_head *list)
2552 {
2553         struct page *head = compound_head(page);
2554         struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2555         struct anon_vma *anon_vma = NULL;
2556         struct address_space *mapping = NULL;
2557         int count, mapcount, extra_pins, ret;
2558         bool mlocked;
2559         unsigned long flags;
2560
2561         VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
2562         VM_BUG_ON_PAGE(!PageLocked(page), page);
2563         VM_BUG_ON_PAGE(!PageCompound(page), page);
2564
2565         if (PageWriteback(page))
2566                 return -EBUSY;
2567
2568         if (PageAnon(head)) {
2569                 /*
2570                  * The caller does not necessarily hold an mmap_sem that would
2571                  * prevent the anon_vma disappearing so we first we take a
2572                  * reference to it and then lock the anon_vma for write. This
2573                  * is similar to page_lock_anon_vma_read except the write lock
2574                  * is taken to serialise against parallel split or collapse
2575                  * operations.
2576                  */
2577                 anon_vma = page_get_anon_vma(head);
2578                 if (!anon_vma) {
2579                         ret = -EBUSY;
2580                         goto out;
2581                 }
2582                 mapping = NULL;
2583                 anon_vma_lock_write(anon_vma);
2584         } else {
2585                 mapping = head->mapping;
2586
2587                 /* Truncated ? */
2588                 if (!mapping) {
2589                         ret = -EBUSY;
2590                         goto out;
2591                 }
2592
2593                 anon_vma = NULL;
2594                 i_mmap_lock_read(mapping);
2595         }
2596
2597         /*
2598          * Racy check if we can split the page, before freeze_page() will
2599          * split PMDs
2600          */
2601         if (!can_split_huge_page(head, &extra_pins)) {
2602                 ret = -EBUSY;
2603                 goto out_unlock;
2604         }
2605
2606         mlocked = PageMlocked(page);
2607         freeze_page(head);
2608         VM_BUG_ON_PAGE(compound_mapcount(head), head);
2609
2610         /* Make sure the page is not on per-CPU pagevec as it takes pin */
2611         if (mlocked)
2612                 lru_add_drain();
2613
2614         /* prevent PageLRU to go away from under us, and freeze lru stats */
2615         spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags);
2616
2617         if (mapping) {
2618                 void **pslot;
2619
2620                 spin_lock(&mapping->tree_lock);
2621                 pslot = radix_tree_lookup_slot(&mapping->page_tree,
2622                                 page_index(head));
2623                 /*
2624                  * Check if the head page is present in radix tree.
2625                  * We assume all tail are present too, if head is there.
2626                  */
2627                 if (radix_tree_deref_slot_protected(pslot,
2628                                         &mapping->tree_lock) != head)
2629                         goto fail;
2630         }
2631
2632         /* Prevent deferred_split_scan() touching ->_refcount */
2633         spin_lock(&pgdata->split_queue_lock);
2634         count = page_count(head);
2635         mapcount = total_mapcount(head);
2636         if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2637                 if (!list_empty(page_deferred_list(head))) {
2638                         pgdata->split_queue_len--;
2639                         list_del(page_deferred_list(head));
2640                 }
2641                 if (mapping)
2642                         __dec_node_page_state(page, NR_SHMEM_THPS);
2643                 spin_unlock(&pgdata->split_queue_lock);
2644                 __split_huge_page(page, list, flags);
2645                 if (PageSwapCache(head)) {
2646                         swp_entry_t entry = { .val = page_private(head) };
2647
2648                         ret = split_swap_cluster(entry);
2649                 } else
2650                         ret = 0;
2651         } else {
2652                 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2653                         pr_alert("total_mapcount: %u, page_count(): %u\n",
2654                                         mapcount, count);
2655                         if (PageTail(page))
2656                                 dump_page(head, NULL);
2657                         dump_page(page, "total_mapcount(head) > 0");
2658                         BUG();
2659                 }
2660                 spin_unlock(&pgdata->split_queue_lock);
2661 fail:           if (mapping)
2662                         spin_unlock(&mapping->tree_lock);
2663                 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
2664                 unfreeze_page(head);
2665                 ret = -EBUSY;
2666         }
2667
2668 out_unlock:
2669         if (anon_vma) {
2670                 anon_vma_unlock_write(anon_vma);
2671                 put_anon_vma(anon_vma);
2672         }
2673         if (mapping)
2674                 i_mmap_unlock_read(mapping);
2675 out:
2676         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2677         return ret;
2678 }
2679
2680 void free_transhuge_page(struct page *page)
2681 {
2682         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2683         unsigned long flags;
2684
2685         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2686         if (!list_empty(page_deferred_list(page))) {
2687                 pgdata->split_queue_len--;
2688                 list_del(page_deferred_list(page));
2689         }
2690         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2691         free_compound_page(page);
2692 }
2693
2694 void deferred_split_huge_page(struct page *page)
2695 {
2696         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2697         unsigned long flags;
2698
2699         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2700
2701         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2702         if (list_empty(page_deferred_list(page))) {
2703                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2704                 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2705                 pgdata->split_queue_len++;
2706         }
2707         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2708 }
2709
2710 static unsigned long deferred_split_count(struct shrinker *shrink,
2711                 struct shrink_control *sc)
2712 {
2713         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2714         return ACCESS_ONCE(pgdata->split_queue_len);
2715 }
2716
2717 static unsigned long deferred_split_scan(struct shrinker *shrink,
2718                 struct shrink_control *sc)
2719 {
2720         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2721         unsigned long flags;
2722         LIST_HEAD(list), *pos, *next;
2723         struct page *page;
2724         int split = 0;
2725
2726         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2727         /* Take pin on all head pages to avoid freeing them under us */
2728         list_for_each_safe(pos, next, &pgdata->split_queue) {
2729                 page = list_entry((void *)pos, struct page, mapping);
2730                 page = compound_head(page);
2731                 if (get_page_unless_zero(page)) {
2732                         list_move(page_deferred_list(page), &list);
2733                 } else {
2734                         /* We lost race with put_compound_page() */
2735                         list_del_init(page_deferred_list(page));
2736                         pgdata->split_queue_len--;
2737                 }
2738                 if (!--sc->nr_to_scan)
2739                         break;
2740         }
2741         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2742
2743         list_for_each_safe(pos, next, &list) {
2744                 page = list_entry((void *)pos, struct page, mapping);
2745                 if (!trylock_page(page))
2746                         goto next;
2747                 /* split_huge_page() removes page from list on success */
2748                 if (!split_huge_page(page))
2749                         split++;
2750                 unlock_page(page);
2751 next:
2752                 put_page(page);
2753         }
2754
2755         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2756         list_splice_tail(&list, &pgdata->split_queue);
2757         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2758
2759         /*
2760          * Stop shrinker if we didn't split any page, but the queue is empty.
2761          * This can happen if pages were freed under us.
2762          */
2763         if (!split && list_empty(&pgdata->split_queue))
2764                 return SHRINK_STOP;
2765         return split;
2766 }
2767
2768 static struct shrinker deferred_split_shrinker = {
2769         .count_objects = deferred_split_count,
2770         .scan_objects = deferred_split_scan,
2771         .seeks = DEFAULT_SEEKS,
2772         .flags = SHRINKER_NUMA_AWARE,
2773 };
2774
2775 #ifdef CONFIG_DEBUG_FS
2776 static int split_huge_pages_set(void *data, u64 val)
2777 {
2778         struct zone *zone;
2779         struct page *page;
2780         unsigned long pfn, max_zone_pfn;
2781         unsigned long total = 0, split = 0;
2782
2783         if (val != 1)
2784                 return -EINVAL;
2785
2786         for_each_populated_zone(zone) {
2787                 max_zone_pfn = zone_end_pfn(zone);
2788                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2789                         if (!pfn_valid(pfn))
2790                                 continue;
2791
2792                         page = pfn_to_page(pfn);
2793                         if (!get_page_unless_zero(page))
2794                                 continue;
2795
2796                         if (zone != page_zone(page))
2797                                 goto next;
2798
2799                         if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2800                                 goto next;
2801
2802                         total++;
2803                         lock_page(page);
2804                         if (!split_huge_page(page))
2805                                 split++;
2806                         unlock_page(page);
2807 next:
2808                         put_page(page);
2809                 }
2810         }
2811
2812         pr_info("%lu of %lu THP split\n", split, total);
2813
2814         return 0;
2815 }
2816 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2817                 "%llu\n");
2818
2819 static int __init split_huge_pages_debugfs(void)
2820 {
2821         void *ret;
2822
2823         ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2824                         &split_huge_pages_fops);
2825         if (!ret)
2826                 pr_warn("Failed to create split_huge_pages in debugfs");
2827         return 0;
2828 }
2829 late_initcall(split_huge_pages_debugfs);
2830 #endif
2831
2832 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2833 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
2834                 struct page *page)
2835 {
2836         struct vm_area_struct *vma = pvmw->vma;
2837         struct mm_struct *mm = vma->vm_mm;
2838         unsigned long address = pvmw->address;
2839         pmd_t pmdval;
2840         swp_entry_t entry;
2841         pmd_t pmdswp;
2842
2843         if (!(pvmw->pmd && !pvmw->pte))
2844                 return;
2845
2846         mmu_notifier_invalidate_range_start(mm, address,
2847                         address + HPAGE_PMD_SIZE);
2848
2849         flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
2850         pmdval = *pvmw->pmd;
2851         pmdp_invalidate(vma, address, pvmw->pmd);
2852         if (pmd_dirty(pmdval))
2853                 set_page_dirty(page);
2854         entry = make_migration_entry(page, pmd_write(pmdval));
2855         pmdswp = swp_entry_to_pmd(entry);
2856         if (pmd_soft_dirty(pmdval))
2857                 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
2858         set_pmd_at(mm, address, pvmw->pmd, pmdswp);
2859         page_remove_rmap(page, true);
2860         put_page(page);
2861
2862         mmu_notifier_invalidate_range_end(mm, address,
2863                         address + HPAGE_PMD_SIZE);
2864 }
2865
2866 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
2867 {
2868         struct vm_area_struct *vma = pvmw->vma;
2869         struct mm_struct *mm = vma->vm_mm;
2870         unsigned long address = pvmw->address;
2871         unsigned long mmun_start = address & HPAGE_PMD_MASK;
2872         pmd_t pmde;
2873         swp_entry_t entry;
2874
2875         if (!(pvmw->pmd && !pvmw->pte))
2876                 return;
2877
2878         entry = pmd_to_swp_entry(*pvmw->pmd);
2879         get_page(new);
2880         pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
2881         if (pmd_swp_soft_dirty(*pvmw->pmd))
2882                 pmde = pmd_mksoft_dirty(pmde);
2883         if (is_write_migration_entry(entry))
2884                 pmde = maybe_pmd_mkwrite(pmde, vma);
2885
2886         flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
2887         page_add_anon_rmap(new, vma, mmun_start, true);
2888         set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
2889         if (vma->vm_flags & VM_LOCKED)
2890                 mlock_vma_page(new);
2891         update_mmu_cache_pmd(vma, address, pvmw->pmd);
2892 }
2893 #endif