mm/huge_memory.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *  Copyright (C) 2009  Red Hat, Inc.
   4  */
   5
   6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   7
   8 #include <linux/mm.h>
   9 #include <linux/sched.h>
  10 #include <linux/sched/coredump.h>
  11 #include <linux/sched/numa_balancing.h>
  12 #include <linux/highmem.h>
  13 #include <linux/hugetlb.h>
  14 #include <linux/mmu_notifier.h>
  15 #include <linux/rmap.h>
  16 #include <linux/swap.h>
  17 #include <linux/shrinker.h>
  18 #include <linux/mm_inline.h>
  19 #include <linux/swapops.h>
  20 #include <linux/dax.h>
  21 #include <linux/khugepaged.h>
  22 #include <linux/freezer.h>
  23 #include <linux/pfn_t.h>
  24 #include <linux/mman.h>
  25 #include <linux/memremap.h>
  26 #include <linux/pagemap.h>
  27 #include <linux/debugfs.h>
  28 #include <linux/migrate.h>
  29 #include <linux/hashtable.h>
  30 #include <linux/userfaultfd_k.h>
  31 #include <linux/page_idle.h>
  32 #include <linux/shmem_fs.h>
  33 #include <linux/oom.h>
  34 #include <linux/numa.h>
  35 #include <linux/page_owner.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 to avoid
  43  * risking an increased memory footprint for applications that are not
  44  * guaranteed to benefit from it. When transparent hugepage support is
  45  * enabled, it is for all mappings, 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 bool transparent_hugepage_enabled(struct vm_area_struct *vma)
  66 {
  67         /* The addr is used to check if the vma size fits */
  68         unsigned long addr = (vma->vm_end & HPAGE_PMD_MASK) - HPAGE_PMD_SIZE;
  69
  70         if (!transhuge_vma_suitable(vma, addr))
  71                 return false;
  72         if (vma_is_anonymous(vma))
  73                 return __transparent_hugepage_enabled(vma);
  74         if (vma_is_shmem(vma))
  75                 return shmem_huge_enabled(vma);
  76
  77         return false;
  78 }
  79
  80 static struct page *get_huge_zero_page(void)
  81 {
  82         struct page *zero_page;
  83 retry:
  84         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
  85                 return READ_ONCE(huge_zero_page);
  86
  87         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
  88                         HPAGE_PMD_ORDER);
  89         if (!zero_page) {
  90                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
  91                 return NULL;
  92         }
  93         count_vm_event(THP_ZERO_PAGE_ALLOC);
  94         preempt_disable();
  95         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
  96                 preempt_enable();
  97                 __free_pages(zero_page, compound_order(zero_page));
  98                 goto retry;
  99         }
 100
 101         /* We take additional reference here. It will be put back by shrinker */
 102         atomic_set(&huge_zero_refcount, 2);
 103         preempt_enable();
 104         return READ_ONCE(huge_zero_page);
 105 }
 106
 107 static void put_huge_zero_page(void)
 108 {
 109         /*
 110          * Counter should never go to zero here. Only shrinker can put
 111          * last reference.
 112          */
 113         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 114 }
 115
 116 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
 117 {
 118         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 119                 return READ_ONCE(huge_zero_page);
 120
 121         if (!get_huge_zero_page())
 122                 return NULL;
 123
 124         if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 125                 put_huge_zero_page();
 126
 127         return READ_ONCE(huge_zero_page);
 128 }
 129
 130 void mm_put_huge_zero_page(struct mm_struct *mm)
 131 {
 132         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 133                 put_huge_zero_page();
 134 }
 135
 136 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 137                                         struct shrink_control *sc)
 138 {
 139         /* we can free zero page only if last reference remains */
 140         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 141 }
 142
 143 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 144                                        struct shrink_control *sc)
 145 {
 146         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 147                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 148                 BUG_ON(zero_page == NULL);
 149                 __free_pages(zero_page, compound_order(zero_page));
 150                 return HPAGE_PMD_NR;
 151         }
 152
 153         return 0;
 154 }
 155
 156 static struct shrinker huge_zero_page_shrinker = {
 157         .count_objects = shrink_huge_zero_page_count,
 158         .scan_objects = shrink_huge_zero_page_scan,
 159         .seeks = DEFAULT_SEEKS,
 160 };
 161
 162 #ifdef CONFIG_SYSFS
 163 static ssize_t enabled_show(struct kobject *kobj,
 164                             struct kobj_attribute *attr, char *buf)
 165 {
 166         if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
 167                 return sprintf(buf, "[always] madvise never\n");
 168         else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
 169                 return sprintf(buf, "always [madvise] never\n");
 170         else
 171                 return sprintf(buf, "always madvise [never]\n");
 172 }
 173
 174 static ssize_t enabled_store(struct kobject *kobj,
 175                              struct kobj_attribute *attr,
 176                              const char *buf, size_t count)
 177 {
 178         ssize_t ret = count;
 179
 180         if (sysfs_streq(buf, "always")) {
 181                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 182                 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 183         } else if (sysfs_streq(buf, "madvise")) {
 184                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 185                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 186         } else if (sysfs_streq(buf, "never")) {
 187                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 188                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 189         } else
 190                 ret = -EINVAL;
 191
 192         if (ret > 0) {
 193                 int err = start_stop_khugepaged();
 194                 if (err)
 195                         ret = err;
 196         }
 197         return ret;
 198 }
 199 static struct kobj_attribute enabled_attr =
 200         __ATTR(enabled, 0644, enabled_show, enabled_store);
 201
 202 ssize_t single_hugepage_flag_show(struct kobject *kobj,
 203                                 struct kobj_attribute *attr, char *buf,
 204                                 enum transparent_hugepage_flag flag)
 205 {
 206         return sprintf(buf, "%d\n",
 207                        !!test_bit(flag, &transparent_hugepage_flags));
 208 }
 209
 210 ssize_t single_hugepage_flag_store(struct kobject *kobj,
 211                                  struct kobj_attribute *attr,
 212                                  const char *buf, size_t count,
 213                                  enum transparent_hugepage_flag flag)
 214 {
 215         unsigned long value;
 216         int ret;
 217
 218         ret = kstrtoul(buf, 10, &value);
 219         if (ret < 0)
 220                 return ret;
 221         if (value > 1)
 222                 return -EINVAL;
 223
 224         if (value)
 225                 set_bit(flag, &transparent_hugepage_flags);
 226         else
 227                 clear_bit(flag, &transparent_hugepage_flags);
 228
 229         return count;
 230 }
 231
 232 static ssize_t defrag_show(struct kobject *kobj,
 233                            struct kobj_attribute *attr, char *buf)
 234 {
 235         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 236                 return sprintf(buf, "[always] defer defer+madvise madvise never\n");
 237         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 238                 return sprintf(buf, "always [defer] defer+madvise madvise never\n");
 239         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 240                 return sprintf(buf, "always defer [defer+madvise] madvise never\n");
 241         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 242                 return sprintf(buf, "always defer defer+madvise [madvise] never\n");
 243         return sprintf(buf, "always defer defer+madvise madvise [never]\n");
 244 }
 245
 246 static ssize_t defrag_store(struct kobject *kobj,
 247                             struct kobj_attribute *attr,
 248                             const char *buf, size_t count)
 249 {
 250         if (sysfs_streq(buf, "always")) {
 251                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 252                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 253                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 254                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 255         } else if (sysfs_streq(buf, "defer+madvise")) {
 256                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 257                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 258                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 259                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 260         } else if (sysfs_streq(buf, "defer")) {
 261                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 262                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 263                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 264                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 265         } else if (sysfs_streq(buf, "madvise")) {
 266                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 267                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 268                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 269                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 270         } else if (sysfs_streq(buf, "never")) {
 271                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 272                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 273                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 274                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 275         } else
 276                 return -EINVAL;
 277
 278         return count;
 279 }
 280 static struct kobj_attribute defrag_attr =
 281         __ATTR(defrag, 0644, defrag_show, defrag_store);
 282
 283 static ssize_t use_zero_page_show(struct kobject *kobj,
 284                 struct kobj_attribute *attr, char *buf)
 285 {
 286         return single_hugepage_flag_show(kobj, attr, buf,
 287                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 288 }
 289 static ssize_t use_zero_page_store(struct kobject *kobj,
 290                 struct kobj_attribute *attr, const char *buf, size_t count)
 291 {
 292         return single_hugepage_flag_store(kobj, attr, buf, count,
 293                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 294 }
 295 static struct kobj_attribute use_zero_page_attr =
 296         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 297
 298 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
 299                 struct kobj_attribute *attr, char *buf)
 300 {
 301         return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE);
 302 }
 303 static struct kobj_attribute hpage_pmd_size_attr =
 304         __ATTR_RO(hpage_pmd_size);
 305
 306 static struct attribute *hugepage_attr[] = {
 307         &enabled_attr.attr,
 308         &defrag_attr.attr,
 309         &use_zero_page_attr.attr,
 310         &hpage_pmd_size_attr.attr,
 311 #ifdef CONFIG_SHMEM
 312         &shmem_enabled_attr.attr,
 313 #endif
 314         NULL,
 315 };
 316
 317 static const struct attribute_group hugepage_attr_group = {
 318         .attrs = hugepage_attr,
 319 };
 320
 321 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 322 {
 323         int err;
 324
 325         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 326         if (unlikely(!*hugepage_kobj)) {
 327                 pr_err("failed to create transparent hugepage kobject\n");
 328                 return -ENOMEM;
 329         }
 330
 331         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 332         if (err) {
 333                 pr_err("failed to register transparent hugepage group\n");
 334                 goto delete_obj;
 335         }
 336
 337         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 338         if (err) {
 339                 pr_err("failed to register transparent hugepage group\n");
 340                 goto remove_hp_group;
 341         }
 342
 343         return 0;
 344
 345 remove_hp_group:
 346         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 347 delete_obj:
 348         kobject_put(*hugepage_kobj);
 349         return err;
 350 }
 351
 352 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 353 {
 354         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 355         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 356         kobject_put(hugepage_kobj);
 357 }
 358 #else
 359 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 360 {
 361         return 0;
 362 }
 363
 364 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 365 {
 366 }
 367 #endif /* CONFIG_SYSFS */
 368
 369 static int __init hugepage_init(void)
 370 {
 371         int err;
 372         struct kobject *hugepage_kobj;
 373
 374         if (!has_transparent_hugepage()) {
 375                 transparent_hugepage_flags = 0;
 376                 return -EINVAL;
 377         }
 378
 379         /*
 380          * hugepages can't be allocated by the buddy allocator
 381          */
 382         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
 383         /*
 384          * we use page->mapping and page->index in second tail page
 385          * as list_head: assuming THP order >= 2
 386          */
 387         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 388
 389         err = hugepage_init_sysfs(&hugepage_kobj);
 390         if (err)
 391                 goto err_sysfs;
 392
 393         err = khugepaged_init();
 394         if (err)
 395                 goto err_slab;
 396
 397         err = register_shrinker(&huge_zero_page_shrinker);
 398         if (err)
 399                 goto err_hzp_shrinker;
 400         err = register_shrinker(&deferred_split_shrinker);
 401         if (err)
 402                 goto err_split_shrinker;
 403
 404         /*
 405          * By default disable transparent hugepages on smaller systems,
 406          * where the extra memory used could hurt more than TLB overhead
 407          * is likely to save.  The admin can still enable it through /sys.
 408          */
 409         if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
 410                 transparent_hugepage_flags = 0;
 411                 return 0;
 412         }
 413
 414         err = start_stop_khugepaged();
 415         if (err)
 416                 goto err_khugepaged;
 417
 418         return 0;
 419 err_khugepaged:
 420         unregister_shrinker(&deferred_split_shrinker);
 421 err_split_shrinker:
 422         unregister_shrinker(&huge_zero_page_shrinker);
 423 err_hzp_shrinker:
 424         khugepaged_destroy();
 425 err_slab:
 426         hugepage_exit_sysfs(hugepage_kobj);
 427 err_sysfs:
 428         return err;
 429 }
 430 subsys_initcall(hugepage_init);
 431
 432 static int __init setup_transparent_hugepage(char *str)
 433 {
 434         int ret = 0;
 435         if (!str)
 436                 goto out;
 437         if (!strcmp(str, "always")) {
 438                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 439                         &transparent_hugepage_flags);
 440                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 441                           &transparent_hugepage_flags);
 442                 ret = 1;
 443         } else if (!strcmp(str, "madvise")) {
 444                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 445                           &transparent_hugepage_flags);
 446                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 447                         &transparent_hugepage_flags);
 448                 ret = 1;
 449         } else if (!strcmp(str, "never")) {
 450                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 451                           &transparent_hugepage_flags);
 452                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 453                           &transparent_hugepage_flags);
 454                 ret = 1;
 455         }
 456 out:
 457         if (!ret)
 458                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 459         return ret;
 460 }
 461 __setup("transparent_hugepage=", setup_transparent_hugepage);
 462
 463 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 464 {
 465         if (likely(vma->vm_flags & VM_WRITE))
 466                 pmd = pmd_mkwrite(pmd);
 467         return pmd;
 468 }
 469
 470 #ifdef CONFIG_MEMCG
 471 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
 472 {
 473         struct mem_cgroup *memcg = compound_head(page)->mem_cgroup;
 474         struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
 475
 476         if (memcg)
 477                 return &memcg->deferred_split_queue;
 478         else
 479                 return &pgdat->deferred_split_queue;
 480 }
 481 #else
 482 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
 483 {
 484         struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
 485
 486         return &pgdat->deferred_split_queue;
 487 }
 488 #endif
 489
 490 void prep_transhuge_page(struct page *page)
 491 {
 492         /*
 493          * we use page->mapping and page->indexlru in second tail page
 494          * as list_head: assuming THP order >= 2
 495          */
 496
 497         INIT_LIST_HEAD(page_deferred_list(page));
 498         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 499 }
 500
 501 bool is_transparent_hugepage(struct page *page)
 502 {
 503         if (!PageCompound(page))
 504                 return false;
 505
 506         page = compound_head(page);
 507         return is_huge_zero_page(page) ||
 508                page[1].compound_dtor == TRANSHUGE_PAGE_DTOR;
 509 }
 510 EXPORT_SYMBOL_GPL(is_transparent_hugepage);
 511
 512 static unsigned long __thp_get_unmapped_area(struct file *filp,
 513                 unsigned long addr, unsigned long len,
 514                 loff_t off, unsigned long flags, unsigned long size)
 515 {
 516         loff_t off_end = off + len;
 517         loff_t off_align = round_up(off, size);
 518         unsigned long len_pad, ret;
 519
 520         if (off_end <= off_align || (off_end - off_align) < size)
 521                 return 0;
 522
 523         len_pad = len + size;
 524         if (len_pad < len || (off + len_pad) < off)
 525                 return 0;
 526
 527         ret = current->mm->get_unmapped_area(filp, addr, len_pad,
 528                                               off >> PAGE_SHIFT, flags);
 529
 530         /*
 531          * The failure might be due to length padding. The caller will retry
 532          * without the padding.
 533          */
 534         if (IS_ERR_VALUE(ret))
 535                 return 0;
 536
 537         /*
 538          * Do not try to align to THP boundary if allocation at the address
 539          * hint succeeds.
 540          */
 541         if (ret == addr)
 542                 return addr;
 543
 544         ret += (off - ret) & (size - 1);
 545         return ret;
 546 }
 547
 548 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 549                 unsigned long len, unsigned long pgoff, unsigned long flags)
 550 {
 551         unsigned long ret;
 552         loff_t off = (loff_t)pgoff << PAGE_SHIFT;
 553
 554         if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
 555                 goto out;
 556
 557         ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE);
 558         if (ret)
 559                 return ret;
 560 out:
 561         return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
 562 }
 563 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
 564
 565 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
 566                         struct page *page, gfp_t gfp)
 567 {
 568         struct vm_area_struct *vma = vmf->vma;
 569         pgtable_t pgtable;
 570         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 571         vm_fault_t ret = 0;
 572
 573         VM_BUG_ON_PAGE(!PageCompound(page), page);
 574
 575         if (mem_cgroup_charge(page, vma->vm_mm, gfp)) {
 576                 put_page(page);
 577                 count_vm_event(THP_FAULT_FALLBACK);
 578                 count_vm_event(THP_FAULT_FALLBACK_CHARGE);
 579                 return VM_FAULT_FALLBACK;
 580         }
 581         cgroup_throttle_swaprate(page, gfp);
 582
 583         pgtable = pte_alloc_one(vma->vm_mm);
 584         if (unlikely(!pgtable)) {
 585                 ret = VM_FAULT_OOM;
 586                 goto release;
 587         }
 588
 589         clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
 590         /*
 591          * The memory barrier inside __SetPageUptodate makes sure that
 592          * clear_huge_page writes become visible before the set_pmd_at()
 593          * write.
 594          */
 595         __SetPageUptodate(page);
 596
 597         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 598         if (unlikely(!pmd_none(*vmf->pmd))) {
 599                 goto unlock_release;
 600         } else {
 601                 pmd_t entry;
 602
 603                 ret = check_stable_address_space(vma->vm_mm);
 604                 if (ret)
 605                         goto unlock_release;
 606
 607                 /* Deliver the page fault to userland */
 608                 if (userfaultfd_missing(vma)) {
 609                         vm_fault_t ret2;
 610
 611                         spin_unlock(vmf->ptl);
 612                         put_page(page);
 613                         pte_free(vma->vm_mm, pgtable);
 614                         ret2 = handle_userfault(vmf, VM_UFFD_MISSING);
 615                         VM_BUG_ON(ret2 & VM_FAULT_FALLBACK);
 616                         return ret2;
 617                 }
 618
 619                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 620                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 621                 page_add_new_anon_rmap(page, vma, haddr, true);
 622                 lru_cache_add_inactive_or_unevictable(page, vma);
 623                 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
 624                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
 625                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 626                 mm_inc_nr_ptes(vma->vm_mm);
 627                 spin_unlock(vmf->ptl);
 628                 count_vm_event(THP_FAULT_ALLOC);
 629                 count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC);
 630         }
 631
 632         return 0;
 633 unlock_release:
 634         spin_unlock(vmf->ptl);
 635 release:
 636         if (pgtable)
 637                 pte_free(vma->vm_mm, pgtable);
 638         put_page(page);
 639         return ret;
 640
 641 }
 642
 643 /*
 644  * always: directly stall for all thp allocations
 645  * defer: wake kswapd and fail if not immediately available
 646  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
 647  *                fail if not immediately available
 648  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
 649  *          available
 650  * never: never stall for any thp allocation
 651  */
 652 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
 653 {
 654         const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
 655
 656         /* Always do synchronous compaction */
 657         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 658                 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
 659
 660         /* Kick kcompactd and fail quickly */
 661         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 662                 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
 663
 664         /* Synchronous compaction if madvised, otherwise kick kcompactd */
 665         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 666                 return GFP_TRANSHUGE_LIGHT |
 667                         (vma_madvised ? __GFP_DIRECT_RECLAIM :
 668                                         __GFP_KSWAPD_RECLAIM);
 669
 670         /* Only do synchronous compaction if madvised */
 671         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 672                 return GFP_TRANSHUGE_LIGHT |
 673                        (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
 674
 675         return GFP_TRANSHUGE_LIGHT;
 676 }
 677
 678 /* Caller must hold page table lock. */
 679 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 680                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 681                 struct page *zero_page)
 682 {
 683         pmd_t entry;
 684         if (!pmd_none(*pmd))
 685                 return false;
 686         entry = mk_pmd(zero_page, vma->vm_page_prot);
 687         entry = pmd_mkhuge(entry);
 688         if (pgtable)
 689                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 690         set_pmd_at(mm, haddr, pmd, entry);
 691         mm_inc_nr_ptes(mm);
 692         return true;
 693 }
 694
 695 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
 696 {
 697         struct vm_area_struct *vma = vmf->vma;
 698         gfp_t gfp;
 699         struct page *page;
 700         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 701
 702         if (!transhuge_vma_suitable(vma, haddr))
 703                 return VM_FAULT_FALLBACK;
 704         if (unlikely(anon_vma_prepare(vma)))
 705                 return VM_FAULT_OOM;
 706         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 707                 return VM_FAULT_OOM;
 708         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
 709                         !mm_forbids_zeropage(vma->vm_mm) &&
 710                         transparent_hugepage_use_zero_page()) {
 711                 pgtable_t pgtable;
 712                 struct page *zero_page;
 713                 bool set;
 714                 vm_fault_t ret;
 715                 pgtable = pte_alloc_one(vma->vm_mm);
 716                 if (unlikely(!pgtable))
 717                         return VM_FAULT_OOM;
 718                 zero_page = mm_get_huge_zero_page(vma->vm_mm);
 719                 if (unlikely(!zero_page)) {
 720                         pte_free(vma->vm_mm, pgtable);
 721                         count_vm_event(THP_FAULT_FALLBACK);
 722                         return VM_FAULT_FALLBACK;
 723                 }
 724                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 725                 ret = 0;
 726                 set = false;
 727                 if (pmd_none(*vmf->pmd)) {
 728                         ret = check_stable_address_space(vma->vm_mm);
 729                         if (ret) {
 730                                 spin_unlock(vmf->ptl);
 731                         } else if (userfaultfd_missing(vma)) {
 732                                 spin_unlock(vmf->ptl);
 733                                 ret = handle_userfault(vmf, VM_UFFD_MISSING);
 734                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 735                         } else {
 736                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
 737                                                    haddr, vmf->pmd, zero_page);
 738                                 spin_unlock(vmf->ptl);
 739                                 set = true;
 740                         }
 741                 } else
 742                         spin_unlock(vmf->ptl);
 743                 if (!set)
 744                         pte_free(vma->vm_mm, pgtable);
 745                 return ret;
 746         }
 747         gfp = alloc_hugepage_direct_gfpmask(vma);
 748         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 749         if (unlikely(!page)) {
 750                 count_vm_event(THP_FAULT_FALLBACK);
 751                 return VM_FAULT_FALLBACK;
 752         }
 753         prep_transhuge_page(page);
 754         return __do_huge_pmd_anonymous_page(vmf, page, gfp);
 755 }
 756
 757 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 758                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
 759                 pgtable_t pgtable)
 760 {
 761         struct mm_struct *mm = vma->vm_mm;
 762         pmd_t entry;
 763         spinlock_t *ptl;
 764
 765         ptl = pmd_lock(mm, pmd);
 766         if (!pmd_none(*pmd)) {
 767                 if (write) {
 768                         if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
 769                                 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
 770                                 goto out_unlock;
 771                         }
 772                         entry = pmd_mkyoung(*pmd);
 773                         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 774                         if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
 775                                 update_mmu_cache_pmd(vma, addr, pmd);
 776                 }
 777
 778                 goto out_unlock;
 779         }
 780
 781         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 782         if (pfn_t_devmap(pfn))
 783                 entry = pmd_mkdevmap(entry);
 784         if (write) {
 785                 entry = pmd_mkyoung(pmd_mkdirty(entry));
 786                 entry = maybe_pmd_mkwrite(entry, vma);
 787         }
 788
 789         if (pgtable) {
 790                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 791                 mm_inc_nr_ptes(mm);
 792                 pgtable = NULL;
 793         }
 794
 795         set_pmd_at(mm, addr, pmd, entry);
 796         update_mmu_cache_pmd(vma, addr, pmd);
 797
 798 out_unlock:
 799         spin_unlock(ptl);
 800         if (pgtable)
 801                 pte_free(mm, pgtable);
 802 }
 803
 804 /**
 805  * vmf_insert_pfn_pmd_prot - insert a pmd size pfn
 806  * @vmf: Structure describing the fault
 807  * @pfn: pfn to insert
 808  * @pgprot: page protection to use
 809  * @write: whether it's a write fault
 810  *
 811  * Insert a pmd size pfn. See vmf_insert_pfn() for additional info and
 812  * also consult the vmf_insert_mixed_prot() documentation when
 813  * @pgprot != @vmf->vma->vm_page_prot.
 814  *
 815  * Return: vm_fault_t value.
 816  */
 817 vm_fault_t vmf_insert_pfn_pmd_prot(struct vm_fault *vmf, pfn_t pfn,
 818                                    pgprot_t pgprot, bool write)
 819 {
 820         unsigned long addr = vmf->address & PMD_MASK;
 821         struct vm_area_struct *vma = vmf->vma;
 822         pgtable_t pgtable = NULL;
 823
 824         /*
 825          * If we had pmd_special, we could avoid all these restrictions,
 826          * but we need to be consistent with PTEs and architectures that
 827          * can't support a 'special' bit.
 828          */
 829         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 830                         !pfn_t_devmap(pfn));
 831         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 832                                                 (VM_PFNMAP|VM_MIXEDMAP));
 833         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 834
 835         if (addr < vma->vm_start || addr >= vma->vm_end)
 836                 return VM_FAULT_SIGBUS;
 837
 838         if (arch_needs_pgtable_deposit()) {
 839                 pgtable = pte_alloc_one(vma->vm_mm);
 840                 if (!pgtable)
 841                         return VM_FAULT_OOM;
 842         }
 843
 844         track_pfn_insert(vma, &pgprot, pfn);
 845
 846         insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
 847         return VM_FAULT_NOPAGE;
 848 }
 849 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd_prot);
 850
 851 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 852 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
 853 {
 854         if (likely(vma->vm_flags & VM_WRITE))
 855                 pud = pud_mkwrite(pud);
 856         return pud;
 857 }
 858
 859 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 860                 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
 861 {
 862         struct mm_struct *mm = vma->vm_mm;
 863         pud_t entry;
 864         spinlock_t *ptl;
 865
 866         ptl = pud_lock(mm, pud);
 867         if (!pud_none(*pud)) {
 868                 if (write) {
 869                         if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
 870                                 WARN_ON_ONCE(!is_huge_zero_pud(*pud));
 871                                 goto out_unlock;
 872                         }
 873                         entry = pud_mkyoung(*pud);
 874                         entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
 875                         if (pudp_set_access_flags(vma, addr, pud, entry, 1))
 876                                 update_mmu_cache_pud(vma, addr, pud);
 877                 }
 878                 goto out_unlock;
 879         }
 880
 881         entry = pud_mkhuge(pfn_t_pud(pfn, prot));
 882         if (pfn_t_devmap(pfn))
 883                 entry = pud_mkdevmap(entry);
 884         if (write) {
 885                 entry = pud_mkyoung(pud_mkdirty(entry));
 886                 entry = maybe_pud_mkwrite(entry, vma);
 887         }
 888         set_pud_at(mm, addr, pud, entry);
 889         update_mmu_cache_pud(vma, addr, pud);
 890
 891 out_unlock:
 892         spin_unlock(ptl);
 893 }
 894
 895 /**
 896  * vmf_insert_pfn_pud_prot - insert a pud size pfn
 897  * @vmf: Structure describing the fault
 898  * @pfn: pfn to insert
 899  * @pgprot: page protection to use
 900  * @write: whether it's a write fault
 901  *
 902  * Insert a pud size pfn. See vmf_insert_pfn() for additional info and
 903  * also consult the vmf_insert_mixed_prot() documentation when
 904  * @pgprot != @vmf->vma->vm_page_prot.
 905  *
 906  * Return: vm_fault_t value.
 907  */
 908 vm_fault_t vmf_insert_pfn_pud_prot(struct vm_fault *vmf, pfn_t pfn,
 909                                    pgprot_t pgprot, bool write)
 910 {
 911         unsigned long addr = vmf->address & PUD_MASK;
 912         struct vm_area_struct *vma = vmf->vma;
 913
 914         /*
 915          * If we had pud_special, we could avoid all these restrictions,
 916          * but we need to be consistent with PTEs and architectures that
 917          * can't support a 'special' bit.
 918          */
 919         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 920                         !pfn_t_devmap(pfn));
 921         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 922                                                 (VM_PFNMAP|VM_MIXEDMAP));
 923         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 924
 925         if (addr < vma->vm_start || addr >= vma->vm_end)
 926                 return VM_FAULT_SIGBUS;
 927
 928         track_pfn_insert(vma, &pgprot, pfn);
 929
 930         insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
 931         return VM_FAULT_NOPAGE;
 932 }
 933 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud_prot);
 934 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
 935
 936 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
 937                 pmd_t *pmd, int flags)
 938 {
 939         pmd_t _pmd;
 940
 941         _pmd = pmd_mkyoung(*pmd);
 942         if (flags & FOLL_WRITE)
 943                 _pmd = pmd_mkdirty(_pmd);
 944         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
 945                                 pmd, _pmd, flags & FOLL_WRITE))
 946                 update_mmu_cache_pmd(vma, addr, pmd);
 947 }
 948
 949 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 950                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
 951 {
 952         unsigned long pfn = pmd_pfn(*pmd);
 953         struct mm_struct *mm = vma->vm_mm;
 954         struct page *page;
 955
 956         assert_spin_locked(pmd_lockptr(mm, pmd));
 957
 958         /*
 959          * When we COW a devmap PMD entry, we split it into PTEs, so we should
 960          * not be in this function with `flags & FOLL_COW` set.
 961          */
 962         WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
 963
 964         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
 965         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
 966                          (FOLL_PIN | FOLL_GET)))
 967                 return NULL;
 968
 969         if (flags & FOLL_WRITE && !pmd_write(*pmd))
 970                 return NULL;
 971
 972         if (pmd_present(*pmd) && pmd_devmap(*pmd))
 973                 /* pass */;
 974         else
 975                 return NULL;
 976
 977         if (flags & FOLL_TOUCH)
 978                 touch_pmd(vma, addr, pmd, flags);
 979
 980         /*
 981          * device mapped pages can only be returned if the
 982          * caller will manage the page reference count.
 983          */
 984         if (!(flags & (FOLL_GET | FOLL_PIN)))
 985                 return ERR_PTR(-EEXIST);
 986
 987         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
 988         *pgmap = get_dev_pagemap(pfn, *pgmap);
 989         if (!*pgmap)
 990                 return ERR_PTR(-EFAULT);
 991         page = pfn_to_page(pfn);
 992         if (!try_grab_page(page, flags))
 993                 page = ERR_PTR(-ENOMEM);
 994
 995         return page;
 996 }
 997
 998 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 999                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1000                   struct vm_area_struct *vma)
1001 {
1002         spinlock_t *dst_ptl, *src_ptl;
1003         struct page *src_page;
1004         pmd_t pmd;
1005         pgtable_t pgtable = NULL;
1006         int ret = -ENOMEM;
1007
1008         /* Skip if can be re-fill on fault */
1009         if (!vma_is_anonymous(vma))
1010                 return 0;
1011
1012         pgtable = pte_alloc_one(dst_mm);
1013         if (unlikely(!pgtable))
1014                 goto out;
1015
1016         dst_ptl = pmd_lock(dst_mm, dst_pmd);
1017         src_ptl = pmd_lockptr(src_mm, src_pmd);
1018         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1019
1020         ret = -EAGAIN;
1021         pmd = *src_pmd;
1022
1023         /*
1024          * Make sure the _PAGE_UFFD_WP bit is cleared if the new VMA
1025          * does not have the VM_UFFD_WP, which means that the uffd
1026          * fork event is not enabled.
1027          */
1028         if (!(vma->vm_flags & VM_UFFD_WP))
1029                 pmd = pmd_clear_uffd_wp(pmd);
1030
1031 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1032         if (unlikely(is_swap_pmd(pmd))) {
1033                 swp_entry_t entry = pmd_to_swp_entry(pmd);
1034
1035                 VM_BUG_ON(!is_pmd_migration_entry(pmd));
1036                 if (is_write_migration_entry(entry)) {
1037                         make_migration_entry_read(&entry);
1038                         pmd = swp_entry_to_pmd(entry);
1039                         if (pmd_swp_soft_dirty(*src_pmd))
1040                                 pmd = pmd_swp_mksoft_dirty(pmd);
1041                         set_pmd_at(src_mm, addr, src_pmd, pmd);
1042                 }
1043                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1044                 mm_inc_nr_ptes(dst_mm);
1045                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1046                 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1047                 ret = 0;
1048                 goto out_unlock;
1049         }
1050 #endif
1051
1052         if (unlikely(!pmd_trans_huge(pmd))) {
1053                 pte_free(dst_mm, pgtable);
1054                 goto out_unlock;
1055         }
1056         /*
1057          * When page table lock is held, the huge zero pmd should not be
1058          * under splitting since we don't split the page itself, only pmd to
1059          * a page table.
1060          */
1061         if (is_huge_zero_pmd(pmd)) {
1062                 struct page *zero_page;
1063                 /*
1064                  * get_huge_zero_page() will never allocate a new page here,
1065                  * since we already have a zero page to copy. It just takes a
1066                  * reference.
1067                  */
1068                 zero_page = mm_get_huge_zero_page(dst_mm);
1069                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1070                                 zero_page);
1071                 ret = 0;
1072                 goto out_unlock;
1073         }
1074
1075         src_page = pmd_page(pmd);
1076         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1077         get_page(src_page);
1078         page_dup_rmap(src_page, true);
1079         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1080         mm_inc_nr_ptes(dst_mm);
1081         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1082
1083         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1084         pmd = pmd_mkold(pmd_wrprotect(pmd));
1085         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1086
1087         ret = 0;
1088 out_unlock:
1089         spin_unlock(src_ptl);
1090         spin_unlock(dst_ptl);
1091 out:
1092         return ret;
1093 }
1094
1095 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1096 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1097                 pud_t *pud, int flags)
1098 {
1099         pud_t _pud;
1100
1101         _pud = pud_mkyoung(*pud);
1102         if (flags & FOLL_WRITE)
1103                 _pud = pud_mkdirty(_pud);
1104         if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1105                                 pud, _pud, flags & FOLL_WRITE))
1106                 update_mmu_cache_pud(vma, addr, pud);
1107 }
1108
1109 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1110                 pud_t *pud, int flags, struct dev_pagemap **pgmap)
1111 {
1112         unsigned long pfn = pud_pfn(*pud);
1113         struct mm_struct *mm = vma->vm_mm;
1114         struct page *page;
1115
1116         assert_spin_locked(pud_lockptr(mm, pud));
1117
1118         if (flags & FOLL_WRITE && !pud_write(*pud))
1119                 return NULL;
1120
1121         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
1122         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
1123                          (FOLL_PIN | FOLL_GET)))
1124                 return NULL;
1125
1126         if (pud_present(*pud) && pud_devmap(*pud))
1127                 /* pass */;
1128         else
1129                 return NULL;
1130
1131         if (flags & FOLL_TOUCH)
1132                 touch_pud(vma, addr, pud, flags);
1133
1134         /*
1135          * device mapped pages can only be returned if the
1136          * caller will manage the page reference count.
1137          *
1138          * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here:
1139          */
1140         if (!(flags & (FOLL_GET | FOLL_PIN)))
1141                 return ERR_PTR(-EEXIST);
1142
1143         pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1144         *pgmap = get_dev_pagemap(pfn, *pgmap);
1145         if (!*pgmap)
1146                 return ERR_PTR(-EFAULT);
1147         page = pfn_to_page(pfn);
1148         if (!try_grab_page(page, flags))
1149                 page = ERR_PTR(-ENOMEM);
1150
1151         return page;
1152 }
1153
1154 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1155                   pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1156                   struct vm_area_struct *vma)
1157 {
1158         spinlock_t *dst_ptl, *src_ptl;
1159         pud_t pud;
1160         int ret;
1161
1162         dst_ptl = pud_lock(dst_mm, dst_pud);
1163         src_ptl = pud_lockptr(src_mm, src_pud);
1164         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1165
1166         ret = -EAGAIN;
1167         pud = *src_pud;
1168         if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1169                 goto out_unlock;
1170
1171         /*
1172          * When page table lock is held, the huge zero pud should not be
1173          * under splitting since we don't split the page itself, only pud to
1174          * a page table.
1175          */
1176         if (is_huge_zero_pud(pud)) {
1177                 /* No huge zero pud yet */
1178         }
1179
1180         pudp_set_wrprotect(src_mm, addr, src_pud);
1181         pud = pud_mkold(pud_wrprotect(pud));
1182         set_pud_at(dst_mm, addr, dst_pud, pud);
1183
1184         ret = 0;
1185 out_unlock:
1186         spin_unlock(src_ptl);
1187         spin_unlock(dst_ptl);
1188         return ret;
1189 }
1190
1191 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1192 {
1193         pud_t entry;
1194         unsigned long haddr;
1195         bool write = vmf->flags & FAULT_FLAG_WRITE;
1196
1197         vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1198         if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1199                 goto unlock;
1200
1201         entry = pud_mkyoung(orig_pud);
1202         if (write)
1203                 entry = pud_mkdirty(entry);
1204         haddr = vmf->address & HPAGE_PUD_MASK;
1205         if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1206                 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1207
1208 unlock:
1209         spin_unlock(vmf->ptl);
1210 }
1211 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1212
1213 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1214 {
1215         pmd_t entry;
1216         unsigned long haddr;
1217         bool write = vmf->flags & FAULT_FLAG_WRITE;
1218
1219         vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1220         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1221                 goto unlock;
1222
1223         entry = pmd_mkyoung(orig_pmd);
1224         if (write)
1225                 entry = pmd_mkdirty(entry);
1226         haddr = vmf->address & HPAGE_PMD_MASK;
1227         if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1228                 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1229
1230 unlock:
1231         spin_unlock(vmf->ptl);
1232 }
1233
1234 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1235 {
1236         struct vm_area_struct *vma = vmf->vma;
1237         struct page *page;
1238         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1239
1240         vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1241         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1242
1243         if (is_huge_zero_pmd(orig_pmd))
1244                 goto fallback;
1245
1246         spin_lock(vmf->ptl);
1247
1248         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1249                 spin_unlock(vmf->ptl);
1250                 return 0;
1251         }
1252
1253         page = pmd_page(orig_pmd);
1254         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1255
1256         /* Lock page for reuse_swap_page() */
1257         if (!trylock_page(page)) {
1258                 get_page(page);
1259                 spin_unlock(vmf->ptl);
1260                 lock_page(page);
1261                 spin_lock(vmf->ptl);
1262                 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1263                         spin_unlock(vmf->ptl);
1264                         unlock_page(page);
1265                         put_page(page);
1266                         return 0;
1267                 }
1268                 put_page(page);
1269         }
1270
1271         /*
1272          * We can only reuse the page if nobody else maps the huge page or it's
1273          * part.
1274          */
1275         if (reuse_swap_page(page, NULL)) {
1276                 pmd_t entry;
1277                 entry = pmd_mkyoung(orig_pmd);
1278                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1279                 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1280                         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1281                 unlock_page(page);
1282                 spin_unlock(vmf->ptl);
1283                 return VM_FAULT_WRITE;
1284         }
1285
1286         unlock_page(page);
1287         spin_unlock(vmf->ptl);
1288 fallback:
1289         __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL);
1290         return VM_FAULT_FALLBACK;
1291 }
1292
1293 /*
1294  * FOLL_FORCE can write to even unwritable pmd's, but only
1295  * after we've gone through a COW cycle and they are dirty.
1296  */
1297 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1298 {
1299         return pmd_write(pmd) ||
1300                ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1301 }
1302
1303 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1304                                    unsigned long addr,
1305                                    pmd_t *pmd,
1306                                    unsigned int flags)
1307 {
1308         struct mm_struct *mm = vma->vm_mm;
1309         struct page *page = NULL;
1310
1311         assert_spin_locked(pmd_lockptr(mm, pmd));
1312
1313         if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1314                 goto out;
1315
1316         /* Avoid dumping huge zero page */
1317         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1318                 return ERR_PTR(-EFAULT);
1319
1320         /* Full NUMA hinting faults to serialise migration in fault paths */
1321         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1322                 goto out;
1323
1324         page = pmd_page(*pmd);
1325         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1326
1327         if (!try_grab_page(page, flags))
1328                 return ERR_PTR(-ENOMEM);
1329
1330         if (flags & FOLL_TOUCH)
1331                 touch_pmd(vma, addr, pmd, flags);
1332
1333         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1334                 /*
1335                  * We don't mlock() pte-mapped THPs. This way we can avoid
1336                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1337                  *
1338                  * For anon THP:
1339                  *
1340                  * In most cases the pmd is the only mapping of the page as we
1341                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1342                  * writable private mappings in populate_vma_page_range().
1343                  *
1344                  * The only scenario when we have the page shared here is if we
1345                  * mlocking read-only mapping shared over fork(). We skip
1346                  * mlocking such pages.
1347                  *
1348                  * For file THP:
1349                  *
1350                  * We can expect PageDoubleMap() to be stable under page lock:
1351                  * for file pages we set it in page_add_file_rmap(), which
1352                  * requires page to be locked.
1353                  */
1354
1355                 if (PageAnon(page) && compound_mapcount(page) != 1)
1356                         goto skip_mlock;
1357                 if (PageDoubleMap(page) || !page->mapping)
1358                         goto skip_mlock;
1359                 if (!trylock_page(page))
1360                         goto skip_mlock;
1361                 if (page->mapping && !PageDoubleMap(page))
1362                         mlock_vma_page(page);
1363                 unlock_page(page);
1364         }
1365 skip_mlock:
1366         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1367         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1368
1369 out:
1370         return page;
1371 }
1372
1373 /* NUMA hinting page fault entry point for trans huge pmds */
1374 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1375 {
1376         struct vm_area_struct *vma = vmf->vma;
1377         struct anon_vma *anon_vma = NULL;
1378         struct page *page;
1379         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1380         int page_nid = NUMA_NO_NODE, this_nid = numa_node_id();
1381         int target_nid, last_cpupid = -1;
1382         bool page_locked;
1383         bool migrated = false;
1384         bool was_writable;
1385         int flags = 0;
1386
1387         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1388         if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1389                 goto out_unlock;
1390
1391         /*
1392          * If there are potential migrations, wait for completion and retry
1393          * without disrupting NUMA hinting information. Do not relock and
1394          * check_same as the page may no longer be mapped.
1395          */
1396         if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1397                 page = pmd_page(*vmf->pmd);
1398                 if (!get_page_unless_zero(page))
1399                         goto out_unlock;
1400                 spin_unlock(vmf->ptl);
1401                 put_and_wait_on_page_locked(page);
1402                 goto out;
1403         }
1404
1405         page = pmd_page(pmd);
1406         BUG_ON(is_huge_zero_page(page));
1407         page_nid = page_to_nid(page);
1408         last_cpupid = page_cpupid_last(page);
1409         count_vm_numa_event(NUMA_HINT_FAULTS);
1410         if (page_nid == this_nid) {
1411                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1412                 flags |= TNF_FAULT_LOCAL;
1413         }
1414
1415         /* See similar comment in do_numa_page for explanation */
1416         if (!pmd_savedwrite(pmd))
1417                 flags |= TNF_NO_GROUP;
1418
1419         /*
1420          * Acquire the page lock to serialise THP migrations but avoid dropping
1421          * page_table_lock if at all possible
1422          */
1423         page_locked = trylock_page(page);
1424         target_nid = mpol_misplaced(page, vma, haddr);
1425         if (target_nid == NUMA_NO_NODE) {
1426                 /* If the page was locked, there are no parallel migrations */
1427                 if (page_locked)
1428                         goto clear_pmdnuma;
1429         }
1430
1431         /* Migration could have started since the pmd_trans_migrating check */
1432         if (!page_locked) {
1433                 page_nid = NUMA_NO_NODE;
1434                 if (!get_page_unless_zero(page))
1435                         goto out_unlock;
1436                 spin_unlock(vmf->ptl);
1437                 put_and_wait_on_page_locked(page);
1438                 goto out;
1439         }
1440
1441         /*
1442          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1443          * to serialises splits
1444          */
1445         get_page(page);
1446         spin_unlock(vmf->ptl);
1447         anon_vma = page_lock_anon_vma_read(page);
1448
1449         /* Confirm the PMD did not change while page_table_lock was released */
1450         spin_lock(vmf->ptl);
1451         if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1452                 unlock_page(page);
1453                 put_page(page);
1454                 page_nid = NUMA_NO_NODE;
1455                 goto out_unlock;
1456         }
1457
1458         /* Bail if we fail to protect against THP splits for any reason */
1459         if (unlikely(!anon_vma)) {
1460                 put_page(page);
1461                 page_nid = NUMA_NO_NODE;
1462                 goto clear_pmdnuma;
1463         }
1464
1465         /*
1466          * Since we took the NUMA fault, we must have observed the !accessible
1467          * bit. Make sure all other CPUs agree with that, to avoid them
1468          * modifying the page we're about to migrate.
1469          *
1470          * Must be done under PTL such that we'll observe the relevant
1471          * inc_tlb_flush_pending().
1472          *
1473          * We are not sure a pending tlb flush here is for a huge page
1474          * mapping or not. Hence use the tlb range variant
1475          */
1476         if (mm_tlb_flush_pending(vma->vm_mm)) {
1477                 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1478                 /*
1479                  * change_huge_pmd() released the pmd lock before
1480                  * invalidating the secondary MMUs sharing the primary
1481                  * MMU pagetables (with ->invalidate_range()). The
1482                  * mmu_notifier_invalidate_range_end() (which
1483                  * internally calls ->invalidate_range()) in
1484                  * change_pmd_range() will run after us, so we can't
1485                  * rely on it here and we need an explicit invalidate.
1486                  */
1487                 mmu_notifier_invalidate_range(vma->vm_mm, haddr,
1488                                               haddr + HPAGE_PMD_SIZE);
1489         }
1490
1491         /*
1492          * Migrate the THP to the requested node, returns with page unlocked
1493          * and access rights restored.
1494          */
1495         spin_unlock(vmf->ptl);
1496
1497         migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1498                                 vmf->pmd, pmd, vmf->address, page, target_nid);
1499         if (migrated) {
1500                 flags |= TNF_MIGRATED;
1501                 page_nid = target_nid;
1502         } else
1503                 flags |= TNF_MIGRATE_FAIL;
1504
1505         goto out;
1506 clear_pmdnuma:
1507         BUG_ON(!PageLocked(page));
1508         was_writable = pmd_savedwrite(pmd);
1509         pmd = pmd_modify(pmd, vma->vm_page_prot);
1510         pmd = pmd_mkyoung(pmd);
1511         if (was_writable)
1512                 pmd = pmd_mkwrite(pmd);
1513         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1514         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1515         unlock_page(page);
1516 out_unlock:
1517         spin_unlock(vmf->ptl);
1518
1519 out:
1520         if (anon_vma)
1521                 page_unlock_anon_vma_read(anon_vma);
1522
1523         if (page_nid != NUMA_NO_NODE)
1524                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1525                                 flags);
1526
1527         return 0;
1528 }
1529
1530 /*
1531  * Return true if we do MADV_FREE successfully on entire pmd page.
1532  * Otherwise, return false.
1533  */
1534 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1535                 pmd_t *pmd, unsigned long addr, unsigned long next)
1536 {
1537         spinlock_t *ptl;
1538         pmd_t orig_pmd;
1539         struct page *page;
1540         struct mm_struct *mm = tlb->mm;
1541         bool ret = false;
1542
1543         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1544
1545         ptl = pmd_trans_huge_lock(pmd, vma);
1546         if (!ptl)
1547                 goto out_unlocked;
1548
1549         orig_pmd = *pmd;
1550         if (is_huge_zero_pmd(orig_pmd))
1551                 goto out;
1552
1553         if (unlikely(!pmd_present(orig_pmd))) {
1554                 VM_BUG_ON(thp_migration_supported() &&
1555                                   !is_pmd_migration_entry(orig_pmd));
1556                 goto out;
1557         }
1558
1559         page = pmd_page(orig_pmd);
1560         /*
1561          * If other processes are mapping this page, we couldn't discard
1562          * the page unless they all do MADV_FREE so let's skip the page.
1563          */
1564         if (page_mapcount(page) != 1)
1565                 goto out;
1566
1567         if (!trylock_page(page))
1568                 goto out;
1569
1570         /*
1571          * If user want to discard part-pages of THP, split it so MADV_FREE
1572          * will deactivate only them.
1573          */
1574         if (next - addr != HPAGE_PMD_SIZE) {
1575                 get_page(page);
1576                 spin_unlock(ptl);
1577                 split_huge_page(page);
1578                 unlock_page(page);
1579                 put_page(page);
1580                 goto out_unlocked;
1581         }
1582
1583         if (PageDirty(page))
1584                 ClearPageDirty(page);
1585         unlock_page(page);
1586
1587         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1588                 pmdp_invalidate(vma, addr, pmd);
1589                 orig_pmd = pmd_mkold(orig_pmd);
1590                 orig_pmd = pmd_mkclean(orig_pmd);
1591
1592                 set_pmd_at(mm, addr, pmd, orig_pmd);
1593                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1594         }
1595
1596         mark_page_lazyfree(page);
1597         ret = true;
1598 out:
1599         spin_unlock(ptl);
1600 out_unlocked:
1601         return ret;
1602 }
1603
1604 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1605 {
1606         pgtable_t pgtable;
1607
1608         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1609         pte_free(mm, pgtable);
1610         mm_dec_nr_ptes(mm);
1611 }
1612
1613 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1614                  pmd_t *pmd, unsigned long addr)
1615 {
1616         pmd_t orig_pmd;
1617         spinlock_t *ptl;
1618
1619         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1620
1621         ptl = __pmd_trans_huge_lock(pmd, vma);
1622         if (!ptl)
1623                 return 0;
1624         /*
1625          * For architectures like ppc64 we look at deposited pgtable
1626          * when calling pmdp_huge_get_and_clear. So do the
1627          * pgtable_trans_huge_withdraw after finishing pmdp related
1628          * operations.
1629          */
1630         orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd,
1631                                                 tlb->fullmm);
1632         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1633         if (vma_is_special_huge(vma)) {
1634                 if (arch_needs_pgtable_deposit())
1635                         zap_deposited_table(tlb->mm, pmd);
1636                 spin_unlock(ptl);
1637                 if (is_huge_zero_pmd(orig_pmd))
1638                         tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1639         } else if (is_huge_zero_pmd(orig_pmd)) {
1640                 zap_deposited_table(tlb->mm, pmd);
1641                 spin_unlock(ptl);
1642                 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1643         } else {
1644                 struct page *page = NULL;
1645                 int flush_needed = 1;
1646
1647                 if (pmd_present(orig_pmd)) {
1648                         page = pmd_page(orig_pmd);
1649                         page_remove_rmap(page, true);
1650                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1651                         VM_BUG_ON_PAGE(!PageHead(page), page);
1652                 } else if (thp_migration_supported()) {
1653                         swp_entry_t entry;
1654
1655                         VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1656                         entry = pmd_to_swp_entry(orig_pmd);
1657                         page = pfn_to_page(swp_offset(entry));
1658                         flush_needed = 0;
1659                 } else
1660                         WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1661
1662                 if (PageAnon(page)) {
1663                         zap_deposited_table(tlb->mm, pmd);
1664                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1665                 } else {
1666                         if (arch_needs_pgtable_deposit())
1667                                 zap_deposited_table(tlb->mm, pmd);
1668                         add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1669                 }
1670
1671                 spin_unlock(ptl);
1672                 if (flush_needed)
1673                         tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1674         }
1675         return 1;
1676 }
1677
1678 #ifndef pmd_move_must_withdraw
1679 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1680                                          spinlock_t *old_pmd_ptl,
1681                                          struct vm_area_struct *vma)
1682 {
1683         /*
1684          * With split pmd lock we also need to move preallocated
1685          * PTE page table if new_pmd is on different PMD page table.
1686          *
1687          * We also don't deposit and withdraw tables for file pages.
1688          */
1689         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1690 }
1691 #endif
1692
1693 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1694 {
1695 #ifdef CONFIG_MEM_SOFT_DIRTY
1696         if (unlikely(is_pmd_migration_entry(pmd)))
1697                 pmd = pmd_swp_mksoft_dirty(pmd);
1698         else if (pmd_present(pmd))
1699                 pmd = pmd_mksoft_dirty(pmd);
1700 #endif
1701         return pmd;
1702 }
1703
1704 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1705                   unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
1706 {
1707         spinlock_t *old_ptl, *new_ptl;
1708         pmd_t pmd;
1709         struct mm_struct *mm = vma->vm_mm;
1710         bool force_flush = false;
1711
1712         /*
1713          * The destination pmd shouldn't be established, free_pgtables()
1714          * should have release it.
1715          */
1716         if (WARN_ON(!pmd_none(*new_pmd))) {
1717                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1718                 return false;
1719         }
1720
1721         /*
1722          * We don't have to worry about the ordering of src and dst
1723          * ptlocks because exclusive mmap_lock prevents deadlock.
1724          */
1725         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1726         if (old_ptl) {
1727                 new_ptl = pmd_lockptr(mm, new_pmd);
1728                 if (new_ptl != old_ptl)
1729                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1730                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1731                 if (pmd_present(pmd))
1732                         force_flush = true;
1733                 VM_BUG_ON(!pmd_none(*new_pmd));
1734
1735                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1736                         pgtable_t pgtable;
1737                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1738                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1739                 }
1740                 pmd = move_soft_dirty_pmd(pmd);
1741                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1742                 if (force_flush)
1743                         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1744                 if (new_ptl != old_ptl)
1745                         spin_unlock(new_ptl);
1746                 spin_unlock(old_ptl);
1747                 return true;
1748         }
1749         return false;
1750 }
1751
1752 /*
1753  * Returns
1754  *  - 0 if PMD could not be locked
1755  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1756  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1757  */
1758 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1759                 unsigned long addr, pgprot_t newprot, unsigned long cp_flags)
1760 {
1761         struct mm_struct *mm = vma->vm_mm;
1762         spinlock_t *ptl;
1763         pmd_t entry;
1764         bool preserve_write;
1765         int ret;
1766         bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
1767         bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
1768         bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
1769
1770         ptl = __pmd_trans_huge_lock(pmd, vma);
1771         if (!ptl)
1772                 return 0;
1773
1774         preserve_write = prot_numa && pmd_write(*pmd);
1775         ret = 1;
1776
1777 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1778         if (is_swap_pmd(*pmd)) {
1779                 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1780
1781                 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1782                 if (is_write_migration_entry(entry)) {
1783                         pmd_t newpmd;
1784                         /*
1785                          * A protection check is difficult so
1786                          * just be safe and disable write
1787                          */
1788                         make_migration_entry_read(&entry);
1789                         newpmd = swp_entry_to_pmd(entry);
1790                         if (pmd_swp_soft_dirty(*pmd))
1791                                 newpmd = pmd_swp_mksoft_dirty(newpmd);
1792                         set_pmd_at(mm, addr, pmd, newpmd);
1793                 }
1794                 goto unlock;
1795         }
1796 #endif
1797
1798         /*
1799          * Avoid trapping faults against the zero page. The read-only
1800          * data is likely to be read-cached on the local CPU and
1801          * local/remote hits to the zero page are not interesting.
1802          */
1803         if (prot_numa && is_huge_zero_pmd(*pmd))
1804                 goto unlock;
1805
1806         if (prot_numa && pmd_protnone(*pmd))
1807                 goto unlock;
1808
1809         /*
1810          * In case prot_numa, we are under mmap_read_lock(mm). It's critical
1811          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1812          * which is also under mmap_read_lock(mm):
1813          *
1814          *      CPU0:                           CPU1:
1815          *                              change_huge_pmd(prot_numa=1)
1816          *                               pmdp_huge_get_and_clear_notify()
1817          * madvise_dontneed()
1818          *  zap_pmd_range()
1819          *   pmd_trans_huge(*pmd) == 0 (without ptl)
1820          *   // skip the pmd
1821          *                               set_pmd_at();
1822          *                               // pmd is re-established
1823          *
1824          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1825          * which may break userspace.
1826          *
1827          * pmdp_invalidate() is required to make sure we don't miss
1828          * dirty/young flags set by hardware.
1829          */
1830         entry = pmdp_invalidate(vma, addr, pmd);
1831
1832         entry = pmd_modify(entry, newprot);
1833         if (preserve_write)
1834                 entry = pmd_mk_savedwrite(entry);
1835         if (uffd_wp) {
1836                 entry = pmd_wrprotect(entry);
1837                 entry = pmd_mkuffd_wp(entry);
1838         } else if (uffd_wp_resolve) {
1839                 /*
1840                  * Leave the write bit to be handled by PF interrupt
1841                  * handler, then things like COW could be properly
1842                  * handled.
1843                  */
1844                 entry = pmd_clear_uffd_wp(entry);
1845         }
1846         ret = HPAGE_PMD_NR;
1847         set_pmd_at(mm, addr, pmd, entry);
1848         BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1849 unlock:
1850         spin_unlock(ptl);
1851         return ret;
1852 }
1853
1854 /*
1855  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1856  *
1857  * Note that if it returns page table lock pointer, this routine returns without
1858  * unlocking page table lock. So callers must unlock it.
1859  */
1860 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1861 {
1862         spinlock_t *ptl;
1863         ptl = pmd_lock(vma->vm_mm, pmd);
1864         if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1865                         pmd_devmap(*pmd)))
1866                 return ptl;
1867         spin_unlock(ptl);
1868         return NULL;
1869 }
1870
1871 /*
1872  * Returns true if a given pud maps a thp, false otherwise.
1873  *
1874  * Note that if it returns true, this routine returns without unlocking page
1875  * table lock. So callers must unlock it.
1876  */
1877 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1878 {
1879         spinlock_t *ptl;
1880
1881         ptl = pud_lock(vma->vm_mm, pud);
1882         if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1883                 return ptl;
1884         spin_unlock(ptl);
1885         return NULL;
1886 }
1887
1888 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1889 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1890                  pud_t *pud, unsigned long addr)
1891 {
1892         spinlock_t *ptl;
1893
1894         ptl = __pud_trans_huge_lock(pud, vma);
1895         if (!ptl)
1896                 return 0;
1897         /*
1898          * For architectures like ppc64 we look at deposited pgtable
1899          * when calling pudp_huge_get_and_clear. So do the
1900          * pgtable_trans_huge_withdraw after finishing pudp related
1901          * operations.
1902          */
1903         pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
1904         tlb_remove_pud_tlb_entry(tlb, pud, addr);
1905         if (vma_is_special_huge(vma)) {
1906                 spin_unlock(ptl);
1907                 /* No zero page support yet */
1908         } else {
1909                 /* No support for anonymous PUD pages yet */
1910                 BUG();
1911         }
1912         return 1;
1913 }
1914
1915 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1916                 unsigned long haddr)
1917 {
1918         VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1919         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1920         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
1921         VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
1922
1923         count_vm_event(THP_SPLIT_PUD);
1924
1925         pudp_huge_clear_flush_notify(vma, haddr, pud);
1926 }
1927
1928 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
1929                 unsigned long address)
1930 {
1931         spinlock_t *ptl;
1932         struct mmu_notifier_range range;
1933
1934         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1935                                 address & HPAGE_PUD_MASK,
1936                                 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
1937         mmu_notifier_invalidate_range_start(&range);
1938         ptl = pud_lock(vma->vm_mm, pud);
1939         if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
1940                 goto out;
1941         __split_huge_pud_locked(vma, pud, range.start);
1942
1943 out:
1944         spin_unlock(ptl);
1945         /*
1946          * No need to double call mmu_notifier->invalidate_range() callback as
1947          * the above pudp_huge_clear_flush_notify() did already call it.
1948          */
1949         mmu_notifier_invalidate_range_only_end(&range);
1950 }
1951 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1952
1953 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1954                 unsigned long haddr, pmd_t *pmd)
1955 {
1956         struct mm_struct *mm = vma->vm_mm;
1957         pgtable_t pgtable;
1958         pmd_t _pmd;
1959         int i;
1960
1961         /*
1962          * Leave pmd empty until pte is filled note that it is fine to delay
1963          * notification until mmu_notifier_invalidate_range_end() as we are
1964          * replacing a zero pmd write protected page with a zero pte write
1965          * protected page.
1966          *
1967          * See Documentation/vm/mmu_notifier.rst
1968          */
1969         pmdp_huge_clear_flush(vma, haddr, pmd);
1970
1971         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1972         pmd_populate(mm, &_pmd, pgtable);
1973
1974         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1975                 pte_t *pte, entry;
1976                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
1977                 entry = pte_mkspecial(entry);
1978                 pte = pte_offset_map(&_pmd, haddr);
1979                 VM_BUG_ON(!pte_none(*pte));
1980                 set_pte_at(mm, haddr, pte, entry);
1981                 pte_unmap(pte);
1982         }
1983         smp_wmb(); /* make pte visible before pmd */
1984         pmd_populate(mm, pmd, pgtable);
1985 }
1986
1987 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
1988                 unsigned long haddr, bool freeze)
1989 {
1990         struct mm_struct *mm = vma->vm_mm;
1991         struct page *page;
1992         pgtable_t pgtable;
1993         pmd_t old_pmd, _pmd;
1994         bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false;
1995         unsigned long addr;
1996         int i;
1997
1998         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
1999         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2000         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2001         VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2002                                 && !pmd_devmap(*pmd));
2003
2004         count_vm_event(THP_SPLIT_PMD);
2005
2006         if (!vma_is_anonymous(vma)) {
2007                 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2008                 /*
2009                  * We are going to unmap this huge page. So
2010                  * just go ahead and zap it
2011                  */
2012                 if (arch_needs_pgtable_deposit())
2013                         zap_deposited_table(mm, pmd);
2014                 if (vma_is_special_huge(vma))
2015                         return;
2016                 page = pmd_page(_pmd);
2017                 if (!PageDirty(page) && pmd_dirty(_pmd))
2018                         set_page_dirty(page);
2019                 if (!PageReferenced(page) && pmd_young(_pmd))
2020                         SetPageReferenced(page);
2021                 page_remove_rmap(page, true);
2022                 put_page(page);
2023                 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2024                 return;
2025         } else if (is_huge_zero_pmd(*pmd)) {
2026                 /*
2027                  * FIXME: Do we want to invalidate secondary mmu by calling
2028                  * mmu_notifier_invalidate_range() see comments below inside
2029                  * __split_huge_pmd() ?
2030                  *
2031                  * We are going from a zero huge page write protected to zero
2032                  * small page also write protected so it does not seems useful
2033                  * to invalidate secondary mmu at this time.
2034                  */
2035                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2036         }
2037
2038         /*
2039          * Up to this point the pmd is present and huge and userland has the
2040          * whole access to the hugepage during the split (which happens in
2041          * place). If we overwrite the pmd with the not-huge version pointing
2042          * to the pte here (which of course we could if all CPUs were bug
2043          * free), userland could trigger a small page size TLB miss on the
2044          * small sized TLB while the hugepage TLB entry is still established in
2045          * the huge TLB. Some CPU doesn't like that.
2046          * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum
2047          * 383 on page 105. Intel should be safe but is also warns that it's
2048          * only safe if the permission and cache attributes of the two entries
2049          * loaded in the two TLB is identical (which should be the case here).
2050          * But it is generally safer to never allow small and huge TLB entries
2051          * for the same virtual address to be loaded simultaneously. So instead
2052          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2053          * current pmd notpresent (atomically because here the pmd_trans_huge
2054          * must remain set at all times on the pmd until the split is complete
2055          * for this pmd), then we flush the SMP TLB and finally we write the
2056          * non-huge version of the pmd entry with pmd_populate.
2057          */
2058         old_pmd = pmdp_invalidate(vma, haddr, pmd);
2059
2060         pmd_migration = is_pmd_migration_entry(old_pmd);
2061         if (unlikely(pmd_migration)) {
2062                 swp_entry_t entry;
2063
2064                 entry = pmd_to_swp_entry(old_pmd);
2065                 page = pfn_to_page(swp_offset(entry));
2066                 write = is_write_migration_entry(entry);
2067                 young = false;
2068                 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2069                 uffd_wp = pmd_swp_uffd_wp(old_pmd);
2070         } else {
2071                 page = pmd_page(old_pmd);
2072                 if (pmd_dirty(old_pmd))
2073                         SetPageDirty(page);
2074                 write = pmd_write(old_pmd);
2075                 young = pmd_young(old_pmd);
2076                 soft_dirty = pmd_soft_dirty(old_pmd);
2077                 uffd_wp = pmd_uffd_wp(old_pmd);
2078         }
2079         VM_BUG_ON_PAGE(!page_count(page), page);
2080         page_ref_add(page, HPAGE_PMD_NR - 1);
2081
2082         /*
2083          * Withdraw the table only after we mark the pmd entry invalid.
2084          * This's critical for some architectures (Power).
2085          */
2086         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2087         pmd_populate(mm, &_pmd, pgtable);
2088
2089         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2090                 pte_t entry, *pte;
2091                 /*
2092                  * Note that NUMA hinting access restrictions are not
2093                  * transferred to avoid any possibility of altering
2094                  * permissions across VMAs.
2095                  */
2096                 if (freeze || pmd_migration) {
2097                         swp_entry_t swp_entry;
2098                         swp_entry = make_migration_entry(page + i, write);
2099                         entry = swp_entry_to_pte(swp_entry);
2100                         if (soft_dirty)
2101                                 entry = pte_swp_mksoft_dirty(entry);
2102                         if (uffd_wp)
2103                                 entry = pte_swp_mkuffd_wp(entry);
2104                 } else {
2105                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2106                         entry = maybe_mkwrite(entry, vma);
2107                         if (!write)
2108                                 entry = pte_wrprotect(entry);
2109                         if (!young)
2110                                 entry = pte_mkold(entry);
2111                         if (soft_dirty)
2112                                 entry = pte_mksoft_dirty(entry);
2113                         if (uffd_wp)
2114                                 entry = pte_mkuffd_wp(entry);
2115                 }
2116                 pte = pte_offset_map(&_pmd, addr);
2117                 BUG_ON(!pte_none(*pte));
2118                 set_pte_at(mm, addr, pte, entry);
2119                 atomic_inc(&page[i]._mapcount);
2120                 pte_unmap(pte);
2121         }
2122
2123         /*
2124          * Set PG_double_map before dropping compound_mapcount to avoid
2125          * false-negative page_mapped().
2126          */
2127         if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2128                 for (i = 0; i < HPAGE_PMD_NR; i++)
2129                         atomic_inc(&page[i]._mapcount);
2130         }
2131
2132         lock_page_memcg(page);
2133         if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2134                 /* Last compound_mapcount is gone. */
2135                 __dec_lruvec_page_state(page, NR_ANON_THPS);
2136                 if (TestClearPageDoubleMap(page)) {
2137                         /* No need in mapcount reference anymore */
2138                         for (i = 0; i < HPAGE_PMD_NR; i++)
2139                                 atomic_dec(&page[i]._mapcount);
2140                 }
2141         }
2142         unlock_page_memcg(page);
2143
2144         smp_wmb(); /* make pte visible before pmd */
2145         pmd_populate(mm, pmd, pgtable);
2146
2147         if (freeze) {
2148                 for (i = 0; i < HPAGE_PMD_NR; i++) {
2149                         page_remove_rmap(page + i, false);
2150                         put_page(page + i);
2151                 }
2152         }
2153 }
2154
2155 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2156                 unsigned long address, bool freeze, struct page *page)
2157 {
2158         spinlock_t *ptl;
2159         struct mmu_notifier_range range;
2160         bool was_locked = false;
2161         pmd_t _pmd;
2162
2163         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2164                                 address & HPAGE_PMD_MASK,
2165                                 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2166         mmu_notifier_invalidate_range_start(&range);
2167         ptl = pmd_lock(vma->vm_mm, pmd);
2168
2169         /*
2170          * If caller asks to setup a migration entries, we need a page to check
2171          * pmd against. Otherwise we can end up replacing wrong page.
2172          */
2173         VM_BUG_ON(freeze && !page);
2174         if (page) {
2175                 VM_WARN_ON_ONCE(!PageLocked(page));
2176                 was_locked = true;
2177                 if (page != pmd_page(*pmd))
2178                         goto out;
2179         }
2180
2181 repeat:
2182         if (pmd_trans_huge(*pmd)) {
2183                 if (!page) {
2184                         page = pmd_page(*pmd);
2185                         if (unlikely(!trylock_page(page))) {
2186                                 get_page(page);
2187                                 _pmd = *pmd;
2188                                 spin_unlock(ptl);
2189                                 lock_page(page);
2190                                 spin_lock(ptl);
2191                                 if (unlikely(!pmd_same(*pmd, _pmd))) {
2192                                         unlock_page(page);
2193                                         put_page(page);
2194                                         page = NULL;
2195                                         goto repeat;
2196                                 }
2197                                 put_page(page);
2198                         }
2199                 }
2200                 if (PageMlocked(page))
2201                         clear_page_mlock(page);
2202         } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2203                 goto out;
2204         __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2205 out:
2206         spin_unlock(ptl);
2207         if (!was_locked && page)
2208                 unlock_page(page);
2209         /*
2210          * No need to double call mmu_notifier->invalidate_range() callback.
2211          * They are 3 cases to consider inside __split_huge_pmd_locked():
2212          *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2213          *  2) __split_huge_zero_page_pmd() read only zero page and any write
2214          *    fault will trigger a flush_notify before pointing to a new page
2215          *    (it is fine if the secondary mmu keeps pointing to the old zero
2216          *    page in the meantime)
2217          *  3) Split a huge pmd into pte pointing to the same page. No need
2218          *     to invalidate secondary tlb entry they are all still valid.
2219          *     any further changes to individual pte will notify. So no need
2220          *     to call mmu_notifier->invalidate_range()
2221          */
2222         mmu_notifier_invalidate_range_only_end(&range);
2223 }
2224
2225 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2226                 bool freeze, struct page *page)
2227 {
2228         pgd_t *pgd;
2229         p4d_t *p4d;
2230         pud_t *pud;
2231         pmd_t *pmd;
2232
2233         pgd = pgd_offset(vma->vm_mm, address);
2234         if (!pgd_present(*pgd))
2235                 return;
2236
2237         p4d = p4d_offset(pgd, address);
2238         if (!p4d_present(*p4d))
2239                 return;
2240
2241         pud = pud_offset(p4d, address);
2242         if (!pud_present(*pud))
2243                 return;
2244
2245         pmd = pmd_offset(pud, address);
2246
2247         __split_huge_pmd(vma, pmd, address, freeze, page);
2248 }
2249
2250 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2251                              unsigned long start,
2252                              unsigned long end,
2253                              long adjust_next)
2254 {
2255         /*
2256          * If the new start address isn't hpage aligned and it could
2257          * previously contain an hugepage: check if we need to split
2258          * an huge pmd.
2259          */
2260         if (start & ~HPAGE_PMD_MASK &&
2261             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2262             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2263                 split_huge_pmd_address(vma, start, false, NULL);
2264
2265         /*
2266          * If the new end address isn't hpage aligned and it could
2267          * previously contain an hugepage: check if we need to split
2268          * an huge pmd.
2269          */
2270         if (end & ~HPAGE_PMD_MASK &&
2271             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2272             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2273                 split_huge_pmd_address(vma, end, false, NULL);
2274
2275         /*
2276          * If we're also updating the vma->vm_next->vm_start, if the new
2277          * vm_next->vm_start isn't page aligned and it could previously
2278          * contain an hugepage: check if we need to split an huge pmd.
2279          */
2280         if (adjust_next > 0) {
2281                 struct vm_area_struct *next = vma->vm_next;
2282                 unsigned long nstart = next->vm_start;
2283                 nstart += adjust_next << PAGE_SHIFT;
2284                 if (nstart & ~HPAGE_PMD_MASK &&
2285                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2286                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2287                         split_huge_pmd_address(next, nstart, false, NULL);
2288         }
2289 }
2290
2291 static void unmap_page(struct page *page)
2292 {
2293         enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
2294                 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2295         bool unmap_success;
2296
2297         VM_BUG_ON_PAGE(!PageHead(page), page);
2298
2299         if (PageAnon(page))
2300                 ttu_flags |= TTU_SPLIT_FREEZE;
2301
2302         unmap_success = try_to_unmap(page, ttu_flags);
2303         VM_BUG_ON_PAGE(!unmap_success, page);
2304 }
2305
2306 static void remap_page(struct page *page)
2307 {
2308         int i;
2309         if (PageTransHuge(page)) {
2310                 remove_migration_ptes(page, page, true);
2311         } else {
2312                 for (i = 0; i < HPAGE_PMD_NR; i++)
2313                         remove_migration_ptes(page + i, page + i, true);
2314         }
2315 }
2316
2317 static void __split_huge_page_tail(struct page *head, int tail,
2318                 struct lruvec *lruvec, struct list_head *list)
2319 {
2320         struct page *page_tail = head + tail;
2321
2322         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2323
2324         /*
2325          * Clone page flags before unfreezing refcount.
2326          *
2327          * After successful get_page_unless_zero() might follow flags change,
2328          * for exmaple lock_page() which set PG_waiters.
2329          */
2330         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2331         page_tail->flags |= (head->flags &
2332                         ((1L << PG_referenced) |
2333                          (1L << PG_swapbacked) |
2334                          (1L << PG_swapcache) |
2335                          (1L << PG_mlocked) |
2336                          (1L << PG_uptodate) |
2337                          (1L << PG_active) |
2338                          (1L << PG_workingset) |
2339                          (1L << PG_locked) |
2340                          (1L << PG_unevictable) |
2341                          (1L << PG_dirty)));
2342
2343         /* ->mapping in first tail page is compound_mapcount */
2344         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2345                         page_tail);
2346         page_tail->mapping = head->mapping;
2347         page_tail->index = head->index + tail;
2348
2349         /* Page flags must be visible before we make the page non-compound. */
2350         smp_wmb();
2351
2352         /*
2353          * Clear PageTail before unfreezing page refcount.
2354          *
2355          * After successful get_page_unless_zero() might follow put_page()
2356          * which needs correct compound_head().
2357          */
2358         clear_compound_head(page_tail);
2359
2360         /* Finally unfreeze refcount. Additional reference from page cache. */
2361         page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2362                                           PageSwapCache(head)));
2363
2364         if (page_is_young(head))
2365                 set_page_young(page_tail);
2366         if (page_is_idle(head))
2367                 set_page_idle(page_tail);
2368
2369         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2370
2371         /*
2372          * always add to the tail because some iterators expect new
2373          * pages to show after the currently processed elements - e.g.
2374          * migrate_pages
2375          */
2376         lru_add_page_tail(head, page_tail, lruvec, list);
2377 }
2378
2379 static void __split_huge_page(struct page *page, struct list_head *list,
2380                 pgoff_t end, unsigned long flags)
2381 {
2382         struct page *head = compound_head(page);
2383         pg_data_t *pgdat = page_pgdat(head);
2384         struct lruvec *lruvec;
2385         struct address_space *swap_cache = NULL;
2386         unsigned long offset = 0;
2387         int i;
2388
2389         lruvec = mem_cgroup_page_lruvec(head, pgdat);
2390
2391         /* complete memcg works before add pages to LRU */
2392         mem_cgroup_split_huge_fixup(head);
2393
2394         if (PageAnon(head) && PageSwapCache(head)) {
2395                 swp_entry_t entry = { .val = page_private(head) };
2396
2397                 offset = swp_offset(entry);
2398                 swap_cache = swap_address_space(entry);
2399                 xa_lock(&swap_cache->i_pages);
2400         }
2401
2402         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
2403                 __split_huge_page_tail(head, i, lruvec, list);
2404                 /* Some pages can be beyond i_size: drop them from page cache */
2405                 if (head[i].index >= end) {
2406                         ClearPageDirty(head + i);
2407                         __delete_from_page_cache(head + i, NULL);
2408                         if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2409                                 shmem_uncharge(head->mapping->host, 1);
2410                         put_page(head + i);
2411                 } else if (!PageAnon(page)) {
2412                         __xa_store(&head->mapping->i_pages, head[i].index,
2413                                         head + i, 0);
2414                 } else if (swap_cache) {
2415                         __xa_store(&swap_cache->i_pages, offset + i,
2416                                         head + i, 0);
2417                 }
2418         }
2419
2420         ClearPageCompound(head);
2421
2422         split_page_owner(head, HPAGE_PMD_ORDER);
2423
2424         /* See comment in __split_huge_page_tail() */
2425         if (PageAnon(head)) {
2426                 /* Additional pin to swap cache */
2427                 if (PageSwapCache(head)) {
2428                         page_ref_add(head, 2);
2429                         xa_unlock(&swap_cache->i_pages);
2430                 } else {
2431                         page_ref_inc(head);
2432                 }
2433         } else {
2434                 /* Additional pin to page cache */
2435                 page_ref_add(head, 2);
2436                 xa_unlock(&head->mapping->i_pages);
2437         }
2438
2439         spin_unlock_irqrestore(&pgdat->lru_lock, flags);
2440
2441         remap_page(head);
2442
2443         for (i = 0; i < HPAGE_PMD_NR; i++) {
2444                 struct page *subpage = head + i;
2445                 if (subpage == page)
2446                         continue;
2447                 unlock_page(subpage);
2448
2449                 /*
2450                  * Subpages may be freed if there wasn't any mapping
2451                  * like if add_to_swap() is running on a lru page that
2452                  * had its mapping zapped. And freeing these pages
2453                  * requires taking the lru_lock so we do the put_page
2454                  * of the tail pages after the split is complete.
2455                  */
2456                 put_page(subpage);
2457         }
2458 }
2459
2460 int total_mapcount(struct page *page)
2461 {
2462         int i, compound, ret;
2463
2464         VM_BUG_ON_PAGE(PageTail(page), page);
2465
2466         if (likely(!PageCompound(page)))
2467                 return atomic_read(&page->_mapcount) + 1;
2468
2469         compound = compound_mapcount(page);
2470         if (PageHuge(page))
2471                 return compound;
2472         ret = compound;
2473         for (i = 0; i < HPAGE_PMD_NR; i++)
2474                 ret += atomic_read(&page[i]._mapcount) + 1;
2475         /* File pages has compound_mapcount included in _mapcount */
2476         if (!PageAnon(page))
2477                 return ret - compound * HPAGE_PMD_NR;
2478         if (PageDoubleMap(page))
2479                 ret -= HPAGE_PMD_NR;
2480         return ret;
2481 }
2482
2483 /*
2484  * This calculates accurately how many mappings a transparent hugepage
2485  * has (unlike page_mapcount() which isn't fully accurate). This full
2486  * accuracy is primarily needed to know if copy-on-write faults can
2487  * reuse the page and change the mapping to read-write instead of
2488  * copying them. At the same time this returns the total_mapcount too.
2489  *
2490  * The function returns the highest mapcount any one of the subpages
2491  * has. If the return value is one, even if different processes are
2492  * mapping different subpages of the transparent hugepage, they can
2493  * all reuse it, because each process is reusing a different subpage.
2494  *
2495  * The total_mapcount is instead counting all virtual mappings of the
2496  * subpages. If the total_mapcount is equal to "one", it tells the
2497  * caller all mappings belong to the same "mm" and in turn the
2498  * anon_vma of the transparent hugepage can become the vma->anon_vma
2499  * local one as no other process may be mapping any of the subpages.
2500  *
2501  * It would be more accurate to replace page_mapcount() with
2502  * page_trans_huge_mapcount(), however we only use
2503  * page_trans_huge_mapcount() in the copy-on-write faults where we
2504  * need full accuracy to avoid breaking page pinning, because
2505  * page_trans_huge_mapcount() is slower than page_mapcount().
2506  */
2507 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2508 {
2509         int i, ret, _total_mapcount, mapcount;
2510
2511         /* hugetlbfs shouldn't call it */
2512         VM_BUG_ON_PAGE(PageHuge(page), page);
2513
2514         if (likely(!PageTransCompound(page))) {
2515                 mapcount = atomic_read(&page->_mapcount) + 1;
2516                 if (total_mapcount)
2517                         *total_mapcount = mapcount;
2518                 return mapcount;
2519         }
2520
2521         page = compound_head(page);
2522
2523         _total_mapcount = ret = 0;
2524         for (i = 0; i < HPAGE_PMD_NR; i++) {
2525                 mapcount = atomic_read(&page[i]._mapcount) + 1;
2526                 ret = max(ret, mapcount);
2527                 _total_mapcount += mapcount;
2528         }
2529         if (PageDoubleMap(page)) {
2530                 ret -= 1;
2531                 _total_mapcount -= HPAGE_PMD_NR;
2532         }
2533         mapcount = compound_mapcount(page);
2534         ret += mapcount;
2535         _total_mapcount += mapcount;
2536         if (total_mapcount)
2537                 *total_mapcount = _total_mapcount;
2538         return ret;
2539 }
2540
2541 /* Racy check whether the huge page can be split */
2542 bool can_split_huge_page(struct page *page, int *pextra_pins)
2543 {
2544         int extra_pins;
2545
2546         /* Additional pins from page cache */
2547         if (PageAnon(page))
2548                 extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0;
2549         else
2550                 extra_pins = HPAGE_PMD_NR;
2551         if (pextra_pins)
2552                 *pextra_pins = extra_pins;
2553         return total_mapcount(page) == page_count(page) - extra_pins - 1;
2554 }
2555
2556 /*
2557  * This function splits huge page into normal pages. @page can point to any
2558  * subpage of huge page to split. Split doesn't change the position of @page.
2559  *
2560  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2561  * The huge page must be locked.
2562  *
2563  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2564  *
2565  * Both head page and tail pages will inherit mapping, flags, and so on from
2566  * the hugepage.
2567  *
2568  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2569  * they are not mapped.
2570  *
2571  * Returns 0 if the hugepage is split successfully.
2572  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2573  * us.
2574  */
2575 int split_huge_page_to_list(struct page *page, struct list_head *list)
2576 {
2577         struct page *head = compound_head(page);
2578         struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2579         struct deferred_split *ds_queue = get_deferred_split_queue(head);
2580         struct anon_vma *anon_vma = NULL;
2581         struct address_space *mapping = NULL;
2582         int count, mapcount, extra_pins, ret;
2583         unsigned long flags;
2584         pgoff_t end;
2585
2586         VM_BUG_ON_PAGE(is_huge_zero_page(head), head);
2587         VM_BUG_ON_PAGE(!PageLocked(head), head);
2588         VM_BUG_ON_PAGE(!PageCompound(head), head);
2589
2590         if (PageWriteback(head))
2591                 return -EBUSY;
2592
2593         if (PageAnon(head)) {
2594                 /*
2595                  * The caller does not necessarily hold an mmap_lock that would
2596                  * prevent the anon_vma disappearing so we first we take a
2597                  * reference to it and then lock the anon_vma for write. This
2598                  * is similar to page_lock_anon_vma_read except the write lock
2599                  * is taken to serialise against parallel split or collapse
2600                  * operations.
2601                  */
2602                 anon_vma = page_get_anon_vma(head);
2603                 if (!anon_vma) {
2604                         ret = -EBUSY;
2605                         goto out;
2606                 }
2607                 end = -1;
2608                 mapping = NULL;
2609                 anon_vma_lock_write(anon_vma);
2610         } else {
2611                 mapping = head->mapping;
2612
2613                 /* Truncated ? */
2614                 if (!mapping) {
2615                         ret = -EBUSY;
2616                         goto out;
2617                 }
2618
2619                 anon_vma = NULL;
2620                 i_mmap_lock_read(mapping);
2621
2622                 /*
2623                  *__split_huge_page() may need to trim off pages beyond EOF:
2624                  * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2625                  * which cannot be nested inside the page tree lock. So note
2626                  * end now: i_size itself may be changed at any moment, but
2627                  * head page lock is good enough to serialize the trimming.
2628                  */
2629                 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2630         }
2631
2632         /*
2633          * Racy check if we can split the page, before unmap_page() will
2634          * split PMDs
2635          */
2636         if (!can_split_huge_page(head, &extra_pins)) {
2637                 ret = -EBUSY;
2638                 goto out_unlock;
2639         }
2640
2641         unmap_page(head);
2642         VM_BUG_ON_PAGE(compound_mapcount(head), head);
2643
2644         /* prevent PageLRU to go away from under us, and freeze lru stats */
2645         spin_lock_irqsave(&pgdata->lru_lock, flags);
2646
2647         if (mapping) {
2648                 XA_STATE(xas, &mapping->i_pages, page_index(head));
2649
2650                 /*
2651                  * Check if the head page is present in page cache.
2652                  * We assume all tail are present too, if head is there.
2653                  */
2654                 xa_lock(&mapping->i_pages);
2655                 if (xas_load(&xas) != head)
2656                         goto fail;
2657         }
2658
2659         /* Prevent deferred_split_scan() touching ->_refcount */
2660         spin_lock(&ds_queue->split_queue_lock);
2661         count = page_count(head);
2662         mapcount = total_mapcount(head);
2663         if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2664                 if (!list_empty(page_deferred_list(head))) {
2665                         ds_queue->split_queue_len--;
2666                         list_del(page_deferred_list(head));
2667                 }
2668                 spin_unlock(&ds_queue->split_queue_lock);
2669                 if (mapping) {
2670                         if (PageSwapBacked(head))
2671                                 __dec_node_page_state(head, NR_SHMEM_THPS);
2672                         else
2673                                 __dec_node_page_state(head, NR_FILE_THPS);
2674                 }
2675
2676                 __split_huge_page(page, list, end, flags);
2677                 if (PageSwapCache(head)) {
2678                         swp_entry_t entry = { .val = page_private(head) };
2679
2680                         ret = split_swap_cluster(entry);
2681                 } else
2682                         ret = 0;
2683         } else {
2684                 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2685                         pr_alert("total_mapcount: %u, page_count(): %u\n",
2686                                         mapcount, count);
2687                         if (PageTail(page))
2688                                 dump_page(head, NULL);
2689                         dump_page(page, "total_mapcount(head) > 0");
2690                         BUG();
2691                 }
2692                 spin_unlock(&ds_queue->split_queue_lock);
2693 fail:           if (mapping)
2694                         xa_unlock(&mapping->i_pages);
2695                 spin_unlock_irqrestore(&pgdata->lru_lock, flags);
2696                 remap_page(head);
2697                 ret = -EBUSY;
2698         }
2699
2700 out_unlock:
2701         if (anon_vma) {
2702                 anon_vma_unlock_write(anon_vma);
2703                 put_anon_vma(anon_vma);
2704         }
2705         if (mapping)
2706                 i_mmap_unlock_read(mapping);
2707 out:
2708         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2709         return ret;
2710 }
2711
2712 void free_transhuge_page(struct page *page)
2713 {
2714         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2715         unsigned long flags;
2716
2717         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2718         if (!list_empty(page_deferred_list(page))) {
2719                 ds_queue->split_queue_len--;
2720                 list_del(page_deferred_list(page));
2721         }
2722         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2723         free_compound_page(page);
2724 }
2725
2726 void deferred_split_huge_page(struct page *page)
2727 {
2728         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2729 #ifdef CONFIG_MEMCG
2730         struct mem_cgroup *memcg = compound_head(page)->mem_cgroup;
2731 #endif
2732         unsigned long flags;
2733
2734         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2735
2736         /*
2737          * The try_to_unmap() in page reclaim path might reach here too,
2738          * this may cause a race condition to corrupt deferred split queue.
2739          * And, if page reclaim is already handling the same page, it is
2740          * unnecessary to handle it again in shrinker.
2741          *
2742          * Check PageSwapCache to determine if the page is being
2743          * handled by page reclaim since THP swap would add the page into
2744          * swap cache before calling try_to_unmap().
2745          */
2746         if (PageSwapCache(page))
2747                 return;
2748
2749         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2750         if (list_empty(page_deferred_list(page))) {
2751                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2752                 list_add_tail(page_deferred_list(page), &ds_queue->split_queue);
2753                 ds_queue->split_queue_len++;
2754 #ifdef CONFIG_MEMCG
2755                 if (memcg)
2756                         memcg_set_shrinker_bit(memcg, page_to_nid(page),
2757                                                deferred_split_shrinker.id);
2758 #endif
2759         }
2760         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2761 }
2762
2763 static unsigned long deferred_split_count(struct shrinker *shrink,
2764                 struct shrink_control *sc)
2765 {
2766         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2767         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2768
2769 #ifdef CONFIG_MEMCG
2770         if (sc->memcg)
2771                 ds_queue = &sc->memcg->deferred_split_queue;
2772 #endif
2773         return READ_ONCE(ds_queue->split_queue_len);
2774 }
2775
2776 static unsigned long deferred_split_scan(struct shrinker *shrink,
2777                 struct shrink_control *sc)
2778 {
2779         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2780         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2781         unsigned long flags;
2782         LIST_HEAD(list), *pos, *next;
2783         struct page *page;
2784         int split = 0;
2785
2786 #ifdef CONFIG_MEMCG
2787         if (sc->memcg)
2788                 ds_queue = &sc->memcg->deferred_split_queue;
2789 #endif
2790
2791         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2792         /* Take pin on all head pages to avoid freeing them under us */
2793         list_for_each_safe(pos, next, &ds_queue->split_queue) {
2794                 page = list_entry((void *)pos, struct page, mapping);
2795                 page = compound_head(page);
2796                 if (get_page_unless_zero(page)) {
2797                         list_move(page_deferred_list(page), &list);
2798                 } else {
2799                         /* We lost race with put_compound_page() */
2800                         list_del_init(page_deferred_list(page));
2801                         ds_queue->split_queue_len--;
2802                 }
2803                 if (!--sc->nr_to_scan)
2804                         break;
2805         }
2806         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2807
2808         list_for_each_safe(pos, next, &list) {
2809                 page = list_entry((void *)pos, struct page, mapping);
2810                 if (!trylock_page(page))
2811                         goto next;
2812                 /* split_huge_page() removes page from list on success */
2813                 if (!split_huge_page(page))
2814                         split++;
2815                 unlock_page(page);
2816 next:
2817                 put_page(page);
2818         }
2819
2820         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2821         list_splice_tail(&list, &ds_queue->split_queue);
2822         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2823
2824         /*
2825          * Stop shrinker if we didn't split any page, but the queue is empty.
2826          * This can happen if pages were freed under us.
2827          */
2828         if (!split && list_empty(&ds_queue->split_queue))
2829                 return SHRINK_STOP;
2830         return split;
2831 }
2832
2833 static struct shrinker deferred_split_shrinker = {
2834         .count_objects = deferred_split_count,
2835         .scan_objects = deferred_split_scan,
2836         .seeks = DEFAULT_SEEKS,
2837         .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
2838                  SHRINKER_NONSLAB,
2839 };
2840
2841 #ifdef CONFIG_DEBUG_FS
2842 static int split_huge_pages_set(void *data, u64 val)
2843 {
2844         struct zone *zone;
2845         struct page *page;
2846         unsigned long pfn, max_zone_pfn;
2847         unsigned long total = 0, split = 0;
2848
2849         if (val != 1)
2850                 return -EINVAL;
2851
2852         for_each_populated_zone(zone) {
2853                 max_zone_pfn = zone_end_pfn(zone);
2854                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2855                         if (!pfn_valid(pfn))
2856                                 continue;
2857
2858                         page = pfn_to_page(pfn);
2859                         if (!get_page_unless_zero(page))
2860                                 continue;
2861
2862                         if (zone != page_zone(page))
2863                                 goto next;
2864
2865                         if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2866                                 goto next;
2867
2868                         total++;
2869                         lock_page(page);
2870                         if (!split_huge_page(page))
2871                                 split++;
2872                         unlock_page(page);
2873 next:
2874                         put_page(page);
2875                 }
2876         }
2877
2878         pr_info("%lu of %lu THP split\n", split, total);
2879
2880         return 0;
2881 }
2882 DEFINE_DEBUGFS_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2883                 "%llu\n");
2884
2885 static int __init split_huge_pages_debugfs(void)
2886 {
2887         debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2888                             &split_huge_pages_fops);
2889         return 0;
2890 }
2891 late_initcall(split_huge_pages_debugfs);
2892 #endif
2893
2894 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2895 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
2896                 struct page *page)
2897 {
2898         struct vm_area_struct *vma = pvmw->vma;
2899         struct mm_struct *mm = vma->vm_mm;
2900         unsigned long address = pvmw->address;
2901         pmd_t pmdval;
2902         swp_entry_t entry;
2903         pmd_t pmdswp;
2904
2905         if (!(pvmw->pmd && !pvmw->pte))
2906                 return;
2907
2908         flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
2909         pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
2910         if (pmd_dirty(pmdval))
2911                 set_page_dirty(page);
2912         entry = make_migration_entry(page, pmd_write(pmdval));
2913         pmdswp = swp_entry_to_pmd(entry);
2914         if (pmd_soft_dirty(pmdval))
2915                 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
2916         set_pmd_at(mm, address, pvmw->pmd, pmdswp);
2917         page_remove_rmap(page, true);
2918         put_page(page);
2919 }
2920
2921 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
2922 {
2923         struct vm_area_struct *vma = pvmw->vma;
2924         struct mm_struct *mm = vma->vm_mm;
2925         unsigned long address = pvmw->address;
2926         unsigned long mmun_start = address & HPAGE_PMD_MASK;
2927         pmd_t pmde;
2928         swp_entry_t entry;
2929
2930         if (!(pvmw->pmd && !pvmw->pte))
2931                 return;
2932
2933         entry = pmd_to_swp_entry(*pvmw->pmd);
2934         get_page(new);
2935         pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
2936         if (pmd_swp_soft_dirty(*pvmw->pmd))
2937                 pmde = pmd_mksoft_dirty(pmde);
2938         if (is_write_migration_entry(entry))
2939                 pmde = maybe_pmd_mkwrite(pmde, vma);
2940
2941         flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
2942         if (PageAnon(new))
2943                 page_add_anon_rmap(new, vma, mmun_start, true);
2944         else
2945                 page_add_file_rmap(new, true);
2946         set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
2947         if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new))
2948                 mlock_vma_page(new);
2949         update_mmu_cache_pmd(vma, address, pvmw->pmd);
2950 }
2951 #endif