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 #ifdef CONFIG_DEBUG_VM
 307 static ssize_t debug_cow_show(struct kobject *kobj,
 308                                 struct kobj_attribute *attr, char *buf)
 309 {
 310         return single_hugepage_flag_show(kobj, attr, buf,
 311                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 312 }
 313 static ssize_t debug_cow_store(struct kobject *kobj,
 314                                struct kobj_attribute *attr,
 315                                const char *buf, size_t count)
 316 {
 317         return single_hugepage_flag_store(kobj, attr, buf, count,
 318                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 319 }
 320 static struct kobj_attribute debug_cow_attr =
 321         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 322 #endif /* CONFIG_DEBUG_VM */
 323
 324 static struct attribute *hugepage_attr[] = {
 325         &enabled_attr.attr,
 326         &defrag_attr.attr,
 327         &use_zero_page_attr.attr,
 328         &hpage_pmd_size_attr.attr,
 329 #ifdef CONFIG_SHMEM
 330         &shmem_enabled_attr.attr,
 331 #endif
 332 #ifdef CONFIG_DEBUG_VM
 333         &debug_cow_attr.attr,
 334 #endif
 335         NULL,
 336 };
 337
 338 static const struct attribute_group hugepage_attr_group = {
 339         .attrs = hugepage_attr,
 340 };
 341
 342 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 343 {
 344         int err;
 345
 346         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 347         if (unlikely(!*hugepage_kobj)) {
 348                 pr_err("failed to create transparent hugepage kobject\n");
 349                 return -ENOMEM;
 350         }
 351
 352         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 353         if (err) {
 354                 pr_err("failed to register transparent hugepage group\n");
 355                 goto delete_obj;
 356         }
 357
 358         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 359         if (err) {
 360                 pr_err("failed to register transparent hugepage group\n");
 361                 goto remove_hp_group;
 362         }
 363
 364         return 0;
 365
 366 remove_hp_group:
 367         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 368 delete_obj:
 369         kobject_put(*hugepage_kobj);
 370         return err;
 371 }
 372
 373 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 374 {
 375         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 376         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 377         kobject_put(hugepage_kobj);
 378 }
 379 #else
 380 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 381 {
 382         return 0;
 383 }
 384
 385 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 386 {
 387 }
 388 #endif /* CONFIG_SYSFS */
 389
 390 static int __init hugepage_init(void)
 391 {
 392         int err;
 393         struct kobject *hugepage_kobj;
 394
 395         if (!has_transparent_hugepage()) {
 396                 transparent_hugepage_flags = 0;
 397                 return -EINVAL;
 398         }
 399
 400         /*
 401          * hugepages can't be allocated by the buddy allocator
 402          */
 403         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
 404         /*
 405          * we use page->mapping and page->index in second tail page
 406          * as list_head: assuming THP order >= 2
 407          */
 408         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 409
 410         err = hugepage_init_sysfs(&hugepage_kobj);
 411         if (err)
 412                 goto err_sysfs;
 413
 414         err = khugepaged_init();
 415         if (err)
 416                 goto err_slab;
 417
 418         err = register_shrinker(&huge_zero_page_shrinker);
 419         if (err)
 420                 goto err_hzp_shrinker;
 421         err = register_shrinker(&deferred_split_shrinker);
 422         if (err)
 423                 goto err_split_shrinker;
 424
 425         /*
 426          * By default disable transparent hugepages on smaller systems,
 427          * where the extra memory used could hurt more than TLB overhead
 428          * is likely to save.  The admin can still enable it through /sys.
 429          */
 430         if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
 431                 transparent_hugepage_flags = 0;
 432                 return 0;
 433         }
 434
 435         err = start_stop_khugepaged();
 436         if (err)
 437                 goto err_khugepaged;
 438
 439         return 0;
 440 err_khugepaged:
 441         unregister_shrinker(&deferred_split_shrinker);
 442 err_split_shrinker:
 443         unregister_shrinker(&huge_zero_page_shrinker);
 444 err_hzp_shrinker:
 445         khugepaged_destroy();
 446 err_slab:
 447         hugepage_exit_sysfs(hugepage_kobj);
 448 err_sysfs:
 449         return err;
 450 }
 451 subsys_initcall(hugepage_init);
 452
 453 static int __init setup_transparent_hugepage(char *str)
 454 {
 455         int ret = 0;
 456         if (!str)
 457                 goto out;
 458         if (!strcmp(str, "always")) {
 459                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 460                         &transparent_hugepage_flags);
 461                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 462                           &transparent_hugepage_flags);
 463                 ret = 1;
 464         } else if (!strcmp(str, "madvise")) {
 465                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 466                           &transparent_hugepage_flags);
 467                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 468                         &transparent_hugepage_flags);
 469                 ret = 1;
 470         } else if (!strcmp(str, "never")) {
 471                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 472                           &transparent_hugepage_flags);
 473                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 474                           &transparent_hugepage_flags);
 475                 ret = 1;
 476         }
 477 out:
 478         if (!ret)
 479                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 480         return ret;
 481 }
 482 __setup("transparent_hugepage=", setup_transparent_hugepage);
 483
 484 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 485 {
 486         if (likely(vma->vm_flags & VM_WRITE))
 487                 pmd = pmd_mkwrite(pmd);
 488         return pmd;
 489 }
 490
 491 #ifdef CONFIG_MEMCG
 492 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
 493 {
 494         struct mem_cgroup *memcg = compound_head(page)->mem_cgroup;
 495         struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
 496
 497         if (memcg)
 498                 return &memcg->deferred_split_queue;
 499         else
 500                 return &pgdat->deferred_split_queue;
 501 }
 502 #else
 503 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
 504 {
 505         struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
 506
 507         return &pgdat->deferred_split_queue;
 508 }
 509 #endif
 510
 511 void prep_transhuge_page(struct page *page)
 512 {
 513         /*
 514          * we use page->mapping and page->indexlru in second tail page
 515          * as list_head: assuming THP order >= 2
 516          */
 517
 518         INIT_LIST_HEAD(page_deferred_list(page));
 519         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 520 }
 521
 522 bool is_transparent_hugepage(struct page *page)
 523 {
 524         if (!PageCompound(page))
 525                 return 0;
 526
 527         page = compound_head(page);
 528         return is_huge_zero_page(page) ||
 529                page[1].compound_dtor == TRANSHUGE_PAGE_DTOR;
 530 }
 531 EXPORT_SYMBOL_GPL(is_transparent_hugepage);
 532
 533 static unsigned long __thp_get_unmapped_area(struct file *filp,
 534                 unsigned long addr, unsigned long len,
 535                 loff_t off, unsigned long flags, unsigned long size)
 536 {
 537         loff_t off_end = off + len;
 538         loff_t off_align = round_up(off, size);
 539         unsigned long len_pad, ret;
 540
 541         if (off_end <= off_align || (off_end - off_align) < size)
 542                 return 0;
 543
 544         len_pad = len + size;
 545         if (len_pad < len || (off + len_pad) < off)
 546                 return 0;
 547
 548         ret = current->mm->get_unmapped_area(filp, addr, len_pad,
 549                                               off >> PAGE_SHIFT, flags);
 550
 551         /*
 552          * The failure might be due to length padding. The caller will retry
 553          * without the padding.
 554          */
 555         if (IS_ERR_VALUE(ret))
 556                 return 0;
 557
 558         /*
 559          * Do not try to align to THP boundary if allocation at the address
 560          * hint succeeds.
 561          */
 562         if (ret == addr)
 563                 return addr;
 564
 565         ret += (off - ret) & (size - 1);
 566         return ret;
 567 }
 568
 569 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 570                 unsigned long len, unsigned long pgoff, unsigned long flags)
 571 {
 572         unsigned long ret;
 573         loff_t off = (loff_t)pgoff << PAGE_SHIFT;
 574
 575         if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
 576                 goto out;
 577
 578         ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE);
 579         if (ret)
 580                 return ret;
 581 out:
 582         return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
 583 }
 584 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
 585
 586 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
 587                         struct page *page, gfp_t gfp)
 588 {
 589         struct vm_area_struct *vma = vmf->vma;
 590         pgtable_t pgtable;
 591         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 592         vm_fault_t ret = 0;
 593
 594         VM_BUG_ON_PAGE(!PageCompound(page), page);
 595
 596         if (mem_cgroup_charge(page, vma->vm_mm, gfp, false)) {
 597                 put_page(page);
 598                 count_vm_event(THP_FAULT_FALLBACK);
 599                 count_vm_event(THP_FAULT_FALLBACK_CHARGE);
 600                 return VM_FAULT_FALLBACK;
 601         }
 602         cgroup_throttle_swaprate(page, gfp);
 603
 604         pgtable = pte_alloc_one(vma->vm_mm);
 605         if (unlikely(!pgtable)) {
 606                 ret = VM_FAULT_OOM;
 607                 goto release;
 608         }
 609
 610         clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
 611         /*
 612          * The memory barrier inside __SetPageUptodate makes sure that
 613          * clear_huge_page writes become visible before the set_pmd_at()
 614          * write.
 615          */
 616         __SetPageUptodate(page);
 617
 618         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 619         if (unlikely(!pmd_none(*vmf->pmd))) {
 620                 goto unlock_release;
 621         } else {
 622                 pmd_t entry;
 623
 624                 ret = check_stable_address_space(vma->vm_mm);
 625                 if (ret)
 626                         goto unlock_release;
 627
 628                 /* Deliver the page fault to userland */
 629                 if (userfaultfd_missing(vma)) {
 630                         vm_fault_t ret2;
 631
 632                         spin_unlock(vmf->ptl);
 633                         put_page(page);
 634                         pte_free(vma->vm_mm, pgtable);
 635                         ret2 = handle_userfault(vmf, VM_UFFD_MISSING);
 636                         VM_BUG_ON(ret2 & VM_FAULT_FALLBACK);
 637                         return ret2;
 638                 }
 639
 640                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 641                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 642                 page_add_new_anon_rmap(page, vma, haddr, true);
 643                 lru_cache_add_active_or_unevictable(page, vma);
 644                 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
 645                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
 646                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 647                 mm_inc_nr_ptes(vma->vm_mm);
 648                 spin_unlock(vmf->ptl);
 649                 count_vm_event(THP_FAULT_ALLOC);
 650                 count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC);
 651         }
 652
 653         return 0;
 654 unlock_release:
 655         spin_unlock(vmf->ptl);
 656 release:
 657         if (pgtable)
 658                 pte_free(vma->vm_mm, pgtable);
 659         put_page(page);
 660         return ret;
 661
 662 }
 663
 664 /*
 665  * always: directly stall for all thp allocations
 666  * defer: wake kswapd and fail if not immediately available
 667  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
 668  *                fail if not immediately available
 669  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
 670  *          available
 671  * never: never stall for any thp allocation
 672  */
 673 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
 674 {
 675         const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
 676
 677         /* Always do synchronous compaction */
 678         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 679                 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
 680
 681         /* Kick kcompactd and fail quickly */
 682         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 683                 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
 684
 685         /* Synchronous compaction if madvised, otherwise kick kcompactd */
 686         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 687                 return GFP_TRANSHUGE_LIGHT |
 688                         (vma_madvised ? __GFP_DIRECT_RECLAIM :
 689                                         __GFP_KSWAPD_RECLAIM);
 690
 691         /* Only do synchronous compaction if madvised */
 692         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 693                 return GFP_TRANSHUGE_LIGHT |
 694                        (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
 695
 696         return GFP_TRANSHUGE_LIGHT;
 697 }
 698
 699 /* Caller must hold page table lock. */
 700 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 701                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 702                 struct page *zero_page)
 703 {
 704         pmd_t entry;
 705         if (!pmd_none(*pmd))
 706                 return false;
 707         entry = mk_pmd(zero_page, vma->vm_page_prot);
 708         entry = pmd_mkhuge(entry);
 709         if (pgtable)
 710                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 711         set_pmd_at(mm, haddr, pmd, entry);
 712         mm_inc_nr_ptes(mm);
 713         return true;
 714 }
 715
 716 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
 717 {
 718         struct vm_area_struct *vma = vmf->vma;
 719         gfp_t gfp;
 720         struct page *page;
 721         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 722
 723         if (!transhuge_vma_suitable(vma, haddr))
 724                 return VM_FAULT_FALLBACK;
 725         if (unlikely(anon_vma_prepare(vma)))
 726                 return VM_FAULT_OOM;
 727         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 728                 return VM_FAULT_OOM;
 729         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
 730                         !mm_forbids_zeropage(vma->vm_mm) &&
 731                         transparent_hugepage_use_zero_page()) {
 732                 pgtable_t pgtable;
 733                 struct page *zero_page;
 734                 bool set;
 735                 vm_fault_t ret;
 736                 pgtable = pte_alloc_one(vma->vm_mm);
 737                 if (unlikely(!pgtable))
 738                         return VM_FAULT_OOM;
 739                 zero_page = mm_get_huge_zero_page(vma->vm_mm);
 740                 if (unlikely(!zero_page)) {
 741                         pte_free(vma->vm_mm, pgtable);
 742                         count_vm_event(THP_FAULT_FALLBACK);
 743                         return VM_FAULT_FALLBACK;
 744                 }
 745                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 746                 ret = 0;
 747                 set = false;
 748                 if (pmd_none(*vmf->pmd)) {
 749                         ret = check_stable_address_space(vma->vm_mm);
 750                         if (ret) {
 751                                 spin_unlock(vmf->ptl);
 752                         } else if (userfaultfd_missing(vma)) {
 753                                 spin_unlock(vmf->ptl);
 754                                 ret = handle_userfault(vmf, VM_UFFD_MISSING);
 755                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 756                         } else {
 757                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
 758                                                    haddr, vmf->pmd, zero_page);
 759                                 spin_unlock(vmf->ptl);
 760                                 set = true;
 761                         }
 762                 } else
 763                         spin_unlock(vmf->ptl);
 764                 if (!set)
 765                         pte_free(vma->vm_mm, pgtable);
 766                 return ret;
 767         }
 768         gfp = alloc_hugepage_direct_gfpmask(vma);
 769         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 770         if (unlikely(!page)) {
 771                 count_vm_event(THP_FAULT_FALLBACK);
 772                 return VM_FAULT_FALLBACK;
 773         }
 774         prep_transhuge_page(page);
 775         return __do_huge_pmd_anonymous_page(vmf, page, gfp);
 776 }
 777
 778 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 779                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
 780                 pgtable_t pgtable)
 781 {
 782         struct mm_struct *mm = vma->vm_mm;
 783         pmd_t entry;
 784         spinlock_t *ptl;
 785
 786         ptl = pmd_lock(mm, pmd);
 787         if (!pmd_none(*pmd)) {
 788                 if (write) {
 789                         if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
 790                                 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
 791                                 goto out_unlock;
 792                         }
 793                         entry = pmd_mkyoung(*pmd);
 794                         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 795                         if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
 796                                 update_mmu_cache_pmd(vma, addr, pmd);
 797                 }
 798
 799                 goto out_unlock;
 800         }
 801
 802         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 803         if (pfn_t_devmap(pfn))
 804                 entry = pmd_mkdevmap(entry);
 805         if (write) {
 806                 entry = pmd_mkyoung(pmd_mkdirty(entry));
 807                 entry = maybe_pmd_mkwrite(entry, vma);
 808         }
 809
 810         if (pgtable) {
 811                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 812                 mm_inc_nr_ptes(mm);
 813                 pgtable = NULL;
 814         }
 815
 816         set_pmd_at(mm, addr, pmd, entry);
 817         update_mmu_cache_pmd(vma, addr, pmd);
 818
 819 out_unlock:
 820         spin_unlock(ptl);
 821         if (pgtable)
 822                 pte_free(mm, pgtable);
 823 }
 824
 825 /**
 826  * vmf_insert_pfn_pmd_prot - insert a pmd size pfn
 827  * @vmf: Structure describing the fault
 828  * @pfn: pfn to insert
 829  * @pgprot: page protection to use
 830  * @write: whether it's a write fault
 831  *
 832  * Insert a pmd size pfn. See vmf_insert_pfn() for additional info and
 833  * also consult the vmf_insert_mixed_prot() documentation when
 834  * @pgprot != @vmf->vma->vm_page_prot.
 835  *
 836  * Return: vm_fault_t value.
 837  */
 838 vm_fault_t vmf_insert_pfn_pmd_prot(struct vm_fault *vmf, pfn_t pfn,
 839                                    pgprot_t pgprot, bool write)
 840 {
 841         unsigned long addr = vmf->address & PMD_MASK;
 842         struct vm_area_struct *vma = vmf->vma;
 843         pgtable_t pgtable = NULL;
 844
 845         /*
 846          * If we had pmd_special, we could avoid all these restrictions,
 847          * but we need to be consistent with PTEs and architectures that
 848          * can't support a 'special' bit.
 849          */
 850         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 851                         !pfn_t_devmap(pfn));
 852         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 853                                                 (VM_PFNMAP|VM_MIXEDMAP));
 854         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 855
 856         if (addr < vma->vm_start || addr >= vma->vm_end)
 857                 return VM_FAULT_SIGBUS;
 858
 859         if (arch_needs_pgtable_deposit()) {
 860                 pgtable = pte_alloc_one(vma->vm_mm);
 861                 if (!pgtable)
 862                         return VM_FAULT_OOM;
 863         }
 864
 865         track_pfn_insert(vma, &pgprot, pfn);
 866
 867         insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
 868         return VM_FAULT_NOPAGE;
 869 }
 870 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd_prot);
 871
 872 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 873 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
 874 {
 875         if (likely(vma->vm_flags & VM_WRITE))
 876                 pud = pud_mkwrite(pud);
 877         return pud;
 878 }
 879
 880 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 881                 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
 882 {
 883         struct mm_struct *mm = vma->vm_mm;
 884         pud_t entry;
 885         spinlock_t *ptl;
 886
 887         ptl = pud_lock(mm, pud);
 888         if (!pud_none(*pud)) {
 889                 if (write) {
 890                         if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
 891                                 WARN_ON_ONCE(!is_huge_zero_pud(*pud));
 892                                 goto out_unlock;
 893                         }
 894                         entry = pud_mkyoung(*pud);
 895                         entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
 896                         if (pudp_set_access_flags(vma, addr, pud, entry, 1))
 897                                 update_mmu_cache_pud(vma, addr, pud);
 898                 }
 899                 goto out_unlock;
 900         }
 901
 902         entry = pud_mkhuge(pfn_t_pud(pfn, prot));
 903         if (pfn_t_devmap(pfn))
 904                 entry = pud_mkdevmap(entry);
 905         if (write) {
 906                 entry = pud_mkyoung(pud_mkdirty(entry));
 907                 entry = maybe_pud_mkwrite(entry, vma);
 908         }
 909         set_pud_at(mm, addr, pud, entry);
 910         update_mmu_cache_pud(vma, addr, pud);
 911
 912 out_unlock:
 913         spin_unlock(ptl);
 914 }
 915
 916 /**
 917  * vmf_insert_pfn_pud_prot - insert a pud size pfn
 918  * @vmf: Structure describing the fault
 919  * @pfn: pfn to insert
 920  * @pgprot: page protection to use
 921  * @write: whether it's a write fault
 922  *
 923  * Insert a pud size pfn. See vmf_insert_pfn() for additional info and
 924  * also consult the vmf_insert_mixed_prot() documentation when
 925  * @pgprot != @vmf->vma->vm_page_prot.
 926  *
 927  * Return: vm_fault_t value.
 928  */
 929 vm_fault_t vmf_insert_pfn_pud_prot(struct vm_fault *vmf, pfn_t pfn,
 930                                    pgprot_t pgprot, bool write)
 931 {
 932         unsigned long addr = vmf->address & PUD_MASK;
 933         struct vm_area_struct *vma = vmf->vma;
 934
 935         /*
 936          * If we had pud_special, we could avoid all these restrictions,
 937          * but we need to be consistent with PTEs and architectures that
 938          * can't support a 'special' bit.
 939          */
 940         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 941                         !pfn_t_devmap(pfn));
 942         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 943                                                 (VM_PFNMAP|VM_MIXEDMAP));
 944         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 945
 946         if (addr < vma->vm_start || addr >= vma->vm_end)
 947                 return VM_FAULT_SIGBUS;
 948
 949         track_pfn_insert(vma, &pgprot, pfn);
 950
 951         insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
 952         return VM_FAULT_NOPAGE;
 953 }
 954 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud_prot);
 955 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
 956
 957 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
 958                 pmd_t *pmd, int flags)
 959 {
 960         pmd_t _pmd;
 961
 962         _pmd = pmd_mkyoung(*pmd);
 963         if (flags & FOLL_WRITE)
 964                 _pmd = pmd_mkdirty(_pmd);
 965         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
 966                                 pmd, _pmd, flags & FOLL_WRITE))
 967                 update_mmu_cache_pmd(vma, addr, pmd);
 968 }
 969
 970 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 971                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
 972 {
 973         unsigned long pfn = pmd_pfn(*pmd);
 974         struct mm_struct *mm = vma->vm_mm;
 975         struct page *page;
 976
 977         assert_spin_locked(pmd_lockptr(mm, pmd));
 978
 979         /*
 980          * When we COW a devmap PMD entry, we split it into PTEs, so we should
 981          * not be in this function with `flags & FOLL_COW` set.
 982          */
 983         WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
 984
 985         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
 986         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
 987                          (FOLL_PIN | FOLL_GET)))
 988                 return NULL;
 989
 990         if (flags & FOLL_WRITE && !pmd_write(*pmd))
 991                 return NULL;
 992
 993         if (pmd_present(*pmd) && pmd_devmap(*pmd))
 994                 /* pass */;
 995         else
 996                 return NULL;
 997
 998         if (flags & FOLL_TOUCH)
 999                 touch_pmd(vma, addr, pmd, flags);
1000
1001         /*
1002          * device mapped pages can only be returned if the
1003          * caller will manage the page reference count.
1004          */
1005         if (!(flags & (FOLL_GET | FOLL_PIN)))
1006                 return ERR_PTR(-EEXIST);
1007
1008         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1009         *pgmap = get_dev_pagemap(pfn, *pgmap);
1010         if (!*pgmap)
1011                 return ERR_PTR(-EFAULT);
1012         page = pfn_to_page(pfn);
1013         if (!try_grab_page(page, flags))
1014                 page = ERR_PTR(-ENOMEM);
1015
1016         return page;
1017 }
1018
1019 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1020                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1021                   struct vm_area_struct *vma)
1022 {
1023         spinlock_t *dst_ptl, *src_ptl;
1024         struct page *src_page;
1025         pmd_t pmd;
1026         pgtable_t pgtable = NULL;
1027         int ret = -ENOMEM;
1028
1029         /* Skip if can be re-fill on fault */
1030         if (!vma_is_anonymous(vma))
1031                 return 0;
1032
1033         pgtable = pte_alloc_one(dst_mm);
1034         if (unlikely(!pgtable))
1035                 goto out;
1036
1037         dst_ptl = pmd_lock(dst_mm, dst_pmd);
1038         src_ptl = pmd_lockptr(src_mm, src_pmd);
1039         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1040
1041         ret = -EAGAIN;
1042         pmd = *src_pmd;
1043
1044         /*
1045          * Make sure the _PAGE_UFFD_WP bit is cleared if the new VMA
1046          * does not have the VM_UFFD_WP, which means that the uffd
1047          * fork event is not enabled.
1048          */
1049         if (!(vma->vm_flags & VM_UFFD_WP))
1050                 pmd = pmd_clear_uffd_wp(pmd);
1051
1052 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1053         if (unlikely(is_swap_pmd(pmd))) {
1054                 swp_entry_t entry = pmd_to_swp_entry(pmd);
1055
1056                 VM_BUG_ON(!is_pmd_migration_entry(pmd));
1057                 if (is_write_migration_entry(entry)) {
1058                         make_migration_entry_read(&entry);
1059                         pmd = swp_entry_to_pmd(entry);
1060                         if (pmd_swp_soft_dirty(*src_pmd))
1061                                 pmd = pmd_swp_mksoft_dirty(pmd);
1062                         set_pmd_at(src_mm, addr, src_pmd, pmd);
1063                 }
1064                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1065                 mm_inc_nr_ptes(dst_mm);
1066                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1067                 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1068                 ret = 0;
1069                 goto out_unlock;
1070         }
1071 #endif
1072
1073         if (unlikely(!pmd_trans_huge(pmd))) {
1074                 pte_free(dst_mm, pgtable);
1075                 goto out_unlock;
1076         }
1077         /*
1078          * When page table lock is held, the huge zero pmd should not be
1079          * under splitting since we don't split the page itself, only pmd to
1080          * a page table.
1081          */
1082         if (is_huge_zero_pmd(pmd)) {
1083                 struct page *zero_page;
1084                 /*
1085                  * get_huge_zero_page() will never allocate a new page here,
1086                  * since we already have a zero page to copy. It just takes a
1087                  * reference.
1088                  */
1089                 zero_page = mm_get_huge_zero_page(dst_mm);
1090                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1091                                 zero_page);
1092                 ret = 0;
1093                 goto out_unlock;
1094         }
1095
1096         src_page = pmd_page(pmd);
1097         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1098         get_page(src_page);
1099         page_dup_rmap(src_page, true);
1100         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1101         mm_inc_nr_ptes(dst_mm);
1102         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1103
1104         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1105         pmd = pmd_mkold(pmd_wrprotect(pmd));
1106         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1107
1108         ret = 0;
1109 out_unlock:
1110         spin_unlock(src_ptl);
1111         spin_unlock(dst_ptl);
1112 out:
1113         return ret;
1114 }
1115
1116 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1117 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1118                 pud_t *pud, int flags)
1119 {
1120         pud_t _pud;
1121
1122         _pud = pud_mkyoung(*pud);
1123         if (flags & FOLL_WRITE)
1124                 _pud = pud_mkdirty(_pud);
1125         if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1126                                 pud, _pud, flags & FOLL_WRITE))
1127                 update_mmu_cache_pud(vma, addr, pud);
1128 }
1129
1130 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1131                 pud_t *pud, int flags, struct dev_pagemap **pgmap)
1132 {
1133         unsigned long pfn = pud_pfn(*pud);
1134         struct mm_struct *mm = vma->vm_mm;
1135         struct page *page;
1136
1137         assert_spin_locked(pud_lockptr(mm, pud));
1138
1139         if (flags & FOLL_WRITE && !pud_write(*pud))
1140                 return NULL;
1141
1142         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
1143         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
1144                          (FOLL_PIN | FOLL_GET)))
1145                 return NULL;
1146
1147         if (pud_present(*pud) && pud_devmap(*pud))
1148                 /* pass */;
1149         else
1150                 return NULL;
1151
1152         if (flags & FOLL_TOUCH)
1153                 touch_pud(vma, addr, pud, flags);
1154
1155         /*
1156          * device mapped pages can only be returned if the
1157          * caller will manage the page reference count.
1158          *
1159          * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here:
1160          */
1161         if (!(flags & (FOLL_GET | FOLL_PIN)))
1162                 return ERR_PTR(-EEXIST);
1163
1164         pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1165         *pgmap = get_dev_pagemap(pfn, *pgmap);
1166         if (!*pgmap)
1167                 return ERR_PTR(-EFAULT);
1168         page = pfn_to_page(pfn);
1169         if (!try_grab_page(page, flags))
1170                 page = ERR_PTR(-ENOMEM);
1171
1172         return page;
1173 }
1174
1175 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1176                   pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1177                   struct vm_area_struct *vma)
1178 {
1179         spinlock_t *dst_ptl, *src_ptl;
1180         pud_t pud;
1181         int ret;
1182
1183         dst_ptl = pud_lock(dst_mm, dst_pud);
1184         src_ptl = pud_lockptr(src_mm, src_pud);
1185         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1186
1187         ret = -EAGAIN;
1188         pud = *src_pud;
1189         if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1190                 goto out_unlock;
1191
1192         /*
1193          * When page table lock is held, the huge zero pud should not be
1194          * under splitting since we don't split the page itself, only pud to
1195          * a page table.
1196          */
1197         if (is_huge_zero_pud(pud)) {
1198                 /* No huge zero pud yet */
1199         }
1200
1201         pudp_set_wrprotect(src_mm, addr, src_pud);
1202         pud = pud_mkold(pud_wrprotect(pud));
1203         set_pud_at(dst_mm, addr, dst_pud, pud);
1204
1205         ret = 0;
1206 out_unlock:
1207         spin_unlock(src_ptl);
1208         spin_unlock(dst_ptl);
1209         return ret;
1210 }
1211
1212 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1213 {
1214         pud_t entry;
1215         unsigned long haddr;
1216         bool write = vmf->flags & FAULT_FLAG_WRITE;
1217
1218         vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1219         if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1220                 goto unlock;
1221
1222         entry = pud_mkyoung(orig_pud);
1223         if (write)
1224                 entry = pud_mkdirty(entry);
1225         haddr = vmf->address & HPAGE_PUD_MASK;
1226         if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1227                 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1228
1229 unlock:
1230         spin_unlock(vmf->ptl);
1231 }
1232 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1233
1234 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1235 {
1236         pmd_t entry;
1237         unsigned long haddr;
1238         bool write = vmf->flags & FAULT_FLAG_WRITE;
1239
1240         vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1241         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1242                 goto unlock;
1243
1244         entry = pmd_mkyoung(orig_pmd);
1245         if (write)
1246                 entry = pmd_mkdirty(entry);
1247         haddr = vmf->address & HPAGE_PMD_MASK;
1248         if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1249                 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1250
1251 unlock:
1252         spin_unlock(vmf->ptl);
1253 }
1254
1255 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1256 {
1257         struct vm_area_struct *vma = vmf->vma;
1258         struct page *page;
1259         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1260
1261         vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1262         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1263
1264         if (is_huge_zero_pmd(orig_pmd))
1265                 goto fallback;
1266
1267         spin_lock(vmf->ptl);
1268
1269         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1270                 spin_unlock(vmf->ptl);
1271                 return 0;
1272         }
1273
1274         page = pmd_page(orig_pmd);
1275         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1276
1277         /* Lock page for reuse_swap_page() */
1278         if (!trylock_page(page)) {
1279                 get_page(page);
1280                 spin_unlock(vmf->ptl);
1281                 lock_page(page);
1282                 spin_lock(vmf->ptl);
1283                 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1284                         spin_unlock(vmf->ptl);
1285                         unlock_page(page);
1286                         put_page(page);
1287                         return 0;
1288                 }
1289                 put_page(page);
1290         }
1291
1292         /*
1293          * We can only reuse the page if nobody else maps the huge page or it's
1294          * part.
1295          */
1296         if (reuse_swap_page(page, NULL)) {
1297                 pmd_t entry;
1298                 entry = pmd_mkyoung(orig_pmd);
1299                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1300                 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1301                         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1302                 unlock_page(page);
1303                 spin_unlock(vmf->ptl);
1304                 return VM_FAULT_WRITE;
1305         }
1306
1307         unlock_page(page);
1308         spin_unlock(vmf->ptl);
1309 fallback:
1310         __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL);
1311         return VM_FAULT_FALLBACK;
1312 }
1313
1314 /*
1315  * FOLL_FORCE or a forced COW break can write even to unwritable pmd's,
1316  * but only after we've gone through a COW cycle and they are dirty.
1317  */
1318 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1319 {
1320         return pmd_write(pmd) || ((flags & FOLL_COW) && pmd_dirty(pmd));
1321 }
1322
1323 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1324                                    unsigned long addr,
1325                                    pmd_t *pmd,
1326                                    unsigned int flags)
1327 {
1328         struct mm_struct *mm = vma->vm_mm;
1329         struct page *page = NULL;
1330
1331         assert_spin_locked(pmd_lockptr(mm, pmd));
1332
1333         if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1334                 goto out;
1335
1336         /* Avoid dumping huge zero page */
1337         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1338                 return ERR_PTR(-EFAULT);
1339
1340         /* Full NUMA hinting faults to serialise migration in fault paths */
1341         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1342                 goto out;
1343
1344         page = pmd_page(*pmd);
1345         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1346
1347         if (!try_grab_page(page, flags))
1348                 return ERR_PTR(-ENOMEM);
1349
1350         if (flags & FOLL_TOUCH)
1351                 touch_pmd(vma, addr, pmd, flags);
1352
1353         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1354                 /*
1355                  * We don't mlock() pte-mapped THPs. This way we can avoid
1356                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1357                  *
1358                  * For anon THP:
1359                  *
1360                  * In most cases the pmd is the only mapping of the page as we
1361                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1362                  * writable private mappings in populate_vma_page_range().
1363                  *
1364                  * The only scenario when we have the page shared here is if we
1365                  * mlocking read-only mapping shared over fork(). We skip
1366                  * mlocking such pages.
1367                  *
1368                  * For file THP:
1369                  *
1370                  * We can expect PageDoubleMap() to be stable under page lock:
1371                  * for file pages we set it in page_add_file_rmap(), which
1372                  * requires page to be locked.
1373                  */
1374
1375                 if (PageAnon(page) && compound_mapcount(page) != 1)
1376                         goto skip_mlock;
1377                 if (PageDoubleMap(page) || !page->mapping)
1378                         goto skip_mlock;
1379                 if (!trylock_page(page))
1380                         goto skip_mlock;
1381                 lru_add_drain();
1382                 if (page->mapping && !PageDoubleMap(page))
1383                         mlock_vma_page(page);
1384                 unlock_page(page);
1385         }
1386 skip_mlock:
1387         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1388         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1389
1390 out:
1391         return page;
1392 }
1393
1394 /* NUMA hinting page fault entry point for trans huge pmds */
1395 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1396 {
1397         struct vm_area_struct *vma = vmf->vma;
1398         struct anon_vma *anon_vma = NULL;
1399         struct page *page;
1400         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1401         int page_nid = NUMA_NO_NODE, this_nid = numa_node_id();
1402         int target_nid, last_cpupid = -1;
1403         bool page_locked;
1404         bool migrated = false;
1405         bool was_writable;
1406         int flags = 0;
1407
1408         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1409         if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1410                 goto out_unlock;
1411
1412         /*
1413          * If there are potential migrations, wait for completion and retry
1414          * without disrupting NUMA hinting information. Do not relock and
1415          * check_same as the page may no longer be mapped.
1416          */
1417         if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1418                 page = pmd_page(*vmf->pmd);
1419                 if (!get_page_unless_zero(page))
1420                         goto out_unlock;
1421                 spin_unlock(vmf->ptl);
1422                 put_and_wait_on_page_locked(page);
1423                 goto out;
1424         }
1425
1426         page = pmd_page(pmd);
1427         BUG_ON(is_huge_zero_page(page));
1428         page_nid = page_to_nid(page);
1429         last_cpupid = page_cpupid_last(page);
1430         count_vm_numa_event(NUMA_HINT_FAULTS);
1431         if (page_nid == this_nid) {
1432                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1433                 flags |= TNF_FAULT_LOCAL;
1434         }
1435
1436         /* See similar comment in do_numa_page for explanation */
1437         if (!pmd_savedwrite(pmd))
1438                 flags |= TNF_NO_GROUP;
1439
1440         /*
1441          * Acquire the page lock to serialise THP migrations but avoid dropping
1442          * page_table_lock if at all possible
1443          */
1444         page_locked = trylock_page(page);
1445         target_nid = mpol_misplaced(page, vma, haddr);
1446         if (target_nid == NUMA_NO_NODE) {
1447                 /* If the page was locked, there are no parallel migrations */
1448                 if (page_locked)
1449                         goto clear_pmdnuma;
1450         }
1451
1452         /* Migration could have started since the pmd_trans_migrating check */
1453         if (!page_locked) {
1454                 page_nid = NUMA_NO_NODE;
1455                 if (!get_page_unless_zero(page))
1456                         goto out_unlock;
1457                 spin_unlock(vmf->ptl);
1458                 put_and_wait_on_page_locked(page);
1459                 goto out;
1460         }
1461
1462         /*
1463          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1464          * to serialises splits
1465          */
1466         get_page(page);
1467         spin_unlock(vmf->ptl);
1468         anon_vma = page_lock_anon_vma_read(page);
1469
1470         /* Confirm the PMD did not change while page_table_lock was released */
1471         spin_lock(vmf->ptl);
1472         if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1473                 unlock_page(page);
1474                 put_page(page);
1475                 page_nid = NUMA_NO_NODE;
1476                 goto out_unlock;
1477         }
1478
1479         /* Bail if we fail to protect against THP splits for any reason */
1480         if (unlikely(!anon_vma)) {
1481                 put_page(page);
1482                 page_nid = NUMA_NO_NODE;
1483                 goto clear_pmdnuma;
1484         }
1485
1486         /*
1487          * Since we took the NUMA fault, we must have observed the !accessible
1488          * bit. Make sure all other CPUs agree with that, to avoid them
1489          * modifying the page we're about to migrate.
1490          *
1491          * Must be done under PTL such that we'll observe the relevant
1492          * inc_tlb_flush_pending().
1493          *
1494          * We are not sure a pending tlb flush here is for a huge page
1495          * mapping or not. Hence use the tlb range variant
1496          */
1497         if (mm_tlb_flush_pending(vma->vm_mm)) {
1498                 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1499                 /*
1500                  * change_huge_pmd() released the pmd lock before
1501                  * invalidating the secondary MMUs sharing the primary
1502                  * MMU pagetables (with ->invalidate_range()). The
1503                  * mmu_notifier_invalidate_range_end() (which
1504                  * internally calls ->invalidate_range()) in
1505                  * change_pmd_range() will run after us, so we can't
1506                  * rely on it here and we need an explicit invalidate.
1507                  */
1508                 mmu_notifier_invalidate_range(vma->vm_mm, haddr,
1509                                               haddr + HPAGE_PMD_SIZE);
1510         }
1511
1512         /*
1513          * Migrate the THP to the requested node, returns with page unlocked
1514          * and access rights restored.
1515          */
1516         spin_unlock(vmf->ptl);
1517
1518         migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1519                                 vmf->pmd, pmd, vmf->address, page, target_nid);
1520         if (migrated) {
1521                 flags |= TNF_MIGRATED;
1522                 page_nid = target_nid;
1523         } else
1524                 flags |= TNF_MIGRATE_FAIL;
1525
1526         goto out;
1527 clear_pmdnuma:
1528         BUG_ON(!PageLocked(page));
1529         was_writable = pmd_savedwrite(pmd);
1530         pmd = pmd_modify(pmd, vma->vm_page_prot);
1531         pmd = pmd_mkyoung(pmd);
1532         if (was_writable)
1533                 pmd = pmd_mkwrite(pmd);
1534         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1535         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1536         unlock_page(page);
1537 out_unlock:
1538         spin_unlock(vmf->ptl);
1539
1540 out:
1541         if (anon_vma)
1542                 page_unlock_anon_vma_read(anon_vma);
1543
1544         if (page_nid != NUMA_NO_NODE)
1545                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1546                                 flags);
1547
1548         return 0;
1549 }
1550
1551 /*
1552  * Return true if we do MADV_FREE successfully on entire pmd page.
1553  * Otherwise, return false.
1554  */
1555 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1556                 pmd_t *pmd, unsigned long addr, unsigned long next)
1557 {
1558         spinlock_t *ptl;
1559         pmd_t orig_pmd;
1560         struct page *page;
1561         struct mm_struct *mm = tlb->mm;
1562         bool ret = false;
1563
1564         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1565
1566         ptl = pmd_trans_huge_lock(pmd, vma);
1567         if (!ptl)
1568                 goto out_unlocked;
1569
1570         orig_pmd = *pmd;
1571         if (is_huge_zero_pmd(orig_pmd))
1572                 goto out;
1573
1574         if (unlikely(!pmd_present(orig_pmd))) {
1575                 VM_BUG_ON(thp_migration_supported() &&
1576                                   !is_pmd_migration_entry(orig_pmd));
1577                 goto out;
1578         }
1579
1580         page = pmd_page(orig_pmd);
1581         /*
1582          * If other processes are mapping this page, we couldn't discard
1583          * the page unless they all do MADV_FREE so let's skip the page.
1584          */
1585         if (page_mapcount(page) != 1)
1586                 goto out;
1587
1588         if (!trylock_page(page))
1589                 goto out;
1590
1591         /*
1592          * If user want to discard part-pages of THP, split it so MADV_FREE
1593          * will deactivate only them.
1594          */
1595         if (next - addr != HPAGE_PMD_SIZE) {
1596                 get_page(page);
1597                 spin_unlock(ptl);
1598                 split_huge_page(page);
1599                 unlock_page(page);
1600                 put_page(page);
1601                 goto out_unlocked;
1602         }
1603
1604         if (PageDirty(page))
1605                 ClearPageDirty(page);
1606         unlock_page(page);
1607
1608         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1609                 pmdp_invalidate(vma, addr, pmd);
1610                 orig_pmd = pmd_mkold(orig_pmd);
1611                 orig_pmd = pmd_mkclean(orig_pmd);
1612
1613                 set_pmd_at(mm, addr, pmd, orig_pmd);
1614                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1615         }
1616
1617         mark_page_lazyfree(page);
1618         ret = true;
1619 out:
1620         spin_unlock(ptl);
1621 out_unlocked:
1622         return ret;
1623 }
1624
1625 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1626 {
1627         pgtable_t pgtable;
1628
1629         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1630         pte_free(mm, pgtable);
1631         mm_dec_nr_ptes(mm);
1632 }
1633
1634 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1635                  pmd_t *pmd, unsigned long addr)
1636 {
1637         pmd_t orig_pmd;
1638         spinlock_t *ptl;
1639
1640         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1641
1642         ptl = __pmd_trans_huge_lock(pmd, vma);
1643         if (!ptl)
1644                 return 0;
1645         /*
1646          * For architectures like ppc64 we look at deposited pgtable
1647          * when calling pmdp_huge_get_and_clear. So do the
1648          * pgtable_trans_huge_withdraw after finishing pmdp related
1649          * operations.
1650          */
1651         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1652                         tlb->fullmm);
1653         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1654         if (vma_is_special_huge(vma)) {
1655                 if (arch_needs_pgtable_deposit())
1656                         zap_deposited_table(tlb->mm, pmd);
1657                 spin_unlock(ptl);
1658                 if (is_huge_zero_pmd(orig_pmd))
1659                         tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1660         } else if (is_huge_zero_pmd(orig_pmd)) {
1661                 zap_deposited_table(tlb->mm, pmd);
1662                 spin_unlock(ptl);
1663                 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1664         } else {
1665                 struct page *page = NULL;
1666                 int flush_needed = 1;
1667
1668                 if (pmd_present(orig_pmd)) {
1669                         page = pmd_page(orig_pmd);
1670                         page_remove_rmap(page, true);
1671                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1672                         VM_BUG_ON_PAGE(!PageHead(page), page);
1673                 } else if (thp_migration_supported()) {
1674                         swp_entry_t entry;
1675
1676                         VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1677                         entry = pmd_to_swp_entry(orig_pmd);
1678                         page = pfn_to_page(swp_offset(entry));
1679                         flush_needed = 0;
1680                 } else
1681                         WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1682
1683                 if (PageAnon(page)) {
1684                         zap_deposited_table(tlb->mm, pmd);
1685                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1686                 } else {
1687                         if (arch_needs_pgtable_deposit())
1688                                 zap_deposited_table(tlb->mm, pmd);
1689                         add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1690                 }
1691
1692                 spin_unlock(ptl);
1693                 if (flush_needed)
1694                         tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1695         }
1696         return 1;
1697 }
1698
1699 #ifndef pmd_move_must_withdraw
1700 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1701                                          spinlock_t *old_pmd_ptl,
1702                                          struct vm_area_struct *vma)
1703 {
1704         /*
1705          * With split pmd lock we also need to move preallocated
1706          * PTE page table if new_pmd is on different PMD page table.
1707          *
1708          * We also don't deposit and withdraw tables for file pages.
1709          */
1710         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1711 }
1712 #endif
1713
1714 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1715 {
1716 #ifdef CONFIG_MEM_SOFT_DIRTY
1717         if (unlikely(is_pmd_migration_entry(pmd)))
1718                 pmd = pmd_swp_mksoft_dirty(pmd);
1719         else if (pmd_present(pmd))
1720                 pmd = pmd_mksoft_dirty(pmd);
1721 #endif
1722         return pmd;
1723 }
1724
1725 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1726                   unsigned long new_addr, unsigned long old_end,
1727                   pmd_t *old_pmd, pmd_t *new_pmd)
1728 {
1729         spinlock_t *old_ptl, *new_ptl;
1730         pmd_t pmd;
1731         struct mm_struct *mm = vma->vm_mm;
1732         bool force_flush = false;
1733
1734         if ((old_addr & ~HPAGE_PMD_MASK) ||
1735             (new_addr & ~HPAGE_PMD_MASK) ||
1736             old_end - old_addr < HPAGE_PMD_SIZE)
1737                 return false;
1738
1739         /*
1740          * The destination pmd shouldn't be established, free_pgtables()
1741          * should have release it.
1742          */
1743         if (WARN_ON(!pmd_none(*new_pmd))) {
1744                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1745                 return false;
1746         }
1747
1748         /*
1749          * We don't have to worry about the ordering of src and dst
1750          * ptlocks because exclusive mmap_sem prevents deadlock.
1751          */
1752         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1753         if (old_ptl) {
1754                 new_ptl = pmd_lockptr(mm, new_pmd);
1755                 if (new_ptl != old_ptl)
1756                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1757                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1758                 if (pmd_present(pmd))
1759                         force_flush = true;
1760                 VM_BUG_ON(!pmd_none(*new_pmd));
1761
1762                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1763                         pgtable_t pgtable;
1764                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1765                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1766                 }
1767                 pmd = move_soft_dirty_pmd(pmd);
1768                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1769                 if (force_flush)
1770                         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1771                 if (new_ptl != old_ptl)
1772                         spin_unlock(new_ptl);
1773                 spin_unlock(old_ptl);
1774                 return true;
1775         }
1776         return false;
1777 }
1778
1779 /*
1780  * Returns
1781  *  - 0 if PMD could not be locked
1782  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1783  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1784  */
1785 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1786                 unsigned long addr, pgprot_t newprot, unsigned long cp_flags)
1787 {
1788         struct mm_struct *mm = vma->vm_mm;
1789         spinlock_t *ptl;
1790         pmd_t entry;
1791         bool preserve_write;
1792         int ret;
1793         bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
1794         bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
1795         bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
1796
1797         ptl = __pmd_trans_huge_lock(pmd, vma);
1798         if (!ptl)
1799                 return 0;
1800
1801         preserve_write = prot_numa && pmd_write(*pmd);
1802         ret = 1;
1803
1804 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1805         if (is_swap_pmd(*pmd)) {
1806                 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1807
1808                 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1809                 if (is_write_migration_entry(entry)) {
1810                         pmd_t newpmd;
1811                         /*
1812                          * A protection check is difficult so
1813                          * just be safe and disable write
1814                          */
1815                         make_migration_entry_read(&entry);
1816                         newpmd = swp_entry_to_pmd(entry);
1817                         if (pmd_swp_soft_dirty(*pmd))
1818                                 newpmd = pmd_swp_mksoft_dirty(newpmd);
1819                         set_pmd_at(mm, addr, pmd, newpmd);
1820                 }
1821                 goto unlock;
1822         }
1823 #endif
1824
1825         /*
1826          * Avoid trapping faults against the zero page. The read-only
1827          * data is likely to be read-cached on the local CPU and
1828          * local/remote hits to the zero page are not interesting.
1829          */
1830         if (prot_numa && is_huge_zero_pmd(*pmd))
1831                 goto unlock;
1832
1833         if (prot_numa && pmd_protnone(*pmd))
1834                 goto unlock;
1835
1836         /*
1837          * In case prot_numa, we are under down_read(mmap_sem). It's critical
1838          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1839          * which is also under down_read(mmap_sem):
1840          *
1841          *      CPU0:                           CPU1:
1842          *                              change_huge_pmd(prot_numa=1)
1843          *                               pmdp_huge_get_and_clear_notify()
1844          * madvise_dontneed()
1845          *  zap_pmd_range()
1846          *   pmd_trans_huge(*pmd) == 0 (without ptl)
1847          *   // skip the pmd
1848          *                               set_pmd_at();
1849          *                               // pmd is re-established
1850          *
1851          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1852          * which may break userspace.
1853          *
1854          * pmdp_invalidate() is required to make sure we don't miss
1855          * dirty/young flags set by hardware.
1856          */
1857         entry = pmdp_invalidate(vma, addr, pmd);
1858
1859         entry = pmd_modify(entry, newprot);
1860         if (preserve_write)
1861                 entry = pmd_mk_savedwrite(entry);
1862         if (uffd_wp) {
1863                 entry = pmd_wrprotect(entry);
1864                 entry = pmd_mkuffd_wp(entry);
1865         } else if (uffd_wp_resolve) {
1866                 /*
1867                  * Leave the write bit to be handled by PF interrupt
1868                  * handler, then things like COW could be properly
1869                  * handled.
1870                  */
1871                 entry = pmd_clear_uffd_wp(entry);
1872         }
1873         ret = HPAGE_PMD_NR;
1874         set_pmd_at(mm, addr, pmd, entry);
1875         BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1876 unlock:
1877         spin_unlock(ptl);
1878         return ret;
1879 }
1880
1881 /*
1882  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1883  *
1884  * Note that if it returns page table lock pointer, this routine returns without
1885  * unlocking page table lock. So callers must unlock it.
1886  */
1887 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1888 {
1889         spinlock_t *ptl;
1890         ptl = pmd_lock(vma->vm_mm, pmd);
1891         if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1892                         pmd_devmap(*pmd)))
1893                 return ptl;
1894         spin_unlock(ptl);
1895         return NULL;
1896 }
1897
1898 /*
1899  * Returns true if a given pud maps a thp, false otherwise.
1900  *
1901  * Note that if it returns true, this routine returns without unlocking page
1902  * table lock. So callers must unlock it.
1903  */
1904 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1905 {
1906         spinlock_t *ptl;
1907
1908         ptl = pud_lock(vma->vm_mm, pud);
1909         if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1910                 return ptl;
1911         spin_unlock(ptl);
1912         return NULL;
1913 }
1914
1915 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1916 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1917                  pud_t *pud, unsigned long addr)
1918 {
1919         spinlock_t *ptl;
1920
1921         ptl = __pud_trans_huge_lock(pud, vma);
1922         if (!ptl)
1923                 return 0;
1924         /*
1925          * For architectures like ppc64 we look at deposited pgtable
1926          * when calling pudp_huge_get_and_clear. So do the
1927          * pgtable_trans_huge_withdraw after finishing pudp related
1928          * operations.
1929          */
1930         pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
1931         tlb_remove_pud_tlb_entry(tlb, pud, addr);
1932         if (vma_is_special_huge(vma)) {
1933                 spin_unlock(ptl);
1934                 /* No zero page support yet */
1935         } else {
1936                 /* No support for anonymous PUD pages yet */
1937                 BUG();
1938         }
1939         return 1;
1940 }
1941
1942 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1943                 unsigned long haddr)
1944 {
1945         VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1946         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1947         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
1948         VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
1949
1950         count_vm_event(THP_SPLIT_PUD);
1951
1952         pudp_huge_clear_flush_notify(vma, haddr, pud);
1953 }
1954
1955 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
1956                 unsigned long address)
1957 {
1958         spinlock_t *ptl;
1959         struct mmu_notifier_range range;
1960
1961         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1962                                 address & HPAGE_PUD_MASK,
1963                                 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
1964         mmu_notifier_invalidate_range_start(&range);
1965         ptl = pud_lock(vma->vm_mm, pud);
1966         if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
1967                 goto out;
1968         __split_huge_pud_locked(vma, pud, range.start);
1969
1970 out:
1971         spin_unlock(ptl);
1972         /*
1973          * No need to double call mmu_notifier->invalidate_range() callback as
1974          * the above pudp_huge_clear_flush_notify() did already call it.
1975          */
1976         mmu_notifier_invalidate_range_only_end(&range);
1977 }
1978 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1979
1980 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1981                 unsigned long haddr, pmd_t *pmd)
1982 {
1983         struct mm_struct *mm = vma->vm_mm;
1984         pgtable_t pgtable;
1985         pmd_t _pmd;
1986         int i;
1987
1988         /*
1989          * Leave pmd empty until pte is filled note that it is fine to delay
1990          * notification until mmu_notifier_invalidate_range_end() as we are
1991          * replacing a zero pmd write protected page with a zero pte write
1992          * protected page.
1993          *
1994          * See Documentation/vm/mmu_notifier.rst
1995          */
1996         pmdp_huge_clear_flush(vma, haddr, pmd);
1997
1998         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1999         pmd_populate(mm, &_pmd, pgtable);
2000
2001         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2002                 pte_t *pte, entry;
2003                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2004                 entry = pte_mkspecial(entry);
2005                 pte = pte_offset_map(&_pmd, haddr);
2006                 VM_BUG_ON(!pte_none(*pte));
2007                 set_pte_at(mm, haddr, pte, entry);
2008                 pte_unmap(pte);
2009         }
2010         smp_wmb(); /* make pte visible before pmd */
2011         pmd_populate(mm, pmd, pgtable);
2012 }
2013
2014 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2015                 unsigned long haddr, bool freeze)
2016 {
2017         struct mm_struct *mm = vma->vm_mm;
2018         struct page *page;
2019         pgtable_t pgtable;
2020         pmd_t old_pmd, _pmd;
2021         bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false;
2022         unsigned long addr;
2023         int i;
2024
2025         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2026         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2027         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2028         VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2029                                 && !pmd_devmap(*pmd));
2030
2031         count_vm_event(THP_SPLIT_PMD);
2032
2033         if (!vma_is_anonymous(vma)) {
2034                 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2035                 /*
2036                  * We are going to unmap this huge page. So
2037                  * just go ahead and zap it
2038                  */
2039                 if (arch_needs_pgtable_deposit())
2040                         zap_deposited_table(mm, pmd);
2041                 if (vma_is_special_huge(vma))
2042                         return;
2043                 page = pmd_page(_pmd);
2044                 if (!PageDirty(page) && pmd_dirty(_pmd))
2045                         set_page_dirty(page);
2046                 if (!PageReferenced(page) && pmd_young(_pmd))
2047                         SetPageReferenced(page);
2048                 page_remove_rmap(page, true);
2049                 put_page(page);
2050                 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2051                 return;
2052         } else if (is_huge_zero_pmd(*pmd)) {
2053                 /*
2054                  * FIXME: Do we want to invalidate secondary mmu by calling
2055                  * mmu_notifier_invalidate_range() see comments below inside
2056                  * __split_huge_pmd() ?
2057                  *
2058                  * We are going from a zero huge page write protected to zero
2059                  * small page also write protected so it does not seems useful
2060                  * to invalidate secondary mmu at this time.
2061                  */
2062                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2063         }
2064
2065         /*
2066          * Up to this point the pmd is present and huge and userland has the
2067          * whole access to the hugepage during the split (which happens in
2068          * place). If we overwrite the pmd with the not-huge version pointing
2069          * to the pte here (which of course we could if all CPUs were bug
2070          * free), userland could trigger a small page size TLB miss on the
2071          * small sized TLB while the hugepage TLB entry is still established in
2072          * the huge TLB. Some CPU doesn't like that.
2073          * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2074          * 383 on page 93. Intel should be safe but is also warns that it's
2075          * only safe if the permission and cache attributes of the two entries
2076          * loaded in the two TLB is identical (which should be the case here).
2077          * But it is generally safer to never allow small and huge TLB entries
2078          * for the same virtual address to be loaded simultaneously. So instead
2079          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2080          * current pmd notpresent (atomically because here the pmd_trans_huge
2081          * must remain set at all times on the pmd until the split is complete
2082          * for this pmd), then we flush the SMP TLB and finally we write the
2083          * non-huge version of the pmd entry with pmd_populate.
2084          */
2085         old_pmd = pmdp_invalidate(vma, haddr, pmd);
2086
2087         pmd_migration = is_pmd_migration_entry(old_pmd);
2088         if (unlikely(pmd_migration)) {
2089                 swp_entry_t entry;
2090
2091                 entry = pmd_to_swp_entry(old_pmd);
2092                 page = pfn_to_page(swp_offset(entry));
2093                 write = is_write_migration_entry(entry);
2094                 young = false;
2095                 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2096                 uffd_wp = pmd_swp_uffd_wp(old_pmd);
2097         } else {
2098                 page = pmd_page(old_pmd);
2099                 if (pmd_dirty(old_pmd))
2100                         SetPageDirty(page);
2101                 write = pmd_write(old_pmd);
2102                 young = pmd_young(old_pmd);
2103                 soft_dirty = pmd_soft_dirty(old_pmd);
2104                 uffd_wp = pmd_uffd_wp(old_pmd);
2105         }
2106         VM_BUG_ON_PAGE(!page_count(page), page);
2107         page_ref_add(page, HPAGE_PMD_NR - 1);
2108
2109         /*
2110          * Withdraw the table only after we mark the pmd entry invalid.
2111          * This's critical for some architectures (Power).
2112          */
2113         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2114         pmd_populate(mm, &_pmd, pgtable);
2115
2116         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2117                 pte_t entry, *pte;
2118                 /*
2119                  * Note that NUMA hinting access restrictions are not
2120                  * transferred to avoid any possibility of altering
2121                  * permissions across VMAs.
2122                  */
2123                 if (freeze || pmd_migration) {
2124                         swp_entry_t swp_entry;
2125                         swp_entry = make_migration_entry(page + i, write);
2126                         entry = swp_entry_to_pte(swp_entry);
2127                         if (soft_dirty)
2128                                 entry = pte_swp_mksoft_dirty(entry);
2129                         if (uffd_wp)
2130                                 entry = pte_swp_mkuffd_wp(entry);
2131                 } else {
2132                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2133                         entry = maybe_mkwrite(entry, vma);
2134                         if (!write)
2135                                 entry = pte_wrprotect(entry);
2136                         if (!young)
2137                                 entry = pte_mkold(entry);
2138                         if (soft_dirty)
2139                                 entry = pte_mksoft_dirty(entry);
2140                         if (uffd_wp)
2141                                 entry = pte_mkuffd_wp(entry);
2142                 }
2143                 pte = pte_offset_map(&_pmd, addr);
2144                 BUG_ON(!pte_none(*pte));
2145                 set_pte_at(mm, addr, pte, entry);
2146                 atomic_inc(&page[i]._mapcount);
2147                 pte_unmap(pte);
2148         }
2149
2150         /*
2151          * Set PG_double_map before dropping compound_mapcount to avoid
2152          * false-negative page_mapped().
2153          */
2154         if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2155                 for (i = 0; i < HPAGE_PMD_NR; i++)
2156                         atomic_inc(&page[i]._mapcount);
2157         }
2158
2159         lock_page_memcg(page);
2160         if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2161                 /* Last compound_mapcount is gone. */
2162                 __dec_lruvec_page_state(page, NR_ANON_THPS);
2163                 if (TestClearPageDoubleMap(page)) {
2164                         /* No need in mapcount reference anymore */
2165                         for (i = 0; i < HPAGE_PMD_NR; i++)
2166                                 atomic_dec(&page[i]._mapcount);
2167                 }
2168         }
2169         unlock_page_memcg(page);
2170
2171         smp_wmb(); /* make pte visible before pmd */
2172         pmd_populate(mm, pmd, pgtable);
2173
2174         if (freeze) {
2175                 for (i = 0; i < HPAGE_PMD_NR; i++) {
2176                         page_remove_rmap(page + i, false);
2177                         put_page(page + i);
2178                 }
2179         }
2180 }
2181
2182 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2183                 unsigned long address, bool freeze, struct page *page)
2184 {
2185         spinlock_t *ptl;
2186         struct mmu_notifier_range range;
2187
2188         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2189                                 address & HPAGE_PMD_MASK,
2190                                 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2191         mmu_notifier_invalidate_range_start(&range);
2192         ptl = pmd_lock(vma->vm_mm, pmd);
2193
2194         /*
2195          * If caller asks to setup a migration entries, we need a page to check
2196          * pmd against. Otherwise we can end up replacing wrong page.
2197          */
2198         VM_BUG_ON(freeze && !page);
2199         if (page && page != pmd_page(*pmd))
2200                 goto out;
2201
2202         if (pmd_trans_huge(*pmd)) {
2203                 page = pmd_page(*pmd);
2204                 if (PageMlocked(page))
2205                         clear_page_mlock(page);
2206         } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2207                 goto out;
2208         __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2209 out:
2210         spin_unlock(ptl);
2211         /*
2212          * No need to double call mmu_notifier->invalidate_range() callback.
2213          * They are 3 cases to consider inside __split_huge_pmd_locked():
2214          *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2215          *  2) __split_huge_zero_page_pmd() read only zero page and any write
2216          *    fault will trigger a flush_notify before pointing to a new page
2217          *    (it is fine if the secondary mmu keeps pointing to the old zero
2218          *    page in the meantime)
2219          *  3) Split a huge pmd into pte pointing to the same page. No need
2220          *     to invalidate secondary tlb entry they are all still valid.
2221          *     any further changes to individual pte will notify. So no need
2222          *     to call mmu_notifier->invalidate_range()
2223          */
2224         mmu_notifier_invalidate_range_only_end(&range);
2225 }
2226
2227 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2228                 bool freeze, struct page *page)
2229 {
2230         pgd_t *pgd;
2231         p4d_t *p4d;
2232         pud_t *pud;
2233         pmd_t *pmd;
2234
2235         pgd = pgd_offset(vma->vm_mm, address);
2236         if (!pgd_present(*pgd))
2237                 return;
2238
2239         p4d = p4d_offset(pgd, address);
2240         if (!p4d_present(*p4d))
2241                 return;
2242
2243         pud = pud_offset(p4d, address);
2244         if (!pud_present(*pud))
2245                 return;
2246
2247         pmd = pmd_offset(pud, address);
2248
2249         __split_huge_pmd(vma, pmd, address, freeze, page);
2250 }
2251
2252 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2253                              unsigned long start,
2254                              unsigned long end,
2255                              long adjust_next)
2256 {
2257         /*
2258          * If the new start address isn't hpage aligned and it could
2259          * previously contain an hugepage: check if we need to split
2260          * an huge pmd.
2261          */
2262         if (start & ~HPAGE_PMD_MASK &&
2263             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2264             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2265                 split_huge_pmd_address(vma, start, false, NULL);
2266
2267         /*
2268          * If the new end address isn't hpage aligned and it could
2269          * previously contain an hugepage: check if we need to split
2270          * an huge pmd.
2271          */
2272         if (end & ~HPAGE_PMD_MASK &&
2273             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2274             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2275                 split_huge_pmd_address(vma, end, false, NULL);
2276
2277         /*
2278          * If we're also updating the vma->vm_next->vm_start, if the new
2279          * vm_next->vm_start isn't page aligned and it could previously
2280          * contain an hugepage: check if we need to split an huge pmd.
2281          */
2282         if (adjust_next > 0) {
2283                 struct vm_area_struct *next = vma->vm_next;
2284                 unsigned long nstart = next->vm_start;
2285                 nstart += adjust_next << PAGE_SHIFT;
2286                 if (nstart & ~HPAGE_PMD_MASK &&
2287                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2288                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2289                         split_huge_pmd_address(next, nstart, false, NULL);
2290         }
2291 }
2292
2293 static void unmap_page(struct page *page)
2294 {
2295         enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
2296                 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2297         bool unmap_success;
2298
2299         VM_BUG_ON_PAGE(!PageHead(page), page);
2300
2301         if (PageAnon(page))
2302                 ttu_flags |= TTU_SPLIT_FREEZE;
2303
2304         unmap_success = try_to_unmap(page, ttu_flags);
2305         VM_BUG_ON_PAGE(!unmap_success, page);
2306 }
2307
2308 static void remap_page(struct page *page)
2309 {
2310         int i;
2311         if (PageTransHuge(page)) {
2312                 remove_migration_ptes(page, page, true);
2313         } else {
2314                 for (i = 0; i < HPAGE_PMD_NR; i++)
2315                         remove_migration_ptes(page + i, page + i, true);
2316         }
2317 }
2318
2319 static void __split_huge_page_tail(struct page *head, int tail,
2320                 struct lruvec *lruvec, struct list_head *list)
2321 {
2322         struct page *page_tail = head + tail;
2323
2324         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2325
2326         /*
2327          * Clone page flags before unfreezing refcount.
2328          *
2329          * After successful get_page_unless_zero() might follow flags change,
2330          * for exmaple lock_page() which set PG_waiters.
2331          */
2332         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2333         page_tail->flags |= (head->flags &
2334                         ((1L << PG_referenced) |
2335                          (1L << PG_swapbacked) |
2336                          (1L << PG_swapcache) |
2337                          (1L << PG_mlocked) |
2338                          (1L << PG_uptodate) |
2339                          (1L << PG_active) |
2340                          (1L << PG_workingset) |
2341                          (1L << PG_locked) |
2342                          (1L << PG_unevictable) |
2343                          (1L << PG_dirty)));
2344
2345         /* ->mapping in first tail page is compound_mapcount */
2346         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2347                         page_tail);
2348         page_tail->mapping = head->mapping;
2349         page_tail->index = head->index + tail;
2350
2351         /* Page flags must be visible before we make the page non-compound. */
2352         smp_wmb();
2353
2354         /*
2355          * Clear PageTail before unfreezing page refcount.
2356          *
2357          * After successful get_page_unless_zero() might follow put_page()
2358          * which needs correct compound_head().
2359          */
2360         clear_compound_head(page_tail);
2361
2362         /* Finally unfreeze refcount. Additional reference from page cache. */
2363         page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2364                                           PageSwapCache(head)));
2365
2366         if (page_is_young(head))
2367                 set_page_young(page_tail);
2368         if (page_is_idle(head))
2369                 set_page_idle(page_tail);
2370
2371         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2372
2373         /*
2374          * always add to the tail because some iterators expect new
2375          * pages to show after the currently processed elements - e.g.
2376          * migrate_pages
2377          */
2378         lru_add_page_tail(head, page_tail, lruvec, list);
2379 }
2380
2381 static void __split_huge_page(struct page *page, struct list_head *list,
2382                 pgoff_t end, unsigned long flags)
2383 {
2384         struct page *head = compound_head(page);
2385         pg_data_t *pgdat = page_pgdat(head);
2386         struct lruvec *lruvec;
2387         struct address_space *swap_cache = NULL;
2388         unsigned long offset = 0;
2389         int i;
2390
2391         lruvec = mem_cgroup_page_lruvec(head, pgdat);
2392
2393         /* complete memcg works before add pages to LRU */
2394         mem_cgroup_split_huge_fixup(head);
2395
2396         if (PageAnon(head) && PageSwapCache(head)) {
2397                 swp_entry_t entry = { .val = page_private(head) };
2398
2399                 offset = swp_offset(entry);
2400                 swap_cache = swap_address_space(entry);
2401                 xa_lock(&swap_cache->i_pages);
2402         }
2403
2404         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
2405                 __split_huge_page_tail(head, i, lruvec, list);
2406                 /* Some pages can be beyond i_size: drop them from page cache */
2407                 if (head[i].index >= end) {
2408                         ClearPageDirty(head + i);
2409                         __delete_from_page_cache(head + i, NULL);
2410                         if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2411                                 shmem_uncharge(head->mapping->host, 1);
2412                         put_page(head + i);
2413                 } else if (!PageAnon(page)) {
2414                         __xa_store(&head->mapping->i_pages, head[i].index,
2415                                         head + i, 0);
2416                 } else if (swap_cache) {
2417                         __xa_store(&swap_cache->i_pages, offset + i,
2418                                         head + i, 0);
2419                 }
2420         }
2421
2422         ClearPageCompound(head);
2423
2424         split_page_owner(head, HPAGE_PMD_ORDER);
2425
2426         /* See comment in __split_huge_page_tail() */
2427         if (PageAnon(head)) {
2428                 /* Additional pin to swap cache */
2429                 if (PageSwapCache(head)) {
2430                         page_ref_add(head, 2);
2431                         xa_unlock(&swap_cache->i_pages);
2432                 } else {
2433                         page_ref_inc(head);
2434                 }
2435         } else {
2436                 /* Additional pin to page cache */
2437                 page_ref_add(head, 2);
2438                 xa_unlock(&head->mapping->i_pages);
2439         }
2440
2441         spin_unlock_irqrestore(&pgdat->lru_lock, flags);
2442
2443         remap_page(head);
2444
2445         for (i = 0; i < HPAGE_PMD_NR; i++) {
2446                 struct page *subpage = head + i;
2447                 if (subpage == page)
2448                         continue;
2449                 unlock_page(subpage);
2450
2451                 /*
2452                  * Subpages may be freed if there wasn't any mapping
2453                  * like if add_to_swap() is running on a lru page that
2454                  * had its mapping zapped. And freeing these pages
2455                  * requires taking the lru_lock so we do the put_page
2456                  * of the tail pages after the split is complete.
2457                  */
2458                 put_page(subpage);
2459         }
2460 }
2461
2462 int total_mapcount(struct page *page)
2463 {
2464         int i, compound, ret;
2465
2466         VM_BUG_ON_PAGE(PageTail(page), page);
2467
2468         if (likely(!PageCompound(page)))
2469                 return atomic_read(&page->_mapcount) + 1;
2470
2471         compound = compound_mapcount(page);
2472         if (PageHuge(page))
2473                 return compound;
2474         ret = compound;
2475         for (i = 0; i < HPAGE_PMD_NR; i++)
2476                 ret += atomic_read(&page[i]._mapcount) + 1;
2477         /* File pages has compound_mapcount included in _mapcount */
2478         if (!PageAnon(page))
2479                 return ret - compound * HPAGE_PMD_NR;
2480         if (PageDoubleMap(page))
2481                 ret -= HPAGE_PMD_NR;
2482         return ret;
2483 }
2484
2485 /*
2486  * This calculates accurately how many mappings a transparent hugepage
2487  * has (unlike page_mapcount() which isn't fully accurate). This full
2488  * accuracy is primarily needed to know if copy-on-write faults can
2489  * reuse the page and change the mapping to read-write instead of
2490  * copying them. At the same time this returns the total_mapcount too.
2491  *
2492  * The function returns the highest mapcount any one of the subpages
2493  * has. If the return value is one, even if different processes are
2494  * mapping different subpages of the transparent hugepage, they can
2495  * all reuse it, because each process is reusing a different subpage.
2496  *
2497  * The total_mapcount is instead counting all virtual mappings of the
2498  * subpages. If the total_mapcount is equal to "one", it tells the
2499  * caller all mappings belong to the same "mm" and in turn the
2500  * anon_vma of the transparent hugepage can become the vma->anon_vma
2501  * local one as no other process may be mapping any of the subpages.
2502  *
2503  * It would be more accurate to replace page_mapcount() with
2504  * page_trans_huge_mapcount(), however we only use
2505  * page_trans_huge_mapcount() in the copy-on-write faults where we
2506  * need full accuracy to avoid breaking page pinning, because
2507  * page_trans_huge_mapcount() is slower than page_mapcount().
2508  */
2509 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2510 {
2511         int i, ret, _total_mapcount, mapcount;
2512
2513         /* hugetlbfs shouldn't call it */
2514         VM_BUG_ON_PAGE(PageHuge(page), page);
2515
2516         if (likely(!PageTransCompound(page))) {
2517                 mapcount = atomic_read(&page->_mapcount) + 1;
2518                 if (total_mapcount)
2519                         *total_mapcount = mapcount;
2520                 return mapcount;
2521         }
2522
2523         page = compound_head(page);
2524
2525         _total_mapcount = ret = 0;
2526         for (i = 0; i < HPAGE_PMD_NR; i++) {
2527                 mapcount = atomic_read(&page[i]._mapcount) + 1;
2528                 ret = max(ret, mapcount);
2529                 _total_mapcount += mapcount;
2530         }
2531         if (PageDoubleMap(page)) {
2532                 ret -= 1;
2533                 _total_mapcount -= HPAGE_PMD_NR;
2534         }
2535         mapcount = compound_mapcount(page);
2536         ret += mapcount;
2537         _total_mapcount += mapcount;
2538         if (total_mapcount)
2539                 *total_mapcount = _total_mapcount;
2540         return ret;
2541 }
2542
2543 /* Racy check whether the huge page can be split */
2544 bool can_split_huge_page(struct page *page, int *pextra_pins)
2545 {
2546         int extra_pins;
2547
2548         /* Additional pins from page cache */
2549         if (PageAnon(page))
2550                 extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0;
2551         else
2552                 extra_pins = HPAGE_PMD_NR;
2553         if (pextra_pins)
2554                 *pextra_pins = extra_pins;
2555         return total_mapcount(page) == page_count(page) - extra_pins - 1;
2556 }
2557
2558 /*
2559  * This function splits huge page into normal pages. @page can point to any
2560  * subpage of huge page to split. Split doesn't change the position of @page.
2561  *
2562  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2563  * The huge page must be locked.
2564  *
2565  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2566  *
2567  * Both head page and tail pages will inherit mapping, flags, and so on from
2568  * the hugepage.
2569  *
2570  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2571  * they are not mapped.
2572  *
2573  * Returns 0 if the hugepage is split successfully.
2574  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2575  * us.
2576  */
2577 int split_huge_page_to_list(struct page *page, struct list_head *list)
2578 {
2579         struct page *head = compound_head(page);
2580         struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2581         struct deferred_split *ds_queue = get_deferred_split_queue(head);
2582         struct anon_vma *anon_vma = NULL;
2583         struct address_space *mapping = NULL;
2584         int count, mapcount, extra_pins, ret;
2585         bool mlocked;
2586         unsigned long flags;
2587         pgoff_t end;
2588
2589         VM_BUG_ON_PAGE(is_huge_zero_page(head), head);
2590         VM_BUG_ON_PAGE(!PageLocked(head), head);
2591         VM_BUG_ON_PAGE(!PageCompound(head), head);
2592
2593         if (PageWriteback(head))
2594                 return -EBUSY;
2595
2596         if (PageAnon(head)) {
2597                 /*
2598                  * The caller does not necessarily hold an mmap_sem that would
2599                  * prevent the anon_vma disappearing so we first we take a
2600                  * reference to it and then lock the anon_vma for write. This
2601                  * is similar to page_lock_anon_vma_read except the write lock
2602                  * is taken to serialise against parallel split or collapse
2603                  * operations.
2604                  */
2605                 anon_vma = page_get_anon_vma(head);
2606                 if (!anon_vma) {
2607                         ret = -EBUSY;
2608                         goto out;
2609                 }
2610                 end = -1;
2611                 mapping = NULL;
2612                 anon_vma_lock_write(anon_vma);
2613         } else {
2614                 mapping = head->mapping;
2615
2616                 /* Truncated ? */
2617                 if (!mapping) {
2618                         ret = -EBUSY;
2619                         goto out;
2620                 }
2621
2622                 anon_vma = NULL;
2623                 i_mmap_lock_read(mapping);
2624
2625                 /*
2626                  *__split_huge_page() may need to trim off pages beyond EOF:
2627                  * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2628                  * which cannot be nested inside the page tree lock. So note
2629                  * end now: i_size itself may be changed at any moment, but
2630                  * head page lock is good enough to serialize the trimming.
2631                  */
2632                 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2633         }
2634
2635         /*
2636          * Racy check if we can split the page, before unmap_page() will
2637          * split PMDs
2638          */
2639         if (!can_split_huge_page(head, &extra_pins)) {
2640                 ret = -EBUSY;
2641                 goto out_unlock;
2642         }
2643
2644         mlocked = PageMlocked(head);
2645         unmap_page(head);
2646         VM_BUG_ON_PAGE(compound_mapcount(head), head);
2647
2648         /* Make sure the page is not on per-CPU pagevec as it takes pin */
2649         if (mlocked)
2650                 lru_add_drain();
2651
2652         /* prevent PageLRU to go away from under us, and freeze lru stats */
2653         spin_lock_irqsave(&pgdata->lru_lock, flags);
2654
2655         if (mapping) {
2656                 XA_STATE(xas, &mapping->i_pages, page_index(head));
2657
2658                 /*
2659                  * Check if the head page is present in page cache.
2660                  * We assume all tail are present too, if head is there.
2661                  */
2662                 xa_lock(&mapping->i_pages);
2663                 if (xas_load(&xas) != head)
2664                         goto fail;
2665         }
2666
2667         /* Prevent deferred_split_scan() touching ->_refcount */
2668         spin_lock(&ds_queue->split_queue_lock);
2669         count = page_count(head);
2670         mapcount = total_mapcount(head);
2671         if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2672                 if (!list_empty(page_deferred_list(head))) {
2673                         ds_queue->split_queue_len--;
2674                         list_del(page_deferred_list(head));
2675                 }
2676                 spin_unlock(&ds_queue->split_queue_lock);
2677                 if (mapping) {
2678                         if (PageSwapBacked(head))
2679                                 __dec_node_page_state(head, NR_SHMEM_THPS);
2680                         else
2681                                 __dec_node_page_state(head, NR_FILE_THPS);
2682                 }
2683
2684                 __split_huge_page(page, list, end, flags);
2685                 if (PageSwapCache(head)) {
2686                         swp_entry_t entry = { .val = page_private(head) };
2687
2688                         ret = split_swap_cluster(entry);
2689                 } else
2690                         ret = 0;
2691         } else {
2692                 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2693                         pr_alert("total_mapcount: %u, page_count(): %u\n",
2694                                         mapcount, count);
2695                         if (PageTail(page))
2696                                 dump_page(head, NULL);
2697                         dump_page(page, "total_mapcount(head) > 0");
2698                         BUG();
2699                 }
2700                 spin_unlock(&ds_queue->split_queue_lock);
2701 fail:           if (mapping)
2702                         xa_unlock(&mapping->i_pages);
2703                 spin_unlock_irqrestore(&pgdata->lru_lock, flags);
2704                 remap_page(head);
2705                 ret = -EBUSY;
2706         }
2707
2708 out_unlock:
2709         if (anon_vma) {
2710                 anon_vma_unlock_write(anon_vma);
2711                 put_anon_vma(anon_vma);
2712         }
2713         if (mapping)
2714                 i_mmap_unlock_read(mapping);
2715 out:
2716         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2717         return ret;
2718 }
2719
2720 void free_transhuge_page(struct page *page)
2721 {
2722         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2723         unsigned long flags;
2724
2725         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2726         if (!list_empty(page_deferred_list(page))) {
2727                 ds_queue->split_queue_len--;
2728                 list_del(page_deferred_list(page));
2729         }
2730         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2731         free_compound_page(page);
2732 }
2733
2734 void deferred_split_huge_page(struct page *page)
2735 {
2736         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2737 #ifdef CONFIG_MEMCG
2738         struct mem_cgroup *memcg = compound_head(page)->mem_cgroup;
2739 #endif
2740         unsigned long flags;
2741
2742         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2743
2744         /*
2745          * The try_to_unmap() in page reclaim path might reach here too,
2746          * this may cause a race condition to corrupt deferred split queue.
2747          * And, if page reclaim is already handling the same page, it is
2748          * unnecessary to handle it again in shrinker.
2749          *
2750          * Check PageSwapCache to determine if the page is being
2751          * handled by page reclaim since THP swap would add the page into
2752          * swap cache before calling try_to_unmap().
2753          */
2754         if (PageSwapCache(page))
2755                 return;
2756
2757         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2758         if (list_empty(page_deferred_list(page))) {
2759                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2760                 list_add_tail(page_deferred_list(page), &ds_queue->split_queue);
2761                 ds_queue->split_queue_len++;
2762 #ifdef CONFIG_MEMCG
2763                 if (memcg)
2764                         memcg_set_shrinker_bit(memcg, page_to_nid(page),
2765                                                deferred_split_shrinker.id);
2766 #endif
2767         }
2768         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2769 }
2770
2771 static unsigned long deferred_split_count(struct shrinker *shrink,
2772                 struct shrink_control *sc)
2773 {
2774         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2775         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2776
2777 #ifdef CONFIG_MEMCG
2778         if (sc->memcg)
2779                 ds_queue = &sc->memcg->deferred_split_queue;
2780 #endif
2781         return READ_ONCE(ds_queue->split_queue_len);
2782 }
2783
2784 static unsigned long deferred_split_scan(struct shrinker *shrink,
2785                 struct shrink_control *sc)
2786 {
2787         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2788         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2789         unsigned long flags;
2790         LIST_HEAD(list), *pos, *next;
2791         struct page *page;
2792         int split = 0;
2793
2794 #ifdef CONFIG_MEMCG
2795         if (sc->memcg)
2796                 ds_queue = &sc->memcg->deferred_split_queue;
2797 #endif
2798
2799         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2800         /* Take pin on all head pages to avoid freeing them under us */
2801         list_for_each_safe(pos, next, &ds_queue->split_queue) {
2802                 page = list_entry((void *)pos, struct page, mapping);
2803                 page = compound_head(page);
2804                 if (get_page_unless_zero(page)) {
2805                         list_move(page_deferred_list(page), &list);
2806                 } else {
2807                         /* We lost race with put_compound_page() */
2808                         list_del_init(page_deferred_list(page));
2809                         ds_queue->split_queue_len--;
2810                 }
2811                 if (!--sc->nr_to_scan)
2812                         break;
2813         }
2814         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2815
2816         list_for_each_safe(pos, next, &list) {
2817                 page = list_entry((void *)pos, struct page, mapping);
2818                 if (!trylock_page(page))
2819                         goto next;
2820                 /* split_huge_page() removes page from list on success */
2821                 if (!split_huge_page(page))
2822                         split++;
2823                 unlock_page(page);
2824 next:
2825                 put_page(page);
2826         }
2827
2828         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2829         list_splice_tail(&list, &ds_queue->split_queue);
2830         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2831
2832         /*
2833          * Stop shrinker if we didn't split any page, but the queue is empty.
2834          * This can happen if pages were freed under us.
2835          */
2836         if (!split && list_empty(&ds_queue->split_queue))
2837                 return SHRINK_STOP;
2838         return split;
2839 }
2840
2841 static struct shrinker deferred_split_shrinker = {
2842         .count_objects = deferred_split_count,
2843         .scan_objects = deferred_split_scan,
2844         .seeks = DEFAULT_SEEKS,
2845         .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
2846                  SHRINKER_NONSLAB,
2847 };
2848
2849 #ifdef CONFIG_DEBUG_FS
2850 static int split_huge_pages_set(void *data, u64 val)
2851 {
2852         struct zone *zone;
2853         struct page *page;
2854         unsigned long pfn, max_zone_pfn;
2855         unsigned long total = 0, split = 0;
2856
2857         if (val != 1)
2858                 return -EINVAL;
2859
2860         for_each_populated_zone(zone) {
2861                 max_zone_pfn = zone_end_pfn(zone);
2862                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2863                         if (!pfn_valid(pfn))
2864                                 continue;
2865
2866                         page = pfn_to_page(pfn);
2867                         if (!get_page_unless_zero(page))
2868                                 continue;
2869
2870                         if (zone != page_zone(page))
2871                                 goto next;
2872
2873                         if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2874                                 goto next;
2875
2876                         total++;
2877                         lock_page(page);
2878                         if (!split_huge_page(page))
2879                                 split++;
2880                         unlock_page(page);
2881 next:
2882                         put_page(page);
2883                 }
2884         }
2885
2886         pr_info("%lu of %lu THP split\n", split, total);
2887
2888         return 0;
2889 }
2890 DEFINE_DEBUGFS_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2891                 "%llu\n");
2892
2893 static int __init split_huge_pages_debugfs(void)
2894 {
2895         debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2896                             &split_huge_pages_fops);
2897         return 0;
2898 }
2899 late_initcall(split_huge_pages_debugfs);
2900 #endif
2901
2902 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2903 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
2904                 struct page *page)
2905 {
2906         struct vm_area_struct *vma = pvmw->vma;
2907         struct mm_struct *mm = vma->vm_mm;
2908         unsigned long address = pvmw->address;
2909         pmd_t pmdval;
2910         swp_entry_t entry;
2911         pmd_t pmdswp;
2912
2913         if (!(pvmw->pmd && !pvmw->pte))
2914                 return;
2915
2916         flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
2917         pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
2918         if (pmd_dirty(pmdval))
2919                 set_page_dirty(page);
2920         entry = make_migration_entry(page, pmd_write(pmdval));
2921         pmdswp = swp_entry_to_pmd(entry);
2922         if (pmd_soft_dirty(pmdval))
2923                 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
2924         set_pmd_at(mm, address, pvmw->pmd, pmdswp);
2925         page_remove_rmap(page, true);
2926         put_page(page);
2927 }
2928
2929 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
2930 {
2931         struct vm_area_struct *vma = pvmw->vma;
2932         struct mm_struct *mm = vma->vm_mm;
2933         unsigned long address = pvmw->address;
2934         unsigned long mmun_start = address & HPAGE_PMD_MASK;
2935         pmd_t pmde;
2936         swp_entry_t entry;
2937
2938         if (!(pvmw->pmd && !pvmw->pte))
2939                 return;
2940
2941         entry = pmd_to_swp_entry(*pvmw->pmd);
2942         get_page(new);
2943         pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
2944         if (pmd_swp_soft_dirty(*pvmw->pmd))
2945                 pmde = pmd_mksoft_dirty(pmde);
2946         if (is_write_migration_entry(entry))
2947                 pmde = maybe_pmd_mkwrite(pmde, vma);
2948
2949         flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
2950         if (PageAnon(new))
2951                 page_add_anon_rmap(new, vma, mmun_start, true);
2952         else
2953                 page_add_file_rmap(new, true);
2954         set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
2955         if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new))
2956                 mlock_vma_page(new);
2957         update_mmu_cache_pmd(vma, address, pvmw->pmd);
2958 }
2959 #endif