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         struct mem_cgroup *memcg;
 591         pgtable_t pgtable;
 592         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 593         vm_fault_t ret = 0;
 594
 595         VM_BUG_ON_PAGE(!PageCompound(page), page);
 596
 597         if (mem_cgroup_try_charge_delay(page, vma->vm_mm, gfp, &memcg)) {
 598                 put_page(page);
 599                 count_vm_event(THP_FAULT_FALLBACK);
 600                 count_vm_event(THP_FAULT_FALLBACK_CHARGE);
 601                 return VM_FAULT_FALLBACK;
 602         }
 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                         mem_cgroup_cancel_charge(page, memcg);
 634                         put_page(page);
 635                         pte_free(vma->vm_mm, pgtable);
 636                         ret2 = handle_userfault(vmf, VM_UFFD_MISSING);
 637                         VM_BUG_ON(ret2 & VM_FAULT_FALLBACK);
 638                         return ret2;
 639                 }
 640
 641                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 642                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 643                 mem_cgroup_commit_charge(page, memcg, false);
 644                 page_add_new_anon_rmap(page, vma, haddr, true);
 645                 lru_cache_add_active_or_unevictable(page, vma);
 646                 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
 647                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
 648                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 649                 mm_inc_nr_ptes(vma->vm_mm);
 650                 spin_unlock(vmf->ptl);
 651                 count_vm_event(THP_FAULT_ALLOC);
 652                 count_memcg_events(memcg, THP_FAULT_ALLOC, 1);
 653         }
 654
 655         return 0;
 656 unlock_release:
 657         spin_unlock(vmf->ptl);
 658 release:
 659         if (pgtable)
 660                 pte_free(vma->vm_mm, pgtable);
 661         mem_cgroup_cancel_charge(page, memcg);
 662         put_page(page);
 663         return ret;
 664
 665 }
 666
 667 /*
 668  * always: directly stall for all thp allocations
 669  * defer: wake kswapd and fail if not immediately available
 670  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
 671  *                fail if not immediately available
 672  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
 673  *          available
 674  * never: never stall for any thp allocation
 675  */
 676 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
 677 {
 678         const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
 679
 680         /* Always do synchronous compaction */
 681         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 682                 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
 683
 684         /* Kick kcompactd and fail quickly */
 685         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 686                 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
 687
 688         /* Synchronous compaction if madvised, otherwise kick kcompactd */
 689         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 690                 return GFP_TRANSHUGE_LIGHT |
 691                         (vma_madvised ? __GFP_DIRECT_RECLAIM :
 692                                         __GFP_KSWAPD_RECLAIM);
 693
 694         /* Only do synchronous compaction if madvised */
 695         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 696                 return GFP_TRANSHUGE_LIGHT |
 697                        (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
 698
 699         return GFP_TRANSHUGE_LIGHT;
 700 }
 701
 702 /* Caller must hold page table lock. */
 703 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 704                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 705                 struct page *zero_page)
 706 {
 707         pmd_t entry;
 708         if (!pmd_none(*pmd))
 709                 return false;
 710         entry = mk_pmd(zero_page, vma->vm_page_prot);
 711         entry = pmd_mkhuge(entry);
 712         if (pgtable)
 713                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 714         set_pmd_at(mm, haddr, pmd, entry);
 715         mm_inc_nr_ptes(mm);
 716         return true;
 717 }
 718
 719 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
 720 {
 721         struct vm_area_struct *vma = vmf->vma;
 722         gfp_t gfp;
 723         struct page *page;
 724         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 725
 726         if (!transhuge_vma_suitable(vma, haddr))
 727                 return VM_FAULT_FALLBACK;
 728         if (unlikely(anon_vma_prepare(vma)))
 729                 return VM_FAULT_OOM;
 730         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 731                 return VM_FAULT_OOM;
 732         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
 733                         !mm_forbids_zeropage(vma->vm_mm) &&
 734                         transparent_hugepage_use_zero_page()) {
 735                 pgtable_t pgtable;
 736                 struct page *zero_page;
 737                 bool set;
 738                 vm_fault_t ret;
 739                 pgtable = pte_alloc_one(vma->vm_mm);
 740                 if (unlikely(!pgtable))
 741                         return VM_FAULT_OOM;
 742                 zero_page = mm_get_huge_zero_page(vma->vm_mm);
 743                 if (unlikely(!zero_page)) {
 744                         pte_free(vma->vm_mm, pgtable);
 745                         count_vm_event(THP_FAULT_FALLBACK);
 746                         return VM_FAULT_FALLBACK;
 747                 }
 748                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 749                 ret = 0;
 750                 set = false;
 751                 if (pmd_none(*vmf->pmd)) {
 752                         ret = check_stable_address_space(vma->vm_mm);
 753                         if (ret) {
 754                                 spin_unlock(vmf->ptl);
 755                         } else if (userfaultfd_missing(vma)) {
 756                                 spin_unlock(vmf->ptl);
 757                                 ret = handle_userfault(vmf, VM_UFFD_MISSING);
 758                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 759                         } else {
 760                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
 761                                                    haddr, vmf->pmd, zero_page);
 762                                 spin_unlock(vmf->ptl);
 763                                 set = true;
 764                         }
 765                 } else
 766                         spin_unlock(vmf->ptl);
 767                 if (!set)
 768                         pte_free(vma->vm_mm, pgtable);
 769                 return ret;
 770         }
 771         gfp = alloc_hugepage_direct_gfpmask(vma);
 772         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 773         if (unlikely(!page)) {
 774                 count_vm_event(THP_FAULT_FALLBACK);
 775                 return VM_FAULT_FALLBACK;
 776         }
 777         prep_transhuge_page(page);
 778         return __do_huge_pmd_anonymous_page(vmf, page, gfp);
 779 }
 780
 781 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 782                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
 783                 pgtable_t pgtable)
 784 {
 785         struct mm_struct *mm = vma->vm_mm;
 786         pmd_t entry;
 787         spinlock_t *ptl;
 788
 789         ptl = pmd_lock(mm, pmd);
 790         if (!pmd_none(*pmd)) {
 791                 if (write) {
 792                         if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
 793                                 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
 794                                 goto out_unlock;
 795                         }
 796                         entry = pmd_mkyoung(*pmd);
 797                         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 798                         if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
 799                                 update_mmu_cache_pmd(vma, addr, pmd);
 800                 }
 801
 802                 goto out_unlock;
 803         }
 804
 805         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 806         if (pfn_t_devmap(pfn))
 807                 entry = pmd_mkdevmap(entry);
 808         if (write) {
 809                 entry = pmd_mkyoung(pmd_mkdirty(entry));
 810                 entry = maybe_pmd_mkwrite(entry, vma);
 811         }
 812
 813         if (pgtable) {
 814                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 815                 mm_inc_nr_ptes(mm);
 816                 pgtable = NULL;
 817         }
 818
 819         set_pmd_at(mm, addr, pmd, entry);
 820         update_mmu_cache_pmd(vma, addr, pmd);
 821
 822 out_unlock:
 823         spin_unlock(ptl);
 824         if (pgtable)
 825                 pte_free(mm, pgtable);
 826 }
 827
 828 /**
 829  * vmf_insert_pfn_pmd_prot - insert a pmd size pfn
 830  * @vmf: Structure describing the fault
 831  * @pfn: pfn to insert
 832  * @pgprot: page protection to use
 833  * @write: whether it's a write fault
 834  *
 835  * Insert a pmd size pfn. See vmf_insert_pfn() for additional info and
 836  * also consult the vmf_insert_mixed_prot() documentation when
 837  * @pgprot != @vmf->vma->vm_page_prot.
 838  *
 839  * Return: vm_fault_t value.
 840  */
 841 vm_fault_t vmf_insert_pfn_pmd_prot(struct vm_fault *vmf, pfn_t pfn,
 842                                    pgprot_t pgprot, bool write)
 843 {
 844         unsigned long addr = vmf->address & PMD_MASK;
 845         struct vm_area_struct *vma = vmf->vma;
 846         pgtable_t pgtable = NULL;
 847
 848         /*
 849          * If we had pmd_special, we could avoid all these restrictions,
 850          * but we need to be consistent with PTEs and architectures that
 851          * can't support a 'special' bit.
 852          */
 853         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 854                         !pfn_t_devmap(pfn));
 855         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 856                                                 (VM_PFNMAP|VM_MIXEDMAP));
 857         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 858
 859         if (addr < vma->vm_start || addr >= vma->vm_end)
 860                 return VM_FAULT_SIGBUS;
 861
 862         if (arch_needs_pgtable_deposit()) {
 863                 pgtable = pte_alloc_one(vma->vm_mm);
 864                 if (!pgtable)
 865                         return VM_FAULT_OOM;
 866         }
 867
 868         track_pfn_insert(vma, &pgprot, pfn);
 869
 870         insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
 871         return VM_FAULT_NOPAGE;
 872 }
 873 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd_prot);
 874
 875 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 876 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
 877 {
 878         if (likely(vma->vm_flags & VM_WRITE))
 879                 pud = pud_mkwrite(pud);
 880         return pud;
 881 }
 882
 883 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 884                 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
 885 {
 886         struct mm_struct *mm = vma->vm_mm;
 887         pud_t entry;
 888         spinlock_t *ptl;
 889
 890         ptl = pud_lock(mm, pud);
 891         if (!pud_none(*pud)) {
 892                 if (write) {
 893                         if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
 894                                 WARN_ON_ONCE(!is_huge_zero_pud(*pud));
 895                                 goto out_unlock;
 896                         }
 897                         entry = pud_mkyoung(*pud);
 898                         entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
 899                         if (pudp_set_access_flags(vma, addr, pud, entry, 1))
 900                                 update_mmu_cache_pud(vma, addr, pud);
 901                 }
 902                 goto out_unlock;
 903         }
 904
 905         entry = pud_mkhuge(pfn_t_pud(pfn, prot));
 906         if (pfn_t_devmap(pfn))
 907                 entry = pud_mkdevmap(entry);
 908         if (write) {
 909                 entry = pud_mkyoung(pud_mkdirty(entry));
 910                 entry = maybe_pud_mkwrite(entry, vma);
 911         }
 912         set_pud_at(mm, addr, pud, entry);
 913         update_mmu_cache_pud(vma, addr, pud);
 914
 915 out_unlock:
 916         spin_unlock(ptl);
 917 }
 918
 919 /**
 920  * vmf_insert_pfn_pud_prot - insert a pud size pfn
 921  * @vmf: Structure describing the fault
 922  * @pfn: pfn to insert
 923  * @pgprot: page protection to use
 924  * @write: whether it's a write fault
 925  *
 926  * Insert a pud size pfn. See vmf_insert_pfn() for additional info and
 927  * also consult the vmf_insert_mixed_prot() documentation when
 928  * @pgprot != @vmf->vma->vm_page_prot.
 929  *
 930  * Return: vm_fault_t value.
 931  */
 932 vm_fault_t vmf_insert_pfn_pud_prot(struct vm_fault *vmf, pfn_t pfn,
 933                                    pgprot_t pgprot, bool write)
 934 {
 935         unsigned long addr = vmf->address & PUD_MASK;
 936         struct vm_area_struct *vma = vmf->vma;
 937
 938         /*
 939          * If we had pud_special, we could avoid all these restrictions,
 940          * but we need to be consistent with PTEs and architectures that
 941          * can't support a 'special' bit.
 942          */
 943         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 944                         !pfn_t_devmap(pfn));
 945         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 946                                                 (VM_PFNMAP|VM_MIXEDMAP));
 947         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 948
 949         if (addr < vma->vm_start || addr >= vma->vm_end)
 950                 return VM_FAULT_SIGBUS;
 951
 952         track_pfn_insert(vma, &pgprot, pfn);
 953
 954         insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
 955         return VM_FAULT_NOPAGE;
 956 }
 957 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud_prot);
 958 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
 959
 960 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
 961                 pmd_t *pmd, int flags)
 962 {
 963         pmd_t _pmd;
 964
 965         _pmd = pmd_mkyoung(*pmd);
 966         if (flags & FOLL_WRITE)
 967                 _pmd = pmd_mkdirty(_pmd);
 968         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
 969                                 pmd, _pmd, flags & FOLL_WRITE))
 970                 update_mmu_cache_pmd(vma, addr, pmd);
 971 }
 972
 973 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 974                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
 975 {
 976         unsigned long pfn = pmd_pfn(*pmd);
 977         struct mm_struct *mm = vma->vm_mm;
 978         struct page *page;
 979
 980         assert_spin_locked(pmd_lockptr(mm, pmd));
 981
 982         /*
 983          * When we COW a devmap PMD entry, we split it into PTEs, so we should
 984          * not be in this function with `flags & FOLL_COW` set.
 985          */
 986         WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
 987
 988         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
 989         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
 990                          (FOLL_PIN | FOLL_GET)))
 991                 return NULL;
 992
 993         if (flags & FOLL_WRITE && !pmd_write(*pmd))
 994                 return NULL;
 995
 996         if (pmd_present(*pmd) && pmd_devmap(*pmd))
 997                 /* pass */;
 998         else
 999                 return NULL;
1000
1001         if (flags & FOLL_TOUCH)
1002                 touch_pmd(vma, addr, pmd, flags);
1003
1004         /*
1005          * device mapped pages can only be returned if the
1006          * caller will manage the page reference count.
1007          */
1008         if (!(flags & (FOLL_GET | FOLL_PIN)))
1009                 return ERR_PTR(-EEXIST);
1010
1011         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1012         *pgmap = get_dev_pagemap(pfn, *pgmap);
1013         if (!*pgmap)
1014                 return ERR_PTR(-EFAULT);
1015         page = pfn_to_page(pfn);
1016         if (!try_grab_page(page, flags))
1017                 page = ERR_PTR(-ENOMEM);
1018
1019         return page;
1020 }
1021
1022 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1023                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1024                   struct vm_area_struct *vma)
1025 {
1026         spinlock_t *dst_ptl, *src_ptl;
1027         struct page *src_page;
1028         pmd_t pmd;
1029         pgtable_t pgtable = NULL;
1030         int ret = -ENOMEM;
1031
1032         /* Skip if can be re-fill on fault */
1033         if (!vma_is_anonymous(vma))
1034                 return 0;
1035
1036         pgtable = pte_alloc_one(dst_mm);
1037         if (unlikely(!pgtable))
1038                 goto out;
1039
1040         dst_ptl = pmd_lock(dst_mm, dst_pmd);
1041         src_ptl = pmd_lockptr(src_mm, src_pmd);
1042         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1043
1044         ret = -EAGAIN;
1045         pmd = *src_pmd;
1046
1047         /*
1048          * Make sure the _PAGE_UFFD_WP bit is cleared if the new VMA
1049          * does not have the VM_UFFD_WP, which means that the uffd
1050          * fork event is not enabled.
1051          */
1052         if (!(vma->vm_flags & VM_UFFD_WP))
1053                 pmd = pmd_clear_uffd_wp(pmd);
1054
1055 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1056         if (unlikely(is_swap_pmd(pmd))) {
1057                 swp_entry_t entry = pmd_to_swp_entry(pmd);
1058
1059                 VM_BUG_ON(!is_pmd_migration_entry(pmd));
1060                 if (is_write_migration_entry(entry)) {
1061                         make_migration_entry_read(&entry);
1062                         pmd = swp_entry_to_pmd(entry);
1063                         if (pmd_swp_soft_dirty(*src_pmd))
1064                                 pmd = pmd_swp_mksoft_dirty(pmd);
1065                         set_pmd_at(src_mm, addr, src_pmd, pmd);
1066                 }
1067                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1068                 mm_inc_nr_ptes(dst_mm);
1069                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1070                 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1071                 ret = 0;
1072                 goto out_unlock;
1073         }
1074 #endif
1075
1076         if (unlikely(!pmd_trans_huge(pmd))) {
1077                 pte_free(dst_mm, pgtable);
1078                 goto out_unlock;
1079         }
1080         /*
1081          * When page table lock is held, the huge zero pmd should not be
1082          * under splitting since we don't split the page itself, only pmd to
1083          * a page table.
1084          */
1085         if (is_huge_zero_pmd(pmd)) {
1086                 struct page *zero_page;
1087                 /*
1088                  * get_huge_zero_page() will never allocate a new page here,
1089                  * since we already have a zero page to copy. It just takes a
1090                  * reference.
1091                  */
1092                 zero_page = mm_get_huge_zero_page(dst_mm);
1093                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1094                                 zero_page);
1095                 ret = 0;
1096                 goto out_unlock;
1097         }
1098
1099         src_page = pmd_page(pmd);
1100         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1101         get_page(src_page);
1102         page_dup_rmap(src_page, true);
1103         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1104         mm_inc_nr_ptes(dst_mm);
1105         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1106
1107         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1108         pmd = pmd_mkold(pmd_wrprotect(pmd));
1109         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1110
1111         ret = 0;
1112 out_unlock:
1113         spin_unlock(src_ptl);
1114         spin_unlock(dst_ptl);
1115 out:
1116         return ret;
1117 }
1118
1119 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1120 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1121                 pud_t *pud, int flags)
1122 {
1123         pud_t _pud;
1124
1125         _pud = pud_mkyoung(*pud);
1126         if (flags & FOLL_WRITE)
1127                 _pud = pud_mkdirty(_pud);
1128         if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1129                                 pud, _pud, flags & FOLL_WRITE))
1130                 update_mmu_cache_pud(vma, addr, pud);
1131 }
1132
1133 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1134                 pud_t *pud, int flags, struct dev_pagemap **pgmap)
1135 {
1136         unsigned long pfn = pud_pfn(*pud);
1137         struct mm_struct *mm = vma->vm_mm;
1138         struct page *page;
1139
1140         assert_spin_locked(pud_lockptr(mm, pud));
1141
1142         if (flags & FOLL_WRITE && !pud_write(*pud))
1143                 return NULL;
1144
1145         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
1146         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
1147                          (FOLL_PIN | FOLL_GET)))
1148                 return NULL;
1149
1150         if (pud_present(*pud) && pud_devmap(*pud))
1151                 /* pass */;
1152         else
1153                 return NULL;
1154
1155         if (flags & FOLL_TOUCH)
1156                 touch_pud(vma, addr, pud, flags);
1157
1158         /*
1159          * device mapped pages can only be returned if the
1160          * caller will manage the page reference count.
1161          *
1162          * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here:
1163          */
1164         if (!(flags & (FOLL_GET | FOLL_PIN)))
1165                 return ERR_PTR(-EEXIST);
1166
1167         pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1168         *pgmap = get_dev_pagemap(pfn, *pgmap);
1169         if (!*pgmap)
1170                 return ERR_PTR(-EFAULT);
1171         page = pfn_to_page(pfn);
1172         if (!try_grab_page(page, flags))
1173                 page = ERR_PTR(-ENOMEM);
1174
1175         return page;
1176 }
1177
1178 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1179                   pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1180                   struct vm_area_struct *vma)
1181 {
1182         spinlock_t *dst_ptl, *src_ptl;
1183         pud_t pud;
1184         int ret;
1185
1186         dst_ptl = pud_lock(dst_mm, dst_pud);
1187         src_ptl = pud_lockptr(src_mm, src_pud);
1188         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1189
1190         ret = -EAGAIN;
1191         pud = *src_pud;
1192         if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1193                 goto out_unlock;
1194
1195         /*
1196          * When page table lock is held, the huge zero pud should not be
1197          * under splitting since we don't split the page itself, only pud to
1198          * a page table.
1199          */
1200         if (is_huge_zero_pud(pud)) {
1201                 /* No huge zero pud yet */
1202         }
1203
1204         pudp_set_wrprotect(src_mm, addr, src_pud);
1205         pud = pud_mkold(pud_wrprotect(pud));
1206         set_pud_at(dst_mm, addr, dst_pud, pud);
1207
1208         ret = 0;
1209 out_unlock:
1210         spin_unlock(src_ptl);
1211         spin_unlock(dst_ptl);
1212         return ret;
1213 }
1214
1215 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1216 {
1217         pud_t entry;
1218         unsigned long haddr;
1219         bool write = vmf->flags & FAULT_FLAG_WRITE;
1220
1221         vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1222         if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1223                 goto unlock;
1224
1225         entry = pud_mkyoung(orig_pud);
1226         if (write)
1227                 entry = pud_mkdirty(entry);
1228         haddr = vmf->address & HPAGE_PUD_MASK;
1229         if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1230                 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1231
1232 unlock:
1233         spin_unlock(vmf->ptl);
1234 }
1235 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1236
1237 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1238 {
1239         pmd_t entry;
1240         unsigned long haddr;
1241         bool write = vmf->flags & FAULT_FLAG_WRITE;
1242
1243         vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1244         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1245                 goto unlock;
1246
1247         entry = pmd_mkyoung(orig_pmd);
1248         if (write)
1249                 entry = pmd_mkdirty(entry);
1250         haddr = vmf->address & HPAGE_PMD_MASK;
1251         if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1252                 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1253
1254 unlock:
1255         spin_unlock(vmf->ptl);
1256 }
1257
1258 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1259 {
1260         struct vm_area_struct *vma = vmf->vma;
1261         struct page *page;
1262         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1263
1264         vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1265         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1266
1267         if (is_huge_zero_pmd(orig_pmd))
1268                 goto fallback;
1269
1270         spin_lock(vmf->ptl);
1271
1272         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1273                 spin_unlock(vmf->ptl);
1274                 return 0;
1275         }
1276
1277         page = pmd_page(orig_pmd);
1278         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1279
1280         /* Lock page for reuse_swap_page() */
1281         if (!trylock_page(page)) {
1282                 get_page(page);
1283                 spin_unlock(vmf->ptl);
1284                 lock_page(page);
1285                 spin_lock(vmf->ptl);
1286                 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1287                         spin_unlock(vmf->ptl);
1288                         unlock_page(page);
1289                         put_page(page);
1290                         return 0;
1291                 }
1292                 put_page(page);
1293         }
1294
1295         /*
1296          * We can only reuse the page if nobody else maps the huge page or it's
1297          * part.
1298          */
1299         if (reuse_swap_page(page, NULL)) {
1300                 pmd_t entry;
1301                 entry = pmd_mkyoung(orig_pmd);
1302                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1303                 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1304                         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1305                 unlock_page(page);
1306                 spin_unlock(vmf->ptl);
1307                 return VM_FAULT_WRITE;
1308         }
1309
1310         unlock_page(page);
1311         spin_unlock(vmf->ptl);
1312 fallback:
1313         __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL);
1314         return VM_FAULT_FALLBACK;
1315 }
1316
1317 /*
1318  * FOLL_FORCE or a forced COW break can write even to unwritable pmd's,
1319  * but only after we've gone through a COW cycle and they are dirty.
1320  */
1321 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1322 {
1323         return pmd_write(pmd) || ((flags & FOLL_COW) && pmd_dirty(pmd));
1324 }
1325
1326 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1327                                    unsigned long addr,
1328                                    pmd_t *pmd,
1329                                    unsigned int flags)
1330 {
1331         struct mm_struct *mm = vma->vm_mm;
1332         struct page *page = NULL;
1333
1334         assert_spin_locked(pmd_lockptr(mm, pmd));
1335
1336         if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1337                 goto out;
1338
1339         /* Avoid dumping huge zero page */
1340         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1341                 return ERR_PTR(-EFAULT);
1342
1343         /* Full NUMA hinting faults to serialise migration in fault paths */
1344         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1345                 goto out;
1346
1347         page = pmd_page(*pmd);
1348         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1349
1350         if (!try_grab_page(page, flags))
1351                 return ERR_PTR(-ENOMEM);
1352
1353         if (flags & FOLL_TOUCH)
1354                 touch_pmd(vma, addr, pmd, flags);
1355
1356         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1357                 /*
1358                  * We don't mlock() pte-mapped THPs. This way we can avoid
1359                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1360                  *
1361                  * For anon THP:
1362                  *
1363                  * In most cases the pmd is the only mapping of the page as we
1364                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1365                  * writable private mappings in populate_vma_page_range().
1366                  *
1367                  * The only scenario when we have the page shared here is if we
1368                  * mlocking read-only mapping shared over fork(). We skip
1369                  * mlocking such pages.
1370                  *
1371                  * For file THP:
1372                  *
1373                  * We can expect PageDoubleMap() to be stable under page lock:
1374                  * for file pages we set it in page_add_file_rmap(), which
1375                  * requires page to be locked.
1376                  */
1377
1378                 if (PageAnon(page) && compound_mapcount(page) != 1)
1379                         goto skip_mlock;
1380                 if (PageDoubleMap(page) || !page->mapping)
1381                         goto skip_mlock;
1382                 if (!trylock_page(page))
1383                         goto skip_mlock;
1384                 lru_add_drain();
1385                 if (page->mapping && !PageDoubleMap(page))
1386                         mlock_vma_page(page);
1387                 unlock_page(page);
1388         }
1389 skip_mlock:
1390         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1391         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1392
1393 out:
1394         return page;
1395 }
1396
1397 /* NUMA hinting page fault entry point for trans huge pmds */
1398 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1399 {
1400         struct vm_area_struct *vma = vmf->vma;
1401         struct anon_vma *anon_vma = NULL;
1402         struct page *page;
1403         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1404         int page_nid = NUMA_NO_NODE, this_nid = numa_node_id();
1405         int target_nid, last_cpupid = -1;
1406         bool page_locked;
1407         bool migrated = false;
1408         bool was_writable;
1409         int flags = 0;
1410
1411         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1412         if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1413                 goto out_unlock;
1414
1415         /*
1416          * If there are potential migrations, wait for completion and retry
1417          * without disrupting NUMA hinting information. Do not relock and
1418          * check_same as the page may no longer be mapped.
1419          */
1420         if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1421                 page = pmd_page(*vmf->pmd);
1422                 if (!get_page_unless_zero(page))
1423                         goto out_unlock;
1424                 spin_unlock(vmf->ptl);
1425                 put_and_wait_on_page_locked(page);
1426                 goto out;
1427         }
1428
1429         page = pmd_page(pmd);
1430         BUG_ON(is_huge_zero_page(page));
1431         page_nid = page_to_nid(page);
1432         last_cpupid = page_cpupid_last(page);
1433         count_vm_numa_event(NUMA_HINT_FAULTS);
1434         if (page_nid == this_nid) {
1435                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1436                 flags |= TNF_FAULT_LOCAL;
1437         }
1438
1439         /* See similar comment in do_numa_page for explanation */
1440         if (!pmd_savedwrite(pmd))
1441                 flags |= TNF_NO_GROUP;
1442
1443         /*
1444          * Acquire the page lock to serialise THP migrations but avoid dropping
1445          * page_table_lock if at all possible
1446          */
1447         page_locked = trylock_page(page);
1448         target_nid = mpol_misplaced(page, vma, haddr);
1449         if (target_nid == NUMA_NO_NODE) {
1450                 /* If the page was locked, there are no parallel migrations */
1451                 if (page_locked)
1452                         goto clear_pmdnuma;
1453         }
1454
1455         /* Migration could have started since the pmd_trans_migrating check */
1456         if (!page_locked) {
1457                 page_nid = NUMA_NO_NODE;
1458                 if (!get_page_unless_zero(page))
1459                         goto out_unlock;
1460                 spin_unlock(vmf->ptl);
1461                 put_and_wait_on_page_locked(page);
1462                 goto out;
1463         }
1464
1465         /*
1466          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1467          * to serialises splits
1468          */
1469         get_page(page);
1470         spin_unlock(vmf->ptl);
1471         anon_vma = page_lock_anon_vma_read(page);
1472
1473         /* Confirm the PMD did not change while page_table_lock was released */
1474         spin_lock(vmf->ptl);
1475         if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1476                 unlock_page(page);
1477                 put_page(page);
1478                 page_nid = NUMA_NO_NODE;
1479                 goto out_unlock;
1480         }
1481
1482         /* Bail if we fail to protect against THP splits for any reason */
1483         if (unlikely(!anon_vma)) {
1484                 put_page(page);
1485                 page_nid = NUMA_NO_NODE;
1486                 goto clear_pmdnuma;
1487         }
1488
1489         /*
1490          * Since we took the NUMA fault, we must have observed the !accessible
1491          * bit. Make sure all other CPUs agree with that, to avoid them
1492          * modifying the page we're about to migrate.
1493          *
1494          * Must be done under PTL such that we'll observe the relevant
1495          * inc_tlb_flush_pending().
1496          *
1497          * We are not sure a pending tlb flush here is for a huge page
1498          * mapping or not. Hence use the tlb range variant
1499          */
1500         if (mm_tlb_flush_pending(vma->vm_mm)) {
1501                 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1502                 /*
1503                  * change_huge_pmd() released the pmd lock before
1504                  * invalidating the secondary MMUs sharing the primary
1505                  * MMU pagetables (with ->invalidate_range()). The
1506                  * mmu_notifier_invalidate_range_end() (which
1507                  * internally calls ->invalidate_range()) in
1508                  * change_pmd_range() will run after us, so we can't
1509                  * rely on it here and we need an explicit invalidate.
1510                  */
1511                 mmu_notifier_invalidate_range(vma->vm_mm, haddr,
1512                                               haddr + HPAGE_PMD_SIZE);
1513         }
1514
1515         /*
1516          * Migrate the THP to the requested node, returns with page unlocked
1517          * and access rights restored.
1518          */
1519         spin_unlock(vmf->ptl);
1520
1521         migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1522                                 vmf->pmd, pmd, vmf->address, page, target_nid);
1523         if (migrated) {
1524                 flags |= TNF_MIGRATED;
1525                 page_nid = target_nid;
1526         } else
1527                 flags |= TNF_MIGRATE_FAIL;
1528
1529         goto out;
1530 clear_pmdnuma:
1531         BUG_ON(!PageLocked(page));
1532         was_writable = pmd_savedwrite(pmd);
1533         pmd = pmd_modify(pmd, vma->vm_page_prot);
1534         pmd = pmd_mkyoung(pmd);
1535         if (was_writable)
1536                 pmd = pmd_mkwrite(pmd);
1537         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1538         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1539         unlock_page(page);
1540 out_unlock:
1541         spin_unlock(vmf->ptl);
1542
1543 out:
1544         if (anon_vma)
1545                 page_unlock_anon_vma_read(anon_vma);
1546
1547         if (page_nid != NUMA_NO_NODE)
1548                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1549                                 flags);
1550
1551         return 0;
1552 }
1553
1554 /*
1555  * Return true if we do MADV_FREE successfully on entire pmd page.
1556  * Otherwise, return false.
1557  */
1558 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1559                 pmd_t *pmd, unsigned long addr, unsigned long next)
1560 {
1561         spinlock_t *ptl;
1562         pmd_t orig_pmd;
1563         struct page *page;
1564         struct mm_struct *mm = tlb->mm;
1565         bool ret = false;
1566
1567         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1568
1569         ptl = pmd_trans_huge_lock(pmd, vma);
1570         if (!ptl)
1571                 goto out_unlocked;
1572
1573         orig_pmd = *pmd;
1574         if (is_huge_zero_pmd(orig_pmd))
1575                 goto out;
1576
1577         if (unlikely(!pmd_present(orig_pmd))) {
1578                 VM_BUG_ON(thp_migration_supported() &&
1579                                   !is_pmd_migration_entry(orig_pmd));
1580                 goto out;
1581         }
1582
1583         page = pmd_page(orig_pmd);
1584         /*
1585          * If other processes are mapping this page, we couldn't discard
1586          * the page unless they all do MADV_FREE so let's skip the page.
1587          */
1588         if (page_mapcount(page) != 1)
1589                 goto out;
1590
1591         if (!trylock_page(page))
1592                 goto out;
1593
1594         /*
1595          * If user want to discard part-pages of THP, split it so MADV_FREE
1596          * will deactivate only them.
1597          */
1598         if (next - addr != HPAGE_PMD_SIZE) {
1599                 get_page(page);
1600                 spin_unlock(ptl);
1601                 split_huge_page(page);
1602                 unlock_page(page);
1603                 put_page(page);
1604                 goto out_unlocked;
1605         }
1606
1607         if (PageDirty(page))
1608                 ClearPageDirty(page);
1609         unlock_page(page);
1610
1611         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1612                 pmdp_invalidate(vma, addr, pmd);
1613                 orig_pmd = pmd_mkold(orig_pmd);
1614                 orig_pmd = pmd_mkclean(orig_pmd);
1615
1616                 set_pmd_at(mm, addr, pmd, orig_pmd);
1617                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1618         }
1619
1620         mark_page_lazyfree(page);
1621         ret = true;
1622 out:
1623         spin_unlock(ptl);
1624 out_unlocked:
1625         return ret;
1626 }
1627
1628 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1629 {
1630         pgtable_t pgtable;
1631
1632         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1633         pte_free(mm, pgtable);
1634         mm_dec_nr_ptes(mm);
1635 }
1636
1637 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1638                  pmd_t *pmd, unsigned long addr)
1639 {
1640         pmd_t orig_pmd;
1641         spinlock_t *ptl;
1642
1643         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1644
1645         ptl = __pmd_trans_huge_lock(pmd, vma);
1646         if (!ptl)
1647                 return 0;
1648         /*
1649          * For architectures like ppc64 we look at deposited pgtable
1650          * when calling pmdp_huge_get_and_clear. So do the
1651          * pgtable_trans_huge_withdraw after finishing pmdp related
1652          * operations.
1653          */
1654         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1655                         tlb->fullmm);
1656         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1657         if (vma_is_special_huge(vma)) {
1658                 if (arch_needs_pgtable_deposit())
1659                         zap_deposited_table(tlb->mm, pmd);
1660                 spin_unlock(ptl);
1661                 if (is_huge_zero_pmd(orig_pmd))
1662                         tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1663         } else if (is_huge_zero_pmd(orig_pmd)) {
1664                 zap_deposited_table(tlb->mm, pmd);
1665                 spin_unlock(ptl);
1666                 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1667         } else {
1668                 struct page *page = NULL;
1669                 int flush_needed = 1;
1670
1671                 if (pmd_present(orig_pmd)) {
1672                         page = pmd_page(orig_pmd);
1673                         page_remove_rmap(page, true);
1674                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1675                         VM_BUG_ON_PAGE(!PageHead(page), page);
1676                 } else if (thp_migration_supported()) {
1677                         swp_entry_t entry;
1678
1679                         VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1680                         entry = pmd_to_swp_entry(orig_pmd);
1681                         page = pfn_to_page(swp_offset(entry));
1682                         flush_needed = 0;
1683                 } else
1684                         WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1685
1686                 if (PageAnon(page)) {
1687                         zap_deposited_table(tlb->mm, pmd);
1688                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1689                 } else {
1690                         if (arch_needs_pgtable_deposit())
1691                                 zap_deposited_table(tlb->mm, pmd);
1692                         add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1693                 }
1694
1695                 spin_unlock(ptl);
1696                 if (flush_needed)
1697                         tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1698         }
1699         return 1;
1700 }
1701
1702 #ifndef pmd_move_must_withdraw
1703 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1704                                          spinlock_t *old_pmd_ptl,
1705                                          struct vm_area_struct *vma)
1706 {
1707         /*
1708          * With split pmd lock we also need to move preallocated
1709          * PTE page table if new_pmd is on different PMD page table.
1710          *
1711          * We also don't deposit and withdraw tables for file pages.
1712          */
1713         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1714 }
1715 #endif
1716
1717 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1718 {
1719 #ifdef CONFIG_MEM_SOFT_DIRTY
1720         if (unlikely(is_pmd_migration_entry(pmd)))
1721                 pmd = pmd_swp_mksoft_dirty(pmd);
1722         else if (pmd_present(pmd))
1723                 pmd = pmd_mksoft_dirty(pmd);
1724 #endif
1725         return pmd;
1726 }
1727
1728 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1729                   unsigned long new_addr, unsigned long old_end,
1730                   pmd_t *old_pmd, pmd_t *new_pmd)
1731 {
1732         spinlock_t *old_ptl, *new_ptl;
1733         pmd_t pmd;
1734         struct mm_struct *mm = vma->vm_mm;
1735         bool force_flush = false;
1736
1737         if ((old_addr & ~HPAGE_PMD_MASK) ||
1738             (new_addr & ~HPAGE_PMD_MASK) ||
1739             old_end - old_addr < HPAGE_PMD_SIZE)
1740                 return false;
1741
1742         /*
1743          * The destination pmd shouldn't be established, free_pgtables()
1744          * should have release it.
1745          */
1746         if (WARN_ON(!pmd_none(*new_pmd))) {
1747                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1748                 return false;
1749         }
1750
1751         /*
1752          * We don't have to worry about the ordering of src and dst
1753          * ptlocks because exclusive mmap_sem prevents deadlock.
1754          */
1755         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1756         if (old_ptl) {
1757                 new_ptl = pmd_lockptr(mm, new_pmd);
1758                 if (new_ptl != old_ptl)
1759                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1760                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1761                 if (pmd_present(pmd))
1762                         force_flush = true;
1763                 VM_BUG_ON(!pmd_none(*new_pmd));
1764
1765                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1766                         pgtable_t pgtable;
1767                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1768                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1769                 }
1770                 pmd = move_soft_dirty_pmd(pmd);
1771                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1772                 if (force_flush)
1773                         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1774                 if (new_ptl != old_ptl)
1775                         spin_unlock(new_ptl);
1776                 spin_unlock(old_ptl);
1777                 return true;
1778         }
1779         return false;
1780 }
1781
1782 /*
1783  * Returns
1784  *  - 0 if PMD could not be locked
1785  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1786  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1787  */
1788 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1789                 unsigned long addr, pgprot_t newprot, unsigned long cp_flags)
1790 {
1791         struct mm_struct *mm = vma->vm_mm;
1792         spinlock_t *ptl;
1793         pmd_t entry;
1794         bool preserve_write;
1795         int ret;
1796         bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
1797         bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
1798         bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
1799
1800         ptl = __pmd_trans_huge_lock(pmd, vma);
1801         if (!ptl)
1802                 return 0;
1803
1804         preserve_write = prot_numa && pmd_write(*pmd);
1805         ret = 1;
1806
1807 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1808         if (is_swap_pmd(*pmd)) {
1809                 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1810
1811                 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1812                 if (is_write_migration_entry(entry)) {
1813                         pmd_t newpmd;
1814                         /*
1815                          * A protection check is difficult so
1816                          * just be safe and disable write
1817                          */
1818                         make_migration_entry_read(&entry);
1819                         newpmd = swp_entry_to_pmd(entry);
1820                         if (pmd_swp_soft_dirty(*pmd))
1821                                 newpmd = pmd_swp_mksoft_dirty(newpmd);
1822                         set_pmd_at(mm, addr, pmd, newpmd);
1823                 }
1824                 goto unlock;
1825         }
1826 #endif
1827
1828         /*
1829          * Avoid trapping faults against the zero page. The read-only
1830          * data is likely to be read-cached on the local CPU and
1831          * local/remote hits to the zero page are not interesting.
1832          */
1833         if (prot_numa && is_huge_zero_pmd(*pmd))
1834                 goto unlock;
1835
1836         if (prot_numa && pmd_protnone(*pmd))
1837                 goto unlock;
1838
1839         /*
1840          * In case prot_numa, we are under down_read(mmap_sem). It's critical
1841          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1842          * which is also under down_read(mmap_sem):
1843          *
1844          *      CPU0:                           CPU1:
1845          *                              change_huge_pmd(prot_numa=1)
1846          *                               pmdp_huge_get_and_clear_notify()
1847          * madvise_dontneed()
1848          *  zap_pmd_range()
1849          *   pmd_trans_huge(*pmd) == 0 (without ptl)
1850          *   // skip the pmd
1851          *                               set_pmd_at();
1852          *                               // pmd is re-established
1853          *
1854          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1855          * which may break userspace.
1856          *
1857          * pmdp_invalidate() is required to make sure we don't miss
1858          * dirty/young flags set by hardware.
1859          */
1860         entry = pmdp_invalidate(vma, addr, pmd);
1861
1862         entry = pmd_modify(entry, newprot);
1863         if (preserve_write)
1864                 entry = pmd_mk_savedwrite(entry);
1865         if (uffd_wp) {
1866                 entry = pmd_wrprotect(entry);
1867                 entry = pmd_mkuffd_wp(entry);
1868         } else if (uffd_wp_resolve) {
1869                 /*
1870                  * Leave the write bit to be handled by PF interrupt
1871                  * handler, then things like COW could be properly
1872                  * handled.
1873                  */
1874                 entry = pmd_clear_uffd_wp(entry);
1875         }
1876         ret = HPAGE_PMD_NR;
1877         set_pmd_at(mm, addr, pmd, entry);
1878         BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1879 unlock:
1880         spin_unlock(ptl);
1881         return ret;
1882 }
1883
1884 /*
1885  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1886  *
1887  * Note that if it returns page table lock pointer, this routine returns without
1888  * unlocking page table lock. So callers must unlock it.
1889  */
1890 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1891 {
1892         spinlock_t *ptl;
1893         ptl = pmd_lock(vma->vm_mm, pmd);
1894         if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1895                         pmd_devmap(*pmd)))
1896                 return ptl;
1897         spin_unlock(ptl);
1898         return NULL;
1899 }
1900
1901 /*
1902  * Returns true if a given pud maps a thp, false otherwise.
1903  *
1904  * Note that if it returns true, this routine returns without unlocking page
1905  * table lock. So callers must unlock it.
1906  */
1907 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1908 {
1909         spinlock_t *ptl;
1910
1911         ptl = pud_lock(vma->vm_mm, pud);
1912         if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1913                 return ptl;
1914         spin_unlock(ptl);
1915         return NULL;
1916 }
1917
1918 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1919 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1920                  pud_t *pud, unsigned long addr)
1921 {
1922         spinlock_t *ptl;
1923
1924         ptl = __pud_trans_huge_lock(pud, vma);
1925         if (!ptl)
1926                 return 0;
1927         /*
1928          * For architectures like ppc64 we look at deposited pgtable
1929          * when calling pudp_huge_get_and_clear. So do the
1930          * pgtable_trans_huge_withdraw after finishing pudp related
1931          * operations.
1932          */
1933         pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
1934         tlb_remove_pud_tlb_entry(tlb, pud, addr);
1935         if (vma_is_special_huge(vma)) {
1936                 spin_unlock(ptl);
1937                 /* No zero page support yet */
1938         } else {
1939                 /* No support for anonymous PUD pages yet */
1940                 BUG();
1941         }
1942         return 1;
1943 }
1944
1945 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1946                 unsigned long haddr)
1947 {
1948         VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1949         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1950         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
1951         VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
1952
1953         count_vm_event(THP_SPLIT_PUD);
1954
1955         pudp_huge_clear_flush_notify(vma, haddr, pud);
1956 }
1957
1958 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
1959                 unsigned long address)
1960 {
1961         spinlock_t *ptl;
1962         struct mmu_notifier_range range;
1963
1964         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1965                                 address & HPAGE_PUD_MASK,
1966                                 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
1967         mmu_notifier_invalidate_range_start(&range);
1968         ptl = pud_lock(vma->vm_mm, pud);
1969         if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
1970                 goto out;
1971         __split_huge_pud_locked(vma, pud, range.start);
1972
1973 out:
1974         spin_unlock(ptl);
1975         /*
1976          * No need to double call mmu_notifier->invalidate_range() callback as
1977          * the above pudp_huge_clear_flush_notify() did already call it.
1978          */
1979         mmu_notifier_invalidate_range_only_end(&range);
1980 }
1981 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1982
1983 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1984                 unsigned long haddr, pmd_t *pmd)
1985 {
1986         struct mm_struct *mm = vma->vm_mm;
1987         pgtable_t pgtable;
1988         pmd_t _pmd;
1989         int i;
1990
1991         /*
1992          * Leave pmd empty until pte is filled note that it is fine to delay
1993          * notification until mmu_notifier_invalidate_range_end() as we are
1994          * replacing a zero pmd write protected page with a zero pte write
1995          * protected page.
1996          *
1997          * See Documentation/vm/mmu_notifier.rst
1998          */
1999         pmdp_huge_clear_flush(vma, haddr, pmd);
2000
2001         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2002         pmd_populate(mm, &_pmd, pgtable);
2003
2004         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2005                 pte_t *pte, entry;
2006                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2007                 entry = pte_mkspecial(entry);
2008                 pte = pte_offset_map(&_pmd, haddr);
2009                 VM_BUG_ON(!pte_none(*pte));
2010                 set_pte_at(mm, haddr, pte, entry);
2011                 pte_unmap(pte);
2012         }
2013         smp_wmb(); /* make pte visible before pmd */
2014         pmd_populate(mm, pmd, pgtable);
2015 }
2016
2017 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2018                 unsigned long haddr, bool freeze)
2019 {
2020         struct mm_struct *mm = vma->vm_mm;
2021         struct page *page;
2022         pgtable_t pgtable;
2023         pmd_t old_pmd, _pmd;
2024         bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false;
2025         unsigned long addr;
2026         int i;
2027
2028         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2029         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2030         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2031         VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2032                                 && !pmd_devmap(*pmd));
2033
2034         count_vm_event(THP_SPLIT_PMD);
2035
2036         if (!vma_is_anonymous(vma)) {
2037                 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2038                 /*
2039                  * We are going to unmap this huge page. So
2040                  * just go ahead and zap it
2041                  */
2042                 if (arch_needs_pgtable_deposit())
2043                         zap_deposited_table(mm, pmd);
2044                 if (vma_is_special_huge(vma))
2045                         return;
2046                 page = pmd_page(_pmd);
2047                 if (!PageDirty(page) && pmd_dirty(_pmd))
2048                         set_page_dirty(page);
2049                 if (!PageReferenced(page) && pmd_young(_pmd))
2050                         SetPageReferenced(page);
2051                 page_remove_rmap(page, true);
2052                 put_page(page);
2053                 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2054                 return;
2055         } else if (is_huge_zero_pmd(*pmd)) {
2056                 /*
2057                  * FIXME: Do we want to invalidate secondary mmu by calling
2058                  * mmu_notifier_invalidate_range() see comments below inside
2059                  * __split_huge_pmd() ?
2060                  *
2061                  * We are going from a zero huge page write protected to zero
2062                  * small page also write protected so it does not seems useful
2063                  * to invalidate secondary mmu at this time.
2064                  */
2065                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2066         }
2067
2068         /*
2069          * Up to this point the pmd is present and huge and userland has the
2070          * whole access to the hugepage during the split (which happens in
2071          * place). If we overwrite the pmd with the not-huge version pointing
2072          * to the pte here (which of course we could if all CPUs were bug
2073          * free), userland could trigger a small page size TLB miss on the
2074          * small sized TLB while the hugepage TLB entry is still established in
2075          * the huge TLB. Some CPU doesn't like that.
2076          * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2077          * 383 on page 93. Intel should be safe but is also warns that it's
2078          * only safe if the permission and cache attributes of the two entries
2079          * loaded in the two TLB is identical (which should be the case here).
2080          * But it is generally safer to never allow small and huge TLB entries
2081          * for the same virtual address to be loaded simultaneously. So instead
2082          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2083          * current pmd notpresent (atomically because here the pmd_trans_huge
2084          * must remain set at all times on the pmd until the split is complete
2085          * for this pmd), then we flush the SMP TLB and finally we write the
2086          * non-huge version of the pmd entry with pmd_populate.
2087          */
2088         old_pmd = pmdp_invalidate(vma, haddr, pmd);
2089
2090         pmd_migration = is_pmd_migration_entry(old_pmd);
2091         if (unlikely(pmd_migration)) {
2092                 swp_entry_t entry;
2093
2094                 entry = pmd_to_swp_entry(old_pmd);
2095                 page = pfn_to_page(swp_offset(entry));
2096                 write = is_write_migration_entry(entry);
2097                 young = false;
2098                 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2099                 uffd_wp = pmd_swp_uffd_wp(old_pmd);
2100         } else {
2101                 page = pmd_page(old_pmd);
2102                 if (pmd_dirty(old_pmd))
2103                         SetPageDirty(page);
2104                 write = pmd_write(old_pmd);
2105                 young = pmd_young(old_pmd);
2106                 soft_dirty = pmd_soft_dirty(old_pmd);
2107                 uffd_wp = pmd_uffd_wp(old_pmd);
2108         }
2109         VM_BUG_ON_PAGE(!page_count(page), page);
2110         page_ref_add(page, HPAGE_PMD_NR - 1);
2111
2112         /*
2113          * Withdraw the table only after we mark the pmd entry invalid.
2114          * This's critical for some architectures (Power).
2115          */
2116         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2117         pmd_populate(mm, &_pmd, pgtable);
2118
2119         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2120                 pte_t entry, *pte;
2121                 /*
2122                  * Note that NUMA hinting access restrictions are not
2123                  * transferred to avoid any possibility of altering
2124                  * permissions across VMAs.
2125                  */
2126                 if (freeze || pmd_migration) {
2127                         swp_entry_t swp_entry;
2128                         swp_entry = make_migration_entry(page + i, write);
2129                         entry = swp_entry_to_pte(swp_entry);
2130                         if (soft_dirty)
2131                                 entry = pte_swp_mksoft_dirty(entry);
2132                         if (uffd_wp)
2133                                 entry = pte_swp_mkuffd_wp(entry);
2134                 } else {
2135                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2136                         entry = maybe_mkwrite(entry, vma);
2137                         if (!write)
2138                                 entry = pte_wrprotect(entry);
2139                         if (!young)
2140                                 entry = pte_mkold(entry);
2141                         if (soft_dirty)
2142                                 entry = pte_mksoft_dirty(entry);
2143                         if (uffd_wp)
2144                                 entry = pte_mkuffd_wp(entry);
2145                 }
2146                 pte = pte_offset_map(&_pmd, addr);
2147                 BUG_ON(!pte_none(*pte));
2148                 set_pte_at(mm, addr, pte, entry);
2149                 atomic_inc(&page[i]._mapcount);
2150                 pte_unmap(pte);
2151         }
2152
2153         /*
2154          * Set PG_double_map before dropping compound_mapcount to avoid
2155          * false-negative page_mapped().
2156          */
2157         if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2158                 for (i = 0; i < HPAGE_PMD_NR; i++)
2159                         atomic_inc(&page[i]._mapcount);
2160         }
2161
2162         lock_page_memcg(page);
2163         if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2164                 /* Last compound_mapcount is gone. */
2165                 __dec_lruvec_page_state(page, NR_ANON_THPS);
2166                 if (TestClearPageDoubleMap(page)) {
2167                         /* No need in mapcount reference anymore */
2168                         for (i = 0; i < HPAGE_PMD_NR; i++)
2169                                 atomic_dec(&page[i]._mapcount);
2170                 }
2171         }
2172         unlock_page_memcg(page);
2173
2174         smp_wmb(); /* make pte visible before pmd */
2175         pmd_populate(mm, pmd, pgtable);
2176
2177         if (freeze) {
2178                 for (i = 0; i < HPAGE_PMD_NR; i++) {
2179                         page_remove_rmap(page + i, false);
2180                         put_page(page + i);
2181                 }
2182         }
2183 }
2184
2185 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2186                 unsigned long address, bool freeze, struct page *page)
2187 {
2188         spinlock_t *ptl;
2189         struct mmu_notifier_range range;
2190
2191         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2192                                 address & HPAGE_PMD_MASK,
2193                                 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2194         mmu_notifier_invalidate_range_start(&range);
2195         ptl = pmd_lock(vma->vm_mm, pmd);
2196
2197         /*
2198          * If caller asks to setup a migration entries, we need a page to check
2199          * pmd against. Otherwise we can end up replacing wrong page.
2200          */
2201         VM_BUG_ON(freeze && !page);
2202         if (page && page != pmd_page(*pmd))
2203                 goto out;
2204
2205         if (pmd_trans_huge(*pmd)) {
2206                 page = pmd_page(*pmd);
2207                 if (PageMlocked(page))
2208                         clear_page_mlock(page);
2209         } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2210                 goto out;
2211         __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2212 out:
2213         spin_unlock(ptl);
2214         /*
2215          * No need to double call mmu_notifier->invalidate_range() callback.
2216          * They are 3 cases to consider inside __split_huge_pmd_locked():
2217          *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2218          *  2) __split_huge_zero_page_pmd() read only zero page and any write
2219          *    fault will trigger a flush_notify before pointing to a new page
2220          *    (it is fine if the secondary mmu keeps pointing to the old zero
2221          *    page in the meantime)
2222          *  3) Split a huge pmd into pte pointing to the same page. No need
2223          *     to invalidate secondary tlb entry they are all still valid.
2224          *     any further changes to individual pte will notify. So no need
2225          *     to call mmu_notifier->invalidate_range()
2226          */
2227         mmu_notifier_invalidate_range_only_end(&range);
2228 }
2229
2230 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2231                 bool freeze, struct page *page)
2232 {
2233         pgd_t *pgd;
2234         p4d_t *p4d;
2235         pud_t *pud;
2236         pmd_t *pmd;
2237
2238         pgd = pgd_offset(vma->vm_mm, address);
2239         if (!pgd_present(*pgd))
2240                 return;
2241
2242         p4d = p4d_offset(pgd, address);
2243         if (!p4d_present(*p4d))
2244                 return;
2245
2246         pud = pud_offset(p4d, address);
2247         if (!pud_present(*pud))
2248                 return;
2249
2250         pmd = pmd_offset(pud, address);
2251
2252         __split_huge_pmd(vma, pmd, address, freeze, page);
2253 }
2254
2255 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2256                              unsigned long start,
2257                              unsigned long end,
2258                              long adjust_next)
2259 {
2260         /*
2261          * If the new start address isn't hpage aligned and it could
2262          * previously contain an hugepage: check if we need to split
2263          * an huge pmd.
2264          */
2265         if (start & ~HPAGE_PMD_MASK &&
2266             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2267             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2268                 split_huge_pmd_address(vma, start, false, NULL);
2269
2270         /*
2271          * If the new end address isn't hpage aligned and it could
2272          * previously contain an hugepage: check if we need to split
2273          * an huge pmd.
2274          */
2275         if (end & ~HPAGE_PMD_MASK &&
2276             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2277             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2278                 split_huge_pmd_address(vma, end, false, NULL);
2279
2280         /*
2281          * If we're also updating the vma->vm_next->vm_start, if the new
2282          * vm_next->vm_start isn't page aligned and it could previously
2283          * contain an hugepage: check if we need to split an huge pmd.
2284          */
2285         if (adjust_next > 0) {
2286                 struct vm_area_struct *next = vma->vm_next;
2287                 unsigned long nstart = next->vm_start;
2288                 nstart += adjust_next << PAGE_SHIFT;
2289                 if (nstart & ~HPAGE_PMD_MASK &&
2290                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2291                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2292                         split_huge_pmd_address(next, nstart, false, NULL);
2293         }
2294 }
2295
2296 static void unmap_page(struct page *page)
2297 {
2298         enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
2299                 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2300         bool unmap_success;
2301
2302         VM_BUG_ON_PAGE(!PageHead(page), page);
2303
2304         if (PageAnon(page))
2305                 ttu_flags |= TTU_SPLIT_FREEZE;
2306
2307         unmap_success = try_to_unmap(page, ttu_flags);
2308         VM_BUG_ON_PAGE(!unmap_success, page);
2309 }
2310
2311 static void remap_page(struct page *page)
2312 {
2313         int i;
2314         if (PageTransHuge(page)) {
2315                 remove_migration_ptes(page, page, true);
2316         } else {
2317                 for (i = 0; i < HPAGE_PMD_NR; i++)
2318                         remove_migration_ptes(page + i, page + i, true);
2319         }
2320 }
2321
2322 static void __split_huge_page_tail(struct page *head, int tail,
2323                 struct lruvec *lruvec, struct list_head *list)
2324 {
2325         struct page *page_tail = head + tail;
2326
2327         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2328
2329         /*
2330          * Clone page flags before unfreezing refcount.
2331          *
2332          * After successful get_page_unless_zero() might follow flags change,
2333          * for exmaple lock_page() which set PG_waiters.
2334          */
2335         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2336         page_tail->flags |= (head->flags &
2337                         ((1L << PG_referenced) |
2338                          (1L << PG_swapbacked) |
2339                          (1L << PG_swapcache) |
2340                          (1L << PG_mlocked) |
2341                          (1L << PG_uptodate) |
2342                          (1L << PG_active) |
2343                          (1L << PG_workingset) |
2344                          (1L << PG_locked) |
2345                          (1L << PG_unevictable) |
2346                          (1L << PG_dirty)));
2347
2348         /* ->mapping in first tail page is compound_mapcount */
2349         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2350                         page_tail);
2351         page_tail->mapping = head->mapping;
2352         page_tail->index = head->index + tail;
2353
2354         /* Page flags must be visible before we make the page non-compound. */
2355         smp_wmb();
2356
2357         /*
2358          * Clear PageTail before unfreezing page refcount.
2359          *
2360          * After successful get_page_unless_zero() might follow put_page()
2361          * which needs correct compound_head().
2362          */
2363         clear_compound_head(page_tail);
2364
2365         /* Finally unfreeze refcount. Additional reference from page cache. */
2366         page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2367                                           PageSwapCache(head)));
2368
2369         if (page_is_young(head))
2370                 set_page_young(page_tail);
2371         if (page_is_idle(head))
2372                 set_page_idle(page_tail);
2373
2374         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2375
2376         /*
2377          * always add to the tail because some iterators expect new
2378          * pages to show after the currently processed elements - e.g.
2379          * migrate_pages
2380          */
2381         lru_add_page_tail(head, page_tail, lruvec, list);
2382 }
2383
2384 static void __split_huge_page(struct page *page, struct list_head *list,
2385                 pgoff_t end, unsigned long flags)
2386 {
2387         struct page *head = compound_head(page);
2388         pg_data_t *pgdat = page_pgdat(head);
2389         struct lruvec *lruvec;
2390         struct address_space *swap_cache = NULL;
2391         unsigned long offset = 0;
2392         int i;
2393
2394         lruvec = mem_cgroup_page_lruvec(head, pgdat);
2395
2396         /* complete memcg works before add pages to LRU */
2397         mem_cgroup_split_huge_fixup(head);
2398
2399         if (PageAnon(head) && PageSwapCache(head)) {
2400                 swp_entry_t entry = { .val = page_private(head) };
2401
2402                 offset = swp_offset(entry);
2403                 swap_cache = swap_address_space(entry);
2404                 xa_lock(&swap_cache->i_pages);
2405         }
2406
2407         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
2408                 __split_huge_page_tail(head, i, lruvec, list);
2409                 /* Some pages can be beyond i_size: drop them from page cache */
2410                 if (head[i].index >= end) {
2411                         ClearPageDirty(head + i);
2412                         __delete_from_page_cache(head + i, NULL);
2413                         if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2414                                 shmem_uncharge(head->mapping->host, 1);
2415                         put_page(head + i);
2416                 } else if (!PageAnon(page)) {
2417                         __xa_store(&head->mapping->i_pages, head[i].index,
2418                                         head + i, 0);
2419                 } else if (swap_cache) {
2420                         __xa_store(&swap_cache->i_pages, offset + i,
2421                                         head + i, 0);
2422                 }
2423         }
2424
2425         ClearPageCompound(head);
2426
2427         split_page_owner(head, HPAGE_PMD_ORDER);
2428
2429         /* See comment in __split_huge_page_tail() */
2430         if (PageAnon(head)) {
2431                 /* Additional pin to swap cache */
2432                 if (PageSwapCache(head)) {
2433                         page_ref_add(head, 2);
2434                         xa_unlock(&swap_cache->i_pages);
2435                 } else {
2436                         page_ref_inc(head);
2437                 }
2438         } else {
2439                 /* Additional pin to page cache */
2440                 page_ref_add(head, 2);
2441                 xa_unlock(&head->mapping->i_pages);
2442         }
2443
2444         spin_unlock_irqrestore(&pgdat->lru_lock, flags);
2445
2446         remap_page(head);
2447
2448         for (i = 0; i < HPAGE_PMD_NR; i++) {
2449                 struct page *subpage = head + i;
2450                 if (subpage == page)
2451                         continue;
2452                 unlock_page(subpage);
2453
2454                 /*
2455                  * Subpages may be freed if there wasn't any mapping
2456                  * like if add_to_swap() is running on a lru page that
2457                  * had its mapping zapped. And freeing these pages
2458                  * requires taking the lru_lock so we do the put_page
2459                  * of the tail pages after the split is complete.
2460                  */
2461                 put_page(subpage);
2462         }
2463 }
2464
2465 int total_mapcount(struct page *page)
2466 {
2467         int i, compound, ret;
2468
2469         VM_BUG_ON_PAGE(PageTail(page), page);
2470
2471         if (likely(!PageCompound(page)))
2472                 return atomic_read(&page->_mapcount) + 1;
2473
2474         compound = compound_mapcount(page);
2475         if (PageHuge(page))
2476                 return compound;
2477         ret = compound;
2478         for (i = 0; i < HPAGE_PMD_NR; i++)
2479                 ret += atomic_read(&page[i]._mapcount) + 1;
2480         /* File pages has compound_mapcount included in _mapcount */
2481         if (!PageAnon(page))
2482                 return ret - compound * HPAGE_PMD_NR;
2483         if (PageDoubleMap(page))
2484                 ret -= HPAGE_PMD_NR;
2485         return ret;
2486 }
2487
2488 /*
2489  * This calculates accurately how many mappings a transparent hugepage
2490  * has (unlike page_mapcount() which isn't fully accurate). This full
2491  * accuracy is primarily needed to know if copy-on-write faults can
2492  * reuse the page and change the mapping to read-write instead of
2493  * copying them. At the same time this returns the total_mapcount too.
2494  *
2495  * The function returns the highest mapcount any one of the subpages
2496  * has. If the return value is one, even if different processes are
2497  * mapping different subpages of the transparent hugepage, they can
2498  * all reuse it, because each process is reusing a different subpage.
2499  *
2500  * The total_mapcount is instead counting all virtual mappings of the
2501  * subpages. If the total_mapcount is equal to "one", it tells the
2502  * caller all mappings belong to the same "mm" and in turn the
2503  * anon_vma of the transparent hugepage can become the vma->anon_vma
2504  * local one as no other process may be mapping any of the subpages.
2505  *
2506  * It would be more accurate to replace page_mapcount() with
2507  * page_trans_huge_mapcount(), however we only use
2508  * page_trans_huge_mapcount() in the copy-on-write faults where we
2509  * need full accuracy to avoid breaking page pinning, because
2510  * page_trans_huge_mapcount() is slower than page_mapcount().
2511  */
2512 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2513 {
2514         int i, ret, _total_mapcount, mapcount;
2515
2516         /* hugetlbfs shouldn't call it */
2517         VM_BUG_ON_PAGE(PageHuge(page), page);
2518
2519         if (likely(!PageTransCompound(page))) {
2520                 mapcount = atomic_read(&page->_mapcount) + 1;
2521                 if (total_mapcount)
2522                         *total_mapcount = mapcount;
2523                 return mapcount;
2524         }
2525
2526         page = compound_head(page);
2527
2528         _total_mapcount = ret = 0;
2529         for (i = 0; i < HPAGE_PMD_NR; i++) {
2530                 mapcount = atomic_read(&page[i]._mapcount) + 1;
2531                 ret = max(ret, mapcount);
2532                 _total_mapcount += mapcount;
2533         }
2534         if (PageDoubleMap(page)) {
2535                 ret -= 1;
2536                 _total_mapcount -= HPAGE_PMD_NR;
2537         }
2538         mapcount = compound_mapcount(page);
2539         ret += mapcount;
2540         _total_mapcount += mapcount;
2541         if (total_mapcount)
2542                 *total_mapcount = _total_mapcount;
2543         return ret;
2544 }
2545
2546 /* Racy check whether the huge page can be split */
2547 bool can_split_huge_page(struct page *page, int *pextra_pins)
2548 {
2549         int extra_pins;
2550
2551         /* Additional pins from page cache */
2552         if (PageAnon(page))
2553                 extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0;
2554         else
2555                 extra_pins = HPAGE_PMD_NR;
2556         if (pextra_pins)
2557                 *pextra_pins = extra_pins;
2558         return total_mapcount(page) == page_count(page) - extra_pins - 1;
2559 }
2560
2561 /*
2562  * This function splits huge page into normal pages. @page can point to any
2563  * subpage of huge page to split. Split doesn't change the position of @page.
2564  *
2565  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2566  * The huge page must be locked.
2567  *
2568  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2569  *
2570  * Both head page and tail pages will inherit mapping, flags, and so on from
2571  * the hugepage.
2572  *
2573  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2574  * they are not mapped.
2575  *
2576  * Returns 0 if the hugepage is split successfully.
2577  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2578  * us.
2579  */
2580 int split_huge_page_to_list(struct page *page, struct list_head *list)
2581 {
2582         struct page *head = compound_head(page);
2583         struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2584         struct deferred_split *ds_queue = get_deferred_split_queue(head);
2585         struct anon_vma *anon_vma = NULL;
2586         struct address_space *mapping = NULL;
2587         int count, mapcount, extra_pins, ret;
2588         bool mlocked;
2589         unsigned long flags;
2590         pgoff_t end;
2591
2592         VM_BUG_ON_PAGE(is_huge_zero_page(head), head);
2593         VM_BUG_ON_PAGE(!PageLocked(head), head);
2594         VM_BUG_ON_PAGE(!PageCompound(head), head);
2595
2596         if (PageWriteback(head))
2597                 return -EBUSY;
2598
2599         if (PageAnon(head)) {
2600                 /*
2601                  * The caller does not necessarily hold an mmap_sem that would
2602                  * prevent the anon_vma disappearing so we first we take a
2603                  * reference to it and then lock the anon_vma for write. This
2604                  * is similar to page_lock_anon_vma_read except the write lock
2605                  * is taken to serialise against parallel split or collapse
2606                  * operations.
2607                  */
2608                 anon_vma = page_get_anon_vma(head);
2609                 if (!anon_vma) {
2610                         ret = -EBUSY;
2611                         goto out;
2612                 }
2613                 end = -1;
2614                 mapping = NULL;
2615                 anon_vma_lock_write(anon_vma);
2616         } else {
2617                 mapping = head->mapping;
2618
2619                 /* Truncated ? */
2620                 if (!mapping) {
2621                         ret = -EBUSY;
2622                         goto out;
2623                 }
2624
2625                 anon_vma = NULL;
2626                 i_mmap_lock_read(mapping);
2627
2628                 /*
2629                  *__split_huge_page() may need to trim off pages beyond EOF:
2630                  * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2631                  * which cannot be nested inside the page tree lock. So note
2632                  * end now: i_size itself may be changed at any moment, but
2633                  * head page lock is good enough to serialize the trimming.
2634                  */
2635                 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2636         }
2637
2638         /*
2639          * Racy check if we can split the page, before unmap_page() will
2640          * split PMDs
2641          */
2642         if (!can_split_huge_page(head, &extra_pins)) {
2643                 ret = -EBUSY;
2644                 goto out_unlock;
2645         }
2646
2647         mlocked = PageMlocked(head);
2648         unmap_page(head);
2649         VM_BUG_ON_PAGE(compound_mapcount(head), head);
2650
2651         /* Make sure the page is not on per-CPU pagevec as it takes pin */
2652         if (mlocked)
2653                 lru_add_drain();
2654
2655         /* prevent PageLRU to go away from under us, and freeze lru stats */
2656         spin_lock_irqsave(&pgdata->lru_lock, flags);
2657
2658         if (mapping) {
2659                 XA_STATE(xas, &mapping->i_pages, page_index(head));
2660
2661                 /*
2662                  * Check if the head page is present in page cache.
2663                  * We assume all tail are present too, if head is there.
2664                  */
2665                 xa_lock(&mapping->i_pages);
2666                 if (xas_load(&xas) != head)
2667                         goto fail;
2668         }
2669
2670         /* Prevent deferred_split_scan() touching ->_refcount */
2671         spin_lock(&ds_queue->split_queue_lock);
2672         count = page_count(head);
2673         mapcount = total_mapcount(head);
2674         if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2675                 if (!list_empty(page_deferred_list(head))) {
2676                         ds_queue->split_queue_len--;
2677                         list_del(page_deferred_list(head));
2678                 }
2679                 spin_unlock(&ds_queue->split_queue_lock);
2680                 if (mapping) {
2681                         if (PageSwapBacked(head))
2682                                 __dec_node_page_state(head, NR_SHMEM_THPS);
2683                         else
2684                                 __dec_node_page_state(head, NR_FILE_THPS);
2685                 }
2686
2687                 __split_huge_page(page, list, end, flags);
2688                 if (PageSwapCache(head)) {
2689                         swp_entry_t entry = { .val = page_private(head) };
2690
2691                         ret = split_swap_cluster(entry);
2692                 } else
2693                         ret = 0;
2694         } else {
2695                 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2696                         pr_alert("total_mapcount: %u, page_count(): %u\n",
2697                                         mapcount, count);
2698                         if (PageTail(page))
2699                                 dump_page(head, NULL);
2700                         dump_page(page, "total_mapcount(head) > 0");
2701                         BUG();
2702                 }
2703                 spin_unlock(&ds_queue->split_queue_lock);
2704 fail:           if (mapping)
2705                         xa_unlock(&mapping->i_pages);
2706                 spin_unlock_irqrestore(&pgdata->lru_lock, flags);
2707                 remap_page(head);
2708                 ret = -EBUSY;
2709         }
2710
2711 out_unlock:
2712         if (anon_vma) {
2713                 anon_vma_unlock_write(anon_vma);
2714                 put_anon_vma(anon_vma);
2715         }
2716         if (mapping)
2717                 i_mmap_unlock_read(mapping);
2718 out:
2719         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2720         return ret;
2721 }
2722
2723 void free_transhuge_page(struct page *page)
2724 {
2725         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2726         unsigned long flags;
2727
2728         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2729         if (!list_empty(page_deferred_list(page))) {
2730                 ds_queue->split_queue_len--;
2731                 list_del(page_deferred_list(page));
2732         }
2733         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2734         free_compound_page(page);
2735 }
2736
2737 void deferred_split_huge_page(struct page *page)
2738 {
2739         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2740 #ifdef CONFIG_MEMCG
2741         struct mem_cgroup *memcg = compound_head(page)->mem_cgroup;
2742 #endif
2743         unsigned long flags;
2744
2745         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2746
2747         /*
2748          * The try_to_unmap() in page reclaim path might reach here too,
2749          * this may cause a race condition to corrupt deferred split queue.
2750          * And, if page reclaim is already handling the same page, it is
2751          * unnecessary to handle it again in shrinker.
2752          *
2753          * Check PageSwapCache to determine if the page is being
2754          * handled by page reclaim since THP swap would add the page into
2755          * swap cache before calling try_to_unmap().
2756          */
2757         if (PageSwapCache(page))
2758                 return;
2759
2760         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2761         if (list_empty(page_deferred_list(page))) {
2762                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2763                 list_add_tail(page_deferred_list(page), &ds_queue->split_queue);
2764                 ds_queue->split_queue_len++;
2765 #ifdef CONFIG_MEMCG
2766                 if (memcg)
2767                         memcg_set_shrinker_bit(memcg, page_to_nid(page),
2768                                                deferred_split_shrinker.id);
2769 #endif
2770         }
2771         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2772 }
2773
2774 static unsigned long deferred_split_count(struct shrinker *shrink,
2775                 struct shrink_control *sc)
2776 {
2777         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2778         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2779
2780 #ifdef CONFIG_MEMCG
2781         if (sc->memcg)
2782                 ds_queue = &sc->memcg->deferred_split_queue;
2783 #endif
2784         return READ_ONCE(ds_queue->split_queue_len);
2785 }
2786
2787 static unsigned long deferred_split_scan(struct shrinker *shrink,
2788                 struct shrink_control *sc)
2789 {
2790         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2791         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2792         unsigned long flags;
2793         LIST_HEAD(list), *pos, *next;
2794         struct page *page;
2795         int split = 0;
2796
2797 #ifdef CONFIG_MEMCG
2798         if (sc->memcg)
2799                 ds_queue = &sc->memcg->deferred_split_queue;
2800 #endif
2801
2802         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2803         /* Take pin on all head pages to avoid freeing them under us */
2804         list_for_each_safe(pos, next, &ds_queue->split_queue) {
2805                 page = list_entry((void *)pos, struct page, mapping);
2806                 page = compound_head(page);
2807                 if (get_page_unless_zero(page)) {
2808                         list_move(page_deferred_list(page), &list);
2809                 } else {
2810                         /* We lost race with put_compound_page() */
2811                         list_del_init(page_deferred_list(page));
2812                         ds_queue->split_queue_len--;
2813                 }
2814                 if (!--sc->nr_to_scan)
2815                         break;
2816         }
2817         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2818
2819         list_for_each_safe(pos, next, &list) {
2820                 page = list_entry((void *)pos, struct page, mapping);
2821                 if (!trylock_page(page))
2822                         goto next;
2823                 /* split_huge_page() removes page from list on success */
2824                 if (!split_huge_page(page))
2825                         split++;
2826                 unlock_page(page);
2827 next:
2828                 put_page(page);
2829         }
2830
2831         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2832         list_splice_tail(&list, &ds_queue->split_queue);
2833         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2834
2835         /*
2836          * Stop shrinker if we didn't split any page, but the queue is empty.
2837          * This can happen if pages were freed under us.
2838          */
2839         if (!split && list_empty(&ds_queue->split_queue))
2840                 return SHRINK_STOP;
2841         return split;
2842 }
2843
2844 static struct shrinker deferred_split_shrinker = {
2845         .count_objects = deferred_split_count,
2846         .scan_objects = deferred_split_scan,
2847         .seeks = DEFAULT_SEEKS,
2848         .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
2849                  SHRINKER_NONSLAB,
2850 };
2851
2852 #ifdef CONFIG_DEBUG_FS
2853 static int split_huge_pages_set(void *data, u64 val)
2854 {
2855         struct zone *zone;
2856         struct page *page;
2857         unsigned long pfn, max_zone_pfn;
2858         unsigned long total = 0, split = 0;
2859
2860         if (val != 1)
2861                 return -EINVAL;
2862
2863         for_each_populated_zone(zone) {
2864                 max_zone_pfn = zone_end_pfn(zone);
2865                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2866                         if (!pfn_valid(pfn))
2867                                 continue;
2868
2869                         page = pfn_to_page(pfn);
2870                         if (!get_page_unless_zero(page))
2871                                 continue;
2872
2873                         if (zone != page_zone(page))
2874                                 goto next;
2875
2876                         if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2877                                 goto next;
2878
2879                         total++;
2880                         lock_page(page);
2881                         if (!split_huge_page(page))
2882                                 split++;
2883                         unlock_page(page);
2884 next:
2885                         put_page(page);
2886                 }
2887         }
2888
2889         pr_info("%lu of %lu THP split\n", split, total);
2890
2891         return 0;
2892 }
2893 DEFINE_DEBUGFS_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2894                 "%llu\n");
2895
2896 static int __init split_huge_pages_debugfs(void)
2897 {
2898         debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2899                             &split_huge_pages_fops);
2900         return 0;
2901 }
2902 late_initcall(split_huge_pages_debugfs);
2903 #endif
2904
2905 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2906 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
2907                 struct page *page)
2908 {
2909         struct vm_area_struct *vma = pvmw->vma;
2910         struct mm_struct *mm = vma->vm_mm;
2911         unsigned long address = pvmw->address;
2912         pmd_t pmdval;
2913         swp_entry_t entry;
2914         pmd_t pmdswp;
2915
2916         if (!(pvmw->pmd && !pvmw->pte))
2917                 return;
2918
2919         flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
2920         pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
2921         if (pmd_dirty(pmdval))
2922                 set_page_dirty(page);
2923         entry = make_migration_entry(page, pmd_write(pmdval));
2924         pmdswp = swp_entry_to_pmd(entry);
2925         if (pmd_soft_dirty(pmdval))
2926                 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
2927         set_pmd_at(mm, address, pvmw->pmd, pmdswp);
2928         page_remove_rmap(page, true);
2929         put_page(page);
2930 }
2931
2932 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
2933 {
2934         struct vm_area_struct *vma = pvmw->vma;
2935         struct mm_struct *mm = vma->vm_mm;
2936         unsigned long address = pvmw->address;
2937         unsigned long mmun_start = address & HPAGE_PMD_MASK;
2938         pmd_t pmde;
2939         swp_entry_t entry;
2940
2941         if (!(pvmw->pmd && !pvmw->pte))
2942                 return;
2943
2944         entry = pmd_to_swp_entry(*pvmw->pmd);
2945         get_page(new);
2946         pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
2947         if (pmd_swp_soft_dirty(*pvmw->pmd))
2948                 pmde = pmd_mksoft_dirty(pmde);
2949         if (is_write_migration_entry(entry))
2950                 pmde = maybe_pmd_mkwrite(pmde, vma);
2951
2952         flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
2953         if (PageAnon(new))
2954                 page_add_anon_rmap(new, vma, mmun_start, true);
2955         else
2956                 page_add_file_rmap(new, true);
2957         set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
2958         if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new))
2959                 mlock_vma_page(new);
2960         update_mmu_cache_pmd(vma, address, pvmw->pmd);
2961 }
2962 #endif