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 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
 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, true)) {
 598                 put_page(page);
 599                 count_vm_event(THP_FAULT_FALLBACK);
 600                 return VM_FAULT_FALLBACK;
 601         }
 602
 603         pgtable = pte_alloc_one(vma->vm_mm);
 604         if (unlikely(!pgtable)) {
 605                 ret = VM_FAULT_OOM;
 606                 goto release;
 607         }
 608
 609         clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
 610         /*
 611          * The memory barrier inside __SetPageUptodate makes sure that
 612          * clear_huge_page writes become visible before the set_pmd_at()
 613          * write.
 614          */
 615         __SetPageUptodate(page);
 616
 617         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 618         if (unlikely(!pmd_none(*vmf->pmd))) {
 619                 goto unlock_release;
 620         } else {
 621                 pmd_t entry;
 622
 623                 ret = check_stable_address_space(vma->vm_mm);
 624                 if (ret)
 625                         goto unlock_release;
 626
 627                 /* Deliver the page fault to userland */
 628                 if (userfaultfd_missing(vma)) {
 629                         vm_fault_t ret2;
 630
 631                         spin_unlock(vmf->ptl);
 632                         mem_cgroup_cancel_charge(page, memcg, true);
 633                         put_page(page);
 634                         pte_free(vma->vm_mm, pgtable);
 635                         ret2 = handle_userfault(vmf, VM_UFFD_MISSING);
 636                         VM_BUG_ON(ret2 & VM_FAULT_FALLBACK);
 637                         return ret2;
 638                 }
 639
 640                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 641                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 642                 page_add_new_anon_rmap(page, vma, haddr, true);
 643                 mem_cgroup_commit_charge(page, memcg, false, true);
 644                 lru_cache_add_active_or_unevictable(page, vma);
 645                 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
 646                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
 647                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 648                 mm_inc_nr_ptes(vma->vm_mm);
 649                 spin_unlock(vmf->ptl);
 650                 count_vm_event(THP_FAULT_ALLOC);
 651                 count_memcg_events(memcg, THP_FAULT_ALLOC, 1);
 652         }
 653
 654         return 0;
 655 unlock_release:
 656         spin_unlock(vmf->ptl);
 657 release:
 658         if (pgtable)
 659                 pte_free(vma->vm_mm, pgtable);
 660         mem_cgroup_cancel_charge(page, memcg, true);
 661         put_page(page);
 662         return ret;
 663
 664 }
 665
 666 /*
 667  * always: directly stall for all thp allocations
 668  * defer: wake kswapd and fail if not immediately available
 669  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
 670  *                fail if not immediately available
 671  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
 672  *          available
 673  * never: never stall for any thp allocation
 674  */
 675 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
 676 {
 677         const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
 678
 679         /* Always do synchronous compaction */
 680         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 681                 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
 682
 683         /* Kick kcompactd and fail quickly */
 684         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 685                 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
 686
 687         /* Synchronous compaction if madvised, otherwise kick kcompactd */
 688         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 689                 return GFP_TRANSHUGE_LIGHT |
 690                         (vma_madvised ? __GFP_DIRECT_RECLAIM :
 691                                         __GFP_KSWAPD_RECLAIM);
 692
 693         /* Only do synchronous compaction if madvised */
 694         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 695                 return GFP_TRANSHUGE_LIGHT |
 696                        (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
 697
 698         return GFP_TRANSHUGE_LIGHT;
 699 }
 700
 701 /* Caller must hold page table lock. */
 702 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 703                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 704                 struct page *zero_page)
 705 {
 706         pmd_t entry;
 707         if (!pmd_none(*pmd))
 708                 return false;
 709         entry = mk_pmd(zero_page, vma->vm_page_prot);
 710         entry = pmd_mkhuge(entry);
 711         if (pgtable)
 712                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 713         set_pmd_at(mm, haddr, pmd, entry);
 714         mm_inc_nr_ptes(mm);
 715         return true;
 716 }
 717
 718 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
 719 {
 720         struct vm_area_struct *vma = vmf->vma;
 721         gfp_t gfp;
 722         struct page *page;
 723         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 724
 725         if (!transhuge_vma_suitable(vma, haddr))
 726                 return VM_FAULT_FALLBACK;
 727         if (unlikely(anon_vma_prepare(vma)))
 728                 return VM_FAULT_OOM;
 729         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 730                 return VM_FAULT_OOM;
 731         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
 732                         !mm_forbids_zeropage(vma->vm_mm) &&
 733                         transparent_hugepage_use_zero_page()) {
 734                 pgtable_t pgtable;
 735                 struct page *zero_page;
 736                 bool set;
 737                 vm_fault_t ret;
 738                 pgtable = pte_alloc_one(vma->vm_mm);
 739                 if (unlikely(!pgtable))
 740                         return VM_FAULT_OOM;
 741                 zero_page = mm_get_huge_zero_page(vma->vm_mm);
 742                 if (unlikely(!zero_page)) {
 743                         pte_free(vma->vm_mm, pgtable);
 744                         count_vm_event(THP_FAULT_FALLBACK);
 745                         return VM_FAULT_FALLBACK;
 746                 }
 747                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 748                 ret = 0;
 749                 set = false;
 750                 if (pmd_none(*vmf->pmd)) {
 751                         ret = check_stable_address_space(vma->vm_mm);
 752                         if (ret) {
 753                                 spin_unlock(vmf->ptl);
 754                         } else if (userfaultfd_missing(vma)) {
 755                                 spin_unlock(vmf->ptl);
 756                                 ret = handle_userfault(vmf, VM_UFFD_MISSING);
 757                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 758                         } else {
 759                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
 760                                                    haddr, vmf->pmd, zero_page);
 761                                 spin_unlock(vmf->ptl);
 762                                 set = true;
 763                         }
 764                 } else
 765                         spin_unlock(vmf->ptl);
 766                 if (!set)
 767                         pte_free(vma->vm_mm, pgtable);
 768                 return ret;
 769         }
 770         gfp = alloc_hugepage_direct_gfpmask(vma);
 771         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 772         if (unlikely(!page)) {
 773                 count_vm_event(THP_FAULT_FALLBACK);
 774                 return VM_FAULT_FALLBACK;
 775         }
 776         prep_transhuge_page(page);
 777         return __do_huge_pmd_anonymous_page(vmf, page, gfp);
 778 }
 779
 780 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 781                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
 782                 pgtable_t pgtable)
 783 {
 784         struct mm_struct *mm = vma->vm_mm;
 785         pmd_t entry;
 786         spinlock_t *ptl;
 787
 788         ptl = pmd_lock(mm, pmd);
 789         if (!pmd_none(*pmd)) {
 790                 if (write) {
 791                         if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
 792                                 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
 793                                 goto out_unlock;
 794                         }
 795                         entry = pmd_mkyoung(*pmd);
 796                         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 797                         if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
 798                                 update_mmu_cache_pmd(vma, addr, pmd);
 799                 }
 800
 801                 goto out_unlock;
 802         }
 803
 804         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 805         if (pfn_t_devmap(pfn))
 806                 entry = pmd_mkdevmap(entry);
 807         if (write) {
 808                 entry = pmd_mkyoung(pmd_mkdirty(entry));
 809                 entry = maybe_pmd_mkwrite(entry, vma);
 810         }
 811
 812         if (pgtable) {
 813                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 814                 mm_inc_nr_ptes(mm);
 815                 pgtable = NULL;
 816         }
 817
 818         set_pmd_at(mm, addr, pmd, entry);
 819         update_mmu_cache_pmd(vma, addr, pmd);
 820
 821 out_unlock:
 822         spin_unlock(ptl);
 823         if (pgtable)
 824                 pte_free(mm, pgtable);
 825 }
 826
 827 /**
 828  * vmf_insert_pfn_pmd_prot - insert a pmd size pfn
 829  * @vmf: Structure describing the fault
 830  * @pfn: pfn to insert
 831  * @pgprot: page protection to use
 832  * @write: whether it's a write fault
 833  *
 834  * Insert a pmd size pfn. See vmf_insert_pfn() for additional info and
 835  * also consult the vmf_insert_mixed_prot() documentation when
 836  * @pgprot != @vmf->vma->vm_page_prot.
 837  *
 838  * Return: vm_fault_t value.
 839  */
 840 vm_fault_t vmf_insert_pfn_pmd_prot(struct vm_fault *vmf, pfn_t pfn,
 841                                    pgprot_t pgprot, bool write)
 842 {
 843         unsigned long addr = vmf->address & PMD_MASK;
 844         struct vm_area_struct *vma = vmf->vma;
 845         pgtable_t pgtable = NULL;
 846
 847         /*
 848          * If we had pmd_special, we could avoid all these restrictions,
 849          * but we need to be consistent with PTEs and architectures that
 850          * can't support a 'special' bit.
 851          */
 852         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 853                         !pfn_t_devmap(pfn));
 854         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 855                                                 (VM_PFNMAP|VM_MIXEDMAP));
 856         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 857
 858         if (addr < vma->vm_start || addr >= vma->vm_end)
 859                 return VM_FAULT_SIGBUS;
 860
 861         if (arch_needs_pgtable_deposit()) {
 862                 pgtable = pte_alloc_one(vma->vm_mm);
 863                 if (!pgtable)
 864                         return VM_FAULT_OOM;
 865         }
 866
 867         track_pfn_insert(vma, &pgprot, pfn);
 868
 869         insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
 870         return VM_FAULT_NOPAGE;
 871 }
 872 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd_prot);
 873
 874 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 875 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
 876 {
 877         if (likely(vma->vm_flags & VM_WRITE))
 878                 pud = pud_mkwrite(pud);
 879         return pud;
 880 }
 881
 882 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 883                 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
 884 {
 885         struct mm_struct *mm = vma->vm_mm;
 886         pud_t entry;
 887         spinlock_t *ptl;
 888
 889         ptl = pud_lock(mm, pud);
 890         if (!pud_none(*pud)) {
 891                 if (write) {
 892                         if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
 893                                 WARN_ON_ONCE(!is_huge_zero_pud(*pud));
 894                                 goto out_unlock;
 895                         }
 896                         entry = pud_mkyoung(*pud);
 897                         entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
 898                         if (pudp_set_access_flags(vma, addr, pud, entry, 1))
 899                                 update_mmu_cache_pud(vma, addr, pud);
 900                 }
 901                 goto out_unlock;
 902         }
 903
 904         entry = pud_mkhuge(pfn_t_pud(pfn, prot));
 905         if (pfn_t_devmap(pfn))
 906                 entry = pud_mkdevmap(entry);
 907         if (write) {
 908                 entry = pud_mkyoung(pud_mkdirty(entry));
 909                 entry = maybe_pud_mkwrite(entry, vma);
 910         }
 911         set_pud_at(mm, addr, pud, entry);
 912         update_mmu_cache_pud(vma, addr, pud);
 913
 914 out_unlock:
 915         spin_unlock(ptl);
 916 }
 917
 918 /**
 919  * vmf_insert_pfn_pud_prot - insert a pud size pfn
 920  * @vmf: Structure describing the fault
 921  * @pfn: pfn to insert
 922  * @pgprot: page protection to use
 923  * @write: whether it's a write fault
 924  *
 925  * Insert a pud size pfn. See vmf_insert_pfn() for additional info and
 926  * also consult the vmf_insert_mixed_prot() documentation when
 927  * @pgprot != @vmf->vma->vm_page_prot.
 928  *
 929  * Return: vm_fault_t value.
 930  */
 931 vm_fault_t vmf_insert_pfn_pud_prot(struct vm_fault *vmf, pfn_t pfn,
 932                                    pgprot_t pgprot, bool write)
 933 {
 934         unsigned long addr = vmf->address & PUD_MASK;
 935         struct vm_area_struct *vma = vmf->vma;
 936
 937         /*
 938          * If we had pud_special, we could avoid all these restrictions,
 939          * but we need to be consistent with PTEs and architectures that
 940          * can't support a 'special' bit.
 941          */
 942         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 943                         !pfn_t_devmap(pfn));
 944         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 945                                                 (VM_PFNMAP|VM_MIXEDMAP));
 946         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 947
 948         if (addr < vma->vm_start || addr >= vma->vm_end)
 949                 return VM_FAULT_SIGBUS;
 950
 951         track_pfn_insert(vma, &pgprot, pfn);
 952
 953         insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
 954         return VM_FAULT_NOPAGE;
 955 }
 956 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud_prot);
 957 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
 958
 959 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
 960                 pmd_t *pmd, int flags)
 961 {
 962         pmd_t _pmd;
 963
 964         _pmd = pmd_mkyoung(*pmd);
 965         if (flags & FOLL_WRITE)
 966                 _pmd = pmd_mkdirty(_pmd);
 967         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
 968                                 pmd, _pmd, flags & FOLL_WRITE))
 969                 update_mmu_cache_pmd(vma, addr, pmd);
 970 }
 971
 972 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 973                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
 974 {
 975         unsigned long pfn = pmd_pfn(*pmd);
 976         struct mm_struct *mm = vma->vm_mm;
 977         struct page *page;
 978
 979         assert_spin_locked(pmd_lockptr(mm, pmd));
 980
 981         /*
 982          * When we COW a devmap PMD entry, we split it into PTEs, so we should
 983          * not be in this function with `flags & FOLL_COW` set.
 984          */
 985         WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
 986
 987         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
 988         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
 989                          (FOLL_PIN | FOLL_GET)))
 990                 return NULL;
 991
 992         if (flags & FOLL_WRITE && !pmd_write(*pmd))
 993                 return NULL;
 994
 995         if (pmd_present(*pmd) && pmd_devmap(*pmd))
 996                 /* pass */;
 997         else
 998                 return NULL;
 999
1000         if (flags & FOLL_TOUCH)
1001                 touch_pmd(vma, addr, pmd, flags);
1002
1003         /*
1004          * device mapped pages can only be returned if the
1005          * caller will manage the page reference count.
1006          */
1007         if (!(flags & (FOLL_GET | FOLL_PIN)))
1008                 return ERR_PTR(-EEXIST);
1009
1010         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1011         *pgmap = get_dev_pagemap(pfn, *pgmap);
1012         if (!*pgmap)
1013                 return ERR_PTR(-EFAULT);
1014         page = pfn_to_page(pfn);
1015         if (!try_grab_page(page, flags))
1016                 page = ERR_PTR(-ENOMEM);
1017
1018         return page;
1019 }
1020
1021 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1022                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1023                   struct vm_area_struct *vma)
1024 {
1025         spinlock_t *dst_ptl, *src_ptl;
1026         struct page *src_page;
1027         pmd_t pmd;
1028         pgtable_t pgtable = NULL;
1029         int ret = -ENOMEM;
1030
1031         /* Skip if can be re-fill on fault */
1032         if (!vma_is_anonymous(vma))
1033                 return 0;
1034
1035         pgtable = pte_alloc_one(dst_mm);
1036         if (unlikely(!pgtable))
1037                 goto out;
1038
1039         dst_ptl = pmd_lock(dst_mm, dst_pmd);
1040         src_ptl = pmd_lockptr(src_mm, src_pmd);
1041         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1042
1043         ret = -EAGAIN;
1044         pmd = *src_pmd;
1045
1046 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1047         if (unlikely(is_swap_pmd(pmd))) {
1048                 swp_entry_t entry = pmd_to_swp_entry(pmd);
1049
1050                 VM_BUG_ON(!is_pmd_migration_entry(pmd));
1051                 if (is_write_migration_entry(entry)) {
1052                         make_migration_entry_read(&entry);
1053                         pmd = swp_entry_to_pmd(entry);
1054                         if (pmd_swp_soft_dirty(*src_pmd))
1055                                 pmd = pmd_swp_mksoft_dirty(pmd);
1056                         set_pmd_at(src_mm, addr, src_pmd, pmd);
1057                 }
1058                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1059                 mm_inc_nr_ptes(dst_mm);
1060                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1061                 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1062                 ret = 0;
1063                 goto out_unlock;
1064         }
1065 #endif
1066
1067         if (unlikely(!pmd_trans_huge(pmd))) {
1068                 pte_free(dst_mm, pgtable);
1069                 goto out_unlock;
1070         }
1071         /*
1072          * When page table lock is held, the huge zero pmd should not be
1073          * under splitting since we don't split the page itself, only pmd to
1074          * a page table.
1075          */
1076         if (is_huge_zero_pmd(pmd)) {
1077                 struct page *zero_page;
1078                 /*
1079                  * get_huge_zero_page() will never allocate a new page here,
1080                  * since we already have a zero page to copy. It just takes a
1081                  * reference.
1082                  */
1083                 zero_page = mm_get_huge_zero_page(dst_mm);
1084                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1085                                 zero_page);
1086                 ret = 0;
1087                 goto out_unlock;
1088         }
1089
1090         src_page = pmd_page(pmd);
1091         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1092         get_page(src_page);
1093         page_dup_rmap(src_page, true);
1094         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1095         mm_inc_nr_ptes(dst_mm);
1096         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1097
1098         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1099         pmd = pmd_mkold(pmd_wrprotect(pmd));
1100         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1101
1102         ret = 0;
1103 out_unlock:
1104         spin_unlock(src_ptl);
1105         spin_unlock(dst_ptl);
1106 out:
1107         return ret;
1108 }
1109
1110 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1111 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1112                 pud_t *pud, int flags)
1113 {
1114         pud_t _pud;
1115
1116         _pud = pud_mkyoung(*pud);
1117         if (flags & FOLL_WRITE)
1118                 _pud = pud_mkdirty(_pud);
1119         if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1120                                 pud, _pud, flags & FOLL_WRITE))
1121                 update_mmu_cache_pud(vma, addr, pud);
1122 }
1123
1124 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1125                 pud_t *pud, int flags, struct dev_pagemap **pgmap)
1126 {
1127         unsigned long pfn = pud_pfn(*pud);
1128         struct mm_struct *mm = vma->vm_mm;
1129         struct page *page;
1130
1131         assert_spin_locked(pud_lockptr(mm, pud));
1132
1133         if (flags & FOLL_WRITE && !pud_write(*pud))
1134                 return NULL;
1135
1136         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
1137         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
1138                          (FOLL_PIN | FOLL_GET)))
1139                 return NULL;
1140
1141         if (pud_present(*pud) && pud_devmap(*pud))
1142                 /* pass */;
1143         else
1144                 return NULL;
1145
1146         if (flags & FOLL_TOUCH)
1147                 touch_pud(vma, addr, pud, flags);
1148
1149         /*
1150          * device mapped pages can only be returned if the
1151          * caller will manage the page reference count.
1152          *
1153          * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here:
1154          */
1155         if (!(flags & (FOLL_GET | FOLL_PIN)))
1156                 return ERR_PTR(-EEXIST);
1157
1158         pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1159         *pgmap = get_dev_pagemap(pfn, *pgmap);
1160         if (!*pgmap)
1161                 return ERR_PTR(-EFAULT);
1162         page = pfn_to_page(pfn);
1163         if (!try_grab_page(page, flags))
1164                 page = ERR_PTR(-ENOMEM);
1165
1166         return page;
1167 }
1168
1169 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1170                   pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1171                   struct vm_area_struct *vma)
1172 {
1173         spinlock_t *dst_ptl, *src_ptl;
1174         pud_t pud;
1175         int ret;
1176
1177         dst_ptl = pud_lock(dst_mm, dst_pud);
1178         src_ptl = pud_lockptr(src_mm, src_pud);
1179         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1180
1181         ret = -EAGAIN;
1182         pud = *src_pud;
1183         if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1184                 goto out_unlock;
1185
1186         /*
1187          * When page table lock is held, the huge zero pud should not be
1188          * under splitting since we don't split the page itself, only pud to
1189          * a page table.
1190          */
1191         if (is_huge_zero_pud(pud)) {
1192                 /* No huge zero pud yet */
1193         }
1194
1195         pudp_set_wrprotect(src_mm, addr, src_pud);
1196         pud = pud_mkold(pud_wrprotect(pud));
1197         set_pud_at(dst_mm, addr, dst_pud, pud);
1198
1199         ret = 0;
1200 out_unlock:
1201         spin_unlock(src_ptl);
1202         spin_unlock(dst_ptl);
1203         return ret;
1204 }
1205
1206 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1207 {
1208         pud_t entry;
1209         unsigned long haddr;
1210         bool write = vmf->flags & FAULT_FLAG_WRITE;
1211
1212         vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1213         if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1214                 goto unlock;
1215
1216         entry = pud_mkyoung(orig_pud);
1217         if (write)
1218                 entry = pud_mkdirty(entry);
1219         haddr = vmf->address & HPAGE_PUD_MASK;
1220         if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1221                 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1222
1223 unlock:
1224         spin_unlock(vmf->ptl);
1225 }
1226 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1227
1228 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1229 {
1230         pmd_t entry;
1231         unsigned long haddr;
1232         bool write = vmf->flags & FAULT_FLAG_WRITE;
1233
1234         vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1235         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1236                 goto unlock;
1237
1238         entry = pmd_mkyoung(orig_pmd);
1239         if (write)
1240                 entry = pmd_mkdirty(entry);
1241         haddr = vmf->address & HPAGE_PMD_MASK;
1242         if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1243                 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1244
1245 unlock:
1246         spin_unlock(vmf->ptl);
1247 }
1248
1249 static vm_fault_t do_huge_pmd_wp_page_fallback(struct vm_fault *vmf,
1250                         pmd_t orig_pmd, struct page *page)
1251 {
1252         struct vm_area_struct *vma = vmf->vma;
1253         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1254         struct mem_cgroup *memcg;
1255         pgtable_t pgtable;
1256         pmd_t _pmd;
1257         int i;
1258         vm_fault_t ret = 0;
1259         struct page **pages;
1260         struct mmu_notifier_range range;
1261
1262         pages = kmalloc_array(HPAGE_PMD_NR, sizeof(struct page *),
1263                               GFP_KERNEL);
1264         if (unlikely(!pages)) {
1265                 ret |= VM_FAULT_OOM;
1266                 goto out;
1267         }
1268
1269         for (i = 0; i < HPAGE_PMD_NR; i++) {
1270                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
1271                                                vmf->address, page_to_nid(page));
1272                 if (unlikely(!pages[i] ||
1273                              mem_cgroup_try_charge_delay(pages[i], vma->vm_mm,
1274                                      GFP_KERNEL, &memcg, false))) {
1275                         if (pages[i])
1276                                 put_page(pages[i]);
1277                         while (--i >= 0) {
1278                                 memcg = (void *)page_private(pages[i]);
1279                                 set_page_private(pages[i], 0);
1280                                 mem_cgroup_cancel_charge(pages[i], memcg,
1281                                                 false);
1282                                 put_page(pages[i]);
1283                         }
1284                         kfree(pages);
1285                         ret |= VM_FAULT_OOM;
1286                         goto out;
1287                 }
1288                 set_page_private(pages[i], (unsigned long)memcg);
1289         }
1290
1291         for (i = 0; i < HPAGE_PMD_NR; i++) {
1292                 copy_user_highpage(pages[i], page + i,
1293                                    haddr + PAGE_SIZE * i, vma);
1294                 __SetPageUptodate(pages[i]);
1295                 cond_resched();
1296         }
1297
1298         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1299                                 haddr, haddr + HPAGE_PMD_SIZE);
1300         mmu_notifier_invalidate_range_start(&range);
1301
1302         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1303         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1304                 goto out_free_pages;
1305         VM_BUG_ON_PAGE(!PageHead(page), page);
1306
1307         /*
1308          * Leave pmd empty until pte is filled note we must notify here as
1309          * concurrent CPU thread might write to new page before the call to
1310          * mmu_notifier_invalidate_range_end() happens which can lead to a
1311          * device seeing memory write in different order than CPU.
1312          *
1313          * See Documentation/vm/mmu_notifier.rst
1314          */
1315         pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1316
1317         pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
1318         pmd_populate(vma->vm_mm, &_pmd, pgtable);
1319
1320         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1321                 pte_t entry;
1322                 entry = mk_pte(pages[i], vma->vm_page_prot);
1323                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1324                 memcg = (void *)page_private(pages[i]);
1325                 set_page_private(pages[i], 0);
1326                 page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
1327                 mem_cgroup_commit_charge(pages[i], memcg, false, false);
1328                 lru_cache_add_active_or_unevictable(pages[i], vma);
1329                 vmf->pte = pte_offset_map(&_pmd, haddr);
1330                 VM_BUG_ON(!pte_none(*vmf->pte));
1331                 set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
1332                 pte_unmap(vmf->pte);
1333         }
1334         kfree(pages);
1335
1336         smp_wmb(); /* make pte visible before pmd */
1337         pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
1338         page_remove_rmap(page, true);
1339         spin_unlock(vmf->ptl);
1340
1341         /*
1342          * No need to double call mmu_notifier->invalidate_range() callback as
1343          * the above pmdp_huge_clear_flush_notify() did already call it.
1344          */
1345         mmu_notifier_invalidate_range_only_end(&range);
1346
1347         ret |= VM_FAULT_WRITE;
1348         put_page(page);
1349
1350 out:
1351         return ret;
1352
1353 out_free_pages:
1354         spin_unlock(vmf->ptl);
1355         mmu_notifier_invalidate_range_end(&range);
1356         for (i = 0; i < HPAGE_PMD_NR; i++) {
1357                 memcg = (void *)page_private(pages[i]);
1358                 set_page_private(pages[i], 0);
1359                 mem_cgroup_cancel_charge(pages[i], memcg, false);
1360                 put_page(pages[i]);
1361         }
1362         kfree(pages);
1363         goto out;
1364 }
1365
1366 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1367 {
1368         struct vm_area_struct *vma = vmf->vma;
1369         struct page *page = NULL, *new_page;
1370         struct mem_cgroup *memcg;
1371         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1372         struct mmu_notifier_range range;
1373         gfp_t huge_gfp;                 /* for allocation and charge */
1374         vm_fault_t ret = 0;
1375
1376         vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1377         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1378         if (is_huge_zero_pmd(orig_pmd))
1379                 goto alloc;
1380         spin_lock(vmf->ptl);
1381         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1382                 goto out_unlock;
1383
1384         page = pmd_page(orig_pmd);
1385         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1386         /*
1387          * We can only reuse the page if nobody else maps the huge page or it's
1388          * part.
1389          */
1390         if (!trylock_page(page)) {
1391                 get_page(page);
1392                 spin_unlock(vmf->ptl);
1393                 lock_page(page);
1394                 spin_lock(vmf->ptl);
1395                 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1396                         unlock_page(page);
1397                         put_page(page);
1398                         goto out_unlock;
1399                 }
1400                 put_page(page);
1401         }
1402         if (reuse_swap_page(page, NULL)) {
1403                 pmd_t entry;
1404                 entry = pmd_mkyoung(orig_pmd);
1405                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1406                 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry,  1))
1407                         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1408                 ret |= VM_FAULT_WRITE;
1409                 unlock_page(page);
1410                 goto out_unlock;
1411         }
1412         unlock_page(page);
1413         get_page(page);
1414         spin_unlock(vmf->ptl);
1415 alloc:
1416         if (__transparent_hugepage_enabled(vma) &&
1417             !transparent_hugepage_debug_cow()) {
1418                 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1419                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1420         } else
1421                 new_page = NULL;
1422
1423         if (likely(new_page)) {
1424                 prep_transhuge_page(new_page);
1425         } else {
1426                 if (!page) {
1427                         split_huge_pmd(vma, vmf->pmd, vmf->address);
1428                         ret |= VM_FAULT_FALLBACK;
1429                 } else {
1430                         ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
1431                         if (ret & VM_FAULT_OOM) {
1432                                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1433                                 ret |= VM_FAULT_FALLBACK;
1434                         }
1435                         put_page(page);
1436                 }
1437                 count_vm_event(THP_FAULT_FALLBACK);
1438                 goto out;
1439         }
1440
1441         if (unlikely(mem_cgroup_try_charge_delay(new_page, vma->vm_mm,
1442                                         huge_gfp, &memcg, true))) {
1443                 put_page(new_page);
1444                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1445                 if (page)
1446                         put_page(page);
1447                 ret |= VM_FAULT_FALLBACK;
1448                 count_vm_event(THP_FAULT_FALLBACK);
1449                 goto out;
1450         }
1451
1452         count_vm_event(THP_FAULT_ALLOC);
1453         count_memcg_events(memcg, THP_FAULT_ALLOC, 1);
1454
1455         if (!page)
1456                 clear_huge_page(new_page, vmf->address, HPAGE_PMD_NR);
1457         else
1458                 copy_user_huge_page(new_page, page, vmf->address,
1459                                     vma, HPAGE_PMD_NR);
1460         __SetPageUptodate(new_page);
1461
1462         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1463                                 haddr, haddr + HPAGE_PMD_SIZE);
1464         mmu_notifier_invalidate_range_start(&range);
1465
1466         spin_lock(vmf->ptl);
1467         if (page)
1468                 put_page(page);
1469         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1470                 spin_unlock(vmf->ptl);
1471                 mem_cgroup_cancel_charge(new_page, memcg, true);
1472                 put_page(new_page);
1473                 goto out_mn;
1474         } else {
1475                 pmd_t entry;
1476                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1477                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1478                 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1479                 page_add_new_anon_rmap(new_page, vma, haddr, true);
1480                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1481                 lru_cache_add_active_or_unevictable(new_page, vma);
1482                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1483                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1484                 if (!page) {
1485                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1486                 } else {
1487                         VM_BUG_ON_PAGE(!PageHead(page), page);
1488                         page_remove_rmap(page, true);
1489                         put_page(page);
1490                 }
1491                 ret |= VM_FAULT_WRITE;
1492         }
1493         spin_unlock(vmf->ptl);
1494 out_mn:
1495         /*
1496          * No need to double call mmu_notifier->invalidate_range() callback as
1497          * the above pmdp_huge_clear_flush_notify() did already call it.
1498          */
1499         mmu_notifier_invalidate_range_only_end(&range);
1500 out:
1501         return ret;
1502 out_unlock:
1503         spin_unlock(vmf->ptl);
1504         return ret;
1505 }
1506
1507 /*
1508  * FOLL_FORCE can write to even unwritable pmd's, but only
1509  * after we've gone through a COW cycle and they are dirty.
1510  */
1511 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1512 {
1513         return pmd_write(pmd) ||
1514                ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1515 }
1516
1517 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1518                                    unsigned long addr,
1519                                    pmd_t *pmd,
1520                                    unsigned int flags)
1521 {
1522         struct mm_struct *mm = vma->vm_mm;
1523         struct page *page = NULL;
1524
1525         assert_spin_locked(pmd_lockptr(mm, pmd));
1526
1527         if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1528                 goto out;
1529
1530         /* Avoid dumping huge zero page */
1531         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1532                 return ERR_PTR(-EFAULT);
1533
1534         /* Full NUMA hinting faults to serialise migration in fault paths */
1535         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1536                 goto out;
1537
1538         page = pmd_page(*pmd);
1539         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1540
1541         if (!try_grab_page(page, flags))
1542                 return ERR_PTR(-ENOMEM);
1543
1544         if (flags & FOLL_TOUCH)
1545                 touch_pmd(vma, addr, pmd, flags);
1546
1547         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1548                 /*
1549                  * We don't mlock() pte-mapped THPs. This way we can avoid
1550                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1551                  *
1552                  * For anon THP:
1553                  *
1554                  * In most cases the pmd is the only mapping of the page as we
1555                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1556                  * writable private mappings in populate_vma_page_range().
1557                  *
1558                  * The only scenario when we have the page shared here is if we
1559                  * mlocking read-only mapping shared over fork(). We skip
1560                  * mlocking such pages.
1561                  *
1562                  * For file THP:
1563                  *
1564                  * We can expect PageDoubleMap() to be stable under page lock:
1565                  * for file pages we set it in page_add_file_rmap(), which
1566                  * requires page to be locked.
1567                  */
1568
1569                 if (PageAnon(page) && compound_mapcount(page) != 1)
1570                         goto skip_mlock;
1571                 if (PageDoubleMap(page) || !page->mapping)
1572                         goto skip_mlock;
1573                 if (!trylock_page(page))
1574                         goto skip_mlock;
1575                 lru_add_drain();
1576                 if (page->mapping && !PageDoubleMap(page))
1577                         mlock_vma_page(page);
1578                 unlock_page(page);
1579         }
1580 skip_mlock:
1581         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1582         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1583
1584 out:
1585         return page;
1586 }
1587
1588 /* NUMA hinting page fault entry point for trans huge pmds */
1589 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1590 {
1591         struct vm_area_struct *vma = vmf->vma;
1592         struct anon_vma *anon_vma = NULL;
1593         struct page *page;
1594         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1595         int page_nid = NUMA_NO_NODE, this_nid = numa_node_id();
1596         int target_nid, last_cpupid = -1;
1597         bool page_locked;
1598         bool migrated = false;
1599         bool was_writable;
1600         int flags = 0;
1601
1602         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1603         if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1604                 goto out_unlock;
1605
1606         /*
1607          * If there are potential migrations, wait for completion and retry
1608          * without disrupting NUMA hinting information. Do not relock and
1609          * check_same as the page may no longer be mapped.
1610          */
1611         if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1612                 page = pmd_page(*vmf->pmd);
1613                 if (!get_page_unless_zero(page))
1614                         goto out_unlock;
1615                 spin_unlock(vmf->ptl);
1616                 put_and_wait_on_page_locked(page);
1617                 goto out;
1618         }
1619
1620         page = pmd_page(pmd);
1621         BUG_ON(is_huge_zero_page(page));
1622         page_nid = page_to_nid(page);
1623         last_cpupid = page_cpupid_last(page);
1624         count_vm_numa_event(NUMA_HINT_FAULTS);
1625         if (page_nid == this_nid) {
1626                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1627                 flags |= TNF_FAULT_LOCAL;
1628         }
1629
1630         /* See similar comment in do_numa_page for explanation */
1631         if (!pmd_savedwrite(pmd))
1632                 flags |= TNF_NO_GROUP;
1633
1634         /*
1635          * Acquire the page lock to serialise THP migrations but avoid dropping
1636          * page_table_lock if at all possible
1637          */
1638         page_locked = trylock_page(page);
1639         target_nid = mpol_misplaced(page, vma, haddr);
1640         if (target_nid == NUMA_NO_NODE) {
1641                 /* If the page was locked, there are no parallel migrations */
1642                 if (page_locked)
1643                         goto clear_pmdnuma;
1644         }
1645
1646         /* Migration could have started since the pmd_trans_migrating check */
1647         if (!page_locked) {
1648                 page_nid = NUMA_NO_NODE;
1649                 if (!get_page_unless_zero(page))
1650                         goto out_unlock;
1651                 spin_unlock(vmf->ptl);
1652                 put_and_wait_on_page_locked(page);
1653                 goto out;
1654         }
1655
1656         /*
1657          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1658          * to serialises splits
1659          */
1660         get_page(page);
1661         spin_unlock(vmf->ptl);
1662         anon_vma = page_lock_anon_vma_read(page);
1663
1664         /* Confirm the PMD did not change while page_table_lock was released */
1665         spin_lock(vmf->ptl);
1666         if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1667                 unlock_page(page);
1668                 put_page(page);
1669                 page_nid = NUMA_NO_NODE;
1670                 goto out_unlock;
1671         }
1672
1673         /* Bail if we fail to protect against THP splits for any reason */
1674         if (unlikely(!anon_vma)) {
1675                 put_page(page);
1676                 page_nid = NUMA_NO_NODE;
1677                 goto clear_pmdnuma;
1678         }
1679
1680         /*
1681          * Since we took the NUMA fault, we must have observed the !accessible
1682          * bit. Make sure all other CPUs agree with that, to avoid them
1683          * modifying the page we're about to migrate.
1684          *
1685          * Must be done under PTL such that we'll observe the relevant
1686          * inc_tlb_flush_pending().
1687          *
1688          * We are not sure a pending tlb flush here is for a huge page
1689          * mapping or not. Hence use the tlb range variant
1690          */
1691         if (mm_tlb_flush_pending(vma->vm_mm)) {
1692                 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1693                 /*
1694                  * change_huge_pmd() released the pmd lock before
1695                  * invalidating the secondary MMUs sharing the primary
1696                  * MMU pagetables (with ->invalidate_range()). The
1697                  * mmu_notifier_invalidate_range_end() (which
1698                  * internally calls ->invalidate_range()) in
1699                  * change_pmd_range() will run after us, so we can't
1700                  * rely on it here and we need an explicit invalidate.
1701                  */
1702                 mmu_notifier_invalidate_range(vma->vm_mm, haddr,
1703                                               haddr + HPAGE_PMD_SIZE);
1704         }
1705
1706         /*
1707          * Migrate the THP to the requested node, returns with page unlocked
1708          * and access rights restored.
1709          */
1710         spin_unlock(vmf->ptl);
1711
1712         migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1713                                 vmf->pmd, pmd, vmf->address, page, target_nid);
1714         if (migrated) {
1715                 flags |= TNF_MIGRATED;
1716                 page_nid = target_nid;
1717         } else
1718                 flags |= TNF_MIGRATE_FAIL;
1719
1720         goto out;
1721 clear_pmdnuma:
1722         BUG_ON(!PageLocked(page));
1723         was_writable = pmd_savedwrite(pmd);
1724         pmd = pmd_modify(pmd, vma->vm_page_prot);
1725         pmd = pmd_mkyoung(pmd);
1726         if (was_writable)
1727                 pmd = pmd_mkwrite(pmd);
1728         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1729         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1730         unlock_page(page);
1731 out_unlock:
1732         spin_unlock(vmf->ptl);
1733
1734 out:
1735         if (anon_vma)
1736                 page_unlock_anon_vma_read(anon_vma);
1737
1738         if (page_nid != NUMA_NO_NODE)
1739                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1740                                 flags);
1741
1742         return 0;
1743 }
1744
1745 /*
1746  * Return true if we do MADV_FREE successfully on entire pmd page.
1747  * Otherwise, return false.
1748  */
1749 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1750                 pmd_t *pmd, unsigned long addr, unsigned long next)
1751 {
1752         spinlock_t *ptl;
1753         pmd_t orig_pmd;
1754         struct page *page;
1755         struct mm_struct *mm = tlb->mm;
1756         bool ret = false;
1757
1758         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1759
1760         ptl = pmd_trans_huge_lock(pmd, vma);
1761         if (!ptl)
1762                 goto out_unlocked;
1763
1764         orig_pmd = *pmd;
1765         if (is_huge_zero_pmd(orig_pmd))
1766                 goto out;
1767
1768         if (unlikely(!pmd_present(orig_pmd))) {
1769                 VM_BUG_ON(thp_migration_supported() &&
1770                                   !is_pmd_migration_entry(orig_pmd));
1771                 goto out;
1772         }
1773
1774         page = pmd_page(orig_pmd);
1775         /*
1776          * If other processes are mapping this page, we couldn't discard
1777          * the page unless they all do MADV_FREE so let's skip the page.
1778          */
1779         if (page_mapcount(page) != 1)
1780                 goto out;
1781
1782         if (!trylock_page(page))
1783                 goto out;
1784
1785         /*
1786          * If user want to discard part-pages of THP, split it so MADV_FREE
1787          * will deactivate only them.
1788          */
1789         if (next - addr != HPAGE_PMD_SIZE) {
1790                 get_page(page);
1791                 spin_unlock(ptl);
1792                 split_huge_page(page);
1793                 unlock_page(page);
1794                 put_page(page);
1795                 goto out_unlocked;
1796         }
1797
1798         if (PageDirty(page))
1799                 ClearPageDirty(page);
1800         unlock_page(page);
1801
1802         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1803                 pmdp_invalidate(vma, addr, pmd);
1804                 orig_pmd = pmd_mkold(orig_pmd);
1805                 orig_pmd = pmd_mkclean(orig_pmd);
1806
1807                 set_pmd_at(mm, addr, pmd, orig_pmd);
1808                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1809         }
1810
1811         mark_page_lazyfree(page);
1812         ret = true;
1813 out:
1814         spin_unlock(ptl);
1815 out_unlocked:
1816         return ret;
1817 }
1818
1819 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1820 {
1821         pgtable_t pgtable;
1822
1823         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1824         pte_free(mm, pgtable);
1825         mm_dec_nr_ptes(mm);
1826 }
1827
1828 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1829                  pmd_t *pmd, unsigned long addr)
1830 {
1831         pmd_t orig_pmd;
1832         spinlock_t *ptl;
1833
1834         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1835
1836         ptl = __pmd_trans_huge_lock(pmd, vma);
1837         if (!ptl)
1838                 return 0;
1839         /*
1840          * For architectures like ppc64 we look at deposited pgtable
1841          * when calling pmdp_huge_get_and_clear. So do the
1842          * pgtable_trans_huge_withdraw after finishing pmdp related
1843          * operations.
1844          */
1845         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1846                         tlb->fullmm);
1847         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1848         if (vma_is_special_huge(vma)) {
1849                 if (arch_needs_pgtable_deposit())
1850                         zap_deposited_table(tlb->mm, pmd);
1851                 spin_unlock(ptl);
1852                 if (is_huge_zero_pmd(orig_pmd))
1853                         tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1854         } else if (is_huge_zero_pmd(orig_pmd)) {
1855                 zap_deposited_table(tlb->mm, pmd);
1856                 spin_unlock(ptl);
1857                 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1858         } else {
1859                 struct page *page = NULL;
1860                 int flush_needed = 1;
1861
1862                 if (pmd_present(orig_pmd)) {
1863                         page = pmd_page(orig_pmd);
1864                         page_remove_rmap(page, true);
1865                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1866                         VM_BUG_ON_PAGE(!PageHead(page), page);
1867                 } else if (thp_migration_supported()) {
1868                         swp_entry_t entry;
1869
1870                         VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1871                         entry = pmd_to_swp_entry(orig_pmd);
1872                         page = pfn_to_page(swp_offset(entry));
1873                         flush_needed = 0;
1874                 } else
1875                         WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1876
1877                 if (PageAnon(page)) {
1878                         zap_deposited_table(tlb->mm, pmd);
1879                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1880                 } else {
1881                         if (arch_needs_pgtable_deposit())
1882                                 zap_deposited_table(tlb->mm, pmd);
1883                         add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1884                 }
1885
1886                 spin_unlock(ptl);
1887                 if (flush_needed)
1888                         tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1889         }
1890         return 1;
1891 }
1892
1893 #ifndef pmd_move_must_withdraw
1894 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1895                                          spinlock_t *old_pmd_ptl,
1896                                          struct vm_area_struct *vma)
1897 {
1898         /*
1899          * With split pmd lock we also need to move preallocated
1900          * PTE page table if new_pmd is on different PMD page table.
1901          *
1902          * We also don't deposit and withdraw tables for file pages.
1903          */
1904         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1905 }
1906 #endif
1907
1908 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1909 {
1910 #ifdef CONFIG_MEM_SOFT_DIRTY
1911         if (unlikely(is_pmd_migration_entry(pmd)))
1912                 pmd = pmd_swp_mksoft_dirty(pmd);
1913         else if (pmd_present(pmd))
1914                 pmd = pmd_mksoft_dirty(pmd);
1915 #endif
1916         return pmd;
1917 }
1918
1919 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1920                   unsigned long new_addr, unsigned long old_end,
1921                   pmd_t *old_pmd, pmd_t *new_pmd)
1922 {
1923         spinlock_t *old_ptl, *new_ptl;
1924         pmd_t pmd;
1925         struct mm_struct *mm = vma->vm_mm;
1926         bool force_flush = false;
1927
1928         if ((old_addr & ~HPAGE_PMD_MASK) ||
1929             (new_addr & ~HPAGE_PMD_MASK) ||
1930             old_end - old_addr < HPAGE_PMD_SIZE)
1931                 return false;
1932
1933         /*
1934          * The destination pmd shouldn't be established, free_pgtables()
1935          * should have release it.
1936          */
1937         if (WARN_ON(!pmd_none(*new_pmd))) {
1938                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1939                 return false;
1940         }
1941
1942         /*
1943          * We don't have to worry about the ordering of src and dst
1944          * ptlocks because exclusive mmap_sem prevents deadlock.
1945          */
1946         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1947         if (old_ptl) {
1948                 new_ptl = pmd_lockptr(mm, new_pmd);
1949                 if (new_ptl != old_ptl)
1950                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1951                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1952                 if (pmd_present(pmd))
1953                         force_flush = true;
1954                 VM_BUG_ON(!pmd_none(*new_pmd));
1955
1956                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1957                         pgtable_t pgtable;
1958                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1959                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1960                 }
1961                 pmd = move_soft_dirty_pmd(pmd);
1962                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1963                 if (force_flush)
1964                         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1965                 if (new_ptl != old_ptl)
1966                         spin_unlock(new_ptl);
1967                 spin_unlock(old_ptl);
1968                 return true;
1969         }
1970         return false;
1971 }
1972
1973 /*
1974  * Returns
1975  *  - 0 if PMD could not be locked
1976  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1977  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1978  */
1979 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1980                 unsigned long addr, pgprot_t newprot, int prot_numa)
1981 {
1982         struct mm_struct *mm = vma->vm_mm;
1983         spinlock_t *ptl;
1984         pmd_t entry;
1985         bool preserve_write;
1986         int ret;
1987
1988         ptl = __pmd_trans_huge_lock(pmd, vma);
1989         if (!ptl)
1990                 return 0;
1991
1992         preserve_write = prot_numa && pmd_write(*pmd);
1993         ret = 1;
1994
1995 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1996         if (is_swap_pmd(*pmd)) {
1997                 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1998
1999                 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
2000                 if (is_write_migration_entry(entry)) {
2001                         pmd_t newpmd;
2002                         /*
2003                          * A protection check is difficult so
2004                          * just be safe and disable write
2005                          */
2006                         make_migration_entry_read(&entry);
2007                         newpmd = swp_entry_to_pmd(entry);
2008                         if (pmd_swp_soft_dirty(*pmd))
2009                                 newpmd = pmd_swp_mksoft_dirty(newpmd);
2010                         set_pmd_at(mm, addr, pmd, newpmd);
2011                 }
2012                 goto unlock;
2013         }
2014 #endif
2015
2016         /*
2017          * Avoid trapping faults against the zero page. The read-only
2018          * data is likely to be read-cached on the local CPU and
2019          * local/remote hits to the zero page are not interesting.
2020          */
2021         if (prot_numa && is_huge_zero_pmd(*pmd))
2022                 goto unlock;
2023
2024         if (prot_numa && pmd_protnone(*pmd))
2025                 goto unlock;
2026
2027         /*
2028          * In case prot_numa, we are under down_read(mmap_sem). It's critical
2029          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
2030          * which is also under down_read(mmap_sem):
2031          *
2032          *      CPU0:                           CPU1:
2033          *                              change_huge_pmd(prot_numa=1)
2034          *                               pmdp_huge_get_and_clear_notify()
2035          * madvise_dontneed()
2036          *  zap_pmd_range()
2037          *   pmd_trans_huge(*pmd) == 0 (without ptl)
2038          *   // skip the pmd
2039          *                               set_pmd_at();
2040          *                               // pmd is re-established
2041          *
2042          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
2043          * which may break userspace.
2044          *
2045          * pmdp_invalidate() is required to make sure we don't miss
2046          * dirty/young flags set by hardware.
2047          */
2048         entry = pmdp_invalidate(vma, addr, pmd);
2049
2050         entry = pmd_modify(entry, newprot);
2051         if (preserve_write)
2052                 entry = pmd_mk_savedwrite(entry);
2053         ret = HPAGE_PMD_NR;
2054         set_pmd_at(mm, addr, pmd, entry);
2055         BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
2056 unlock:
2057         spin_unlock(ptl);
2058         return ret;
2059 }
2060
2061 /*
2062  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
2063  *
2064  * Note that if it returns page table lock pointer, this routine returns without
2065  * unlocking page table lock. So callers must unlock it.
2066  */
2067 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
2068 {
2069         spinlock_t *ptl;
2070         ptl = pmd_lock(vma->vm_mm, pmd);
2071         if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
2072                         pmd_devmap(*pmd)))
2073                 return ptl;
2074         spin_unlock(ptl);
2075         return NULL;
2076 }
2077
2078 /*
2079  * Returns true if a given pud maps a thp, false otherwise.
2080  *
2081  * Note that if it returns true, this routine returns without unlocking page
2082  * table lock. So callers must unlock it.
2083  */
2084 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
2085 {
2086         spinlock_t *ptl;
2087
2088         ptl = pud_lock(vma->vm_mm, pud);
2089         if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
2090                 return ptl;
2091         spin_unlock(ptl);
2092         return NULL;
2093 }
2094
2095 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
2096 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
2097                  pud_t *pud, unsigned long addr)
2098 {
2099         spinlock_t *ptl;
2100
2101         ptl = __pud_trans_huge_lock(pud, vma);
2102         if (!ptl)
2103                 return 0;
2104         /*
2105          * For architectures like ppc64 we look at deposited pgtable
2106          * when calling pudp_huge_get_and_clear. So do the
2107          * pgtable_trans_huge_withdraw after finishing pudp related
2108          * operations.
2109          */
2110         pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
2111         tlb_remove_pud_tlb_entry(tlb, pud, addr);
2112         if (vma_is_special_huge(vma)) {
2113                 spin_unlock(ptl);
2114                 /* No zero page support yet */
2115         } else {
2116                 /* No support for anonymous PUD pages yet */
2117                 BUG();
2118         }
2119         return 1;
2120 }
2121
2122 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
2123                 unsigned long haddr)
2124 {
2125         VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
2126         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2127         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
2128         VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
2129
2130         count_vm_event(THP_SPLIT_PUD);
2131
2132         pudp_huge_clear_flush_notify(vma, haddr, pud);
2133 }
2134
2135 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2136                 unsigned long address)
2137 {
2138         spinlock_t *ptl;
2139         struct mmu_notifier_range range;
2140
2141         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2142                                 address & HPAGE_PUD_MASK,
2143                                 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
2144         mmu_notifier_invalidate_range_start(&range);
2145         ptl = pud_lock(vma->vm_mm, pud);
2146         if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2147                 goto out;
2148         __split_huge_pud_locked(vma, pud, range.start);
2149
2150 out:
2151         spin_unlock(ptl);
2152         /*
2153          * No need to double call mmu_notifier->invalidate_range() callback as
2154          * the above pudp_huge_clear_flush_notify() did already call it.
2155          */
2156         mmu_notifier_invalidate_range_only_end(&range);
2157 }
2158 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2159
2160 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2161                 unsigned long haddr, pmd_t *pmd)
2162 {
2163         struct mm_struct *mm = vma->vm_mm;
2164         pgtable_t pgtable;
2165         pmd_t _pmd;
2166         int i;
2167
2168         /*
2169          * Leave pmd empty until pte is filled note that it is fine to delay
2170          * notification until mmu_notifier_invalidate_range_end() as we are
2171          * replacing a zero pmd write protected page with a zero pte write
2172          * protected page.
2173          *
2174          * See Documentation/vm/mmu_notifier.rst
2175          */
2176         pmdp_huge_clear_flush(vma, haddr, pmd);
2177
2178         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2179         pmd_populate(mm, &_pmd, pgtable);
2180
2181         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2182                 pte_t *pte, entry;
2183                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2184                 entry = pte_mkspecial(entry);
2185                 pte = pte_offset_map(&_pmd, haddr);
2186                 VM_BUG_ON(!pte_none(*pte));
2187                 set_pte_at(mm, haddr, pte, entry);
2188                 pte_unmap(pte);
2189         }
2190         smp_wmb(); /* make pte visible before pmd */
2191         pmd_populate(mm, pmd, pgtable);
2192 }
2193
2194 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2195                 unsigned long haddr, bool freeze)
2196 {
2197         struct mm_struct *mm = vma->vm_mm;
2198         struct page *page;
2199         pgtable_t pgtable;
2200         pmd_t old_pmd, _pmd;
2201         bool young, write, soft_dirty, pmd_migration = false;
2202         unsigned long addr;
2203         int i;
2204
2205         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2206         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2207         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2208         VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2209                                 && !pmd_devmap(*pmd));
2210
2211         count_vm_event(THP_SPLIT_PMD);
2212
2213         if (!vma_is_anonymous(vma)) {
2214                 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2215                 /*
2216                  * We are going to unmap this huge page. So
2217                  * just go ahead and zap it
2218                  */
2219                 if (arch_needs_pgtable_deposit())
2220                         zap_deposited_table(mm, pmd);
2221                 if (vma_is_special_huge(vma))
2222                         return;
2223                 page = pmd_page(_pmd);
2224                 if (!PageDirty(page) && pmd_dirty(_pmd))
2225                         set_page_dirty(page);
2226                 if (!PageReferenced(page) && pmd_young(_pmd))
2227                         SetPageReferenced(page);
2228                 page_remove_rmap(page, true);
2229                 put_page(page);
2230                 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2231                 return;
2232         } else if (is_huge_zero_pmd(*pmd)) {
2233                 /*
2234                  * FIXME: Do we want to invalidate secondary mmu by calling
2235                  * mmu_notifier_invalidate_range() see comments below inside
2236                  * __split_huge_pmd() ?
2237                  *
2238                  * We are going from a zero huge page write protected to zero
2239                  * small page also write protected so it does not seems useful
2240                  * to invalidate secondary mmu at this time.
2241                  */
2242                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2243         }
2244
2245         /*
2246          * Up to this point the pmd is present and huge and userland has the
2247          * whole access to the hugepage during the split (which happens in
2248          * place). If we overwrite the pmd with the not-huge version pointing
2249          * to the pte here (which of course we could if all CPUs were bug
2250          * free), userland could trigger a small page size TLB miss on the
2251          * small sized TLB while the hugepage TLB entry is still established in
2252          * the huge TLB. Some CPU doesn't like that.
2253          * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2254          * 383 on page 93. Intel should be safe but is also warns that it's
2255          * only safe if the permission and cache attributes of the two entries
2256          * loaded in the two TLB is identical (which should be the case here).
2257          * But it is generally safer to never allow small and huge TLB entries
2258          * for the same virtual address to be loaded simultaneously. So instead
2259          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2260          * current pmd notpresent (atomically because here the pmd_trans_huge
2261          * must remain set at all times on the pmd until the split is complete
2262          * for this pmd), then we flush the SMP TLB and finally we write the
2263          * non-huge version of the pmd entry with pmd_populate.
2264          */
2265         old_pmd = pmdp_invalidate(vma, haddr, pmd);
2266
2267         pmd_migration = is_pmd_migration_entry(old_pmd);
2268         if (unlikely(pmd_migration)) {
2269                 swp_entry_t entry;
2270
2271                 entry = pmd_to_swp_entry(old_pmd);
2272                 page = pfn_to_page(swp_offset(entry));
2273                 write = is_write_migration_entry(entry);
2274                 young = false;
2275                 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2276         } else {
2277                 page = pmd_page(old_pmd);
2278                 if (pmd_dirty(old_pmd))
2279                         SetPageDirty(page);
2280                 write = pmd_write(old_pmd);
2281                 young = pmd_young(old_pmd);
2282                 soft_dirty = pmd_soft_dirty(old_pmd);
2283         }
2284         VM_BUG_ON_PAGE(!page_count(page), page);
2285         page_ref_add(page, HPAGE_PMD_NR - 1);
2286
2287         /*
2288          * Withdraw the table only after we mark the pmd entry invalid.
2289          * This's critical for some architectures (Power).
2290          */
2291         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2292         pmd_populate(mm, &_pmd, pgtable);
2293
2294         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2295                 pte_t entry, *pte;
2296                 /*
2297                  * Note that NUMA hinting access restrictions are not
2298                  * transferred to avoid any possibility of altering
2299                  * permissions across VMAs.
2300                  */
2301                 if (freeze || pmd_migration) {
2302                         swp_entry_t swp_entry;
2303                         swp_entry = make_migration_entry(page + i, write);
2304                         entry = swp_entry_to_pte(swp_entry);
2305                         if (soft_dirty)
2306                                 entry = pte_swp_mksoft_dirty(entry);
2307                 } else {
2308                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2309                         entry = maybe_mkwrite(entry, vma);
2310                         if (!write)
2311                                 entry = pte_wrprotect(entry);
2312                         if (!young)
2313                                 entry = pte_mkold(entry);
2314                         if (soft_dirty)
2315                                 entry = pte_mksoft_dirty(entry);
2316                 }
2317                 pte = pte_offset_map(&_pmd, addr);
2318                 BUG_ON(!pte_none(*pte));
2319                 set_pte_at(mm, addr, pte, entry);
2320                 atomic_inc(&page[i]._mapcount);
2321                 pte_unmap(pte);
2322         }
2323
2324         /*
2325          * Set PG_double_map before dropping compound_mapcount to avoid
2326          * false-negative page_mapped().
2327          */
2328         if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2329                 for (i = 0; i < HPAGE_PMD_NR; i++)
2330                         atomic_inc(&page[i]._mapcount);
2331         }
2332
2333         if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2334                 /* Last compound_mapcount is gone. */
2335                 __dec_node_page_state(page, NR_ANON_THPS);
2336                 if (TestClearPageDoubleMap(page)) {
2337                         /* No need in mapcount reference anymore */
2338                         for (i = 0; i < HPAGE_PMD_NR; i++)
2339                                 atomic_dec(&page[i]._mapcount);
2340                 }
2341         }
2342
2343         smp_wmb(); /* make pte visible before pmd */
2344         pmd_populate(mm, pmd, pgtable);
2345
2346         if (freeze) {
2347                 for (i = 0; i < HPAGE_PMD_NR; i++) {
2348                         page_remove_rmap(page + i, false);
2349                         put_page(page + i);
2350                 }
2351         }
2352 }
2353
2354 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2355                 unsigned long address, bool freeze, struct page *page)
2356 {
2357         spinlock_t *ptl;
2358         struct mmu_notifier_range range;
2359
2360         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2361                                 address & HPAGE_PMD_MASK,
2362                                 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2363         mmu_notifier_invalidate_range_start(&range);
2364         ptl = pmd_lock(vma->vm_mm, pmd);
2365
2366         /*
2367          * If caller asks to setup a migration entries, we need a page to check
2368          * pmd against. Otherwise we can end up replacing wrong page.
2369          */
2370         VM_BUG_ON(freeze && !page);
2371         if (page && page != pmd_page(*pmd))
2372                 goto out;
2373
2374         if (pmd_trans_huge(*pmd)) {
2375                 page = pmd_page(*pmd);
2376                 if (PageMlocked(page))
2377                         clear_page_mlock(page);
2378         } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2379                 goto out;
2380         __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2381 out:
2382         spin_unlock(ptl);
2383         /*
2384          * No need to double call mmu_notifier->invalidate_range() callback.
2385          * They are 3 cases to consider inside __split_huge_pmd_locked():
2386          *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2387          *  2) __split_huge_zero_page_pmd() read only zero page and any write
2388          *    fault will trigger a flush_notify before pointing to a new page
2389          *    (it is fine if the secondary mmu keeps pointing to the old zero
2390          *    page in the meantime)
2391          *  3) Split a huge pmd into pte pointing to the same page. No need
2392          *     to invalidate secondary tlb entry they are all still valid.
2393          *     any further changes to individual pte will notify. So no need
2394          *     to call mmu_notifier->invalidate_range()
2395          */
2396         mmu_notifier_invalidate_range_only_end(&range);
2397 }
2398
2399 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2400                 bool freeze, struct page *page)
2401 {
2402         pgd_t *pgd;
2403         p4d_t *p4d;
2404         pud_t *pud;
2405         pmd_t *pmd;
2406
2407         pgd = pgd_offset(vma->vm_mm, address);
2408         if (!pgd_present(*pgd))
2409                 return;
2410
2411         p4d = p4d_offset(pgd, address);
2412         if (!p4d_present(*p4d))
2413                 return;
2414
2415         pud = pud_offset(p4d, address);
2416         if (!pud_present(*pud))
2417                 return;
2418
2419         pmd = pmd_offset(pud, address);
2420
2421         __split_huge_pmd(vma, pmd, address, freeze, page);
2422 }
2423
2424 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2425                              unsigned long start,
2426                              unsigned long end,
2427                              long adjust_next)
2428 {
2429         /*
2430          * If the new start address isn't hpage aligned and it could
2431          * previously contain an hugepage: check if we need to split
2432          * an huge pmd.
2433          */
2434         if (start & ~HPAGE_PMD_MASK &&
2435             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2436             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2437                 split_huge_pmd_address(vma, start, false, NULL);
2438
2439         /*
2440          * If the new end address isn't hpage aligned and it could
2441          * previously contain an hugepage: check if we need to split
2442          * an huge pmd.
2443          */
2444         if (end & ~HPAGE_PMD_MASK &&
2445             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2446             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2447                 split_huge_pmd_address(vma, end, false, NULL);
2448
2449         /*
2450          * If we're also updating the vma->vm_next->vm_start, if the new
2451          * vm_next->vm_start isn't page aligned and it could previously
2452          * contain an hugepage: check if we need to split an huge pmd.
2453          */
2454         if (adjust_next > 0) {
2455                 struct vm_area_struct *next = vma->vm_next;
2456                 unsigned long nstart = next->vm_start;
2457                 nstart += adjust_next << PAGE_SHIFT;
2458                 if (nstart & ~HPAGE_PMD_MASK &&
2459                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2460                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2461                         split_huge_pmd_address(next, nstart, false, NULL);
2462         }
2463 }
2464
2465 static void unmap_page(struct page *page)
2466 {
2467         enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
2468                 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2469         bool unmap_success;
2470
2471         VM_BUG_ON_PAGE(!PageHead(page), page);
2472
2473         if (PageAnon(page))
2474                 ttu_flags |= TTU_SPLIT_FREEZE;
2475
2476         unmap_success = try_to_unmap(page, ttu_flags);
2477         VM_BUG_ON_PAGE(!unmap_success, page);
2478 }
2479
2480 static void remap_page(struct page *page)
2481 {
2482         int i;
2483         if (PageTransHuge(page)) {
2484                 remove_migration_ptes(page, page, true);
2485         } else {
2486                 for (i = 0; i < HPAGE_PMD_NR; i++)
2487                         remove_migration_ptes(page + i, page + i, true);
2488         }
2489 }
2490
2491 static void __split_huge_page_tail(struct page *head, int tail,
2492                 struct lruvec *lruvec, struct list_head *list)
2493 {
2494         struct page *page_tail = head + tail;
2495
2496         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2497
2498         /*
2499          * Clone page flags before unfreezing refcount.
2500          *
2501          * After successful get_page_unless_zero() might follow flags change,
2502          * for exmaple lock_page() which set PG_waiters.
2503          */
2504         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2505         page_tail->flags |= (head->flags &
2506                         ((1L << PG_referenced) |
2507                          (1L << PG_swapbacked) |
2508                          (1L << PG_swapcache) |
2509                          (1L << PG_mlocked) |
2510                          (1L << PG_uptodate) |
2511                          (1L << PG_active) |
2512                          (1L << PG_workingset) |
2513                          (1L << PG_locked) |
2514                          (1L << PG_unevictable) |
2515                          (1L << PG_dirty)));
2516
2517         /* ->mapping in first tail page is compound_mapcount */
2518         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2519                         page_tail);
2520         page_tail->mapping = head->mapping;
2521         page_tail->index = head->index + tail;
2522
2523         /* Page flags must be visible before we make the page non-compound. */
2524         smp_wmb();
2525
2526         /*
2527          * Clear PageTail before unfreezing page refcount.
2528          *
2529          * After successful get_page_unless_zero() might follow put_page()
2530          * which needs correct compound_head().
2531          */
2532         clear_compound_head(page_tail);
2533
2534         /* Finally unfreeze refcount. Additional reference from page cache. */
2535         page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2536                                           PageSwapCache(head)));
2537
2538         if (page_is_young(head))
2539                 set_page_young(page_tail);
2540         if (page_is_idle(head))
2541                 set_page_idle(page_tail);
2542
2543         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2544
2545         /*
2546          * always add to the tail because some iterators expect new
2547          * pages to show after the currently processed elements - e.g.
2548          * migrate_pages
2549          */
2550         lru_add_page_tail(head, page_tail, lruvec, list);
2551 }
2552
2553 static void __split_huge_page(struct page *page, struct list_head *list,
2554                 pgoff_t end, unsigned long flags)
2555 {
2556         struct page *head = compound_head(page);
2557         pg_data_t *pgdat = page_pgdat(head);
2558         struct lruvec *lruvec;
2559         struct address_space *swap_cache = NULL;
2560         unsigned long offset = 0;
2561         int i;
2562
2563         lruvec = mem_cgroup_page_lruvec(head, pgdat);
2564
2565         /* complete memcg works before add pages to LRU */
2566         mem_cgroup_split_huge_fixup(head);
2567
2568         if (PageAnon(head) && PageSwapCache(head)) {
2569                 swp_entry_t entry = { .val = page_private(head) };
2570
2571                 offset = swp_offset(entry);
2572                 swap_cache = swap_address_space(entry);
2573                 xa_lock(&swap_cache->i_pages);
2574         }
2575
2576         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
2577                 __split_huge_page_tail(head, i, lruvec, list);
2578                 /* Some pages can be beyond i_size: drop them from page cache */
2579                 if (head[i].index >= end) {
2580                         ClearPageDirty(head + i);
2581                         __delete_from_page_cache(head + i, NULL);
2582                         if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2583                                 shmem_uncharge(head->mapping->host, 1);
2584                         put_page(head + i);
2585                 } else if (!PageAnon(page)) {
2586                         __xa_store(&head->mapping->i_pages, head[i].index,
2587                                         head + i, 0);
2588                 } else if (swap_cache) {
2589                         __xa_store(&swap_cache->i_pages, offset + i,
2590                                         head + i, 0);
2591                 }
2592         }
2593
2594         ClearPageCompound(head);
2595
2596         split_page_owner(head, HPAGE_PMD_ORDER);
2597
2598         /* See comment in __split_huge_page_tail() */
2599         if (PageAnon(head)) {
2600                 /* Additional pin to swap cache */
2601                 if (PageSwapCache(head)) {
2602                         page_ref_add(head, 2);
2603                         xa_unlock(&swap_cache->i_pages);
2604                 } else {
2605                         page_ref_inc(head);
2606                 }
2607         } else {
2608                 /* Additional pin to page cache */
2609                 page_ref_add(head, 2);
2610                 xa_unlock(&head->mapping->i_pages);
2611         }
2612
2613         spin_unlock_irqrestore(&pgdat->lru_lock, flags);
2614
2615         remap_page(head);
2616
2617         for (i = 0; i < HPAGE_PMD_NR; i++) {
2618                 struct page *subpage = head + i;
2619                 if (subpage == page)
2620                         continue;
2621                 unlock_page(subpage);
2622
2623                 /*
2624                  * Subpages may be freed if there wasn't any mapping
2625                  * like if add_to_swap() is running on a lru page that
2626                  * had its mapping zapped. And freeing these pages
2627                  * requires taking the lru_lock so we do the put_page
2628                  * of the tail pages after the split is complete.
2629                  */
2630                 put_page(subpage);
2631         }
2632 }
2633
2634 int total_mapcount(struct page *page)
2635 {
2636         int i, compound, ret;
2637
2638         VM_BUG_ON_PAGE(PageTail(page), page);
2639
2640         if (likely(!PageCompound(page)))
2641                 return atomic_read(&page->_mapcount) + 1;
2642
2643         compound = compound_mapcount(page);
2644         if (PageHuge(page))
2645                 return compound;
2646         ret = compound;
2647         for (i = 0; i < HPAGE_PMD_NR; i++)
2648                 ret += atomic_read(&page[i]._mapcount) + 1;
2649         /* File pages has compound_mapcount included in _mapcount */
2650         if (!PageAnon(page))
2651                 return ret - compound * HPAGE_PMD_NR;
2652         if (PageDoubleMap(page))
2653                 ret -= HPAGE_PMD_NR;
2654         return ret;
2655 }
2656
2657 /*
2658  * This calculates accurately how many mappings a transparent hugepage
2659  * has (unlike page_mapcount() which isn't fully accurate). This full
2660  * accuracy is primarily needed to know if copy-on-write faults can
2661  * reuse the page and change the mapping to read-write instead of
2662  * copying them. At the same time this returns the total_mapcount too.
2663  *
2664  * The function returns the highest mapcount any one of the subpages
2665  * has. If the return value is one, even if different processes are
2666  * mapping different subpages of the transparent hugepage, they can
2667  * all reuse it, because each process is reusing a different subpage.
2668  *
2669  * The total_mapcount is instead counting all virtual mappings of the
2670  * subpages. If the total_mapcount is equal to "one", it tells the
2671  * caller all mappings belong to the same "mm" and in turn the
2672  * anon_vma of the transparent hugepage can become the vma->anon_vma
2673  * local one as no other process may be mapping any of the subpages.
2674  *
2675  * It would be more accurate to replace page_mapcount() with
2676  * page_trans_huge_mapcount(), however we only use
2677  * page_trans_huge_mapcount() in the copy-on-write faults where we
2678  * need full accuracy to avoid breaking page pinning, because
2679  * page_trans_huge_mapcount() is slower than page_mapcount().
2680  */
2681 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2682 {
2683         int i, ret, _total_mapcount, mapcount;
2684
2685         /* hugetlbfs shouldn't call it */
2686         VM_BUG_ON_PAGE(PageHuge(page), page);
2687
2688         if (likely(!PageTransCompound(page))) {
2689                 mapcount = atomic_read(&page->_mapcount) + 1;
2690                 if (total_mapcount)
2691                         *total_mapcount = mapcount;
2692                 return mapcount;
2693         }
2694
2695         page = compound_head(page);
2696
2697         _total_mapcount = ret = 0;
2698         for (i = 0; i < HPAGE_PMD_NR; i++) {
2699                 mapcount = atomic_read(&page[i]._mapcount) + 1;
2700                 ret = max(ret, mapcount);
2701                 _total_mapcount += mapcount;
2702         }
2703         if (PageDoubleMap(page)) {
2704                 ret -= 1;
2705                 _total_mapcount -= HPAGE_PMD_NR;
2706         }
2707         mapcount = compound_mapcount(page);
2708         ret += mapcount;
2709         _total_mapcount += mapcount;
2710         if (total_mapcount)
2711                 *total_mapcount = _total_mapcount;
2712         return ret;
2713 }
2714
2715 /* Racy check whether the huge page can be split */
2716 bool can_split_huge_page(struct page *page, int *pextra_pins)
2717 {
2718         int extra_pins;
2719
2720         /* Additional pins from page cache */
2721         if (PageAnon(page))
2722                 extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0;
2723         else
2724                 extra_pins = HPAGE_PMD_NR;
2725         if (pextra_pins)
2726                 *pextra_pins = extra_pins;
2727         return total_mapcount(page) == page_count(page) - extra_pins - 1;
2728 }
2729
2730 /*
2731  * This function splits huge page into normal pages. @page can point to any
2732  * subpage of huge page to split. Split doesn't change the position of @page.
2733  *
2734  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2735  * The huge page must be locked.
2736  *
2737  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2738  *
2739  * Both head page and tail pages will inherit mapping, flags, and so on from
2740  * the hugepage.
2741  *
2742  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2743  * they are not mapped.
2744  *
2745  * Returns 0 if the hugepage is split successfully.
2746  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2747  * us.
2748  */
2749 int split_huge_page_to_list(struct page *page, struct list_head *list)
2750 {
2751         struct page *head = compound_head(page);
2752         struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2753         struct deferred_split *ds_queue = get_deferred_split_queue(head);
2754         struct anon_vma *anon_vma = NULL;
2755         struct address_space *mapping = NULL;
2756         int count, mapcount, extra_pins, ret;
2757         bool mlocked;
2758         unsigned long flags;
2759         pgoff_t end;
2760
2761         VM_BUG_ON_PAGE(is_huge_zero_page(head), head);
2762         VM_BUG_ON_PAGE(!PageLocked(head), head);
2763         VM_BUG_ON_PAGE(!PageCompound(head), head);
2764
2765         if (PageWriteback(head))
2766                 return -EBUSY;
2767
2768         if (PageAnon(head)) {
2769                 /*
2770                  * The caller does not necessarily hold an mmap_sem that would
2771                  * prevent the anon_vma disappearing so we first we take a
2772                  * reference to it and then lock the anon_vma for write. This
2773                  * is similar to page_lock_anon_vma_read except the write lock
2774                  * is taken to serialise against parallel split or collapse
2775                  * operations.
2776                  */
2777                 anon_vma = page_get_anon_vma(head);
2778                 if (!anon_vma) {
2779                         ret = -EBUSY;
2780                         goto out;
2781                 }
2782                 end = -1;
2783                 mapping = NULL;
2784                 anon_vma_lock_write(anon_vma);
2785         } else {
2786                 mapping = head->mapping;
2787
2788                 /* Truncated ? */
2789                 if (!mapping) {
2790                         ret = -EBUSY;
2791                         goto out;
2792                 }
2793
2794                 anon_vma = NULL;
2795                 i_mmap_lock_read(mapping);
2796
2797                 /*
2798                  *__split_huge_page() may need to trim off pages beyond EOF:
2799                  * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2800                  * which cannot be nested inside the page tree lock. So note
2801                  * end now: i_size itself may be changed at any moment, but
2802                  * head page lock is good enough to serialize the trimming.
2803                  */
2804                 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2805         }
2806
2807         /*
2808          * Racy check if we can split the page, before unmap_page() will
2809          * split PMDs
2810          */
2811         if (!can_split_huge_page(head, &extra_pins)) {
2812                 ret = -EBUSY;
2813                 goto out_unlock;
2814         }
2815
2816         mlocked = PageMlocked(head);
2817         unmap_page(head);
2818         VM_BUG_ON_PAGE(compound_mapcount(head), head);
2819
2820         /* Make sure the page is not on per-CPU pagevec as it takes pin */
2821         if (mlocked)
2822                 lru_add_drain();
2823
2824         /* prevent PageLRU to go away from under us, and freeze lru stats */
2825         spin_lock_irqsave(&pgdata->lru_lock, flags);
2826
2827         if (mapping) {
2828                 XA_STATE(xas, &mapping->i_pages, page_index(head));
2829
2830                 /*
2831                  * Check if the head page is present in page cache.
2832                  * We assume all tail are present too, if head is there.
2833                  */
2834                 xa_lock(&mapping->i_pages);
2835                 if (xas_load(&xas) != head)
2836                         goto fail;
2837         }
2838
2839         /* Prevent deferred_split_scan() touching ->_refcount */
2840         spin_lock(&ds_queue->split_queue_lock);
2841         count = page_count(head);
2842         mapcount = total_mapcount(head);
2843         if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2844                 if (!list_empty(page_deferred_list(head))) {
2845                         ds_queue->split_queue_len--;
2846                         list_del(page_deferred_list(head));
2847                 }
2848                 spin_unlock(&ds_queue->split_queue_lock);
2849                 if (mapping) {
2850                         if (PageSwapBacked(head))
2851                                 __dec_node_page_state(head, NR_SHMEM_THPS);
2852                         else
2853                                 __dec_node_page_state(head, NR_FILE_THPS);
2854                 }
2855
2856                 __split_huge_page(page, list, end, flags);
2857                 if (PageSwapCache(head)) {
2858                         swp_entry_t entry = { .val = page_private(head) };
2859
2860                         ret = split_swap_cluster(entry);
2861                 } else
2862                         ret = 0;
2863         } else {
2864                 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2865                         pr_alert("total_mapcount: %u, page_count(): %u\n",
2866                                         mapcount, count);
2867                         if (PageTail(page))
2868                                 dump_page(head, NULL);
2869                         dump_page(page, "total_mapcount(head) > 0");
2870                         BUG();
2871                 }
2872                 spin_unlock(&ds_queue->split_queue_lock);
2873 fail:           if (mapping)
2874                         xa_unlock(&mapping->i_pages);
2875                 spin_unlock_irqrestore(&pgdata->lru_lock, flags);
2876                 remap_page(head);
2877                 ret = -EBUSY;
2878         }
2879
2880 out_unlock:
2881         if (anon_vma) {
2882                 anon_vma_unlock_write(anon_vma);
2883                 put_anon_vma(anon_vma);
2884         }
2885         if (mapping)
2886                 i_mmap_unlock_read(mapping);
2887 out:
2888         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2889         return ret;
2890 }
2891
2892 void free_transhuge_page(struct page *page)
2893 {
2894         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2895         unsigned long flags;
2896
2897         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2898         if (!list_empty(page_deferred_list(page))) {
2899                 ds_queue->split_queue_len--;
2900                 list_del(page_deferred_list(page));
2901         }
2902         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2903         free_compound_page(page);
2904 }
2905
2906 void deferred_split_huge_page(struct page *page)
2907 {
2908         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2909 #ifdef CONFIG_MEMCG
2910         struct mem_cgroup *memcg = compound_head(page)->mem_cgroup;
2911 #endif
2912         unsigned long flags;
2913
2914         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2915
2916         /*
2917          * The try_to_unmap() in page reclaim path might reach here too,
2918          * this may cause a race condition to corrupt deferred split queue.
2919          * And, if page reclaim is already handling the same page, it is
2920          * unnecessary to handle it again in shrinker.
2921          *
2922          * Check PageSwapCache to determine if the page is being
2923          * handled by page reclaim since THP swap would add the page into
2924          * swap cache before calling try_to_unmap().
2925          */
2926         if (PageSwapCache(page))
2927                 return;
2928
2929         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2930         if (list_empty(page_deferred_list(page))) {
2931                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2932                 list_add_tail(page_deferred_list(page), &ds_queue->split_queue);
2933                 ds_queue->split_queue_len++;
2934 #ifdef CONFIG_MEMCG
2935                 if (memcg)
2936                         memcg_set_shrinker_bit(memcg, page_to_nid(page),
2937                                                deferred_split_shrinker.id);
2938 #endif
2939         }
2940         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2941 }
2942
2943 static unsigned long deferred_split_count(struct shrinker *shrink,
2944                 struct shrink_control *sc)
2945 {
2946         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2947         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2948
2949 #ifdef CONFIG_MEMCG
2950         if (sc->memcg)
2951                 ds_queue = &sc->memcg->deferred_split_queue;
2952 #endif
2953         return READ_ONCE(ds_queue->split_queue_len);
2954 }
2955
2956 static unsigned long deferred_split_scan(struct shrinker *shrink,
2957                 struct shrink_control *sc)
2958 {
2959         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2960         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2961         unsigned long flags;
2962         LIST_HEAD(list), *pos, *next;
2963         struct page *page;
2964         int split = 0;
2965
2966 #ifdef CONFIG_MEMCG
2967         if (sc->memcg)
2968                 ds_queue = &sc->memcg->deferred_split_queue;
2969 #endif
2970
2971         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2972         /* Take pin on all head pages to avoid freeing them under us */
2973         list_for_each_safe(pos, next, &ds_queue->split_queue) {
2974                 page = list_entry((void *)pos, struct page, mapping);
2975                 page = compound_head(page);
2976                 if (get_page_unless_zero(page)) {
2977                         list_move(page_deferred_list(page), &list);
2978                 } else {
2979                         /* We lost race with put_compound_page() */
2980                         list_del_init(page_deferred_list(page));
2981                         ds_queue->split_queue_len--;
2982                 }
2983                 if (!--sc->nr_to_scan)
2984                         break;
2985         }
2986         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2987
2988         list_for_each_safe(pos, next, &list) {
2989                 page = list_entry((void *)pos, struct page, mapping);
2990                 if (!trylock_page(page))
2991                         goto next;
2992                 /* split_huge_page() removes page from list on success */
2993                 if (!split_huge_page(page))
2994                         split++;
2995                 unlock_page(page);
2996 next:
2997                 put_page(page);
2998         }
2999
3000         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
3001         list_splice_tail(&list, &ds_queue->split_queue);
3002         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
3003
3004         /*
3005          * Stop shrinker if we didn't split any page, but the queue is empty.
3006          * This can happen if pages were freed under us.
3007          */
3008         if (!split && list_empty(&ds_queue->split_queue))
3009                 return SHRINK_STOP;
3010         return split;
3011 }
3012
3013 static struct shrinker deferred_split_shrinker = {
3014         .count_objects = deferred_split_count,
3015         .scan_objects = deferred_split_scan,
3016         .seeks = DEFAULT_SEEKS,
3017         .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
3018                  SHRINKER_NONSLAB,
3019 };
3020
3021 #ifdef CONFIG_DEBUG_FS
3022 static int split_huge_pages_set(void *data, u64 val)
3023 {
3024         struct zone *zone;
3025         struct page *page;
3026         unsigned long pfn, max_zone_pfn;
3027         unsigned long total = 0, split = 0;
3028
3029         if (val != 1)
3030                 return -EINVAL;
3031
3032         for_each_populated_zone(zone) {
3033                 max_zone_pfn = zone_end_pfn(zone);
3034                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
3035                         if (!pfn_valid(pfn))
3036                                 continue;
3037
3038                         page = pfn_to_page(pfn);
3039                         if (!get_page_unless_zero(page))
3040                                 continue;
3041
3042                         if (zone != page_zone(page))
3043                                 goto next;
3044
3045                         if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
3046                                 goto next;
3047
3048                         total++;
3049                         lock_page(page);
3050                         if (!split_huge_page(page))
3051                                 split++;
3052                         unlock_page(page);
3053 next:
3054                         put_page(page);
3055                 }
3056         }
3057
3058         pr_info("%lu of %lu THP split\n", split, total);
3059
3060         return 0;
3061 }
3062 DEFINE_DEBUGFS_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
3063                 "%llu\n");
3064
3065 static int __init split_huge_pages_debugfs(void)
3066 {
3067         debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3068                             &split_huge_pages_fops);
3069         return 0;
3070 }
3071 late_initcall(split_huge_pages_debugfs);
3072 #endif
3073
3074 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
3075 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
3076                 struct page *page)
3077 {
3078         struct vm_area_struct *vma = pvmw->vma;
3079         struct mm_struct *mm = vma->vm_mm;
3080         unsigned long address = pvmw->address;
3081         pmd_t pmdval;
3082         swp_entry_t entry;
3083         pmd_t pmdswp;
3084
3085         if (!(pvmw->pmd && !pvmw->pte))
3086                 return;
3087
3088         flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
3089         pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
3090         if (pmd_dirty(pmdval))
3091                 set_page_dirty(page);
3092         entry = make_migration_entry(page, pmd_write(pmdval));
3093         pmdswp = swp_entry_to_pmd(entry);
3094         if (pmd_soft_dirty(pmdval))
3095                 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
3096         set_pmd_at(mm, address, pvmw->pmd, pmdswp);
3097         page_remove_rmap(page, true);
3098         put_page(page);
3099 }
3100
3101 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
3102 {
3103         struct vm_area_struct *vma = pvmw->vma;
3104         struct mm_struct *mm = vma->vm_mm;
3105         unsigned long address = pvmw->address;
3106         unsigned long mmun_start = address & HPAGE_PMD_MASK;
3107         pmd_t pmde;
3108         swp_entry_t entry;
3109
3110         if (!(pvmw->pmd && !pvmw->pte))
3111                 return;
3112
3113         entry = pmd_to_swp_entry(*pvmw->pmd);
3114         get_page(new);
3115         pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
3116         if (pmd_swp_soft_dirty(*pvmw->pmd))
3117                 pmde = pmd_mksoft_dirty(pmde);
3118         if (is_write_migration_entry(entry))
3119                 pmde = maybe_pmd_mkwrite(pmde, vma);
3120
3121         flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
3122         if (PageAnon(new))
3123                 page_add_anon_rmap(new, vma, mmun_start, true);
3124         else
3125                 page_add_file_rmap(new, true);
3126         set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
3127         if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new))
3128                 mlock_vma_page(new);
3129         update_mmu_cache_pmd(vma, address, pvmw->pmd);
3130 }
3131 #endif