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