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                 count_vm_event(THP_FAULT_FALLBACK_CHARGE);
 601                 return VM_FAULT_FALLBACK;
 602         }
 603
 604         pgtable = pte_alloc_one(vma->vm_mm);
 605         if (unlikely(!pgtable)) {
 606                 ret = VM_FAULT_OOM;
 607                 goto release;
 608         }
 609
 610         clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
 611         /*
 612          * The memory barrier inside __SetPageUptodate makes sure that
 613          * clear_huge_page writes become visible before the set_pmd_at()
 614          * write.
 615          */
 616         __SetPageUptodate(page);
 617
 618         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 619         if (unlikely(!pmd_none(*vmf->pmd))) {
 620                 goto unlock_release;
 621         } else {
 622                 pmd_t entry;
 623
 624                 ret = check_stable_address_space(vma->vm_mm);
 625                 if (ret)
 626                         goto unlock_release;
 627
 628                 /* Deliver the page fault to userland */
 629                 if (userfaultfd_missing(vma)) {
 630                         vm_fault_t ret2;
 631
 632                         spin_unlock(vmf->ptl);
 633                         mem_cgroup_cancel_charge(page, memcg, true);
 634                         put_page(page);
 635                         pte_free(vma->vm_mm, pgtable);
 636                         ret2 = handle_userfault(vmf, VM_UFFD_MISSING);
 637                         VM_BUG_ON(ret2 & VM_FAULT_FALLBACK);
 638                         return ret2;
 639                 }
 640
 641                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 642                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 643                 page_add_new_anon_rmap(page, vma, haddr, true);
 644                 mem_cgroup_commit_charge(page, memcg, false, true);
 645                 lru_cache_add_active_or_unevictable(page, vma);
 646                 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
 647                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
 648                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 649                 mm_inc_nr_ptes(vma->vm_mm);
 650                 spin_unlock(vmf->ptl);
 651                 count_vm_event(THP_FAULT_ALLOC);
 652                 count_memcg_events(memcg, THP_FAULT_ALLOC, 1);
 653         }
 654
 655         return 0;
 656 unlock_release:
 657         spin_unlock(vmf->ptl);
 658 release:
 659         if (pgtable)
 660                 pte_free(vma->vm_mm, pgtable);
 661         mem_cgroup_cancel_charge(page, memcg, true);
 662         put_page(page);
 663         return ret;
 664
 665 }
 666
 667 /*
 668  * always: directly stall for all thp allocations
 669  * defer: wake kswapd and fail if not immediately available
 670  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
 671  *                fail if not immediately available
 672  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
 673  *          available
 674  * never: never stall for any thp allocation
 675  */
 676 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
 677 {
 678         const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
 679
 680         /* Always do synchronous compaction */
 681         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 682                 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
 683
 684         /* Kick kcompactd and fail quickly */
 685         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 686                 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
 687
 688         /* Synchronous compaction if madvised, otherwise kick kcompactd */
 689         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 690                 return GFP_TRANSHUGE_LIGHT |
 691                         (vma_madvised ? __GFP_DIRECT_RECLAIM :
 692                                         __GFP_KSWAPD_RECLAIM);
 693
 694         /* Only do synchronous compaction if madvised */
 695         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 696                 return GFP_TRANSHUGE_LIGHT |
 697                        (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
 698
 699         return GFP_TRANSHUGE_LIGHT;
 700 }
 701
 702 /* Caller must hold page table lock. */
 703 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 704                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 705                 struct page *zero_page)
 706 {
 707         pmd_t entry;
 708         if (!pmd_none(*pmd))
 709                 return false;
 710         entry = mk_pmd(zero_page, vma->vm_page_prot);
 711         entry = pmd_mkhuge(entry);
 712         if (pgtable)
 713                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 714         set_pmd_at(mm, haddr, pmd, entry);
 715         mm_inc_nr_ptes(mm);
 716         return true;
 717 }
 718
 719 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
 720 {
 721         struct vm_area_struct *vma = vmf->vma;
 722         gfp_t gfp;
 723         struct page *page;
 724         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 725
 726         if (!transhuge_vma_suitable(vma, haddr))
 727                 return VM_FAULT_FALLBACK;
 728         if (unlikely(anon_vma_prepare(vma)))
 729                 return VM_FAULT_OOM;
 730         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 731                 return VM_FAULT_OOM;
 732         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
 733                         !mm_forbids_zeropage(vma->vm_mm) &&
 734                         transparent_hugepage_use_zero_page()) {
 735                 pgtable_t pgtable;
 736                 struct page *zero_page;
 737                 bool set;
 738                 vm_fault_t ret;
 739                 pgtable = pte_alloc_one(vma->vm_mm);
 740                 if (unlikely(!pgtable))
 741                         return VM_FAULT_OOM;
 742                 zero_page = mm_get_huge_zero_page(vma->vm_mm);
 743                 if (unlikely(!zero_page)) {
 744                         pte_free(vma->vm_mm, pgtable);
 745                         count_vm_event(THP_FAULT_FALLBACK);
 746                         return VM_FAULT_FALLBACK;
 747                 }
 748                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 749                 ret = 0;
 750                 set = false;
 751                 if (pmd_none(*vmf->pmd)) {
 752                         ret = check_stable_address_space(vma->vm_mm);
 753                         if (ret) {
 754                                 spin_unlock(vmf->ptl);
 755                         } else if (userfaultfd_missing(vma)) {
 756                                 spin_unlock(vmf->ptl);
 757                                 ret = handle_userfault(vmf, VM_UFFD_MISSING);
 758                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 759                         } else {
 760                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
 761                                                    haddr, vmf->pmd, zero_page);
 762                                 spin_unlock(vmf->ptl);
 763                                 set = true;
 764                         }
 765                 } else
 766                         spin_unlock(vmf->ptl);
 767                 if (!set)
 768                         pte_free(vma->vm_mm, pgtable);
 769                 return ret;
 770         }
 771         gfp = alloc_hugepage_direct_gfpmask(vma);
 772         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 773         if (unlikely(!page)) {
 774                 count_vm_event(THP_FAULT_FALLBACK);
 775                 return VM_FAULT_FALLBACK;
 776         }
 777         prep_transhuge_page(page);
 778         return __do_huge_pmd_anonymous_page(vmf, page, gfp);
 779 }
 780
 781 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 782                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
 783                 pgtable_t pgtable)
 784 {
 785         struct mm_struct *mm = vma->vm_mm;
 786         pmd_t entry;
 787         spinlock_t *ptl;
 788
 789         ptl = pmd_lock(mm, pmd);
 790         if (!pmd_none(*pmd)) {
 791                 if (write) {
 792                         if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
 793                                 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
 794                                 goto out_unlock;
 795                         }
 796                         entry = pmd_mkyoung(*pmd);
 797                         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 798                         if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
 799                                 update_mmu_cache_pmd(vma, addr, pmd);
 800                 }
 801
 802                 goto out_unlock;
 803         }
 804
 805         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 806         if (pfn_t_devmap(pfn))
 807                 entry = pmd_mkdevmap(entry);
 808         if (write) {
 809                 entry = pmd_mkyoung(pmd_mkdirty(entry));
 810                 entry = maybe_pmd_mkwrite(entry, vma);
 811         }
 812
 813         if (pgtable) {
 814                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 815                 mm_inc_nr_ptes(mm);
 816                 pgtable = NULL;
 817         }
 818
 819         set_pmd_at(mm, addr, pmd, entry);
 820         update_mmu_cache_pmd(vma, addr, pmd);
 821
 822 out_unlock:
 823         spin_unlock(ptl);
 824         if (pgtable)
 825                 pte_free(mm, pgtable);
 826 }
 827
 828 /**
 829  * vmf_insert_pfn_pmd_prot - insert a pmd size pfn
 830  * @vmf: Structure describing the fault
 831  * @pfn: pfn to insert
 832  * @pgprot: page protection to use
 833  * @write: whether it's a write fault
 834  *
 835  * Insert a pmd size pfn. See vmf_insert_pfn() for additional info and
 836  * also consult the vmf_insert_mixed_prot() documentation when
 837  * @pgprot != @vmf->vma->vm_page_prot.
 838  *
 839  * Return: vm_fault_t value.
 840  */
 841 vm_fault_t vmf_insert_pfn_pmd_prot(struct vm_fault *vmf, pfn_t pfn,
 842                                    pgprot_t pgprot, bool write)
 843 {
 844         unsigned long addr = vmf->address & PMD_MASK;
 845         struct vm_area_struct *vma = vmf->vma;
 846         pgtable_t pgtable = NULL;
 847
 848         /*
 849          * If we had pmd_special, we could avoid all these restrictions,
 850          * but we need to be consistent with PTEs and architectures that
 851          * can't support a 'special' bit.
 852          */
 853         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 854                         !pfn_t_devmap(pfn));
 855         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 856                                                 (VM_PFNMAP|VM_MIXEDMAP));
 857         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 858
 859         if (addr < vma->vm_start || addr >= vma->vm_end)
 860                 return VM_FAULT_SIGBUS;
 861
 862         if (arch_needs_pgtable_deposit()) {
 863                 pgtable = pte_alloc_one(vma->vm_mm);
 864                 if (!pgtable)
 865                         return VM_FAULT_OOM;
 866         }
 867
 868         track_pfn_insert(vma, &pgprot, pfn);
 869
 870         insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
 871         return VM_FAULT_NOPAGE;
 872 }
 873 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd_prot);
 874
 875 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 876 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
 877 {
 878         if (likely(vma->vm_flags & VM_WRITE))
 879                 pud = pud_mkwrite(pud);
 880         return pud;
 881 }
 882
 883 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 884                 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
 885 {
 886         struct mm_struct *mm = vma->vm_mm;
 887         pud_t entry;
 888         spinlock_t *ptl;
 889
 890         ptl = pud_lock(mm, pud);
 891         if (!pud_none(*pud)) {
 892                 if (write) {
 893                         if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
 894                                 WARN_ON_ONCE(!is_huge_zero_pud(*pud));
 895                                 goto out_unlock;
 896                         }
 897                         entry = pud_mkyoung(*pud);
 898                         entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
 899                         if (pudp_set_access_flags(vma, addr, pud, entry, 1))
 900                                 update_mmu_cache_pud(vma, addr, pud);
 901                 }
 902                 goto out_unlock;
 903         }
 904
 905         entry = pud_mkhuge(pfn_t_pud(pfn, prot));
 906         if (pfn_t_devmap(pfn))
 907                 entry = pud_mkdevmap(entry);
 908         if (write) {
 909                 entry = pud_mkyoung(pud_mkdirty(entry));
 910                 entry = maybe_pud_mkwrite(entry, vma);
 911         }
 912         set_pud_at(mm, addr, pud, entry);
 913         update_mmu_cache_pud(vma, addr, pud);
 914
 915 out_unlock:
 916         spin_unlock(ptl);
 917 }
 918
 919 /**
 920  * vmf_insert_pfn_pud_prot - insert a pud size pfn
 921  * @vmf: Structure describing the fault
 922  * @pfn: pfn to insert
 923  * @pgprot: page protection to use
 924  * @write: whether it's a write fault
 925  *
 926  * Insert a pud size pfn. See vmf_insert_pfn() for additional info and
 927  * also consult the vmf_insert_mixed_prot() documentation when
 928  * @pgprot != @vmf->vma->vm_page_prot.
 929  *
 930  * Return: vm_fault_t value.
 931  */
 932 vm_fault_t vmf_insert_pfn_pud_prot(struct vm_fault *vmf, pfn_t pfn,
 933                                    pgprot_t pgprot, bool write)
 934 {
 935         unsigned long addr = vmf->address & PUD_MASK;
 936         struct vm_area_struct *vma = vmf->vma;
 937
 938         /*
 939          * If we had pud_special, we could avoid all these restrictions,
 940          * but we need to be consistent with PTEs and architectures that
 941          * can't support a 'special' bit.
 942          */
 943         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 944                         !pfn_t_devmap(pfn));
 945         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 946                                                 (VM_PFNMAP|VM_MIXEDMAP));
 947         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 948
 949         if (addr < vma->vm_start || addr >= vma->vm_end)
 950                 return VM_FAULT_SIGBUS;
 951
 952         track_pfn_insert(vma, &pgprot, pfn);
 953
 954         insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
 955         return VM_FAULT_NOPAGE;
 956 }
 957 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud_prot);
 958 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
 959
 960 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
 961                 pmd_t *pmd, int flags)
 962 {
 963         pmd_t _pmd;
 964
 965         _pmd = pmd_mkyoung(*pmd);
 966         if (flags & FOLL_WRITE)
 967                 _pmd = pmd_mkdirty(_pmd);
 968         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
 969                                 pmd, _pmd, flags & FOLL_WRITE))
 970                 update_mmu_cache_pmd(vma, addr, pmd);
 971 }
 972
 973 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 974                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
 975 {
 976         unsigned long pfn = pmd_pfn(*pmd);
 977         struct mm_struct *mm = vma->vm_mm;
 978         struct page *page;
 979
 980         assert_spin_locked(pmd_lockptr(mm, pmd));
 981
 982         /*
 983          * When we COW a devmap PMD entry, we split it into PTEs, so we should
 984          * not be in this function with `flags & FOLL_COW` set.
 985          */
 986         WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
 987
 988         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
 989         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
 990                          (FOLL_PIN | FOLL_GET)))
 991                 return NULL;
 992
 993         if (flags & FOLL_WRITE && !pmd_write(*pmd))
 994                 return NULL;
 995
 996         if (pmd_present(*pmd) && pmd_devmap(*pmd))
 997                 /* pass */;
 998         else
 999                 return NULL;
1000
1001         if (flags & FOLL_TOUCH)
1002                 touch_pmd(vma, addr, pmd, flags);
1003
1004         /*
1005          * device mapped pages can only be returned if the
1006          * caller will manage the page reference count.
1007          */
1008         if (!(flags & (FOLL_GET | FOLL_PIN)))
1009                 return ERR_PTR(-EEXIST);
1010
1011         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1012         *pgmap = get_dev_pagemap(pfn, *pgmap);
1013         if (!*pgmap)
1014                 return ERR_PTR(-EFAULT);
1015         page = pfn_to_page(pfn);
1016         if (!try_grab_page(page, flags))
1017                 page = ERR_PTR(-ENOMEM);
1018
1019         return page;
1020 }
1021
1022 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1023                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1024                   struct vm_area_struct *vma)
1025 {
1026         spinlock_t *dst_ptl, *src_ptl;
1027         struct page *src_page;
1028         pmd_t pmd;
1029         pgtable_t pgtable = NULL;
1030         int ret = -ENOMEM;
1031
1032         /* Skip if can be re-fill on fault */
1033         if (!vma_is_anonymous(vma))
1034                 return 0;
1035
1036         pgtable = pte_alloc_one(dst_mm);
1037         if (unlikely(!pgtable))
1038                 goto out;
1039
1040         dst_ptl = pmd_lock(dst_mm, dst_pmd);
1041         src_ptl = pmd_lockptr(src_mm, src_pmd);
1042         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1043
1044         ret = -EAGAIN;
1045         pmd = *src_pmd;
1046
1047 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1048         if (unlikely(is_swap_pmd(pmd))) {
1049                 swp_entry_t entry = pmd_to_swp_entry(pmd);
1050
1051                 VM_BUG_ON(!is_pmd_migration_entry(pmd));
1052                 if (is_write_migration_entry(entry)) {
1053                         make_migration_entry_read(&entry);
1054                         pmd = swp_entry_to_pmd(entry);
1055                         if (pmd_swp_soft_dirty(*src_pmd))
1056                                 pmd = pmd_swp_mksoft_dirty(pmd);
1057                         set_pmd_at(src_mm, addr, src_pmd, pmd);
1058                 }
1059                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1060                 mm_inc_nr_ptes(dst_mm);
1061                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1062                 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1063                 ret = 0;
1064                 goto out_unlock;
1065         }
1066 #endif
1067
1068         if (unlikely(!pmd_trans_huge(pmd))) {
1069                 pte_free(dst_mm, pgtable);
1070                 goto out_unlock;
1071         }
1072         /*
1073          * When page table lock is held, the huge zero pmd should not be
1074          * under splitting since we don't split the page itself, only pmd to
1075          * a page table.
1076          */
1077         if (is_huge_zero_pmd(pmd)) {
1078                 struct page *zero_page;
1079                 /*
1080                  * get_huge_zero_page() will never allocate a new page here,
1081                  * since we already have a zero page to copy. It just takes a
1082                  * reference.
1083                  */
1084                 zero_page = mm_get_huge_zero_page(dst_mm);
1085                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1086                                 zero_page);
1087                 ret = 0;
1088                 goto out_unlock;
1089         }
1090
1091         src_page = pmd_page(pmd);
1092         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1093         get_page(src_page);
1094         page_dup_rmap(src_page, true);
1095         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1096         mm_inc_nr_ptes(dst_mm);
1097         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1098
1099         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1100         pmd = pmd_mkold(pmd_wrprotect(pmd));
1101         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1102
1103         ret = 0;
1104 out_unlock:
1105         spin_unlock(src_ptl);
1106         spin_unlock(dst_ptl);
1107 out:
1108         return ret;
1109 }
1110
1111 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1112 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1113                 pud_t *pud, int flags)
1114 {
1115         pud_t _pud;
1116
1117         _pud = pud_mkyoung(*pud);
1118         if (flags & FOLL_WRITE)
1119                 _pud = pud_mkdirty(_pud);
1120         if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1121                                 pud, _pud, flags & FOLL_WRITE))
1122                 update_mmu_cache_pud(vma, addr, pud);
1123 }
1124
1125 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1126                 pud_t *pud, int flags, struct dev_pagemap **pgmap)
1127 {
1128         unsigned long pfn = pud_pfn(*pud);
1129         struct mm_struct *mm = vma->vm_mm;
1130         struct page *page;
1131
1132         assert_spin_locked(pud_lockptr(mm, pud));
1133
1134         if (flags & FOLL_WRITE && !pud_write(*pud))
1135                 return NULL;
1136
1137         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
1138         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
1139                          (FOLL_PIN | FOLL_GET)))
1140                 return NULL;
1141
1142         if (pud_present(*pud) && pud_devmap(*pud))
1143                 /* pass */;
1144         else
1145                 return NULL;
1146
1147         if (flags & FOLL_TOUCH)
1148                 touch_pud(vma, addr, pud, flags);
1149
1150         /*
1151          * device mapped pages can only be returned if the
1152          * caller will manage the page reference count.
1153          *
1154          * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here:
1155          */
1156         if (!(flags & (FOLL_GET | FOLL_PIN)))
1157                 return ERR_PTR(-EEXIST);
1158
1159         pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1160         *pgmap = get_dev_pagemap(pfn, *pgmap);
1161         if (!*pgmap)
1162                 return ERR_PTR(-EFAULT);
1163         page = pfn_to_page(pfn);
1164         if (!try_grab_page(page, flags))
1165                 page = ERR_PTR(-ENOMEM);
1166
1167         return page;
1168 }
1169
1170 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1171                   pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1172                   struct vm_area_struct *vma)
1173 {
1174         spinlock_t *dst_ptl, *src_ptl;
1175         pud_t pud;
1176         int ret;
1177
1178         dst_ptl = pud_lock(dst_mm, dst_pud);
1179         src_ptl = pud_lockptr(src_mm, src_pud);
1180         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1181
1182         ret = -EAGAIN;
1183         pud = *src_pud;
1184         if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1185                 goto out_unlock;
1186
1187         /*
1188          * When page table lock is held, the huge zero pud should not be
1189          * under splitting since we don't split the page itself, only pud to
1190          * a page table.
1191          */
1192         if (is_huge_zero_pud(pud)) {
1193                 /* No huge zero pud yet */
1194         }
1195
1196         pudp_set_wrprotect(src_mm, addr, src_pud);
1197         pud = pud_mkold(pud_wrprotect(pud));
1198         set_pud_at(dst_mm, addr, dst_pud, pud);
1199
1200         ret = 0;
1201 out_unlock:
1202         spin_unlock(src_ptl);
1203         spin_unlock(dst_ptl);
1204         return ret;
1205 }
1206
1207 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1208 {
1209         pud_t entry;
1210         unsigned long haddr;
1211         bool write = vmf->flags & FAULT_FLAG_WRITE;
1212
1213         vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1214         if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1215                 goto unlock;
1216
1217         entry = pud_mkyoung(orig_pud);
1218         if (write)
1219                 entry = pud_mkdirty(entry);
1220         haddr = vmf->address & HPAGE_PUD_MASK;
1221         if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1222                 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1223
1224 unlock:
1225         spin_unlock(vmf->ptl);
1226 }
1227 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1228
1229 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1230 {
1231         pmd_t entry;
1232         unsigned long haddr;
1233         bool write = vmf->flags & FAULT_FLAG_WRITE;
1234
1235         vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1236         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1237                 goto unlock;
1238
1239         entry = pmd_mkyoung(orig_pmd);
1240         if (write)
1241                 entry = pmd_mkdirty(entry);
1242         haddr = vmf->address & HPAGE_PMD_MASK;
1243         if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1244                 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1245
1246 unlock:
1247         spin_unlock(vmf->ptl);
1248 }
1249
1250 static vm_fault_t do_huge_pmd_wp_page_fallback(struct vm_fault *vmf,
1251                         pmd_t orig_pmd, struct page *page)
1252 {
1253         struct vm_area_struct *vma = vmf->vma;
1254         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1255         struct mem_cgroup *memcg;
1256         pgtable_t pgtable;
1257         pmd_t _pmd;
1258         int i;
1259         vm_fault_t ret = 0;
1260         struct page **pages;
1261         struct mmu_notifier_range range;
1262
1263         pages = kmalloc_array(HPAGE_PMD_NR, sizeof(struct page *),
1264                               GFP_KERNEL);
1265         if (unlikely(!pages)) {
1266                 ret |= VM_FAULT_OOM;
1267                 goto out;
1268         }
1269
1270         for (i = 0; i < HPAGE_PMD_NR; i++) {
1271                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
1272                                                vmf->address, page_to_nid(page));
1273                 if (unlikely(!pages[i] ||
1274                              mem_cgroup_try_charge_delay(pages[i], vma->vm_mm,
1275                                      GFP_KERNEL, &memcg, false))) {
1276                         if (pages[i])
1277                                 put_page(pages[i]);
1278                         while (--i >= 0) {
1279                                 memcg = (void *)page_private(pages[i]);
1280                                 set_page_private(pages[i], 0);
1281                                 mem_cgroup_cancel_charge(pages[i], memcg,
1282                                                 false);
1283                                 put_page(pages[i]);
1284                         }
1285                         kfree(pages);
1286                         ret |= VM_FAULT_OOM;
1287                         goto out;
1288                 }
1289                 set_page_private(pages[i], (unsigned long)memcg);
1290         }
1291
1292         for (i = 0; i < HPAGE_PMD_NR; i++) {
1293                 copy_user_highpage(pages[i], page + i,
1294                                    haddr + PAGE_SIZE * i, vma);
1295                 __SetPageUptodate(pages[i]);
1296                 cond_resched();
1297         }
1298
1299         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1300                                 haddr, haddr + HPAGE_PMD_SIZE);
1301         mmu_notifier_invalidate_range_start(&range);
1302
1303         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1304         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1305                 goto out_free_pages;
1306         VM_BUG_ON_PAGE(!PageHead(page), page);
1307
1308         /*
1309          * Leave pmd empty until pte is filled note we must notify here as
1310          * concurrent CPU thread might write to new page before the call to
1311          * mmu_notifier_invalidate_range_end() happens which can lead to a
1312          * device seeing memory write in different order than CPU.
1313          *
1314          * See Documentation/vm/mmu_notifier.rst
1315          */
1316         pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1317
1318         pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
1319         pmd_populate(vma->vm_mm, &_pmd, pgtable);
1320
1321         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1322                 pte_t entry;
1323                 entry = mk_pte(pages[i], vma->vm_page_prot);
1324                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1325                 memcg = (void *)page_private(pages[i]);
1326                 set_page_private(pages[i], 0);
1327                 page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
1328                 mem_cgroup_commit_charge(pages[i], memcg, false, false);
1329                 lru_cache_add_active_or_unevictable(pages[i], vma);
1330                 vmf->pte = pte_offset_map(&_pmd, haddr);
1331                 VM_BUG_ON(!pte_none(*vmf->pte));
1332                 set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
1333                 pte_unmap(vmf->pte);
1334         }
1335         kfree(pages);
1336
1337         smp_wmb(); /* make pte visible before pmd */
1338         pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
1339         page_remove_rmap(page, true);
1340         spin_unlock(vmf->ptl);
1341
1342         /*
1343          * No need to double call mmu_notifier->invalidate_range() callback as
1344          * the above pmdp_huge_clear_flush_notify() did already call it.
1345          */
1346         mmu_notifier_invalidate_range_only_end(&range);
1347
1348         ret |= VM_FAULT_WRITE;
1349         put_page(page);
1350
1351 out:
1352         return ret;
1353
1354 out_free_pages:
1355         spin_unlock(vmf->ptl);
1356         mmu_notifier_invalidate_range_end(&range);
1357         for (i = 0; i < HPAGE_PMD_NR; i++) {
1358                 memcg = (void *)page_private(pages[i]);
1359                 set_page_private(pages[i], 0);
1360                 mem_cgroup_cancel_charge(pages[i], memcg, false);
1361                 put_page(pages[i]);
1362         }
1363         kfree(pages);
1364         goto out;
1365 }
1366
1367 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1368 {
1369         struct vm_area_struct *vma = vmf->vma;
1370         struct page *page = NULL, *new_page;
1371         struct mem_cgroup *memcg;
1372         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1373         struct mmu_notifier_range range;
1374         gfp_t huge_gfp;                 /* for allocation and charge */
1375         vm_fault_t ret = 0;
1376
1377         vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1378         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1379         if (is_huge_zero_pmd(orig_pmd))
1380                 goto alloc;
1381         spin_lock(vmf->ptl);
1382         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1383                 goto out_unlock;
1384
1385         page = pmd_page(orig_pmd);
1386         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1387         /*
1388          * We can only reuse the page if nobody else maps the huge page or it's
1389          * part.
1390          */
1391         if (!trylock_page(page)) {
1392                 get_page(page);
1393                 spin_unlock(vmf->ptl);
1394                 lock_page(page);
1395                 spin_lock(vmf->ptl);
1396                 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1397                         unlock_page(page);
1398                         put_page(page);
1399                         goto out_unlock;
1400                 }
1401                 put_page(page);
1402         }
1403         if (reuse_swap_page(page, NULL)) {
1404                 pmd_t entry;
1405                 entry = pmd_mkyoung(orig_pmd);
1406                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1407                 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry,  1))
1408                         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1409                 ret |= VM_FAULT_WRITE;
1410                 unlock_page(page);
1411                 goto out_unlock;
1412         }
1413         unlock_page(page);
1414         get_page(page);
1415         spin_unlock(vmf->ptl);
1416 alloc:
1417         if (__transparent_hugepage_enabled(vma) &&
1418             !transparent_hugepage_debug_cow()) {
1419                 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1420                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1421         } else
1422                 new_page = NULL;
1423
1424         if (likely(new_page)) {
1425                 prep_transhuge_page(new_page);
1426         } else {
1427                 if (!page) {
1428                         split_huge_pmd(vma, vmf->pmd, vmf->address);
1429                         ret |= VM_FAULT_FALLBACK;
1430                 } else {
1431                         ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
1432                         if (ret & VM_FAULT_OOM) {
1433                                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1434                                 ret |= VM_FAULT_FALLBACK;
1435                         }
1436                         put_page(page);
1437                 }
1438                 count_vm_event(THP_FAULT_FALLBACK);
1439                 goto out;
1440         }
1441
1442         if (unlikely(mem_cgroup_try_charge_delay(new_page, vma->vm_mm,
1443                                         huge_gfp, &memcg, true))) {
1444                 put_page(new_page);
1445                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1446                 if (page)
1447                         put_page(page);
1448                 ret |= VM_FAULT_FALLBACK;
1449                 count_vm_event(THP_FAULT_FALLBACK);
1450                 count_vm_event(THP_FAULT_FALLBACK_CHARGE);
1451                 goto out;
1452         }
1453
1454         count_vm_event(THP_FAULT_ALLOC);
1455         count_memcg_events(memcg, THP_FAULT_ALLOC, 1);
1456
1457         if (!page)
1458                 clear_huge_page(new_page, vmf->address, HPAGE_PMD_NR);
1459         else
1460                 copy_user_huge_page(new_page, page, vmf->address,
1461                                     vma, HPAGE_PMD_NR);
1462         __SetPageUptodate(new_page);
1463
1464         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1465                                 haddr, haddr + HPAGE_PMD_SIZE);
1466         mmu_notifier_invalidate_range_start(&range);
1467
1468         spin_lock(vmf->ptl);
1469         if (page)
1470                 put_page(page);
1471         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1472                 spin_unlock(vmf->ptl);
1473                 mem_cgroup_cancel_charge(new_page, memcg, true);
1474                 put_page(new_page);
1475                 goto out_mn;
1476         } else {
1477                 pmd_t entry;
1478                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1479                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1480                 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1481                 page_add_new_anon_rmap(new_page, vma, haddr, true);
1482                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1483                 lru_cache_add_active_or_unevictable(new_page, vma);
1484                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1485                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1486                 if (!page) {
1487                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1488                 } else {
1489                         VM_BUG_ON_PAGE(!PageHead(page), page);
1490                         page_remove_rmap(page, true);
1491                         put_page(page);
1492                 }
1493                 ret |= VM_FAULT_WRITE;
1494         }
1495         spin_unlock(vmf->ptl);
1496 out_mn:
1497         /*
1498          * No need to double call mmu_notifier->invalidate_range() callback as
1499          * the above pmdp_huge_clear_flush_notify() did already call it.
1500          */
1501         mmu_notifier_invalidate_range_only_end(&range);
1502 out:
1503         return ret;
1504 out_unlock:
1505         spin_unlock(vmf->ptl);
1506         return ret;
1507 }
1508
1509 /*
1510  * FOLL_FORCE can write to even unwritable pmd's, but only
1511  * after we've gone through a COW cycle and they are dirty.
1512  */
1513 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1514 {
1515         return pmd_write(pmd) ||
1516                ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1517 }
1518
1519 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1520                                    unsigned long addr,
1521                                    pmd_t *pmd,
1522                                    unsigned int flags)
1523 {
1524         struct mm_struct *mm = vma->vm_mm;
1525         struct page *page = NULL;
1526
1527         assert_spin_locked(pmd_lockptr(mm, pmd));
1528
1529         if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1530                 goto out;
1531
1532         /* Avoid dumping huge zero page */
1533         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1534                 return ERR_PTR(-EFAULT);
1535
1536         /* Full NUMA hinting faults to serialise migration in fault paths */
1537         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1538                 goto out;
1539
1540         page = pmd_page(*pmd);
1541         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1542
1543         if (!try_grab_page(page, flags))
1544                 return ERR_PTR(-ENOMEM);
1545
1546         if (flags & FOLL_TOUCH)
1547                 touch_pmd(vma, addr, pmd, flags);
1548
1549         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1550                 /*
1551                  * We don't mlock() pte-mapped THPs. This way we can avoid
1552                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1553                  *
1554                  * For anon THP:
1555                  *
1556                  * In most cases the pmd is the only mapping of the page as we
1557                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1558                  * writable private mappings in populate_vma_page_range().
1559                  *
1560                  * The only scenario when we have the page shared here is if we
1561                  * mlocking read-only mapping shared over fork(). We skip
1562                  * mlocking such pages.
1563                  *
1564                  * For file THP:
1565                  *
1566                  * We can expect PageDoubleMap() to be stable under page lock:
1567                  * for file pages we set it in page_add_file_rmap(), which
1568                  * requires page to be locked.
1569                  */
1570
1571                 if (PageAnon(page) && compound_mapcount(page) != 1)
1572                         goto skip_mlock;
1573                 if (PageDoubleMap(page) || !page->mapping)
1574                         goto skip_mlock;
1575                 if (!trylock_page(page))
1576                         goto skip_mlock;
1577                 lru_add_drain();
1578                 if (page->mapping && !PageDoubleMap(page))
1579                         mlock_vma_page(page);
1580                 unlock_page(page);
1581         }
1582 skip_mlock:
1583         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1584         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1585
1586 out:
1587         return page;
1588 }
1589
1590 /* NUMA hinting page fault entry point for trans huge pmds */
1591 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1592 {
1593         struct vm_area_struct *vma = vmf->vma;
1594         struct anon_vma *anon_vma = NULL;
1595         struct page *page;
1596         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1597         int page_nid = NUMA_NO_NODE, this_nid = numa_node_id();
1598         int target_nid, last_cpupid = -1;
1599         bool page_locked;
1600         bool migrated = false;
1601         bool was_writable;
1602         int flags = 0;
1603
1604         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1605         if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1606                 goto out_unlock;
1607
1608         /*
1609          * If there are potential migrations, wait for completion and retry
1610          * without disrupting NUMA hinting information. Do not relock and
1611          * check_same as the page may no longer be mapped.
1612          */
1613         if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1614                 page = pmd_page(*vmf->pmd);
1615                 if (!get_page_unless_zero(page))
1616                         goto out_unlock;
1617                 spin_unlock(vmf->ptl);
1618                 put_and_wait_on_page_locked(page);
1619                 goto out;
1620         }
1621
1622         page = pmd_page(pmd);
1623         BUG_ON(is_huge_zero_page(page));
1624         page_nid = page_to_nid(page);
1625         last_cpupid = page_cpupid_last(page);
1626         count_vm_numa_event(NUMA_HINT_FAULTS);
1627         if (page_nid == this_nid) {
1628                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1629                 flags |= TNF_FAULT_LOCAL;
1630         }
1631
1632         /* See similar comment in do_numa_page for explanation */
1633         if (!pmd_savedwrite(pmd))
1634                 flags |= TNF_NO_GROUP;
1635
1636         /*
1637          * Acquire the page lock to serialise THP migrations but avoid dropping
1638          * page_table_lock if at all possible
1639          */
1640         page_locked = trylock_page(page);
1641         target_nid = mpol_misplaced(page, vma, haddr);
1642         if (target_nid == NUMA_NO_NODE) {
1643                 /* If the page was locked, there are no parallel migrations */
1644                 if (page_locked)
1645                         goto clear_pmdnuma;
1646         }
1647
1648         /* Migration could have started since the pmd_trans_migrating check */
1649         if (!page_locked) {
1650                 page_nid = NUMA_NO_NODE;
1651                 if (!get_page_unless_zero(page))
1652                         goto out_unlock;
1653                 spin_unlock(vmf->ptl);
1654                 put_and_wait_on_page_locked(page);
1655                 goto out;
1656         }
1657
1658         /*
1659          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1660          * to serialises splits
1661          */
1662         get_page(page);
1663         spin_unlock(vmf->ptl);
1664         anon_vma = page_lock_anon_vma_read(page);
1665
1666         /* Confirm the PMD did not change while page_table_lock was released */
1667         spin_lock(vmf->ptl);
1668         if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1669                 unlock_page(page);
1670                 put_page(page);
1671                 page_nid = NUMA_NO_NODE;
1672                 goto out_unlock;
1673         }
1674
1675         /* Bail if we fail to protect against THP splits for any reason */
1676         if (unlikely(!anon_vma)) {
1677                 put_page(page);
1678                 page_nid = NUMA_NO_NODE;
1679                 goto clear_pmdnuma;
1680         }
1681
1682         /*
1683          * Since we took the NUMA fault, we must have observed the !accessible
1684          * bit. Make sure all other CPUs agree with that, to avoid them
1685          * modifying the page we're about to migrate.
1686          *
1687          * Must be done under PTL such that we'll observe the relevant
1688          * inc_tlb_flush_pending().
1689          *
1690          * We are not sure a pending tlb flush here is for a huge page
1691          * mapping or not. Hence use the tlb range variant
1692          */
1693         if (mm_tlb_flush_pending(vma->vm_mm)) {
1694                 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1695                 /*
1696                  * change_huge_pmd() released the pmd lock before
1697                  * invalidating the secondary MMUs sharing the primary
1698                  * MMU pagetables (with ->invalidate_range()). The
1699                  * mmu_notifier_invalidate_range_end() (which
1700                  * internally calls ->invalidate_range()) in
1701                  * change_pmd_range() will run after us, so we can't
1702                  * rely on it here and we need an explicit invalidate.
1703                  */
1704                 mmu_notifier_invalidate_range(vma->vm_mm, haddr,
1705                                               haddr + HPAGE_PMD_SIZE);
1706         }
1707
1708         /*
1709          * Migrate the THP to the requested node, returns with page unlocked
1710          * and access rights restored.
1711          */
1712         spin_unlock(vmf->ptl);
1713
1714         migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1715                                 vmf->pmd, pmd, vmf->address, page, target_nid);
1716         if (migrated) {
1717                 flags |= TNF_MIGRATED;
1718                 page_nid = target_nid;
1719         } else
1720                 flags |= TNF_MIGRATE_FAIL;
1721
1722         goto out;
1723 clear_pmdnuma:
1724         BUG_ON(!PageLocked(page));
1725         was_writable = pmd_savedwrite(pmd);
1726         pmd = pmd_modify(pmd, vma->vm_page_prot);
1727         pmd = pmd_mkyoung(pmd);
1728         if (was_writable)
1729                 pmd = pmd_mkwrite(pmd);
1730         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1731         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1732         unlock_page(page);
1733 out_unlock:
1734         spin_unlock(vmf->ptl);
1735
1736 out:
1737         if (anon_vma)
1738                 page_unlock_anon_vma_read(anon_vma);
1739
1740         if (page_nid != NUMA_NO_NODE)
1741                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1742                                 flags);
1743
1744         return 0;
1745 }
1746
1747 /*
1748  * Return true if we do MADV_FREE successfully on entire pmd page.
1749  * Otherwise, return false.
1750  */
1751 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1752                 pmd_t *pmd, unsigned long addr, unsigned long next)
1753 {
1754         spinlock_t *ptl;
1755         pmd_t orig_pmd;
1756         struct page *page;
1757         struct mm_struct *mm = tlb->mm;
1758         bool ret = false;
1759
1760         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1761
1762         ptl = pmd_trans_huge_lock(pmd, vma);
1763         if (!ptl)
1764                 goto out_unlocked;
1765
1766         orig_pmd = *pmd;
1767         if (is_huge_zero_pmd(orig_pmd))
1768                 goto out;
1769
1770         if (unlikely(!pmd_present(orig_pmd))) {
1771                 VM_BUG_ON(thp_migration_supported() &&
1772                                   !is_pmd_migration_entry(orig_pmd));
1773                 goto out;
1774         }
1775
1776         page = pmd_page(orig_pmd);
1777         /*
1778          * If other processes are mapping this page, we couldn't discard
1779          * the page unless they all do MADV_FREE so let's skip the page.
1780          */
1781         if (page_mapcount(page) != 1)
1782                 goto out;
1783
1784         if (!trylock_page(page))
1785                 goto out;
1786
1787         /*
1788          * If user want to discard part-pages of THP, split it so MADV_FREE
1789          * will deactivate only them.
1790          */
1791         if (next - addr != HPAGE_PMD_SIZE) {
1792                 get_page(page);
1793                 spin_unlock(ptl);
1794                 split_huge_page(page);
1795                 unlock_page(page);
1796                 put_page(page);
1797                 goto out_unlocked;
1798         }
1799
1800         if (PageDirty(page))
1801                 ClearPageDirty(page);
1802         unlock_page(page);
1803
1804         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1805                 pmdp_invalidate(vma, addr, pmd);
1806                 orig_pmd = pmd_mkold(orig_pmd);
1807                 orig_pmd = pmd_mkclean(orig_pmd);
1808
1809                 set_pmd_at(mm, addr, pmd, orig_pmd);
1810                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1811         }
1812
1813         mark_page_lazyfree(page);
1814         ret = true;
1815 out:
1816         spin_unlock(ptl);
1817 out_unlocked:
1818         return ret;
1819 }
1820
1821 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1822 {
1823         pgtable_t pgtable;
1824
1825         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1826         pte_free(mm, pgtable);
1827         mm_dec_nr_ptes(mm);
1828 }
1829
1830 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1831                  pmd_t *pmd, unsigned long addr)
1832 {
1833         pmd_t orig_pmd;
1834         spinlock_t *ptl;
1835
1836         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1837
1838         ptl = __pmd_trans_huge_lock(pmd, vma);
1839         if (!ptl)
1840                 return 0;
1841         /*
1842          * For architectures like ppc64 we look at deposited pgtable
1843          * when calling pmdp_huge_get_and_clear. So do the
1844          * pgtable_trans_huge_withdraw after finishing pmdp related
1845          * operations.
1846          */
1847         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1848                         tlb->fullmm);
1849         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1850         if (vma_is_special_huge(vma)) {
1851                 if (arch_needs_pgtable_deposit())
1852                         zap_deposited_table(tlb->mm, pmd);
1853                 spin_unlock(ptl);
1854                 if (is_huge_zero_pmd(orig_pmd))
1855                         tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1856         } else if (is_huge_zero_pmd(orig_pmd)) {
1857                 zap_deposited_table(tlb->mm, pmd);
1858                 spin_unlock(ptl);
1859                 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1860         } else {
1861                 struct page *page = NULL;
1862                 int flush_needed = 1;
1863
1864                 if (pmd_present(orig_pmd)) {
1865                         page = pmd_page(orig_pmd);
1866                         page_remove_rmap(page, true);
1867                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1868                         VM_BUG_ON_PAGE(!PageHead(page), page);
1869                 } else if (thp_migration_supported()) {
1870                         swp_entry_t entry;
1871
1872                         VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1873                         entry = pmd_to_swp_entry(orig_pmd);
1874                         page = pfn_to_page(swp_offset(entry));
1875                         flush_needed = 0;
1876                 } else
1877                         WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1878
1879                 if (PageAnon(page)) {
1880                         zap_deposited_table(tlb->mm, pmd);
1881                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1882                 } else {
1883                         if (arch_needs_pgtable_deposit())
1884                                 zap_deposited_table(tlb->mm, pmd);
1885                         add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1886                 }
1887
1888                 spin_unlock(ptl);
1889                 if (flush_needed)
1890                         tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1891         }
1892         return 1;
1893 }
1894
1895 #ifndef pmd_move_must_withdraw
1896 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1897                                          spinlock_t *old_pmd_ptl,
1898                                          struct vm_area_struct *vma)
1899 {
1900         /*
1901          * With split pmd lock we also need to move preallocated
1902          * PTE page table if new_pmd is on different PMD page table.
1903          *
1904          * We also don't deposit and withdraw tables for file pages.
1905          */
1906         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1907 }
1908 #endif
1909
1910 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1911 {
1912 #ifdef CONFIG_MEM_SOFT_DIRTY
1913         if (unlikely(is_pmd_migration_entry(pmd)))
1914                 pmd = pmd_swp_mksoft_dirty(pmd);
1915         else if (pmd_present(pmd))
1916                 pmd = pmd_mksoft_dirty(pmd);
1917 #endif
1918         return pmd;
1919 }
1920
1921 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1922                   unsigned long new_addr, unsigned long old_end,
1923                   pmd_t *old_pmd, pmd_t *new_pmd)
1924 {
1925         spinlock_t *old_ptl, *new_ptl;
1926         pmd_t pmd;
1927         struct mm_struct *mm = vma->vm_mm;
1928         bool force_flush = false;
1929
1930         if ((old_addr & ~HPAGE_PMD_MASK) ||
1931             (new_addr & ~HPAGE_PMD_MASK) ||
1932             old_end - old_addr < HPAGE_PMD_SIZE)
1933                 return false;
1934
1935         /*
1936          * The destination pmd shouldn't be established, free_pgtables()
1937          * should have release it.
1938          */
1939         if (WARN_ON(!pmd_none(*new_pmd))) {
1940                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1941                 return false;
1942         }
1943
1944         /*
1945          * We don't have to worry about the ordering of src and dst
1946          * ptlocks because exclusive mmap_sem prevents deadlock.
1947          */
1948         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1949         if (old_ptl) {
1950                 new_ptl = pmd_lockptr(mm, new_pmd);
1951                 if (new_ptl != old_ptl)
1952                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1953                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1954                 if (pmd_present(pmd))
1955                         force_flush = true;
1956                 VM_BUG_ON(!pmd_none(*new_pmd));
1957
1958                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1959                         pgtable_t pgtable;
1960                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1961                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1962                 }
1963                 pmd = move_soft_dirty_pmd(pmd);
1964                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1965                 if (force_flush)
1966                         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1967                 if (new_ptl != old_ptl)
1968                         spin_unlock(new_ptl);
1969                 spin_unlock(old_ptl);
1970                 return true;
1971         }
1972         return false;
1973 }
1974
1975 /*
1976  * Returns
1977  *  - 0 if PMD could not be locked
1978  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1979  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1980  */
1981 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1982                 unsigned long addr, pgprot_t newprot, int prot_numa)
1983 {
1984         struct mm_struct *mm = vma->vm_mm;
1985         spinlock_t *ptl;
1986         pmd_t entry;
1987         bool preserve_write;
1988         int ret;
1989
1990         ptl = __pmd_trans_huge_lock(pmd, vma);
1991         if (!ptl)
1992                 return 0;
1993
1994         preserve_write = prot_numa && pmd_write(*pmd);
1995         ret = 1;
1996
1997 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1998         if (is_swap_pmd(*pmd)) {
1999                 swp_entry_t entry = pmd_to_swp_entry(*pmd);
2000
2001                 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
2002                 if (is_write_migration_entry(entry)) {
2003                         pmd_t newpmd;
2004                         /*
2005                          * A protection check is difficult so
2006                          * just be safe and disable write
2007                          */
2008                         make_migration_entry_read(&entry);
2009                         newpmd = swp_entry_to_pmd(entry);
2010                         if (pmd_swp_soft_dirty(*pmd))
2011                                 newpmd = pmd_swp_mksoft_dirty(newpmd);
2012                         set_pmd_at(mm, addr, pmd, newpmd);
2013                 }
2014                 goto unlock;
2015         }
2016 #endif
2017
2018         /*
2019          * Avoid trapping faults against the zero page. The read-only
2020          * data is likely to be read-cached on the local CPU and
2021          * local/remote hits to the zero page are not interesting.
2022          */
2023         if (prot_numa && is_huge_zero_pmd(*pmd))
2024                 goto unlock;
2025
2026         if (prot_numa && pmd_protnone(*pmd))
2027                 goto unlock;
2028
2029         /*
2030          * In case prot_numa, we are under down_read(mmap_sem). It's critical
2031          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
2032          * which is also under down_read(mmap_sem):
2033          *
2034          *      CPU0:                           CPU1:
2035          *                              change_huge_pmd(prot_numa=1)
2036          *                               pmdp_huge_get_and_clear_notify()
2037          * madvise_dontneed()
2038          *  zap_pmd_range()
2039          *   pmd_trans_huge(*pmd) == 0 (without ptl)
2040          *   // skip the pmd
2041          *                               set_pmd_at();
2042          *                               // pmd is re-established
2043          *
2044          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
2045          * which may break userspace.
2046          *
2047          * pmdp_invalidate() is required to make sure we don't miss
2048          * dirty/young flags set by hardware.
2049          */
2050         entry = pmdp_invalidate(vma, addr, pmd);
2051
2052         entry = pmd_modify(entry, newprot);
2053         if (preserve_write)
2054                 entry = pmd_mk_savedwrite(entry);
2055         ret = HPAGE_PMD_NR;
2056         set_pmd_at(mm, addr, pmd, entry);
2057         BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
2058 unlock:
2059         spin_unlock(ptl);
2060         return ret;
2061 }
2062
2063 /*
2064  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
2065  *
2066  * Note that if it returns page table lock pointer, this routine returns without
2067  * unlocking page table lock. So callers must unlock it.
2068  */
2069 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
2070 {
2071         spinlock_t *ptl;
2072         ptl = pmd_lock(vma->vm_mm, pmd);
2073         if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
2074                         pmd_devmap(*pmd)))
2075                 return ptl;
2076         spin_unlock(ptl);
2077         return NULL;
2078 }
2079
2080 /*
2081  * Returns true if a given pud maps a thp, false otherwise.
2082  *
2083  * Note that if it returns true, this routine returns without unlocking page
2084  * table lock. So callers must unlock it.
2085  */
2086 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
2087 {
2088         spinlock_t *ptl;
2089
2090         ptl = pud_lock(vma->vm_mm, pud);
2091         if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
2092                 return ptl;
2093         spin_unlock(ptl);
2094         return NULL;
2095 }
2096
2097 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
2098 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
2099                  pud_t *pud, unsigned long addr)
2100 {
2101         spinlock_t *ptl;
2102
2103         ptl = __pud_trans_huge_lock(pud, vma);
2104         if (!ptl)
2105                 return 0;
2106         /*
2107          * For architectures like ppc64 we look at deposited pgtable
2108          * when calling pudp_huge_get_and_clear. So do the
2109          * pgtable_trans_huge_withdraw after finishing pudp related
2110          * operations.
2111          */
2112         pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
2113         tlb_remove_pud_tlb_entry(tlb, pud, addr);
2114         if (vma_is_special_huge(vma)) {
2115                 spin_unlock(ptl);
2116                 /* No zero page support yet */
2117         } else {
2118                 /* No support for anonymous PUD pages yet */
2119                 BUG();
2120         }
2121         return 1;
2122 }
2123
2124 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
2125                 unsigned long haddr)
2126 {
2127         VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
2128         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2129         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
2130         VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
2131
2132         count_vm_event(THP_SPLIT_PUD);
2133
2134         pudp_huge_clear_flush_notify(vma, haddr, pud);
2135 }
2136
2137 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2138                 unsigned long address)
2139 {
2140         spinlock_t *ptl;
2141         struct mmu_notifier_range range;
2142
2143         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2144                                 address & HPAGE_PUD_MASK,
2145                                 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
2146         mmu_notifier_invalidate_range_start(&range);
2147         ptl = pud_lock(vma->vm_mm, pud);
2148         if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2149                 goto out;
2150         __split_huge_pud_locked(vma, pud, range.start);
2151
2152 out:
2153         spin_unlock(ptl);
2154         /*
2155          * No need to double call mmu_notifier->invalidate_range() callback as
2156          * the above pudp_huge_clear_flush_notify() did already call it.
2157          */
2158         mmu_notifier_invalidate_range_only_end(&range);
2159 }
2160 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2161
2162 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2163                 unsigned long haddr, pmd_t *pmd)
2164 {
2165         struct mm_struct *mm = vma->vm_mm;
2166         pgtable_t pgtable;
2167         pmd_t _pmd;
2168         int i;
2169
2170         /*
2171          * Leave pmd empty until pte is filled note that it is fine to delay
2172          * notification until mmu_notifier_invalidate_range_end() as we are
2173          * replacing a zero pmd write protected page with a zero pte write
2174          * protected page.
2175          *
2176          * See Documentation/vm/mmu_notifier.rst
2177          */
2178         pmdp_huge_clear_flush(vma, haddr, pmd);
2179
2180         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2181         pmd_populate(mm, &_pmd, pgtable);
2182
2183         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2184                 pte_t *pte, entry;
2185                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2186                 entry = pte_mkspecial(entry);
2187                 pte = pte_offset_map(&_pmd, haddr);
2188                 VM_BUG_ON(!pte_none(*pte));
2189                 set_pte_at(mm, haddr, pte, entry);
2190                 pte_unmap(pte);
2191         }
2192         smp_wmb(); /* make pte visible before pmd */
2193         pmd_populate(mm, pmd, pgtable);
2194 }
2195
2196 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2197                 unsigned long haddr, bool freeze)
2198 {
2199         struct mm_struct *mm = vma->vm_mm;
2200         struct page *page;
2201         pgtable_t pgtable;
2202         pmd_t old_pmd, _pmd;
2203         bool young, write, soft_dirty, pmd_migration = false;
2204         unsigned long addr;
2205         int i;
2206
2207         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2208         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2209         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2210         VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2211                                 && !pmd_devmap(*pmd));
2212
2213         count_vm_event(THP_SPLIT_PMD);
2214
2215         if (!vma_is_anonymous(vma)) {
2216                 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2217                 /*
2218                  * We are going to unmap this huge page. So
2219                  * just go ahead and zap it
2220                  */
2221                 if (arch_needs_pgtable_deposit())
2222                         zap_deposited_table(mm, pmd);
2223                 if (vma_is_special_huge(vma))
2224                         return;
2225                 page = pmd_page(_pmd);
2226                 if (!PageDirty(page) && pmd_dirty(_pmd))
2227                         set_page_dirty(page);
2228                 if (!PageReferenced(page) && pmd_young(_pmd))
2229                         SetPageReferenced(page);
2230                 page_remove_rmap(page, true);
2231                 put_page(page);
2232                 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2233                 return;
2234         } else if (is_huge_zero_pmd(*pmd)) {
2235                 /*
2236                  * FIXME: Do we want to invalidate secondary mmu by calling
2237                  * mmu_notifier_invalidate_range() see comments below inside
2238                  * __split_huge_pmd() ?
2239                  *
2240                  * We are going from a zero huge page write protected to zero
2241                  * small page also write protected so it does not seems useful
2242                  * to invalidate secondary mmu at this time.
2243                  */
2244                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2245         }
2246
2247         /*
2248          * Up to this point the pmd is present and huge and userland has the
2249          * whole access to the hugepage during the split (which happens in
2250          * place). If we overwrite the pmd with the not-huge version pointing
2251          * to the pte here (which of course we could if all CPUs were bug
2252          * free), userland could trigger a small page size TLB miss on the
2253          * small sized TLB while the hugepage TLB entry is still established in
2254          * the huge TLB. Some CPU doesn't like that.
2255          * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2256          * 383 on page 93. Intel should be safe but is also warns that it's
2257          * only safe if the permission and cache attributes of the two entries
2258          * loaded in the two TLB is identical (which should be the case here).
2259          * But it is generally safer to never allow small and huge TLB entries
2260          * for the same virtual address to be loaded simultaneously. So instead
2261          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2262          * current pmd notpresent (atomically because here the pmd_trans_huge
2263          * must remain set at all times on the pmd until the split is complete
2264          * for this pmd), then we flush the SMP TLB and finally we write the
2265          * non-huge version of the pmd entry with pmd_populate.
2266          */
2267         old_pmd = pmdp_invalidate(vma, haddr, pmd);
2268
2269         pmd_migration = is_pmd_migration_entry(old_pmd);
2270         if (unlikely(pmd_migration)) {
2271                 swp_entry_t entry;
2272
2273                 entry = pmd_to_swp_entry(old_pmd);
2274                 page = pfn_to_page(swp_offset(entry));
2275                 write = is_write_migration_entry(entry);
2276                 young = false;
2277                 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2278         } else {
2279                 page = pmd_page(old_pmd);
2280                 if (pmd_dirty(old_pmd))
2281                         SetPageDirty(page);
2282                 write = pmd_write(old_pmd);
2283                 young = pmd_young(old_pmd);
2284                 soft_dirty = pmd_soft_dirty(old_pmd);
2285         }
2286         VM_BUG_ON_PAGE(!page_count(page), page);
2287         page_ref_add(page, HPAGE_PMD_NR - 1);
2288
2289         /*
2290          * Withdraw the table only after we mark the pmd entry invalid.
2291          * This's critical for some architectures (Power).
2292          */
2293         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2294         pmd_populate(mm, &_pmd, pgtable);
2295
2296         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2297                 pte_t entry, *pte;
2298                 /*
2299                  * Note that NUMA hinting access restrictions are not
2300                  * transferred to avoid any possibility of altering
2301                  * permissions across VMAs.
2302                  */
2303                 if (freeze || pmd_migration) {
2304                         swp_entry_t swp_entry;
2305                         swp_entry = make_migration_entry(page + i, write);
2306                         entry = swp_entry_to_pte(swp_entry);
2307                         if (soft_dirty)
2308                                 entry = pte_swp_mksoft_dirty(entry);
2309                 } else {
2310                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2311                         entry = maybe_mkwrite(entry, vma);
2312                         if (!write)
2313                                 entry = pte_wrprotect(entry);
2314                         if (!young)
2315                                 entry = pte_mkold(entry);
2316                         if (soft_dirty)
2317                                 entry = pte_mksoft_dirty(entry);
2318                 }
2319                 pte = pte_offset_map(&_pmd, addr);
2320                 BUG_ON(!pte_none(*pte));
2321                 set_pte_at(mm, addr, pte, entry);
2322                 atomic_inc(&page[i]._mapcount);
2323                 pte_unmap(pte);
2324         }
2325
2326         /*
2327          * Set PG_double_map before dropping compound_mapcount to avoid
2328          * false-negative page_mapped().
2329          */
2330         if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2331                 for (i = 0; i < HPAGE_PMD_NR; i++)
2332                         atomic_inc(&page[i]._mapcount);
2333         }
2334
2335         if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2336                 /* Last compound_mapcount is gone. */
2337                 __dec_node_page_state(page, NR_ANON_THPS);
2338                 if (TestClearPageDoubleMap(page)) {
2339                         /* No need in mapcount reference anymore */
2340                         for (i = 0; i < HPAGE_PMD_NR; i++)
2341                                 atomic_dec(&page[i]._mapcount);
2342                 }
2343         }
2344
2345         smp_wmb(); /* make pte visible before pmd */
2346         pmd_populate(mm, pmd, pgtable);
2347
2348         if (freeze) {
2349                 for (i = 0; i < HPAGE_PMD_NR; i++) {
2350                         page_remove_rmap(page + i, false);
2351                         put_page(page + i);
2352                 }
2353         }
2354 }
2355
2356 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2357                 unsigned long address, bool freeze, struct page *page)
2358 {
2359         spinlock_t *ptl;
2360         struct mmu_notifier_range range;
2361
2362         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2363                                 address & HPAGE_PMD_MASK,
2364                                 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2365         mmu_notifier_invalidate_range_start(&range);
2366         ptl = pmd_lock(vma->vm_mm, pmd);
2367
2368         /*
2369          * If caller asks to setup a migration entries, we need a page to check
2370          * pmd against. Otherwise we can end up replacing wrong page.
2371          */
2372         VM_BUG_ON(freeze && !page);
2373         if (page && page != pmd_page(*pmd))
2374                 goto out;
2375
2376         if (pmd_trans_huge(*pmd)) {
2377                 page = pmd_page(*pmd);
2378                 if (PageMlocked(page))
2379                         clear_page_mlock(page);
2380         } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2381                 goto out;
2382         __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2383 out:
2384         spin_unlock(ptl);
2385         /*
2386          * No need to double call mmu_notifier->invalidate_range() callback.
2387          * They are 3 cases to consider inside __split_huge_pmd_locked():
2388          *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2389          *  2) __split_huge_zero_page_pmd() read only zero page and any write
2390          *    fault will trigger a flush_notify before pointing to a new page
2391          *    (it is fine if the secondary mmu keeps pointing to the old zero
2392          *    page in the meantime)
2393          *  3) Split a huge pmd into pte pointing to the same page. No need
2394          *     to invalidate secondary tlb entry they are all still valid.
2395          *     any further changes to individual pte will notify. So no need
2396          *     to call mmu_notifier->invalidate_range()
2397          */
2398         mmu_notifier_invalidate_range_only_end(&range);
2399 }
2400
2401 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2402                 bool freeze, struct page *page)
2403 {
2404         pgd_t *pgd;
2405         p4d_t *p4d;
2406         pud_t *pud;
2407         pmd_t *pmd;
2408
2409         pgd = pgd_offset(vma->vm_mm, address);
2410         if (!pgd_present(*pgd))
2411                 return;
2412
2413         p4d = p4d_offset(pgd, address);
2414         if (!p4d_present(*p4d))
2415                 return;
2416
2417         pud = pud_offset(p4d, address);
2418         if (!pud_present(*pud))
2419                 return;
2420
2421         pmd = pmd_offset(pud, address);
2422
2423         __split_huge_pmd(vma, pmd, address, freeze, page);
2424 }
2425
2426 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2427                              unsigned long start,
2428                              unsigned long end,
2429                              long adjust_next)
2430 {
2431         /*
2432          * If the new start address isn't hpage aligned and it could
2433          * previously contain an hugepage: check if we need to split
2434          * an huge pmd.
2435          */
2436         if (start & ~HPAGE_PMD_MASK &&
2437             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2438             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2439                 split_huge_pmd_address(vma, start, false, NULL);
2440
2441         /*
2442          * If the new end address isn't hpage aligned and it could
2443          * previously contain an hugepage: check if we need to split
2444          * an huge pmd.
2445          */
2446         if (end & ~HPAGE_PMD_MASK &&
2447             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2448             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2449                 split_huge_pmd_address(vma, end, false, NULL);
2450
2451         /*
2452          * If we're also updating the vma->vm_next->vm_start, if the new
2453          * vm_next->vm_start isn't page aligned and it could previously
2454          * contain an hugepage: check if we need to split an huge pmd.
2455          */
2456         if (adjust_next > 0) {
2457                 struct vm_area_struct *next = vma->vm_next;
2458                 unsigned long nstart = next->vm_start;
2459                 nstart += adjust_next << PAGE_SHIFT;
2460                 if (nstart & ~HPAGE_PMD_MASK &&
2461                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2462                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2463                         split_huge_pmd_address(next, nstart, false, NULL);
2464         }
2465 }
2466
2467 static void unmap_page(struct page *page)
2468 {
2469         enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
2470                 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2471         bool unmap_success;
2472
2473         VM_BUG_ON_PAGE(!PageHead(page), page);
2474
2475         if (PageAnon(page))
2476                 ttu_flags |= TTU_SPLIT_FREEZE;
2477
2478         unmap_success = try_to_unmap(page, ttu_flags);
2479         VM_BUG_ON_PAGE(!unmap_success, page);
2480 }
2481
2482 static void remap_page(struct page *page)
2483 {
2484         int i;
2485         if (PageTransHuge(page)) {
2486                 remove_migration_ptes(page, page, true);
2487         } else {
2488                 for (i = 0; i < HPAGE_PMD_NR; i++)
2489                         remove_migration_ptes(page + i, page + i, true);
2490         }
2491 }
2492
2493 static void __split_huge_page_tail(struct page *head, int tail,
2494                 struct lruvec *lruvec, struct list_head *list)
2495 {
2496         struct page *page_tail = head + tail;
2497
2498         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2499
2500         /*
2501          * Clone page flags before unfreezing refcount.
2502          *
2503          * After successful get_page_unless_zero() might follow flags change,
2504          * for exmaple lock_page() which set PG_waiters.
2505          */
2506         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2507         page_tail->flags |= (head->flags &
2508                         ((1L << PG_referenced) |
2509                          (1L << PG_swapbacked) |
2510                          (1L << PG_swapcache) |
2511                          (1L << PG_mlocked) |
2512                          (1L << PG_uptodate) |
2513                          (1L << PG_active) |
2514                          (1L << PG_workingset) |
2515                          (1L << PG_locked) |
2516                          (1L << PG_unevictable) |
2517                          (1L << PG_dirty)));
2518
2519         /* ->mapping in first tail page is compound_mapcount */
2520         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2521                         page_tail);
2522         page_tail->mapping = head->mapping;
2523         page_tail->index = head->index + tail;
2524
2525         /* Page flags must be visible before we make the page non-compound. */
2526         smp_wmb();
2527
2528         /*
2529          * Clear PageTail before unfreezing page refcount.
2530          *
2531          * After successful get_page_unless_zero() might follow put_page()
2532          * which needs correct compound_head().
2533          */
2534         clear_compound_head(page_tail);
2535
2536         /* Finally unfreeze refcount. Additional reference from page cache. */
2537         page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2538                                           PageSwapCache(head)));
2539
2540         if (page_is_young(head))
2541                 set_page_young(page_tail);
2542         if (page_is_idle(head))
2543                 set_page_idle(page_tail);
2544
2545         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2546
2547         /*
2548          * always add to the tail because some iterators expect new
2549          * pages to show after the currently processed elements - e.g.
2550          * migrate_pages
2551          */
2552         lru_add_page_tail(head, page_tail, lruvec, list);
2553 }
2554
2555 static void __split_huge_page(struct page *page, struct list_head *list,
2556                 pgoff_t end, unsigned long flags)
2557 {
2558         struct page *head = compound_head(page);
2559         pg_data_t *pgdat = page_pgdat(head);
2560         struct lruvec *lruvec;
2561         struct address_space *swap_cache = NULL;
2562         unsigned long offset = 0;
2563         int i;
2564
2565         lruvec = mem_cgroup_page_lruvec(head, pgdat);
2566
2567         /* complete memcg works before add pages to LRU */
2568         mem_cgroup_split_huge_fixup(head);
2569
2570         if (PageAnon(head) && PageSwapCache(head)) {
2571                 swp_entry_t entry = { .val = page_private(head) };
2572
2573                 offset = swp_offset(entry);
2574                 swap_cache = swap_address_space(entry);
2575                 xa_lock(&swap_cache->i_pages);
2576         }
2577
2578         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
2579                 __split_huge_page_tail(head, i, lruvec, list);
2580                 /* Some pages can be beyond i_size: drop them from page cache */
2581                 if (head[i].index >= end) {
2582                         ClearPageDirty(head + i);
2583                         __delete_from_page_cache(head + i, NULL);
2584                         if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2585                                 shmem_uncharge(head->mapping->host, 1);
2586                         put_page(head + i);
2587                 } else if (!PageAnon(page)) {
2588                         __xa_store(&head->mapping->i_pages, head[i].index,
2589                                         head + i, 0);
2590                 } else if (swap_cache) {
2591                         __xa_store(&swap_cache->i_pages, offset + i,
2592                                         head + i, 0);
2593                 }
2594         }
2595
2596         ClearPageCompound(head);
2597
2598         split_page_owner(head, HPAGE_PMD_ORDER);
2599
2600         /* See comment in __split_huge_page_tail() */
2601         if (PageAnon(head)) {
2602                 /* Additional pin to swap cache */
2603                 if (PageSwapCache(head)) {
2604                         page_ref_add(head, 2);
2605                         xa_unlock(&swap_cache->i_pages);
2606                 } else {
2607                         page_ref_inc(head);
2608                 }
2609         } else {
2610                 /* Additional pin to page cache */
2611                 page_ref_add(head, 2);
2612                 xa_unlock(&head->mapping->i_pages);
2613         }
2614
2615         spin_unlock_irqrestore(&pgdat->lru_lock, flags);
2616
2617         remap_page(head);
2618
2619         for (i = 0; i < HPAGE_PMD_NR; i++) {
2620                 struct page *subpage = head + i;
2621                 if (subpage == page)
2622                         continue;
2623                 unlock_page(subpage);
2624
2625                 /*
2626                  * Subpages may be freed if there wasn't any mapping
2627                  * like if add_to_swap() is running on a lru page that
2628                  * had its mapping zapped. And freeing these pages
2629                  * requires taking the lru_lock so we do the put_page
2630                  * of the tail pages after the split is complete.
2631                  */
2632                 put_page(subpage);
2633         }
2634 }
2635
2636 int total_mapcount(struct page *page)
2637 {
2638         int i, compound, ret;
2639
2640         VM_BUG_ON_PAGE(PageTail(page), page);
2641
2642         if (likely(!PageCompound(page)))
2643                 return atomic_read(&page->_mapcount) + 1;
2644
2645         compound = compound_mapcount(page);
2646         if (PageHuge(page))
2647                 return compound;
2648         ret = compound;
2649         for (i = 0; i < HPAGE_PMD_NR; i++)
2650                 ret += atomic_read(&page[i]._mapcount) + 1;
2651         /* File pages has compound_mapcount included in _mapcount */
2652         if (!PageAnon(page))
2653                 return ret - compound * HPAGE_PMD_NR;
2654         if (PageDoubleMap(page))
2655                 ret -= HPAGE_PMD_NR;
2656         return ret;
2657 }
2658
2659 /*
2660  * This calculates accurately how many mappings a transparent hugepage
2661  * has (unlike page_mapcount() which isn't fully accurate). This full
2662  * accuracy is primarily needed to know if copy-on-write faults can
2663  * reuse the page and change the mapping to read-write instead of
2664  * copying them. At the same time this returns the total_mapcount too.
2665  *
2666  * The function returns the highest mapcount any one of the subpages
2667  * has. If the return value is one, even if different processes are
2668  * mapping different subpages of the transparent hugepage, they can
2669  * all reuse it, because each process is reusing a different subpage.
2670  *
2671  * The total_mapcount is instead counting all virtual mappings of the
2672  * subpages. If the total_mapcount is equal to "one", it tells the
2673  * caller all mappings belong to the same "mm" and in turn the
2674  * anon_vma of the transparent hugepage can become the vma->anon_vma
2675  * local one as no other process may be mapping any of the subpages.
2676  *
2677  * It would be more accurate to replace page_mapcount() with
2678  * page_trans_huge_mapcount(), however we only use
2679  * page_trans_huge_mapcount() in the copy-on-write faults where we
2680  * need full accuracy to avoid breaking page pinning, because
2681  * page_trans_huge_mapcount() is slower than page_mapcount().
2682  */
2683 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2684 {
2685         int i, ret, _total_mapcount, mapcount;
2686
2687         /* hugetlbfs shouldn't call it */
2688         VM_BUG_ON_PAGE(PageHuge(page), page);
2689
2690         if (likely(!PageTransCompound(page))) {
2691                 mapcount = atomic_read(&page->_mapcount) + 1;
2692                 if (total_mapcount)
2693                         *total_mapcount = mapcount;
2694                 return mapcount;
2695         }
2696
2697         page = compound_head(page);
2698
2699         _total_mapcount = ret = 0;
2700         for (i = 0; i < HPAGE_PMD_NR; i++) {
2701                 mapcount = atomic_read(&page[i]._mapcount) + 1;
2702                 ret = max(ret, mapcount);
2703                 _total_mapcount += mapcount;
2704         }
2705         if (PageDoubleMap(page)) {
2706                 ret -= 1;
2707                 _total_mapcount -= HPAGE_PMD_NR;
2708         }
2709         mapcount = compound_mapcount(page);
2710         ret += mapcount;
2711         _total_mapcount += mapcount;
2712         if (total_mapcount)
2713                 *total_mapcount = _total_mapcount;
2714         return ret;
2715 }
2716
2717 /* Racy check whether the huge page can be split */
2718 bool can_split_huge_page(struct page *page, int *pextra_pins)
2719 {
2720         int extra_pins;
2721
2722         /* Additional pins from page cache */
2723         if (PageAnon(page))
2724                 extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0;
2725         else
2726                 extra_pins = HPAGE_PMD_NR;
2727         if (pextra_pins)
2728                 *pextra_pins = extra_pins;
2729         return total_mapcount(page) == page_count(page) - extra_pins - 1;
2730 }
2731
2732 /*
2733  * This function splits huge page into normal pages. @page can point to any
2734  * subpage of huge page to split. Split doesn't change the position of @page.
2735  *
2736  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2737  * The huge page must be locked.
2738  *
2739  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2740  *
2741  * Both head page and tail pages will inherit mapping, flags, and so on from
2742  * the hugepage.
2743  *
2744  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2745  * they are not mapped.
2746  *
2747  * Returns 0 if the hugepage is split successfully.
2748  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2749  * us.
2750  */
2751 int split_huge_page_to_list(struct page *page, struct list_head *list)
2752 {
2753         struct page *head = compound_head(page);
2754         struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2755         struct deferred_split *ds_queue = get_deferred_split_queue(head);
2756         struct anon_vma *anon_vma = NULL;
2757         struct address_space *mapping = NULL;
2758         int count, mapcount, extra_pins, ret;
2759         bool mlocked;
2760         unsigned long flags;
2761         pgoff_t end;
2762
2763         VM_BUG_ON_PAGE(is_huge_zero_page(head), head);
2764         VM_BUG_ON_PAGE(!PageLocked(head), head);
2765         VM_BUG_ON_PAGE(!PageCompound(head), head);
2766
2767         if (PageWriteback(head))
2768                 return -EBUSY;
2769
2770         if (PageAnon(head)) {
2771                 /*
2772                  * The caller does not necessarily hold an mmap_sem that would
2773                  * prevent the anon_vma disappearing so we first we take a
2774                  * reference to it and then lock the anon_vma for write. This
2775                  * is similar to page_lock_anon_vma_read except the write lock
2776                  * is taken to serialise against parallel split or collapse
2777                  * operations.
2778                  */
2779                 anon_vma = page_get_anon_vma(head);
2780                 if (!anon_vma) {
2781                         ret = -EBUSY;
2782                         goto out;
2783                 }
2784                 end = -1;
2785                 mapping = NULL;
2786                 anon_vma_lock_write(anon_vma);
2787         } else {
2788                 mapping = head->mapping;
2789
2790                 /* Truncated ? */
2791                 if (!mapping) {
2792                         ret = -EBUSY;
2793                         goto out;
2794                 }
2795
2796                 anon_vma = NULL;
2797                 i_mmap_lock_read(mapping);
2798
2799                 /*
2800                  *__split_huge_page() may need to trim off pages beyond EOF:
2801                  * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2802                  * which cannot be nested inside the page tree lock. So note
2803                  * end now: i_size itself may be changed at any moment, but
2804                  * head page lock is good enough to serialize the trimming.
2805                  */
2806                 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2807         }
2808
2809         /*
2810          * Racy check if we can split the page, before unmap_page() will
2811          * split PMDs
2812          */
2813         if (!can_split_huge_page(head, &extra_pins)) {
2814                 ret = -EBUSY;
2815                 goto out_unlock;
2816         }
2817
2818         mlocked = PageMlocked(head);
2819         unmap_page(head);
2820         VM_BUG_ON_PAGE(compound_mapcount(head), head);
2821
2822         /* Make sure the page is not on per-CPU pagevec as it takes pin */
2823         if (mlocked)
2824                 lru_add_drain();
2825
2826         /* prevent PageLRU to go away from under us, and freeze lru stats */
2827         spin_lock_irqsave(&pgdata->lru_lock, flags);
2828
2829         if (mapping) {
2830                 XA_STATE(xas, &mapping->i_pages, page_index(head));
2831
2832                 /*
2833                  * Check if the head page is present in page cache.
2834                  * We assume all tail are present too, if head is there.
2835                  */
2836                 xa_lock(&mapping->i_pages);
2837                 if (xas_load(&xas) != head)
2838                         goto fail;
2839         }
2840
2841         /* Prevent deferred_split_scan() touching ->_refcount */
2842         spin_lock(&ds_queue->split_queue_lock);
2843         count = page_count(head);
2844         mapcount = total_mapcount(head);
2845         if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2846                 if (!list_empty(page_deferred_list(head))) {
2847                         ds_queue->split_queue_len--;
2848                         list_del(page_deferred_list(head));
2849                 }
2850                 spin_unlock(&ds_queue->split_queue_lock);
2851                 if (mapping) {
2852                         if (PageSwapBacked(head))
2853                                 __dec_node_page_state(head, NR_SHMEM_THPS);
2854                         else
2855                                 __dec_node_page_state(head, NR_FILE_THPS);
2856                 }
2857
2858                 __split_huge_page(page, list, end, flags);
2859                 if (PageSwapCache(head)) {
2860                         swp_entry_t entry = { .val = page_private(head) };
2861
2862                         ret = split_swap_cluster(entry);
2863                 } else
2864                         ret = 0;
2865         } else {
2866                 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2867                         pr_alert("total_mapcount: %u, page_count(): %u\n",
2868                                         mapcount, count);
2869                         if (PageTail(page))
2870                                 dump_page(head, NULL);
2871                         dump_page(page, "total_mapcount(head) > 0");
2872                         BUG();
2873                 }
2874                 spin_unlock(&ds_queue->split_queue_lock);
2875 fail:           if (mapping)
2876                         xa_unlock(&mapping->i_pages);
2877                 spin_unlock_irqrestore(&pgdata->lru_lock, flags);
2878                 remap_page(head);
2879                 ret = -EBUSY;
2880         }
2881
2882 out_unlock:
2883         if (anon_vma) {
2884                 anon_vma_unlock_write(anon_vma);
2885                 put_anon_vma(anon_vma);
2886         }
2887         if (mapping)
2888                 i_mmap_unlock_read(mapping);
2889 out:
2890         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2891         return ret;
2892 }
2893
2894 void free_transhuge_page(struct page *page)
2895 {
2896         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2897         unsigned long flags;
2898
2899         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2900         if (!list_empty(page_deferred_list(page))) {
2901                 ds_queue->split_queue_len--;
2902                 list_del(page_deferred_list(page));
2903         }
2904         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2905         free_compound_page(page);
2906 }
2907
2908 void deferred_split_huge_page(struct page *page)
2909 {
2910         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2911 #ifdef CONFIG_MEMCG
2912         struct mem_cgroup *memcg = compound_head(page)->mem_cgroup;
2913 #endif
2914         unsigned long flags;
2915
2916         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2917
2918         /*
2919          * The try_to_unmap() in page reclaim path might reach here too,
2920          * this may cause a race condition to corrupt deferred split queue.
2921          * And, if page reclaim is already handling the same page, it is
2922          * unnecessary to handle it again in shrinker.
2923          *
2924          * Check PageSwapCache to determine if the page is being
2925          * handled by page reclaim since THP swap would add the page into
2926          * swap cache before calling try_to_unmap().
2927          */
2928         if (PageSwapCache(page))
2929                 return;
2930
2931         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2932         if (list_empty(page_deferred_list(page))) {
2933                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2934                 list_add_tail(page_deferred_list(page), &ds_queue->split_queue);
2935                 ds_queue->split_queue_len++;
2936 #ifdef CONFIG_MEMCG
2937                 if (memcg)
2938                         memcg_set_shrinker_bit(memcg, page_to_nid(page),
2939                                                deferred_split_shrinker.id);
2940 #endif
2941         }
2942         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2943 }
2944
2945 static unsigned long deferred_split_count(struct shrinker *shrink,
2946                 struct shrink_control *sc)
2947 {
2948         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2949         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2950
2951 #ifdef CONFIG_MEMCG
2952         if (sc->memcg)
2953                 ds_queue = &sc->memcg->deferred_split_queue;
2954 #endif
2955         return READ_ONCE(ds_queue->split_queue_len);
2956 }
2957
2958 static unsigned long deferred_split_scan(struct shrinker *shrink,
2959                 struct shrink_control *sc)
2960 {
2961         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2962         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2963         unsigned long flags;
2964         LIST_HEAD(list), *pos, *next;
2965         struct page *page;
2966         int split = 0;
2967
2968 #ifdef CONFIG_MEMCG
2969         if (sc->memcg)
2970                 ds_queue = &sc->memcg->deferred_split_queue;
2971 #endif
2972
2973         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2974         /* Take pin on all head pages to avoid freeing them under us */
2975         list_for_each_safe(pos, next, &ds_queue->split_queue) {
2976                 page = list_entry((void *)pos, struct page, mapping);
2977                 page = compound_head(page);
2978                 if (get_page_unless_zero(page)) {
2979                         list_move(page_deferred_list(page), &list);
2980                 } else {
2981                         /* We lost race with put_compound_page() */
2982                         list_del_init(page_deferred_list(page));
2983                         ds_queue->split_queue_len--;
2984                 }
2985                 if (!--sc->nr_to_scan)
2986                         break;
2987         }
2988         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2989
2990         list_for_each_safe(pos, next, &list) {
2991                 page = list_entry((void *)pos, struct page, mapping);
2992                 if (!trylock_page(page))
2993                         goto next;
2994                 /* split_huge_page() removes page from list on success */
2995                 if (!split_huge_page(page))
2996                         split++;
2997                 unlock_page(page);
2998 next:
2999                 put_page(page);
3000         }
3001
3002         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
3003         list_splice_tail(&list, &ds_queue->split_queue);
3004         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
3005
3006         /*
3007          * Stop shrinker if we didn't split any page, but the queue is empty.
3008          * This can happen if pages were freed under us.
3009          */
3010         if (!split && list_empty(&ds_queue->split_queue))
3011                 return SHRINK_STOP;
3012         return split;
3013 }
3014
3015 static struct shrinker deferred_split_shrinker = {
3016         .count_objects = deferred_split_count,
3017         .scan_objects = deferred_split_scan,
3018         .seeks = DEFAULT_SEEKS,
3019         .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
3020                  SHRINKER_NONSLAB,
3021 };
3022
3023 #ifdef CONFIG_DEBUG_FS
3024 static int split_huge_pages_set(void *data, u64 val)
3025 {
3026         struct zone *zone;
3027         struct page *page;
3028         unsigned long pfn, max_zone_pfn;
3029         unsigned long total = 0, split = 0;
3030
3031         if (val != 1)
3032                 return -EINVAL;
3033
3034         for_each_populated_zone(zone) {
3035                 max_zone_pfn = zone_end_pfn(zone);
3036                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
3037                         if (!pfn_valid(pfn))
3038                                 continue;
3039
3040                         page = pfn_to_page(pfn);
3041                         if (!get_page_unless_zero(page))
3042                                 continue;
3043
3044                         if (zone != page_zone(page))
3045                                 goto next;
3046
3047                         if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
3048                                 goto next;
3049
3050                         total++;
3051                         lock_page(page);
3052                         if (!split_huge_page(page))
3053                                 split++;
3054                         unlock_page(page);
3055 next:
3056                         put_page(page);
3057                 }
3058         }
3059
3060         pr_info("%lu of %lu THP split\n", split, total);
3061
3062         return 0;
3063 }
3064 DEFINE_DEBUGFS_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
3065                 "%llu\n");
3066
3067 static int __init split_huge_pages_debugfs(void)
3068 {
3069         debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3070                             &split_huge_pages_fops);
3071         return 0;
3072 }
3073 late_initcall(split_huge_pages_debugfs);
3074 #endif
3075
3076 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
3077 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
3078                 struct page *page)
3079 {
3080         struct vm_area_struct *vma = pvmw->vma;
3081         struct mm_struct *mm = vma->vm_mm;
3082         unsigned long address = pvmw->address;
3083         pmd_t pmdval;
3084         swp_entry_t entry;
3085         pmd_t pmdswp;
3086
3087         if (!(pvmw->pmd && !pvmw->pte))
3088                 return;
3089
3090         flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
3091         pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
3092         if (pmd_dirty(pmdval))
3093                 set_page_dirty(page);
3094         entry = make_migration_entry(page, pmd_write(pmdval));
3095         pmdswp = swp_entry_to_pmd(entry);
3096         if (pmd_soft_dirty(pmdval))
3097                 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
3098         set_pmd_at(mm, address, pvmw->pmd, pmdswp);
3099         page_remove_rmap(page, true);
3100         put_page(page);
3101 }
3102
3103 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
3104 {
3105         struct vm_area_struct *vma = pvmw->vma;
3106         struct mm_struct *mm = vma->vm_mm;
3107         unsigned long address = pvmw->address;
3108         unsigned long mmun_start = address & HPAGE_PMD_MASK;
3109         pmd_t pmde;
3110         swp_entry_t entry;
3111
3112         if (!(pvmw->pmd && !pvmw->pte))
3113                 return;
3114
3115         entry = pmd_to_swp_entry(*pvmw->pmd);
3116         get_page(new);
3117         pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
3118         if (pmd_swp_soft_dirty(*pvmw->pmd))
3119                 pmde = pmd_mksoft_dirty(pmde);
3120         if (is_write_migration_entry(entry))
3121                 pmde = maybe_pmd_mkwrite(pmde, vma);
3122
3123         flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
3124         if (PageAnon(new))
3125                 page_add_anon_rmap(new, vma, mmun_start, true);
3126         else
3127                 page_add_file_rmap(new, true);
3128         set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
3129         if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new))
3130                 mlock_vma_page(new);
3131         update_mmu_cache_pmd(vma, address, pvmw->pmd);
3132 }
3133 #endif