mm/huge_memory.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *  Copyright (C) 2009  Red Hat, Inc.
   4  */
   5
   6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   7
   8 #include <linux/mm.h>
   9 #include <linux/sched.h>
  10 #include <linux/sched/coredump.h>
  11 #include <linux/sched/numa_balancing.h>
  12 #include <linux/highmem.h>
  13 #include <linux/hugetlb.h>
  14 #include <linux/mmu_notifier.h>
  15 #include <linux/rmap.h>
  16 #include <linux/swap.h>
  17 #include <linux/shrinker.h>
  18 #include <linux/mm_inline.h>
  19 #include <linux/swapops.h>
  20 #include <linux/dax.h>
  21 #include <linux/khugepaged.h>
  22 #include <linux/freezer.h>
  23 #include <linux/pfn_t.h>
  24 #include <linux/mman.h>
  25 #include <linux/memremap.h>
  26 #include <linux/pagemap.h>
  27 #include <linux/debugfs.h>
  28 #include <linux/migrate.h>
  29 #include <linux/hashtable.h>
  30 #include <linux/userfaultfd_k.h>
  31 #include <linux/page_idle.h>
  32 #include <linux/shmem_fs.h>
  33 #include <linux/oom.h>
  34 #include <linux/numa.h>
  35 #include <linux/page_owner.h>
  36
  37 #include <asm/tlb.h>
  38 #include <asm/pgalloc.h>
  39 #include "internal.h"
  40
  41 /*
  42  * By default, transparent hugepage support is disabled in order to avoid
  43  * risking an increased memory footprint for applications that are not
  44  * guaranteed to benefit from it. When transparent hugepage support is
  45  * enabled, it is for all mappings, and khugepaged scans all mappings.
  46  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  47  * for all hugepage allocations.
  48  */
  49 unsigned long transparent_hugepage_flags __read_mostly =
  50 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  51         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  52 #endif
  53 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  54         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  55 #endif
  56         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
  57         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  58         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  59
  60 static struct shrinker deferred_split_shrinker;
  61
  62 static atomic_t huge_zero_refcount;
  63 struct page *huge_zero_page __read_mostly;
  64
  65 bool transparent_hugepage_enabled(struct vm_area_struct *vma)
  66 {
  67         /* The addr is used to check if the vma size fits */
  68         unsigned long addr = (vma->vm_end & HPAGE_PMD_MASK) - HPAGE_PMD_SIZE;
  69
  70         if (!transhuge_vma_suitable(vma, addr))
  71                 return false;
  72         if (vma_is_anonymous(vma))
  73                 return __transparent_hugepage_enabled(vma);
  74         if (vma_is_shmem(vma))
  75                 return shmem_huge_enabled(vma);
  76
  77         return false;
  78 }
  79
  80 static struct page *get_huge_zero_page(void)
  81 {
  82         struct page *zero_page;
  83 retry:
  84         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
  85                 return READ_ONCE(huge_zero_page);
  86
  87         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
  88                         HPAGE_PMD_ORDER);
  89         if (!zero_page) {
  90                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
  91                 return NULL;
  92         }
  93         count_vm_event(THP_ZERO_PAGE_ALLOC);
  94         preempt_disable();
  95         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
  96                 preempt_enable();
  97                 __free_pages(zero_page, compound_order(zero_page));
  98                 goto retry;
  99         }
 100
 101         /* We take additional reference here. It will be put back by shrinker */
 102         atomic_set(&huge_zero_refcount, 2);
 103         preempt_enable();
 104         return READ_ONCE(huge_zero_page);
 105 }
 106
 107 static void put_huge_zero_page(void)
 108 {
 109         /*
 110          * Counter should never go to zero here. Only shrinker can put
 111          * last reference.
 112          */
 113         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 114 }
 115
 116 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
 117 {
 118         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 119                 return READ_ONCE(huge_zero_page);
 120
 121         if (!get_huge_zero_page())
 122                 return NULL;
 123
 124         if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 125                 put_huge_zero_page();
 126
 127         return READ_ONCE(huge_zero_page);
 128 }
 129
 130 void mm_put_huge_zero_page(struct mm_struct *mm)
 131 {
 132         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 133                 put_huge_zero_page();
 134 }
 135
 136 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 137                                         struct shrink_control *sc)
 138 {
 139         /* we can free zero page only if last reference remains */
 140         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 141 }
 142
 143 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 144                                        struct shrink_control *sc)
 145 {
 146         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 147                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 148                 BUG_ON(zero_page == NULL);
 149                 __free_pages(zero_page, compound_order(zero_page));
 150                 return HPAGE_PMD_NR;
 151         }
 152
 153         return 0;
 154 }
 155
 156 static struct shrinker huge_zero_page_shrinker = {
 157         .count_objects = shrink_huge_zero_page_count,
 158         .scan_objects = shrink_huge_zero_page_scan,
 159         .seeks = DEFAULT_SEEKS,
 160 };
 161
 162 #ifdef CONFIG_SYSFS
 163 static ssize_t enabled_show(struct kobject *kobj,
 164                             struct kobj_attribute *attr, char *buf)
 165 {
 166         if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
 167                 return sprintf(buf, "[always] madvise never\n");
 168         else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
 169                 return sprintf(buf, "always [madvise] never\n");
 170         else
 171                 return sprintf(buf, "always madvise [never]\n");
 172 }
 173
 174 static ssize_t enabled_store(struct kobject *kobj,
 175                              struct kobj_attribute *attr,
 176                              const char *buf, size_t count)
 177 {
 178         ssize_t ret = count;
 179
 180         if (sysfs_streq(buf, "always")) {
 181                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 182                 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 183         } else if (sysfs_streq(buf, "madvise")) {
 184                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 185                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 186         } else if (sysfs_streq(buf, "never")) {
 187                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 188                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 189         } else
 190                 ret = -EINVAL;
 191
 192         if (ret > 0) {
 193                 int err = start_stop_khugepaged();
 194                 if (err)
 195                         ret = err;
 196         }
 197         return ret;
 198 }
 199 static struct kobj_attribute enabled_attr =
 200         __ATTR(enabled, 0644, enabled_show, enabled_store);
 201
 202 ssize_t single_hugepage_flag_show(struct kobject *kobj,
 203                                 struct kobj_attribute *attr, char *buf,
 204                                 enum transparent_hugepage_flag flag)
 205 {
 206         return sprintf(buf, "%d\n",
 207                        !!test_bit(flag, &transparent_hugepage_flags));
 208 }
 209
 210 ssize_t single_hugepage_flag_store(struct kobject *kobj,
 211                                  struct kobj_attribute *attr,
 212                                  const char *buf, size_t count,
 213                                  enum transparent_hugepage_flag flag)
 214 {
 215         unsigned long value;
 216         int ret;
 217
 218         ret = kstrtoul(buf, 10, &value);
 219         if (ret < 0)
 220                 return ret;
 221         if (value > 1)
 222                 return -EINVAL;
 223
 224         if (value)
 225                 set_bit(flag, &transparent_hugepage_flags);
 226         else
 227                 clear_bit(flag, &transparent_hugepage_flags);
 228
 229         return count;
 230 }
 231
 232 static ssize_t defrag_show(struct kobject *kobj,
 233                            struct kobj_attribute *attr, char *buf)
 234 {
 235         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 236                 return sprintf(buf, "[always] defer defer+madvise madvise never\n");
 237         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 238                 return sprintf(buf, "always [defer] defer+madvise madvise never\n");
 239         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 240                 return sprintf(buf, "always defer [defer+madvise] madvise never\n");
 241         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 242                 return sprintf(buf, "always defer defer+madvise [madvise] never\n");
 243         return sprintf(buf, "always defer defer+madvise madvise [never]\n");
 244 }
 245
 246 static ssize_t defrag_store(struct kobject *kobj,
 247                             struct kobj_attribute *attr,
 248                             const char *buf, size_t count)
 249 {
 250         if (sysfs_streq(buf, "always")) {
 251                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 252                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 253                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 254                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 255         } else if (sysfs_streq(buf, "defer+madvise")) {
 256                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 257                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 258                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 259                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 260         } else if (sysfs_streq(buf, "defer")) {
 261                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 262                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 263                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 264                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 265         } else if (sysfs_streq(buf, "madvise")) {
 266                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 267                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 268                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 269                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 270         } else if (sysfs_streq(buf, "never")) {
 271                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 272                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 273                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 274                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 275         } else
 276                 return -EINVAL;
 277
 278         return count;
 279 }
 280 static struct kobj_attribute defrag_attr =
 281         __ATTR(defrag, 0644, defrag_show, defrag_store);
 282
 283 static ssize_t use_zero_page_show(struct kobject *kobj,
 284                 struct kobj_attribute *attr, char *buf)
 285 {
 286         return single_hugepage_flag_show(kobj, attr, buf,
 287                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 288 }
 289 static ssize_t use_zero_page_store(struct kobject *kobj,
 290                 struct kobj_attribute *attr, const char *buf, size_t count)
 291 {
 292         return single_hugepage_flag_store(kobj, attr, buf, count,
 293                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 294 }
 295 static struct kobj_attribute use_zero_page_attr =
 296         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 297
 298 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
 299                 struct kobj_attribute *attr, char *buf)
 300 {
 301         return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE);
 302 }
 303 static struct kobj_attribute hpage_pmd_size_attr =
 304         __ATTR_RO(hpage_pmd_size);
 305
 306 #ifdef CONFIG_DEBUG_VM
 307 static ssize_t debug_cow_show(struct kobject *kobj,
 308                                 struct kobj_attribute *attr, char *buf)
 309 {
 310         return single_hugepage_flag_show(kobj, attr, buf,
 311                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 312 }
 313 static ssize_t debug_cow_store(struct kobject *kobj,
 314                                struct kobj_attribute *attr,
 315                                const char *buf, size_t count)
 316 {
 317         return single_hugepage_flag_store(kobj, attr, buf, count,
 318                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 319 }
 320 static struct kobj_attribute debug_cow_attr =
 321         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 322 #endif /* CONFIG_DEBUG_VM */
 323
 324 static struct attribute *hugepage_attr[] = {
 325         &enabled_attr.attr,
 326         &defrag_attr.attr,
 327         &use_zero_page_attr.attr,
 328         &hpage_pmd_size_attr.attr,
 329 #ifdef CONFIG_SHMEM
 330         &shmem_enabled_attr.attr,
 331 #endif
 332 #ifdef CONFIG_DEBUG_VM
 333         &debug_cow_attr.attr,
 334 #endif
 335         NULL,
 336 };
 337
 338 static const struct attribute_group hugepage_attr_group = {
 339         .attrs = hugepage_attr,
 340 };
 341
 342 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 343 {
 344         int err;
 345
 346         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 347         if (unlikely(!*hugepage_kobj)) {
 348                 pr_err("failed to create transparent hugepage kobject\n");
 349                 return -ENOMEM;
 350         }
 351
 352         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 353         if (err) {
 354                 pr_err("failed to register transparent hugepage group\n");
 355                 goto delete_obj;
 356         }
 357
 358         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 359         if (err) {
 360                 pr_err("failed to register transparent hugepage group\n");
 361                 goto remove_hp_group;
 362         }
 363
 364         return 0;
 365
 366 remove_hp_group:
 367         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 368 delete_obj:
 369         kobject_put(*hugepage_kobj);
 370         return err;
 371 }
 372
 373 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 374 {
 375         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 376         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 377         kobject_put(hugepage_kobj);
 378 }
 379 #else
 380 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 381 {
 382         return 0;
 383 }
 384
 385 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 386 {
 387 }
 388 #endif /* CONFIG_SYSFS */
 389
 390 static int __init hugepage_init(void)
 391 {
 392         int err;
 393         struct kobject *hugepage_kobj;
 394
 395         if (!has_transparent_hugepage()) {
 396                 transparent_hugepage_flags = 0;
 397                 return -EINVAL;
 398         }
 399
 400         /*
 401          * hugepages can't be allocated by the buddy allocator
 402          */
 403         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
 404         /*
 405          * we use page->mapping and page->index in second tail page
 406          * as list_head: assuming THP order >= 2
 407          */
 408         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 409
 410         err = hugepage_init_sysfs(&hugepage_kobj);
 411         if (err)
 412                 goto err_sysfs;
 413
 414         err = khugepaged_init();
 415         if (err)
 416                 goto err_slab;
 417
 418         err = register_shrinker(&huge_zero_page_shrinker);
 419         if (err)
 420                 goto err_hzp_shrinker;
 421         err = register_shrinker(&deferred_split_shrinker);
 422         if (err)
 423                 goto err_split_shrinker;
 424
 425         /*
 426          * By default disable transparent hugepages on smaller systems,
 427          * where the extra memory used could hurt more than TLB overhead
 428          * is likely to save.  The admin can still enable it through /sys.
 429          */
 430         if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
 431                 transparent_hugepage_flags = 0;
 432                 return 0;
 433         }
 434
 435         err = start_stop_khugepaged();
 436         if (err)
 437                 goto err_khugepaged;
 438
 439         return 0;
 440 err_khugepaged:
 441         unregister_shrinker(&deferred_split_shrinker);
 442 err_split_shrinker:
 443         unregister_shrinker(&huge_zero_page_shrinker);
 444 err_hzp_shrinker:
 445         khugepaged_destroy();
 446 err_slab:
 447         hugepage_exit_sysfs(hugepage_kobj);
 448 err_sysfs:
 449         return err;
 450 }
 451 subsys_initcall(hugepage_init);
 452
 453 static int __init setup_transparent_hugepage(char *str)
 454 {
 455         int ret = 0;
 456         if (!str)
 457                 goto out;
 458         if (!strcmp(str, "always")) {
 459                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 460                         &transparent_hugepage_flags);
 461                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 462                           &transparent_hugepage_flags);
 463                 ret = 1;
 464         } else if (!strcmp(str, "madvise")) {
 465                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 466                           &transparent_hugepage_flags);
 467                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 468                         &transparent_hugepage_flags);
 469                 ret = 1;
 470         } else if (!strcmp(str, "never")) {
 471                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 472                           &transparent_hugepage_flags);
 473                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 474                           &transparent_hugepage_flags);
 475                 ret = 1;
 476         }
 477 out:
 478         if (!ret)
 479                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 480         return ret;
 481 }
 482 __setup("transparent_hugepage=", setup_transparent_hugepage);
 483
 484 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 485 {
 486         if (likely(vma->vm_flags & VM_WRITE))
 487                 pmd = pmd_mkwrite(pmd);
 488         return pmd;
 489 }
 490
 491 #ifdef CONFIG_MEMCG
 492 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
 493 {
 494         struct mem_cgroup *memcg = compound_head(page)->mem_cgroup;
 495         struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
 496
 497         if (memcg)
 498                 return &memcg->deferred_split_queue;
 499         else
 500                 return &pgdat->deferred_split_queue;
 501 }
 502 #else
 503 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
 504 {
 505         struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
 506
 507         return &pgdat->deferred_split_queue;
 508 }
 509 #endif
 510
 511 void prep_transhuge_page(struct page *page)
 512 {
 513         /*
 514          * we use page->mapping and page->indexlru in second tail page
 515          * as list_head: assuming THP order >= 2
 516          */
 517
 518         INIT_LIST_HEAD(page_deferred_list(page));
 519         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 520 }
 521
 522 bool is_transparent_hugepage(struct page *page)
 523 {
 524         if (!PageCompound(page))
 525                 return 0;
 526
 527         page = compound_head(page);
 528         return is_huge_zero_page(page) ||
 529                page[1].compound_dtor == TRANSHUGE_PAGE_DTOR;
 530 }
 531 EXPORT_SYMBOL_GPL(is_transparent_hugepage);
 532
 533 static unsigned long __thp_get_unmapped_area(struct file *filp,
 534                 unsigned long addr, unsigned long len,
 535                 loff_t off, unsigned long flags, unsigned long size)
 536 {
 537         loff_t off_end = off + len;
 538         loff_t off_align = round_up(off, size);
 539         unsigned long len_pad, ret;
 540
 541         if (off_end <= off_align || (off_end - off_align) < size)
 542                 return 0;
 543
 544         len_pad = len + size;
 545         if (len_pad < len || (off + len_pad) < off)
 546                 return 0;
 547
 548         ret = current->mm->get_unmapped_area(filp, addr, len_pad,
 549                                               off >> PAGE_SHIFT, flags);
 550
 551         /*
 552          * The failure might be due to length padding. The caller will retry
 553          * without the padding.
 554          */
 555         if (IS_ERR_VALUE(ret))
 556                 return 0;
 557
 558         /*
 559          * Do not try to align to THP boundary if allocation at the address
 560          * hint succeeds.
 561          */
 562         if (ret == addr)
 563                 return addr;
 564
 565         ret += (off - ret) & (size - 1);
 566         return ret;
 567 }
 568
 569 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 570                 unsigned long len, unsigned long pgoff, unsigned long flags)
 571 {
 572         unsigned long ret;
 573         loff_t off = (loff_t)pgoff << PAGE_SHIFT;
 574
 575         if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
 576                 goto out;
 577
 578         ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE);
 579         if (ret)
 580                 return ret;
 581 out:
 582         return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
 583 }
 584 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
 585
 586 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
 587                         struct page *page, gfp_t gfp)
 588 {
 589         struct vm_area_struct *vma = vmf->vma;
 590         struct mem_cgroup *memcg;
 591         pgtable_t pgtable;
 592         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 593         vm_fault_t ret = 0;
 594
 595         VM_BUG_ON_PAGE(!PageCompound(page), page);
 596
 597         if (mem_cgroup_try_charge_delay(page, vma->vm_mm, gfp, &memcg, 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         /*
1048          * Make sure the _PAGE_UFFD_WP bit is cleared if the new VMA
1049          * does not have the VM_UFFD_WP, which means that the uffd
1050          * fork event is not enabled.
1051          */
1052         if (!(vma->vm_flags & VM_UFFD_WP))
1053                 pmd = pmd_clear_uffd_wp(pmd);
1054
1055 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1056         if (unlikely(is_swap_pmd(pmd))) {
1057                 swp_entry_t entry = pmd_to_swp_entry(pmd);
1058
1059                 VM_BUG_ON(!is_pmd_migration_entry(pmd));
1060                 if (is_write_migration_entry(entry)) {
1061                         make_migration_entry_read(&entry);
1062                         pmd = swp_entry_to_pmd(entry);
1063                         if (pmd_swp_soft_dirty(*src_pmd))
1064                                 pmd = pmd_swp_mksoft_dirty(pmd);
1065                         set_pmd_at(src_mm, addr, src_pmd, pmd);
1066                 }
1067                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1068                 mm_inc_nr_ptes(dst_mm);
1069                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1070                 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1071                 ret = 0;
1072                 goto out_unlock;
1073         }
1074 #endif
1075
1076         if (unlikely(!pmd_trans_huge(pmd))) {
1077                 pte_free(dst_mm, pgtable);
1078                 goto out_unlock;
1079         }
1080         /*
1081          * When page table lock is held, the huge zero pmd should not be
1082          * under splitting since we don't split the page itself, only pmd to
1083          * a page table.
1084          */
1085         if (is_huge_zero_pmd(pmd)) {
1086                 struct page *zero_page;
1087                 /*
1088                  * get_huge_zero_page() will never allocate a new page here,
1089                  * since we already have a zero page to copy. It just takes a
1090                  * reference.
1091                  */
1092                 zero_page = mm_get_huge_zero_page(dst_mm);
1093                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1094                                 zero_page);
1095                 ret = 0;
1096                 goto out_unlock;
1097         }
1098
1099         src_page = pmd_page(pmd);
1100         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1101         get_page(src_page);
1102         page_dup_rmap(src_page, true);
1103         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1104         mm_inc_nr_ptes(dst_mm);
1105         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1106
1107         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1108         pmd = pmd_mkold(pmd_wrprotect(pmd));
1109         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1110
1111         ret = 0;
1112 out_unlock:
1113         spin_unlock(src_ptl);
1114         spin_unlock(dst_ptl);
1115 out:
1116         return ret;
1117 }
1118
1119 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1120 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1121                 pud_t *pud, int flags)
1122 {
1123         pud_t _pud;
1124
1125         _pud = pud_mkyoung(*pud);
1126         if (flags & FOLL_WRITE)
1127                 _pud = pud_mkdirty(_pud);
1128         if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1129                                 pud, _pud, flags & FOLL_WRITE))
1130                 update_mmu_cache_pud(vma, addr, pud);
1131 }
1132
1133 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1134                 pud_t *pud, int flags, struct dev_pagemap **pgmap)
1135 {
1136         unsigned long pfn = pud_pfn(*pud);
1137         struct mm_struct *mm = vma->vm_mm;
1138         struct page *page;
1139
1140         assert_spin_locked(pud_lockptr(mm, pud));
1141
1142         if (flags & FOLL_WRITE && !pud_write(*pud))
1143                 return NULL;
1144
1145         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
1146         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
1147                          (FOLL_PIN | FOLL_GET)))
1148                 return NULL;
1149
1150         if (pud_present(*pud) && pud_devmap(*pud))
1151                 /* pass */;
1152         else
1153                 return NULL;
1154
1155         if (flags & FOLL_TOUCH)
1156                 touch_pud(vma, addr, pud, flags);
1157
1158         /*
1159          * device mapped pages can only be returned if the
1160          * caller will manage the page reference count.
1161          *
1162          * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here:
1163          */
1164         if (!(flags & (FOLL_GET | FOLL_PIN)))
1165                 return ERR_PTR(-EEXIST);
1166
1167         pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1168         *pgmap = get_dev_pagemap(pfn, *pgmap);
1169         if (!*pgmap)
1170                 return ERR_PTR(-EFAULT);
1171         page = pfn_to_page(pfn);
1172         if (!try_grab_page(page, flags))
1173                 page = ERR_PTR(-ENOMEM);
1174
1175         return page;
1176 }
1177
1178 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1179                   pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1180                   struct vm_area_struct *vma)
1181 {
1182         spinlock_t *dst_ptl, *src_ptl;
1183         pud_t pud;
1184         int ret;
1185
1186         dst_ptl = pud_lock(dst_mm, dst_pud);
1187         src_ptl = pud_lockptr(src_mm, src_pud);
1188         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1189
1190         ret = -EAGAIN;
1191         pud = *src_pud;
1192         if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1193                 goto out_unlock;
1194
1195         /*
1196          * When page table lock is held, the huge zero pud should not be
1197          * under splitting since we don't split the page itself, only pud to
1198          * a page table.
1199          */
1200         if (is_huge_zero_pud(pud)) {
1201                 /* No huge zero pud yet */
1202         }
1203
1204         pudp_set_wrprotect(src_mm, addr, src_pud);
1205         pud = pud_mkold(pud_wrprotect(pud));
1206         set_pud_at(dst_mm, addr, dst_pud, pud);
1207
1208         ret = 0;
1209 out_unlock:
1210         spin_unlock(src_ptl);
1211         spin_unlock(dst_ptl);
1212         return ret;
1213 }
1214
1215 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1216 {
1217         pud_t entry;
1218         unsigned long haddr;
1219         bool write = vmf->flags & FAULT_FLAG_WRITE;
1220
1221         vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1222         if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1223                 goto unlock;
1224
1225         entry = pud_mkyoung(orig_pud);
1226         if (write)
1227                 entry = pud_mkdirty(entry);
1228         haddr = vmf->address & HPAGE_PUD_MASK;
1229         if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1230                 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1231
1232 unlock:
1233         spin_unlock(vmf->ptl);
1234 }
1235 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1236
1237 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1238 {
1239         pmd_t entry;
1240         unsigned long haddr;
1241         bool write = vmf->flags & FAULT_FLAG_WRITE;
1242
1243         vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1244         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1245                 goto unlock;
1246
1247         entry = pmd_mkyoung(orig_pmd);
1248         if (write)
1249                 entry = pmd_mkdirty(entry);
1250         haddr = vmf->address & HPAGE_PMD_MASK;
1251         if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1252                 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1253
1254 unlock:
1255         spin_unlock(vmf->ptl);
1256 }
1257
1258 static vm_fault_t do_huge_pmd_wp_page_fallback(struct vm_fault *vmf,
1259                         pmd_t orig_pmd, struct page *page)
1260 {
1261         struct vm_area_struct *vma = vmf->vma;
1262         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1263         struct mem_cgroup *memcg;
1264         pgtable_t pgtable;
1265         pmd_t _pmd;
1266         int i;
1267         vm_fault_t ret = 0;
1268         struct page **pages;
1269         struct mmu_notifier_range range;
1270
1271         pages = kmalloc_array(HPAGE_PMD_NR, sizeof(struct page *),
1272                               GFP_KERNEL);
1273         if (unlikely(!pages)) {
1274                 ret |= VM_FAULT_OOM;
1275                 goto out;
1276         }
1277
1278         for (i = 0; i < HPAGE_PMD_NR; i++) {
1279                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
1280                                                vmf->address, page_to_nid(page));
1281                 if (unlikely(!pages[i] ||
1282                              mem_cgroup_try_charge_delay(pages[i], vma->vm_mm,
1283                                      GFP_KERNEL, &memcg, false))) {
1284                         if (pages[i])
1285                                 put_page(pages[i]);
1286                         while (--i >= 0) {
1287                                 memcg = (void *)page_private(pages[i]);
1288                                 set_page_private(pages[i], 0);
1289                                 mem_cgroup_cancel_charge(pages[i], memcg,
1290                                                 false);
1291                                 put_page(pages[i]);
1292                         }
1293                         kfree(pages);
1294                         ret |= VM_FAULT_OOM;
1295                         goto out;
1296                 }
1297                 set_page_private(pages[i], (unsigned long)memcg);
1298         }
1299
1300         for (i = 0; i < HPAGE_PMD_NR; i++) {
1301                 copy_user_highpage(pages[i], page + i,
1302                                    haddr + PAGE_SIZE * i, vma);
1303                 __SetPageUptodate(pages[i]);
1304                 cond_resched();
1305         }
1306
1307         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1308                                 haddr, haddr + HPAGE_PMD_SIZE);
1309         mmu_notifier_invalidate_range_start(&range);
1310
1311         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1312         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1313                 goto out_free_pages;
1314         VM_BUG_ON_PAGE(!PageHead(page), page);
1315
1316         /*
1317          * Leave pmd empty until pte is filled note we must notify here as
1318          * concurrent CPU thread might write to new page before the call to
1319          * mmu_notifier_invalidate_range_end() happens which can lead to a
1320          * device seeing memory write in different order than CPU.
1321          *
1322          * See Documentation/vm/mmu_notifier.rst
1323          */
1324         pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1325
1326         pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
1327         pmd_populate(vma->vm_mm, &_pmd, pgtable);
1328
1329         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1330                 pte_t entry;
1331                 entry = mk_pte(pages[i], vma->vm_page_prot);
1332                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1333                 memcg = (void *)page_private(pages[i]);
1334                 set_page_private(pages[i], 0);
1335                 page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
1336                 mem_cgroup_commit_charge(pages[i], memcg, false, false);
1337                 lru_cache_add_active_or_unevictable(pages[i], vma);
1338                 vmf->pte = pte_offset_map(&_pmd, haddr);
1339                 VM_BUG_ON(!pte_none(*vmf->pte));
1340                 set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
1341                 pte_unmap(vmf->pte);
1342         }
1343         kfree(pages);
1344
1345         smp_wmb(); /* make pte visible before pmd */
1346         pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
1347         page_remove_rmap(page, true);
1348         spin_unlock(vmf->ptl);
1349
1350         /*
1351          * No need to double call mmu_notifier->invalidate_range() callback as
1352          * the above pmdp_huge_clear_flush_notify() did already call it.
1353          */
1354         mmu_notifier_invalidate_range_only_end(&range);
1355
1356         ret |= VM_FAULT_WRITE;
1357         put_page(page);
1358
1359 out:
1360         return ret;
1361
1362 out_free_pages:
1363         spin_unlock(vmf->ptl);
1364         mmu_notifier_invalidate_range_end(&range);
1365         for (i = 0; i < HPAGE_PMD_NR; i++) {
1366                 memcg = (void *)page_private(pages[i]);
1367                 set_page_private(pages[i], 0);
1368                 mem_cgroup_cancel_charge(pages[i], memcg, false);
1369                 put_page(pages[i]);
1370         }
1371         kfree(pages);
1372         goto out;
1373 }
1374
1375 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1376 {
1377         struct vm_area_struct *vma = vmf->vma;
1378         struct page *page = NULL, *new_page;
1379         struct mem_cgroup *memcg;
1380         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1381         struct mmu_notifier_range range;
1382         gfp_t huge_gfp;                 /* for allocation and charge */
1383         vm_fault_t ret = 0;
1384
1385         vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1386         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1387         if (is_huge_zero_pmd(orig_pmd))
1388                 goto alloc;
1389         spin_lock(vmf->ptl);
1390         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1391                 goto out_unlock;
1392
1393         page = pmd_page(orig_pmd);
1394         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1395         /*
1396          * We can only reuse the page if nobody else maps the huge page or it's
1397          * part.
1398          */
1399         if (!trylock_page(page)) {
1400                 get_page(page);
1401                 spin_unlock(vmf->ptl);
1402                 lock_page(page);
1403                 spin_lock(vmf->ptl);
1404                 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1405                         unlock_page(page);
1406                         put_page(page);
1407                         goto out_unlock;
1408                 }
1409                 put_page(page);
1410         }
1411         if (reuse_swap_page(page, NULL)) {
1412                 pmd_t entry;
1413                 entry = pmd_mkyoung(orig_pmd);
1414                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1415                 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry,  1))
1416                         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1417                 ret |= VM_FAULT_WRITE;
1418                 unlock_page(page);
1419                 goto out_unlock;
1420         }
1421         unlock_page(page);
1422         get_page(page);
1423         spin_unlock(vmf->ptl);
1424 alloc:
1425         if (__transparent_hugepage_enabled(vma) &&
1426             !transparent_hugepage_debug_cow()) {
1427                 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1428                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1429         } else
1430                 new_page = NULL;
1431
1432         if (likely(new_page)) {
1433                 prep_transhuge_page(new_page);
1434         } else {
1435                 if (!page) {
1436                         split_huge_pmd(vma, vmf->pmd, vmf->address);
1437                         ret |= VM_FAULT_FALLBACK;
1438                 } else {
1439                         ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
1440                         if (ret & VM_FAULT_OOM) {
1441                                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1442                                 ret |= VM_FAULT_FALLBACK;
1443                         }
1444                         put_page(page);
1445                 }
1446                 count_vm_event(THP_FAULT_FALLBACK);
1447                 goto out;
1448         }
1449
1450         if (unlikely(mem_cgroup_try_charge_delay(new_page, vma->vm_mm,
1451                                         huge_gfp, &memcg, true))) {
1452                 put_page(new_page);
1453                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1454                 if (page)
1455                         put_page(page);
1456                 ret |= VM_FAULT_FALLBACK;
1457                 count_vm_event(THP_FAULT_FALLBACK);
1458                 count_vm_event(THP_FAULT_FALLBACK_CHARGE);
1459                 goto out;
1460         }
1461
1462         count_vm_event(THP_FAULT_ALLOC);
1463         count_memcg_events(memcg, THP_FAULT_ALLOC, 1);
1464
1465         if (!page)
1466                 clear_huge_page(new_page, vmf->address, HPAGE_PMD_NR);
1467         else
1468                 copy_user_huge_page(new_page, page, vmf->address,
1469                                     vma, HPAGE_PMD_NR);
1470         __SetPageUptodate(new_page);
1471
1472         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1473                                 haddr, haddr + HPAGE_PMD_SIZE);
1474         mmu_notifier_invalidate_range_start(&range);
1475
1476         spin_lock(vmf->ptl);
1477         if (page)
1478                 put_page(page);
1479         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1480                 spin_unlock(vmf->ptl);
1481                 mem_cgroup_cancel_charge(new_page, memcg, true);
1482                 put_page(new_page);
1483                 goto out_mn;
1484         } else {
1485                 pmd_t entry;
1486                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1487                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1488                 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1489                 page_add_new_anon_rmap(new_page, vma, haddr, true);
1490                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1491                 lru_cache_add_active_or_unevictable(new_page, vma);
1492                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1493                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1494                 if (!page) {
1495                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1496                 } else {
1497                         VM_BUG_ON_PAGE(!PageHead(page), page);
1498                         page_remove_rmap(page, true);
1499                         put_page(page);
1500                 }
1501                 ret |= VM_FAULT_WRITE;
1502         }
1503         spin_unlock(vmf->ptl);
1504 out_mn:
1505         /*
1506          * No need to double call mmu_notifier->invalidate_range() callback as
1507          * the above pmdp_huge_clear_flush_notify() did already call it.
1508          */
1509         mmu_notifier_invalidate_range_only_end(&range);
1510 out:
1511         return ret;
1512 out_unlock:
1513         spin_unlock(vmf->ptl);
1514         return ret;
1515 }
1516
1517 /*
1518  * FOLL_FORCE or a forced COW break can write even to unwritable pmd's,
1519  * but only after we've gone through a COW cycle and they are dirty.
1520  */
1521 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1522 {
1523         return pmd_write(pmd) || ((flags & FOLL_COW) && pmd_dirty(pmd));
1524 }
1525
1526 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1527                                    unsigned long addr,
1528                                    pmd_t *pmd,
1529                                    unsigned int flags)
1530 {
1531         struct mm_struct *mm = vma->vm_mm;
1532         struct page *page = NULL;
1533
1534         assert_spin_locked(pmd_lockptr(mm, pmd));
1535
1536         if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1537                 goto out;
1538
1539         /* Avoid dumping huge zero page */
1540         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1541                 return ERR_PTR(-EFAULT);
1542
1543         /* Full NUMA hinting faults to serialise migration in fault paths */
1544         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1545                 goto out;
1546
1547         page = pmd_page(*pmd);
1548         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1549
1550         if (!try_grab_page(page, flags))
1551                 return ERR_PTR(-ENOMEM);
1552
1553         if (flags & FOLL_TOUCH)
1554                 touch_pmd(vma, addr, pmd, flags);
1555
1556         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1557                 /*
1558                  * We don't mlock() pte-mapped THPs. This way we can avoid
1559                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1560                  *
1561                  * For anon THP:
1562                  *
1563                  * In most cases the pmd is the only mapping of the page as we
1564                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1565                  * writable private mappings in populate_vma_page_range().
1566                  *
1567                  * The only scenario when we have the page shared here is if we
1568                  * mlocking read-only mapping shared over fork(). We skip
1569                  * mlocking such pages.
1570                  *
1571                  * For file THP:
1572                  *
1573                  * We can expect PageDoubleMap() to be stable under page lock:
1574                  * for file pages we set it in page_add_file_rmap(), which
1575                  * requires page to be locked.
1576                  */
1577
1578                 if (PageAnon(page) && compound_mapcount(page) != 1)
1579                         goto skip_mlock;
1580                 if (PageDoubleMap(page) || !page->mapping)
1581                         goto skip_mlock;
1582                 if (!trylock_page(page))
1583                         goto skip_mlock;
1584                 lru_add_drain();
1585                 if (page->mapping && !PageDoubleMap(page))
1586                         mlock_vma_page(page);
1587                 unlock_page(page);
1588         }
1589 skip_mlock:
1590         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1591         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1592
1593 out:
1594         return page;
1595 }
1596
1597 /* NUMA hinting page fault entry point for trans huge pmds */
1598 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1599 {
1600         struct vm_area_struct *vma = vmf->vma;
1601         struct anon_vma *anon_vma = NULL;
1602         struct page *page;
1603         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1604         int page_nid = NUMA_NO_NODE, this_nid = numa_node_id();
1605         int target_nid, last_cpupid = -1;
1606         bool page_locked;
1607         bool migrated = false;
1608         bool was_writable;
1609         int flags = 0;
1610
1611         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1612         if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1613                 goto out_unlock;
1614
1615         /*
1616          * If there are potential migrations, wait for completion and retry
1617          * without disrupting NUMA hinting information. Do not relock and
1618          * check_same as the page may no longer be mapped.
1619          */
1620         if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1621                 page = pmd_page(*vmf->pmd);
1622                 if (!get_page_unless_zero(page))
1623                         goto out_unlock;
1624                 spin_unlock(vmf->ptl);
1625                 put_and_wait_on_page_locked(page);
1626                 goto out;
1627         }
1628
1629         page = pmd_page(pmd);
1630         BUG_ON(is_huge_zero_page(page));
1631         page_nid = page_to_nid(page);
1632         last_cpupid = page_cpupid_last(page);
1633         count_vm_numa_event(NUMA_HINT_FAULTS);
1634         if (page_nid == this_nid) {
1635                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1636                 flags |= TNF_FAULT_LOCAL;
1637         }
1638
1639         /* See similar comment in do_numa_page for explanation */
1640         if (!pmd_savedwrite(pmd))
1641                 flags |= TNF_NO_GROUP;
1642
1643         /*
1644          * Acquire the page lock to serialise THP migrations but avoid dropping
1645          * page_table_lock if at all possible
1646          */
1647         page_locked = trylock_page(page);
1648         target_nid = mpol_misplaced(page, vma, haddr);
1649         if (target_nid == NUMA_NO_NODE) {
1650                 /* If the page was locked, there are no parallel migrations */
1651                 if (page_locked)
1652                         goto clear_pmdnuma;
1653         }
1654
1655         /* Migration could have started since the pmd_trans_migrating check */
1656         if (!page_locked) {
1657                 page_nid = NUMA_NO_NODE;
1658                 if (!get_page_unless_zero(page))
1659                         goto out_unlock;
1660                 spin_unlock(vmf->ptl);
1661                 put_and_wait_on_page_locked(page);
1662                 goto out;
1663         }
1664
1665         /*
1666          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1667          * to serialises splits
1668          */
1669         get_page(page);
1670         spin_unlock(vmf->ptl);
1671         anon_vma = page_lock_anon_vma_read(page);
1672
1673         /* Confirm the PMD did not change while page_table_lock was released */
1674         spin_lock(vmf->ptl);
1675         if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1676                 unlock_page(page);
1677                 put_page(page);
1678                 page_nid = NUMA_NO_NODE;
1679                 goto out_unlock;
1680         }
1681
1682         /* Bail if we fail to protect against THP splits for any reason */
1683         if (unlikely(!anon_vma)) {
1684                 put_page(page);
1685                 page_nid = NUMA_NO_NODE;
1686                 goto clear_pmdnuma;
1687         }
1688
1689         /*
1690          * Since we took the NUMA fault, we must have observed the !accessible
1691          * bit. Make sure all other CPUs agree with that, to avoid them
1692          * modifying the page we're about to migrate.
1693          *
1694          * Must be done under PTL such that we'll observe the relevant
1695          * inc_tlb_flush_pending().
1696          *
1697          * We are not sure a pending tlb flush here is for a huge page
1698          * mapping or not. Hence use the tlb range variant
1699          */
1700         if (mm_tlb_flush_pending(vma->vm_mm)) {
1701                 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1702                 /*
1703                  * change_huge_pmd() released the pmd lock before
1704                  * invalidating the secondary MMUs sharing the primary
1705                  * MMU pagetables (with ->invalidate_range()). The
1706                  * mmu_notifier_invalidate_range_end() (which
1707                  * internally calls ->invalidate_range()) in
1708                  * change_pmd_range() will run after us, so we can't
1709                  * rely on it here and we need an explicit invalidate.
1710                  */
1711                 mmu_notifier_invalidate_range(vma->vm_mm, haddr,
1712                                               haddr + HPAGE_PMD_SIZE);
1713         }
1714
1715         /*
1716          * Migrate the THP to the requested node, returns with page unlocked
1717          * and access rights restored.
1718          */
1719         spin_unlock(vmf->ptl);
1720
1721         migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1722                                 vmf->pmd, pmd, vmf->address, page, target_nid);
1723         if (migrated) {
1724                 flags |= TNF_MIGRATED;
1725                 page_nid = target_nid;
1726         } else
1727                 flags |= TNF_MIGRATE_FAIL;
1728
1729         goto out;
1730 clear_pmdnuma:
1731         BUG_ON(!PageLocked(page));
1732         was_writable = pmd_savedwrite(pmd);
1733         pmd = pmd_modify(pmd, vma->vm_page_prot);
1734         pmd = pmd_mkyoung(pmd);
1735         if (was_writable)
1736                 pmd = pmd_mkwrite(pmd);
1737         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1738         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1739         unlock_page(page);
1740 out_unlock:
1741         spin_unlock(vmf->ptl);
1742
1743 out:
1744         if (anon_vma)
1745                 page_unlock_anon_vma_read(anon_vma);
1746
1747         if (page_nid != NUMA_NO_NODE)
1748                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1749                                 flags);
1750
1751         return 0;
1752 }
1753
1754 /*
1755  * Return true if we do MADV_FREE successfully on entire pmd page.
1756  * Otherwise, return false.
1757  */
1758 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1759                 pmd_t *pmd, unsigned long addr, unsigned long next)
1760 {
1761         spinlock_t *ptl;
1762         pmd_t orig_pmd;
1763         struct page *page;
1764         struct mm_struct *mm = tlb->mm;
1765         bool ret = false;
1766
1767         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1768
1769         ptl = pmd_trans_huge_lock(pmd, vma);
1770         if (!ptl)
1771                 goto out_unlocked;
1772
1773         orig_pmd = *pmd;
1774         if (is_huge_zero_pmd(orig_pmd))
1775                 goto out;
1776
1777         if (unlikely(!pmd_present(orig_pmd))) {
1778                 VM_BUG_ON(thp_migration_supported() &&
1779                                   !is_pmd_migration_entry(orig_pmd));
1780                 goto out;
1781         }
1782
1783         page = pmd_page(orig_pmd);
1784         /*
1785          * If other processes are mapping this page, we couldn't discard
1786          * the page unless they all do MADV_FREE so let's skip the page.
1787          */
1788         if (page_mapcount(page) != 1)
1789                 goto out;
1790
1791         if (!trylock_page(page))
1792                 goto out;
1793
1794         /*
1795          * If user want to discard part-pages of THP, split it so MADV_FREE
1796          * will deactivate only them.
1797          */
1798         if (next - addr != HPAGE_PMD_SIZE) {
1799                 get_page(page);
1800                 spin_unlock(ptl);
1801                 split_huge_page(page);
1802                 unlock_page(page);
1803                 put_page(page);
1804                 goto out_unlocked;
1805         }
1806
1807         if (PageDirty(page))
1808                 ClearPageDirty(page);
1809         unlock_page(page);
1810
1811         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1812                 pmdp_invalidate(vma, addr, pmd);
1813                 orig_pmd = pmd_mkold(orig_pmd);
1814                 orig_pmd = pmd_mkclean(orig_pmd);
1815
1816                 set_pmd_at(mm, addr, pmd, orig_pmd);
1817                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1818         }
1819
1820         mark_page_lazyfree(page);
1821         ret = true;
1822 out:
1823         spin_unlock(ptl);
1824 out_unlocked:
1825         return ret;
1826 }
1827
1828 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1829 {
1830         pgtable_t pgtable;
1831
1832         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1833         pte_free(mm, pgtable);
1834         mm_dec_nr_ptes(mm);
1835 }
1836
1837 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1838                  pmd_t *pmd, unsigned long addr)
1839 {
1840         pmd_t orig_pmd;
1841         spinlock_t *ptl;
1842
1843         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1844
1845         ptl = __pmd_trans_huge_lock(pmd, vma);
1846         if (!ptl)
1847                 return 0;
1848         /*
1849          * For architectures like ppc64 we look at deposited pgtable
1850          * when calling pmdp_huge_get_and_clear. So do the
1851          * pgtable_trans_huge_withdraw after finishing pmdp related
1852          * operations.
1853          */
1854         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1855                         tlb->fullmm);
1856         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1857         if (vma_is_special_huge(vma)) {
1858                 if (arch_needs_pgtable_deposit())
1859                         zap_deposited_table(tlb->mm, pmd);
1860                 spin_unlock(ptl);
1861                 if (is_huge_zero_pmd(orig_pmd))
1862                         tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1863         } else if (is_huge_zero_pmd(orig_pmd)) {
1864                 zap_deposited_table(tlb->mm, pmd);
1865                 spin_unlock(ptl);
1866                 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1867         } else {
1868                 struct page *page = NULL;
1869                 int flush_needed = 1;
1870
1871                 if (pmd_present(orig_pmd)) {
1872                         page = pmd_page(orig_pmd);
1873                         page_remove_rmap(page, true);
1874                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1875                         VM_BUG_ON_PAGE(!PageHead(page), page);
1876                 } else if (thp_migration_supported()) {
1877                         swp_entry_t entry;
1878
1879                         VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1880                         entry = pmd_to_swp_entry(orig_pmd);
1881                         page = pfn_to_page(swp_offset(entry));
1882                         flush_needed = 0;
1883                 } else
1884                         WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1885
1886                 if (PageAnon(page)) {
1887                         zap_deposited_table(tlb->mm, pmd);
1888                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1889                 } else {
1890                         if (arch_needs_pgtable_deposit())
1891                                 zap_deposited_table(tlb->mm, pmd);
1892                         add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1893                 }
1894
1895                 spin_unlock(ptl);
1896                 if (flush_needed)
1897                         tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1898         }
1899         return 1;
1900 }
1901
1902 #ifndef pmd_move_must_withdraw
1903 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1904                                          spinlock_t *old_pmd_ptl,
1905                                          struct vm_area_struct *vma)
1906 {
1907         /*
1908          * With split pmd lock we also need to move preallocated
1909          * PTE page table if new_pmd is on different PMD page table.
1910          *
1911          * We also don't deposit and withdraw tables for file pages.
1912          */
1913         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1914 }
1915 #endif
1916
1917 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1918 {
1919 #ifdef CONFIG_MEM_SOFT_DIRTY
1920         if (unlikely(is_pmd_migration_entry(pmd)))
1921                 pmd = pmd_swp_mksoft_dirty(pmd);
1922         else if (pmd_present(pmd))
1923                 pmd = pmd_mksoft_dirty(pmd);
1924 #endif
1925         return pmd;
1926 }
1927
1928 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1929                   unsigned long new_addr, unsigned long old_end,
1930                   pmd_t *old_pmd, pmd_t *new_pmd)
1931 {
1932         spinlock_t *old_ptl, *new_ptl;
1933         pmd_t pmd;
1934         struct mm_struct *mm = vma->vm_mm;
1935         bool force_flush = false;
1936
1937         if ((old_addr & ~HPAGE_PMD_MASK) ||
1938             (new_addr & ~HPAGE_PMD_MASK) ||
1939             old_end - old_addr < HPAGE_PMD_SIZE)
1940                 return false;
1941
1942         /*
1943          * The destination pmd shouldn't be established, free_pgtables()
1944          * should have release it.
1945          */
1946         if (WARN_ON(!pmd_none(*new_pmd))) {
1947                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1948                 return false;
1949         }
1950
1951         /*
1952          * We don't have to worry about the ordering of src and dst
1953          * ptlocks because exclusive mmap_sem prevents deadlock.
1954          */
1955         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1956         if (old_ptl) {
1957                 new_ptl = pmd_lockptr(mm, new_pmd);
1958                 if (new_ptl != old_ptl)
1959                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1960                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1961                 if (pmd_present(pmd))
1962                         force_flush = true;
1963                 VM_BUG_ON(!pmd_none(*new_pmd));
1964
1965                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1966                         pgtable_t pgtable;
1967                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1968                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1969                 }
1970                 pmd = move_soft_dirty_pmd(pmd);
1971                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1972                 if (force_flush)
1973                         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1974                 if (new_ptl != old_ptl)
1975                         spin_unlock(new_ptl);
1976                 spin_unlock(old_ptl);
1977                 return true;
1978         }
1979         return false;
1980 }
1981
1982 /*
1983  * Returns
1984  *  - 0 if PMD could not be locked
1985  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1986  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1987  */
1988 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1989                 unsigned long addr, pgprot_t newprot, unsigned long cp_flags)
1990 {
1991         struct mm_struct *mm = vma->vm_mm;
1992         spinlock_t *ptl;
1993         pmd_t entry;
1994         bool preserve_write;
1995         int ret;
1996         bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
1997         bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
1998         bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
1999
2000         ptl = __pmd_trans_huge_lock(pmd, vma);
2001         if (!ptl)
2002                 return 0;
2003
2004         preserve_write = prot_numa && pmd_write(*pmd);
2005         ret = 1;
2006
2007 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2008         if (is_swap_pmd(*pmd)) {
2009                 swp_entry_t entry = pmd_to_swp_entry(*pmd);
2010
2011                 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
2012                 if (is_write_migration_entry(entry)) {
2013                         pmd_t newpmd;
2014                         /*
2015                          * A protection check is difficult so
2016                          * just be safe and disable write
2017                          */
2018                         make_migration_entry_read(&entry);
2019                         newpmd = swp_entry_to_pmd(entry);
2020                         if (pmd_swp_soft_dirty(*pmd))
2021                                 newpmd = pmd_swp_mksoft_dirty(newpmd);
2022                         set_pmd_at(mm, addr, pmd, newpmd);
2023                 }
2024                 goto unlock;
2025         }
2026 #endif
2027
2028         /*
2029          * Avoid trapping faults against the zero page. The read-only
2030          * data is likely to be read-cached on the local CPU and
2031          * local/remote hits to the zero page are not interesting.
2032          */
2033         if (prot_numa && is_huge_zero_pmd(*pmd))
2034                 goto unlock;
2035
2036         if (prot_numa && pmd_protnone(*pmd))
2037                 goto unlock;
2038
2039         /*
2040          * In case prot_numa, we are under down_read(mmap_sem). It's critical
2041          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
2042          * which is also under down_read(mmap_sem):
2043          *
2044          *      CPU0:                           CPU1:
2045          *                              change_huge_pmd(prot_numa=1)
2046          *                               pmdp_huge_get_and_clear_notify()
2047          * madvise_dontneed()
2048          *  zap_pmd_range()
2049          *   pmd_trans_huge(*pmd) == 0 (without ptl)
2050          *   // skip the pmd
2051          *                               set_pmd_at();
2052          *                               // pmd is re-established
2053          *
2054          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
2055          * which may break userspace.
2056          *
2057          * pmdp_invalidate() is required to make sure we don't miss
2058          * dirty/young flags set by hardware.
2059          */
2060         entry = pmdp_invalidate(vma, addr, pmd);
2061
2062         entry = pmd_modify(entry, newprot);
2063         if (preserve_write)
2064                 entry = pmd_mk_savedwrite(entry);
2065         if (uffd_wp) {
2066                 entry = pmd_wrprotect(entry);
2067                 entry = pmd_mkuffd_wp(entry);
2068         } else if (uffd_wp_resolve) {
2069                 /*
2070                  * Leave the write bit to be handled by PF interrupt
2071                  * handler, then things like COW could be properly
2072                  * handled.
2073                  */
2074                 entry = pmd_clear_uffd_wp(entry);
2075         }
2076         ret = HPAGE_PMD_NR;
2077         set_pmd_at(mm, addr, pmd, entry);
2078         BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
2079 unlock:
2080         spin_unlock(ptl);
2081         return ret;
2082 }
2083
2084 /*
2085  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
2086  *
2087  * Note that if it returns page table lock pointer, this routine returns without
2088  * unlocking page table lock. So callers must unlock it.
2089  */
2090 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
2091 {
2092         spinlock_t *ptl;
2093         ptl = pmd_lock(vma->vm_mm, pmd);
2094         if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
2095                         pmd_devmap(*pmd)))
2096                 return ptl;
2097         spin_unlock(ptl);
2098         return NULL;
2099 }
2100
2101 /*
2102  * Returns true if a given pud maps a thp, false otherwise.
2103  *
2104  * Note that if it returns true, this routine returns without unlocking page
2105  * table lock. So callers must unlock it.
2106  */
2107 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
2108 {
2109         spinlock_t *ptl;
2110
2111         ptl = pud_lock(vma->vm_mm, pud);
2112         if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
2113                 return ptl;
2114         spin_unlock(ptl);
2115         return NULL;
2116 }
2117
2118 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
2119 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
2120                  pud_t *pud, unsigned long addr)
2121 {
2122         spinlock_t *ptl;
2123
2124         ptl = __pud_trans_huge_lock(pud, vma);
2125         if (!ptl)
2126                 return 0;
2127         /*
2128          * For architectures like ppc64 we look at deposited pgtable
2129          * when calling pudp_huge_get_and_clear. So do the
2130          * pgtable_trans_huge_withdraw after finishing pudp related
2131          * operations.
2132          */
2133         pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
2134         tlb_remove_pud_tlb_entry(tlb, pud, addr);
2135         if (vma_is_special_huge(vma)) {
2136                 spin_unlock(ptl);
2137                 /* No zero page support yet */
2138         } else {
2139                 /* No support for anonymous PUD pages yet */
2140                 BUG();
2141         }
2142         return 1;
2143 }
2144
2145 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
2146                 unsigned long haddr)
2147 {
2148         VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
2149         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2150         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
2151         VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
2152
2153         count_vm_event(THP_SPLIT_PUD);
2154
2155         pudp_huge_clear_flush_notify(vma, haddr, pud);
2156 }
2157
2158 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2159                 unsigned long address)
2160 {
2161         spinlock_t *ptl;
2162         struct mmu_notifier_range range;
2163
2164         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2165                                 address & HPAGE_PUD_MASK,
2166                                 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
2167         mmu_notifier_invalidate_range_start(&range);
2168         ptl = pud_lock(vma->vm_mm, pud);
2169         if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2170                 goto out;
2171         __split_huge_pud_locked(vma, pud, range.start);
2172
2173 out:
2174         spin_unlock(ptl);
2175         /*
2176          * No need to double call mmu_notifier->invalidate_range() callback as
2177          * the above pudp_huge_clear_flush_notify() did already call it.
2178          */
2179         mmu_notifier_invalidate_range_only_end(&range);
2180 }
2181 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2182
2183 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2184                 unsigned long haddr, pmd_t *pmd)
2185 {
2186         struct mm_struct *mm = vma->vm_mm;
2187         pgtable_t pgtable;
2188         pmd_t _pmd;
2189         int i;
2190
2191         /*
2192          * Leave pmd empty until pte is filled note that it is fine to delay
2193          * notification until mmu_notifier_invalidate_range_end() as we are
2194          * replacing a zero pmd write protected page with a zero pte write
2195          * protected page.
2196          *
2197          * See Documentation/vm/mmu_notifier.rst
2198          */
2199         pmdp_huge_clear_flush(vma, haddr, pmd);
2200
2201         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2202         pmd_populate(mm, &_pmd, pgtable);
2203
2204         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2205                 pte_t *pte, entry;
2206                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2207                 entry = pte_mkspecial(entry);
2208                 pte = pte_offset_map(&_pmd, haddr);
2209                 VM_BUG_ON(!pte_none(*pte));
2210                 set_pte_at(mm, haddr, pte, entry);
2211                 pte_unmap(pte);
2212         }
2213         smp_wmb(); /* make pte visible before pmd */
2214         pmd_populate(mm, pmd, pgtable);
2215 }
2216
2217 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2218                 unsigned long haddr, bool freeze)
2219 {
2220         struct mm_struct *mm = vma->vm_mm;
2221         struct page *page;
2222         pgtable_t pgtable;
2223         pmd_t old_pmd, _pmd;
2224         bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false;
2225         unsigned long addr;
2226         int i;
2227
2228         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2229         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2230         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2231         VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2232                                 && !pmd_devmap(*pmd));
2233
2234         count_vm_event(THP_SPLIT_PMD);
2235
2236         if (!vma_is_anonymous(vma)) {
2237                 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2238                 /*
2239                  * We are going to unmap this huge page. So
2240                  * just go ahead and zap it
2241                  */
2242                 if (arch_needs_pgtable_deposit())
2243                         zap_deposited_table(mm, pmd);
2244                 if (vma_is_special_huge(vma))
2245                         return;
2246                 page = pmd_page(_pmd);
2247                 if (!PageDirty(page) && pmd_dirty(_pmd))
2248                         set_page_dirty(page);
2249                 if (!PageReferenced(page) && pmd_young(_pmd))
2250                         SetPageReferenced(page);
2251                 page_remove_rmap(page, true);
2252                 put_page(page);
2253                 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2254                 return;
2255         } else if (is_huge_zero_pmd(*pmd)) {
2256                 /*
2257                  * FIXME: Do we want to invalidate secondary mmu by calling
2258                  * mmu_notifier_invalidate_range() see comments below inside
2259                  * __split_huge_pmd() ?
2260                  *
2261                  * We are going from a zero huge page write protected to zero
2262                  * small page also write protected so it does not seems useful
2263                  * to invalidate secondary mmu at this time.
2264                  */
2265                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2266         }
2267
2268         /*
2269          * Up to this point the pmd is present and huge and userland has the
2270          * whole access to the hugepage during the split (which happens in
2271          * place). If we overwrite the pmd with the not-huge version pointing
2272          * to the pte here (which of course we could if all CPUs were bug
2273          * free), userland could trigger a small page size TLB miss on the
2274          * small sized TLB while the hugepage TLB entry is still established in
2275          * the huge TLB. Some CPU doesn't like that.
2276          * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2277          * 383 on page 93. Intel should be safe but is also warns that it's
2278          * only safe if the permission and cache attributes of the two entries
2279          * loaded in the two TLB is identical (which should be the case here).
2280          * But it is generally safer to never allow small and huge TLB entries
2281          * for the same virtual address to be loaded simultaneously. So instead
2282          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2283          * current pmd notpresent (atomically because here the pmd_trans_huge
2284          * must remain set at all times on the pmd until the split is complete
2285          * for this pmd), then we flush the SMP TLB and finally we write the
2286          * non-huge version of the pmd entry with pmd_populate.
2287          */
2288         old_pmd = pmdp_invalidate(vma, haddr, pmd);
2289
2290         pmd_migration = is_pmd_migration_entry(old_pmd);
2291         if (unlikely(pmd_migration)) {
2292                 swp_entry_t entry;
2293
2294                 entry = pmd_to_swp_entry(old_pmd);
2295                 page = pfn_to_page(swp_offset(entry));
2296                 write = is_write_migration_entry(entry);
2297                 young = false;
2298                 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2299                 uffd_wp = pmd_swp_uffd_wp(old_pmd);
2300         } else {
2301                 page = pmd_page(old_pmd);
2302                 if (pmd_dirty(old_pmd))
2303                         SetPageDirty(page);
2304                 write = pmd_write(old_pmd);
2305                 young = pmd_young(old_pmd);
2306                 soft_dirty = pmd_soft_dirty(old_pmd);
2307                 uffd_wp = pmd_uffd_wp(old_pmd);
2308         }
2309         VM_BUG_ON_PAGE(!page_count(page), page);
2310         page_ref_add(page, HPAGE_PMD_NR - 1);
2311
2312         /*
2313          * Withdraw the table only after we mark the pmd entry invalid.
2314          * This's critical for some architectures (Power).
2315          */
2316         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2317         pmd_populate(mm, &_pmd, pgtable);
2318
2319         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2320                 pte_t entry, *pte;
2321                 /*
2322                  * Note that NUMA hinting access restrictions are not
2323                  * transferred to avoid any possibility of altering
2324                  * permissions across VMAs.
2325                  */
2326                 if (freeze || pmd_migration) {
2327                         swp_entry_t swp_entry;
2328                         swp_entry = make_migration_entry(page + i, write);
2329                         entry = swp_entry_to_pte(swp_entry);
2330                         if (soft_dirty)
2331                                 entry = pte_swp_mksoft_dirty(entry);
2332                         if (uffd_wp)
2333                                 entry = pte_swp_mkuffd_wp(entry);
2334                 } else {
2335                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2336                         entry = maybe_mkwrite(entry, vma);
2337                         if (!write)
2338                                 entry = pte_wrprotect(entry);
2339                         if (!young)
2340                                 entry = pte_mkold(entry);
2341                         if (soft_dirty)
2342                                 entry = pte_mksoft_dirty(entry);
2343                         if (uffd_wp)
2344                                 entry = pte_mkuffd_wp(entry);
2345                 }
2346                 pte = pte_offset_map(&_pmd, addr);
2347                 BUG_ON(!pte_none(*pte));
2348                 set_pte_at(mm, addr, pte, entry);
2349                 atomic_inc(&page[i]._mapcount);
2350                 pte_unmap(pte);
2351         }
2352
2353         /*
2354          * Set PG_double_map before dropping compound_mapcount to avoid
2355          * false-negative page_mapped().
2356          */
2357         if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2358                 for (i = 0; i < HPAGE_PMD_NR; i++)
2359                         atomic_inc(&page[i]._mapcount);
2360         }
2361
2362         if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2363                 /* Last compound_mapcount is gone. */
2364                 __dec_node_page_state(page, NR_ANON_THPS);
2365                 if (TestClearPageDoubleMap(page)) {
2366                         /* No need in mapcount reference anymore */
2367                         for (i = 0; i < HPAGE_PMD_NR; i++)
2368                                 atomic_dec(&page[i]._mapcount);
2369                 }
2370         }
2371
2372         smp_wmb(); /* make pte visible before pmd */
2373         pmd_populate(mm, pmd, pgtable);
2374
2375         if (freeze) {
2376                 for (i = 0; i < HPAGE_PMD_NR; i++) {
2377                         page_remove_rmap(page + i, false);
2378                         put_page(page + i);
2379                 }
2380         }
2381 }
2382
2383 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2384                 unsigned long address, bool freeze, struct page *page)
2385 {
2386         spinlock_t *ptl;
2387         struct mmu_notifier_range range;
2388
2389         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2390                                 address & HPAGE_PMD_MASK,
2391                                 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2392         mmu_notifier_invalidate_range_start(&range);
2393         ptl = pmd_lock(vma->vm_mm, pmd);
2394
2395         /*
2396          * If caller asks to setup a migration entries, we need a page to check
2397          * pmd against. Otherwise we can end up replacing wrong page.
2398          */
2399         VM_BUG_ON(freeze && !page);
2400         if (page && page != pmd_page(*pmd))
2401                 goto out;
2402
2403         if (pmd_trans_huge(*pmd)) {
2404                 page = pmd_page(*pmd);
2405                 if (PageMlocked(page))
2406                         clear_page_mlock(page);
2407         } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2408                 goto out;
2409         __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2410 out:
2411         spin_unlock(ptl);
2412         /*
2413          * No need to double call mmu_notifier->invalidate_range() callback.
2414          * They are 3 cases to consider inside __split_huge_pmd_locked():
2415          *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2416          *  2) __split_huge_zero_page_pmd() read only zero page and any write
2417          *    fault will trigger a flush_notify before pointing to a new page
2418          *    (it is fine if the secondary mmu keeps pointing to the old zero
2419          *    page in the meantime)
2420          *  3) Split a huge pmd into pte pointing to the same page. No need
2421          *     to invalidate secondary tlb entry they are all still valid.
2422          *     any further changes to individual pte will notify. So no need
2423          *     to call mmu_notifier->invalidate_range()
2424          */
2425         mmu_notifier_invalidate_range_only_end(&range);
2426 }
2427
2428 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2429                 bool freeze, struct page *page)
2430 {
2431         pgd_t *pgd;
2432         p4d_t *p4d;
2433         pud_t *pud;
2434         pmd_t *pmd;
2435
2436         pgd = pgd_offset(vma->vm_mm, address);
2437         if (!pgd_present(*pgd))
2438                 return;
2439
2440         p4d = p4d_offset(pgd, address);
2441         if (!p4d_present(*p4d))
2442                 return;
2443
2444         pud = pud_offset(p4d, address);
2445         if (!pud_present(*pud))
2446                 return;
2447
2448         pmd = pmd_offset(pud, address);
2449
2450         __split_huge_pmd(vma, pmd, address, freeze, page);
2451 }
2452
2453 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2454                              unsigned long start,
2455                              unsigned long end,
2456                              long adjust_next)
2457 {
2458         /*
2459          * If the new start address isn't hpage aligned and it could
2460          * previously contain an hugepage: check if we need to split
2461          * an huge pmd.
2462          */
2463         if (start & ~HPAGE_PMD_MASK &&
2464             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2465             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2466                 split_huge_pmd_address(vma, start, false, NULL);
2467
2468         /*
2469          * If the new end address isn't hpage aligned and it could
2470          * previously contain an hugepage: check if we need to split
2471          * an huge pmd.
2472          */
2473         if (end & ~HPAGE_PMD_MASK &&
2474             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2475             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2476                 split_huge_pmd_address(vma, end, false, NULL);
2477
2478         /*
2479          * If we're also updating the vma->vm_next->vm_start, if the new
2480          * vm_next->vm_start isn't page aligned and it could previously
2481          * contain an hugepage: check if we need to split an huge pmd.
2482          */
2483         if (adjust_next > 0) {
2484                 struct vm_area_struct *next = vma->vm_next;
2485                 unsigned long nstart = next->vm_start;
2486                 nstart += adjust_next << PAGE_SHIFT;
2487                 if (nstart & ~HPAGE_PMD_MASK &&
2488                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2489                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2490                         split_huge_pmd_address(next, nstart, false, NULL);
2491         }
2492 }
2493
2494 static void unmap_page(struct page *page)
2495 {
2496         enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
2497                 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2498         bool unmap_success;
2499
2500         VM_BUG_ON_PAGE(!PageHead(page), page);
2501
2502         if (PageAnon(page))
2503                 ttu_flags |= TTU_SPLIT_FREEZE;
2504
2505         unmap_success = try_to_unmap(page, ttu_flags);
2506         VM_BUG_ON_PAGE(!unmap_success, page);
2507 }
2508
2509 static void remap_page(struct page *page)
2510 {
2511         int i;
2512         if (PageTransHuge(page)) {
2513                 remove_migration_ptes(page, page, true);
2514         } else {
2515                 for (i = 0; i < HPAGE_PMD_NR; i++)
2516                         remove_migration_ptes(page + i, page + i, true);
2517         }
2518 }
2519
2520 static void __split_huge_page_tail(struct page *head, int tail,
2521                 struct lruvec *lruvec, struct list_head *list)
2522 {
2523         struct page *page_tail = head + tail;
2524
2525         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2526
2527         /*
2528          * Clone page flags before unfreezing refcount.
2529          *
2530          * After successful get_page_unless_zero() might follow flags change,
2531          * for exmaple lock_page() which set PG_waiters.
2532          */
2533         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2534         page_tail->flags |= (head->flags &
2535                         ((1L << PG_referenced) |
2536                          (1L << PG_swapbacked) |
2537                          (1L << PG_swapcache) |
2538                          (1L << PG_mlocked) |
2539                          (1L << PG_uptodate) |
2540                          (1L << PG_active) |
2541                          (1L << PG_workingset) |
2542                          (1L << PG_locked) |
2543                          (1L << PG_unevictable) |
2544                          (1L << PG_dirty)));
2545
2546         /* ->mapping in first tail page is compound_mapcount */
2547         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2548                         page_tail);
2549         page_tail->mapping = head->mapping;
2550         page_tail->index = head->index + tail;
2551
2552         /* Page flags must be visible before we make the page non-compound. */
2553         smp_wmb();
2554
2555         /*
2556          * Clear PageTail before unfreezing page refcount.
2557          *
2558          * After successful get_page_unless_zero() might follow put_page()
2559          * which needs correct compound_head().
2560          */
2561         clear_compound_head(page_tail);
2562
2563         /* Finally unfreeze refcount. Additional reference from page cache. */
2564         page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2565                                           PageSwapCache(head)));
2566
2567         if (page_is_young(head))
2568                 set_page_young(page_tail);
2569         if (page_is_idle(head))
2570                 set_page_idle(page_tail);
2571
2572         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2573
2574         /*
2575          * always add to the tail because some iterators expect new
2576          * pages to show after the currently processed elements - e.g.
2577          * migrate_pages
2578          */
2579         lru_add_page_tail(head, page_tail, lruvec, list);
2580 }
2581
2582 static void __split_huge_page(struct page *page, struct list_head *list,
2583                 pgoff_t end, unsigned long flags)
2584 {
2585         struct page *head = compound_head(page);
2586         pg_data_t *pgdat = page_pgdat(head);
2587         struct lruvec *lruvec;
2588         struct address_space *swap_cache = NULL;
2589         unsigned long offset = 0;
2590         int i;
2591
2592         lruvec = mem_cgroup_page_lruvec(head, pgdat);
2593
2594         /* complete memcg works before add pages to LRU */
2595         mem_cgroup_split_huge_fixup(head);
2596
2597         if (PageAnon(head) && PageSwapCache(head)) {
2598                 swp_entry_t entry = { .val = page_private(head) };
2599
2600                 offset = swp_offset(entry);
2601                 swap_cache = swap_address_space(entry);
2602                 xa_lock(&swap_cache->i_pages);
2603         }
2604
2605         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
2606                 __split_huge_page_tail(head, i, lruvec, list);
2607                 /* Some pages can be beyond i_size: drop them from page cache */
2608                 if (head[i].index >= end) {
2609                         ClearPageDirty(head + i);
2610                         __delete_from_page_cache(head + i, NULL);
2611                         if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2612                                 shmem_uncharge(head->mapping->host, 1);
2613                         put_page(head + i);
2614                 } else if (!PageAnon(page)) {
2615                         __xa_store(&head->mapping->i_pages, head[i].index,
2616                                         head + i, 0);
2617                 } else if (swap_cache) {
2618                         __xa_store(&swap_cache->i_pages, offset + i,
2619                                         head + i, 0);
2620                 }
2621         }
2622
2623         ClearPageCompound(head);
2624
2625         split_page_owner(head, HPAGE_PMD_ORDER);
2626
2627         /* See comment in __split_huge_page_tail() */
2628         if (PageAnon(head)) {
2629                 /* Additional pin to swap cache */
2630                 if (PageSwapCache(head)) {
2631                         page_ref_add(head, 2);
2632                         xa_unlock(&swap_cache->i_pages);
2633                 } else {
2634                         page_ref_inc(head);
2635                 }
2636         } else {
2637                 /* Additional pin to page cache */
2638                 page_ref_add(head, 2);
2639                 xa_unlock(&head->mapping->i_pages);
2640         }
2641
2642         spin_unlock_irqrestore(&pgdat->lru_lock, flags);
2643
2644         remap_page(head);
2645
2646         for (i = 0; i < HPAGE_PMD_NR; i++) {
2647                 struct page *subpage = head + i;
2648                 if (subpage == page)
2649                         continue;
2650                 unlock_page(subpage);
2651
2652                 /*
2653                  * Subpages may be freed if there wasn't any mapping
2654                  * like if add_to_swap() is running on a lru page that
2655                  * had its mapping zapped. And freeing these pages
2656                  * requires taking the lru_lock so we do the put_page
2657                  * of the tail pages after the split is complete.
2658                  */
2659                 put_page(subpage);
2660         }
2661 }
2662
2663 int total_mapcount(struct page *page)
2664 {
2665         int i, compound, ret;
2666
2667         VM_BUG_ON_PAGE(PageTail(page), page);
2668
2669         if (likely(!PageCompound(page)))
2670                 return atomic_read(&page->_mapcount) + 1;
2671
2672         compound = compound_mapcount(page);
2673         if (PageHuge(page))
2674                 return compound;
2675         ret = compound;
2676         for (i = 0; i < HPAGE_PMD_NR; i++)
2677                 ret += atomic_read(&page[i]._mapcount) + 1;
2678         /* File pages has compound_mapcount included in _mapcount */
2679         if (!PageAnon(page))
2680                 return ret - compound * HPAGE_PMD_NR;
2681         if (PageDoubleMap(page))
2682                 ret -= HPAGE_PMD_NR;
2683         return ret;
2684 }
2685
2686 /*
2687  * This calculates accurately how many mappings a transparent hugepage
2688  * has (unlike page_mapcount() which isn't fully accurate). This full
2689  * accuracy is primarily needed to know if copy-on-write faults can
2690  * reuse the page and change the mapping to read-write instead of
2691  * copying them. At the same time this returns the total_mapcount too.
2692  *
2693  * The function returns the highest mapcount any one of the subpages
2694  * has. If the return value is one, even if different processes are
2695  * mapping different subpages of the transparent hugepage, they can
2696  * all reuse it, because each process is reusing a different subpage.
2697  *
2698  * The total_mapcount is instead counting all virtual mappings of the
2699  * subpages. If the total_mapcount is equal to "one", it tells the
2700  * caller all mappings belong to the same "mm" and in turn the
2701  * anon_vma of the transparent hugepage can become the vma->anon_vma
2702  * local one as no other process may be mapping any of the subpages.
2703  *
2704  * It would be more accurate to replace page_mapcount() with
2705  * page_trans_huge_mapcount(), however we only use
2706  * page_trans_huge_mapcount() in the copy-on-write faults where we
2707  * need full accuracy to avoid breaking page pinning, because
2708  * page_trans_huge_mapcount() is slower than page_mapcount().
2709  */
2710 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2711 {
2712         int i, ret, _total_mapcount, mapcount;
2713
2714         /* hugetlbfs shouldn't call it */
2715         VM_BUG_ON_PAGE(PageHuge(page), page);
2716
2717         if (likely(!PageTransCompound(page))) {
2718                 mapcount = atomic_read(&page->_mapcount) + 1;
2719                 if (total_mapcount)
2720                         *total_mapcount = mapcount;
2721                 return mapcount;
2722         }
2723
2724         page = compound_head(page);
2725
2726         _total_mapcount = ret = 0;
2727         for (i = 0; i < HPAGE_PMD_NR; i++) {
2728                 mapcount = atomic_read(&page[i]._mapcount) + 1;
2729                 ret = max(ret, mapcount);
2730                 _total_mapcount += mapcount;
2731         }
2732         if (PageDoubleMap(page)) {
2733                 ret -= 1;
2734                 _total_mapcount -= HPAGE_PMD_NR;
2735         }
2736         mapcount = compound_mapcount(page);
2737         ret += mapcount;
2738         _total_mapcount += mapcount;
2739         if (total_mapcount)
2740                 *total_mapcount = _total_mapcount;
2741         return ret;
2742 }
2743
2744 /* Racy check whether the huge page can be split */
2745 bool can_split_huge_page(struct page *page, int *pextra_pins)
2746 {
2747         int extra_pins;
2748
2749         /* Additional pins from page cache */
2750         if (PageAnon(page))
2751                 extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0;
2752         else
2753                 extra_pins = HPAGE_PMD_NR;
2754         if (pextra_pins)
2755                 *pextra_pins = extra_pins;
2756         return total_mapcount(page) == page_count(page) - extra_pins - 1;
2757 }
2758
2759 /*
2760  * This function splits huge page into normal pages. @page can point to any
2761  * subpage of huge page to split. Split doesn't change the position of @page.
2762  *
2763  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2764  * The huge page must be locked.
2765  *
2766  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2767  *
2768  * Both head page and tail pages will inherit mapping, flags, and so on from
2769  * the hugepage.
2770  *
2771  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2772  * they are not mapped.
2773  *
2774  * Returns 0 if the hugepage is split successfully.
2775  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2776  * us.
2777  */
2778 int split_huge_page_to_list(struct page *page, struct list_head *list)
2779 {
2780         struct page *head = compound_head(page);
2781         struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2782         struct deferred_split *ds_queue = get_deferred_split_queue(head);
2783         struct anon_vma *anon_vma = NULL;
2784         struct address_space *mapping = NULL;
2785         int count, mapcount, extra_pins, ret;
2786         bool mlocked;
2787         unsigned long flags;
2788         pgoff_t end;
2789
2790         VM_BUG_ON_PAGE(is_huge_zero_page(head), head);
2791         VM_BUG_ON_PAGE(!PageLocked(head), head);
2792         VM_BUG_ON_PAGE(!PageCompound(head), head);
2793
2794         if (PageWriteback(head))
2795                 return -EBUSY;
2796
2797         if (PageAnon(head)) {
2798                 /*
2799                  * The caller does not necessarily hold an mmap_sem that would
2800                  * prevent the anon_vma disappearing so we first we take a
2801                  * reference to it and then lock the anon_vma for write. This
2802                  * is similar to page_lock_anon_vma_read except the write lock
2803                  * is taken to serialise against parallel split or collapse
2804                  * operations.
2805                  */
2806                 anon_vma = page_get_anon_vma(head);
2807                 if (!anon_vma) {
2808                         ret = -EBUSY;
2809                         goto out;
2810                 }
2811                 end = -1;
2812                 mapping = NULL;
2813                 anon_vma_lock_write(anon_vma);
2814         } else {
2815                 mapping = head->mapping;
2816
2817                 /* Truncated ? */
2818                 if (!mapping) {
2819                         ret = -EBUSY;
2820                         goto out;
2821                 }
2822
2823                 anon_vma = NULL;
2824                 i_mmap_lock_read(mapping);
2825
2826                 /*
2827                  *__split_huge_page() may need to trim off pages beyond EOF:
2828                  * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2829                  * which cannot be nested inside the page tree lock. So note
2830                  * end now: i_size itself may be changed at any moment, but
2831                  * head page lock is good enough to serialize the trimming.
2832                  */
2833                 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2834         }
2835
2836         /*
2837          * Racy check if we can split the page, before unmap_page() will
2838          * split PMDs
2839          */
2840         if (!can_split_huge_page(head, &extra_pins)) {
2841                 ret = -EBUSY;
2842                 goto out_unlock;
2843         }
2844
2845         mlocked = PageMlocked(head);
2846         unmap_page(head);
2847         VM_BUG_ON_PAGE(compound_mapcount(head), head);
2848
2849         /* Make sure the page is not on per-CPU pagevec as it takes pin */
2850         if (mlocked)
2851                 lru_add_drain();
2852
2853         /* prevent PageLRU to go away from under us, and freeze lru stats */
2854         spin_lock_irqsave(&pgdata->lru_lock, flags);
2855
2856         if (mapping) {
2857                 XA_STATE(xas, &mapping->i_pages, page_index(head));
2858
2859                 /*
2860                  * Check if the head page is present in page cache.
2861                  * We assume all tail are present too, if head is there.
2862                  */
2863                 xa_lock(&mapping->i_pages);
2864                 if (xas_load(&xas) != head)
2865                         goto fail;
2866         }
2867
2868         /* Prevent deferred_split_scan() touching ->_refcount */
2869         spin_lock(&ds_queue->split_queue_lock);
2870         count = page_count(head);
2871         mapcount = total_mapcount(head);
2872         if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2873                 if (!list_empty(page_deferred_list(head))) {
2874                         ds_queue->split_queue_len--;
2875                         list_del(page_deferred_list(head));
2876                 }
2877                 spin_unlock(&ds_queue->split_queue_lock);
2878                 if (mapping) {
2879                         if (PageSwapBacked(head))
2880                                 __dec_node_page_state(head, NR_SHMEM_THPS);
2881                         else
2882                                 __dec_node_page_state(head, NR_FILE_THPS);
2883                 }
2884
2885                 __split_huge_page(page, list, end, flags);
2886                 if (PageSwapCache(head)) {
2887                         swp_entry_t entry = { .val = page_private(head) };
2888
2889                         ret = split_swap_cluster(entry);
2890                 } else
2891                         ret = 0;
2892         } else {
2893                 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2894                         pr_alert("total_mapcount: %u, page_count(): %u\n",
2895                                         mapcount, count);
2896                         if (PageTail(page))
2897                                 dump_page(head, NULL);
2898                         dump_page(page, "total_mapcount(head) > 0");
2899                         BUG();
2900                 }
2901                 spin_unlock(&ds_queue->split_queue_lock);
2902 fail:           if (mapping)
2903                         xa_unlock(&mapping->i_pages);
2904                 spin_unlock_irqrestore(&pgdata->lru_lock, flags);
2905                 remap_page(head);
2906                 ret = -EBUSY;
2907         }
2908
2909 out_unlock:
2910         if (anon_vma) {
2911                 anon_vma_unlock_write(anon_vma);
2912                 put_anon_vma(anon_vma);
2913         }
2914         if (mapping)
2915                 i_mmap_unlock_read(mapping);
2916 out:
2917         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2918         return ret;
2919 }
2920
2921 void free_transhuge_page(struct page *page)
2922 {
2923         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2924         unsigned long flags;
2925
2926         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2927         if (!list_empty(page_deferred_list(page))) {
2928                 ds_queue->split_queue_len--;
2929                 list_del(page_deferred_list(page));
2930         }
2931         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2932         free_compound_page(page);
2933 }
2934
2935 void deferred_split_huge_page(struct page *page)
2936 {
2937         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2938 #ifdef CONFIG_MEMCG
2939         struct mem_cgroup *memcg = compound_head(page)->mem_cgroup;
2940 #endif
2941         unsigned long flags;
2942
2943         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2944
2945         /*
2946          * The try_to_unmap() in page reclaim path might reach here too,
2947          * this may cause a race condition to corrupt deferred split queue.
2948          * And, if page reclaim is already handling the same page, it is
2949          * unnecessary to handle it again in shrinker.
2950          *
2951          * Check PageSwapCache to determine if the page is being
2952          * handled by page reclaim since THP swap would add the page into
2953          * swap cache before calling try_to_unmap().
2954          */
2955         if (PageSwapCache(page))
2956                 return;
2957
2958         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2959         if (list_empty(page_deferred_list(page))) {
2960                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2961                 list_add_tail(page_deferred_list(page), &ds_queue->split_queue);
2962                 ds_queue->split_queue_len++;
2963 #ifdef CONFIG_MEMCG
2964                 if (memcg)
2965                         memcg_set_shrinker_bit(memcg, page_to_nid(page),
2966                                                deferred_split_shrinker.id);
2967 #endif
2968         }
2969         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2970 }
2971
2972 static unsigned long deferred_split_count(struct shrinker *shrink,
2973                 struct shrink_control *sc)
2974 {
2975         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2976         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2977
2978 #ifdef CONFIG_MEMCG
2979         if (sc->memcg)
2980                 ds_queue = &sc->memcg->deferred_split_queue;
2981 #endif
2982         return READ_ONCE(ds_queue->split_queue_len);
2983 }
2984
2985 static unsigned long deferred_split_scan(struct shrinker *shrink,
2986                 struct shrink_control *sc)
2987 {
2988         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2989         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2990         unsigned long flags;
2991         LIST_HEAD(list), *pos, *next;
2992         struct page *page;
2993         int split = 0;
2994
2995 #ifdef CONFIG_MEMCG
2996         if (sc->memcg)
2997                 ds_queue = &sc->memcg->deferred_split_queue;
2998 #endif
2999
3000         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
3001         /* Take pin on all head pages to avoid freeing them under us */
3002         list_for_each_safe(pos, next, &ds_queue->split_queue) {
3003                 page = list_entry((void *)pos, struct page, mapping);
3004                 page = compound_head(page);
3005                 if (get_page_unless_zero(page)) {
3006                         list_move(page_deferred_list(page), &list);
3007                 } else {
3008                         /* We lost race with put_compound_page() */
3009                         list_del_init(page_deferred_list(page));
3010                         ds_queue->split_queue_len--;
3011                 }
3012                 if (!--sc->nr_to_scan)
3013                         break;
3014         }
3015         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
3016
3017         list_for_each_safe(pos, next, &list) {
3018                 page = list_entry((void *)pos, struct page, mapping);
3019                 if (!trylock_page(page))
3020                         goto next;
3021                 /* split_huge_page() removes page from list on success */
3022                 if (!split_huge_page(page))
3023                         split++;
3024                 unlock_page(page);
3025 next:
3026                 put_page(page);
3027         }
3028
3029         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
3030         list_splice_tail(&list, &ds_queue->split_queue);
3031         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
3032
3033         /*
3034          * Stop shrinker if we didn't split any page, but the queue is empty.
3035          * This can happen if pages were freed under us.
3036          */
3037         if (!split && list_empty(&ds_queue->split_queue))
3038                 return SHRINK_STOP;
3039         return split;
3040 }
3041
3042 static struct shrinker deferred_split_shrinker = {
3043         .count_objects = deferred_split_count,
3044         .scan_objects = deferred_split_scan,
3045         .seeks = DEFAULT_SEEKS,
3046         .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
3047                  SHRINKER_NONSLAB,
3048 };
3049
3050 #ifdef CONFIG_DEBUG_FS
3051 static int split_huge_pages_set(void *data, u64 val)
3052 {
3053         struct zone *zone;
3054         struct page *page;
3055         unsigned long pfn, max_zone_pfn;
3056         unsigned long total = 0, split = 0;
3057
3058         if (val != 1)
3059                 return -EINVAL;
3060
3061         for_each_populated_zone(zone) {
3062                 max_zone_pfn = zone_end_pfn(zone);
3063                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
3064                         if (!pfn_valid(pfn))
3065                                 continue;
3066
3067                         page = pfn_to_page(pfn);
3068                         if (!get_page_unless_zero(page))
3069                                 continue;
3070
3071                         if (zone != page_zone(page))
3072                                 goto next;
3073
3074                         if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
3075                                 goto next;
3076
3077                         total++;
3078                         lock_page(page);
3079                         if (!split_huge_page(page))
3080                                 split++;
3081                         unlock_page(page);
3082 next:
3083                         put_page(page);
3084                 }
3085         }
3086
3087         pr_info("%lu of %lu THP split\n", split, total);
3088
3089         return 0;
3090 }
3091 DEFINE_DEBUGFS_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
3092                 "%llu\n");
3093
3094 static int __init split_huge_pages_debugfs(void)
3095 {
3096         debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3097                             &split_huge_pages_fops);
3098         return 0;
3099 }
3100 late_initcall(split_huge_pages_debugfs);
3101 #endif
3102
3103 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
3104 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
3105                 struct page *page)
3106 {
3107         struct vm_area_struct *vma = pvmw->vma;
3108         struct mm_struct *mm = vma->vm_mm;
3109         unsigned long address = pvmw->address;
3110         pmd_t pmdval;
3111         swp_entry_t entry;
3112         pmd_t pmdswp;
3113
3114         if (!(pvmw->pmd && !pvmw->pte))
3115                 return;
3116
3117         flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
3118         pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
3119         if (pmd_dirty(pmdval))
3120                 set_page_dirty(page);
3121         entry = make_migration_entry(page, pmd_write(pmdval));
3122         pmdswp = swp_entry_to_pmd(entry);
3123         if (pmd_soft_dirty(pmdval))
3124                 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
3125         set_pmd_at(mm, address, pvmw->pmd, pmdswp);
3126         page_remove_rmap(page, true);
3127         put_page(page);
3128 }
3129
3130 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
3131 {
3132         struct vm_area_struct *vma = pvmw->vma;
3133         struct mm_struct *mm = vma->vm_mm;
3134         unsigned long address = pvmw->address;
3135         unsigned long mmun_start = address & HPAGE_PMD_MASK;
3136         pmd_t pmde;
3137         swp_entry_t entry;
3138
3139         if (!(pvmw->pmd && !pvmw->pte))
3140                 return;
3141
3142         entry = pmd_to_swp_entry(*pvmw->pmd);
3143         get_page(new);
3144         pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
3145         if (pmd_swp_soft_dirty(*pvmw->pmd))
3146                 pmde = pmd_mksoft_dirty(pmde);
3147         if (is_write_migration_entry(entry))
3148                 pmde = maybe_pmd_mkwrite(pmde, vma);
3149
3150         flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
3151         if (PageAnon(new))
3152                 page_add_anon_rmap(new, vma, mmun_start, true);
3153         else
3154                 page_add_file_rmap(new, true);
3155         set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
3156         if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new))
3157                 mlock_vma_page(new);
3158         update_mmu_cache_pmd(vma, address, pvmw->pmd);
3159 }
3160 #endif