mm/huge_memory.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *  Copyright (C) 2009  Red Hat, Inc.
   4  */
   5
   6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   7
   8 #include <linux/mm.h>
   9 #include <linux/sched.h>
  10 #include <linux/sched/coredump.h>
  11 #include <linux/sched/numa_balancing.h>
  12 #include <linux/highmem.h>
  13 #include <linux/hugetlb.h>
  14 #include <linux/mmu_notifier.h>
  15 #include <linux/rmap.h>
  16 #include <linux/swap.h>
  17 #include <linux/shrinker.h>
  18 #include <linux/mm_inline.h>
  19 #include <linux/swapops.h>
  20 #include <linux/dax.h>
  21 #include <linux/khugepaged.h>
  22 #include <linux/freezer.h>
  23 #include <linux/pfn_t.h>
  24 #include <linux/mman.h>
  25 #include <linux/memremap.h>
  26 #include <linux/pagemap.h>
  27 #include <linux/debugfs.h>
  28 #include <linux/migrate.h>
  29 #include <linux/hashtable.h>
  30 #include <linux/userfaultfd_k.h>
  31 #include <linux/page_idle.h>
  32 #include <linux/shmem_fs.h>
  33 #include <linux/oom.h>
  34 #include <linux/numa.h>
  35 #include <linux/page_owner.h>
  36
  37 #include <asm/tlb.h>
  38 #include <asm/pgalloc.h>
  39 #include "internal.h"
  40
  41 /*
  42  * By default, transparent hugepage support is disabled in order to avoid
  43  * risking an increased memory footprint for applications that are not
  44  * guaranteed to benefit from it. When transparent hugepage support is
  45  * enabled, it is for all mappings, and khugepaged scans all mappings.
  46  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  47  * for all hugepage allocations.
  48  */
  49 unsigned long transparent_hugepage_flags __read_mostly =
  50 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  51         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  52 #endif
  53 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  54         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  55 #endif
  56         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
  57         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  58         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  59
  60 static struct shrinker deferred_split_shrinker;
  61
  62 static atomic_t huge_zero_refcount;
  63 struct page *huge_zero_page __read_mostly;
  64
  65 bool transparent_hugepage_enabled(struct vm_area_struct *vma)
  66 {
  67         /* The addr is used to check if the vma size fits */
  68         unsigned long addr = (vma->vm_end & HPAGE_PMD_MASK) - HPAGE_PMD_SIZE;
  69
  70         if (!transhuge_vma_suitable(vma, addr))
  71                 return false;
  72         if (vma_is_anonymous(vma))
  73                 return __transparent_hugepage_enabled(vma);
  74         if (vma_is_shmem(vma))
  75                 return shmem_huge_enabled(vma);
  76
  77         return false;
  78 }
  79
  80 static struct page *get_huge_zero_page(void)
  81 {
  82         struct page *zero_page;
  83 retry:
  84         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
  85                 return READ_ONCE(huge_zero_page);
  86
  87         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
  88                         HPAGE_PMD_ORDER);
  89         if (!zero_page) {
  90                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
  91                 return NULL;
  92         }
  93         count_vm_event(THP_ZERO_PAGE_ALLOC);
  94         preempt_disable();
  95         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
  96                 preempt_enable();
  97                 __free_pages(zero_page, compound_order(zero_page));
  98                 goto retry;
  99         }
 100
 101         /* We take additional reference here. It will be put back by shrinker */
 102         atomic_set(&huge_zero_refcount, 2);
 103         preempt_enable();
 104         return READ_ONCE(huge_zero_page);
 105 }
 106
 107 static void put_huge_zero_page(void)
 108 {
 109         /*
 110          * Counter should never go to zero here. Only shrinker can put
 111          * last reference.
 112          */
 113         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 114 }
 115
 116 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
 117 {
 118         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 119                 return READ_ONCE(huge_zero_page);
 120
 121         if (!get_huge_zero_page())
 122                 return NULL;
 123
 124         if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 125                 put_huge_zero_page();
 126
 127         return READ_ONCE(huge_zero_page);
 128 }
 129
 130 void mm_put_huge_zero_page(struct mm_struct *mm)
 131 {
 132         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 133                 put_huge_zero_page();
 134 }
 135
 136 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 137                                         struct shrink_control *sc)
 138 {
 139         /* we can free zero page only if last reference remains */
 140         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 141 }
 142
 143 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 144                                        struct shrink_control *sc)
 145 {
 146         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 147                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 148                 BUG_ON(zero_page == NULL);
 149                 __free_pages(zero_page, compound_order(zero_page));
 150                 return HPAGE_PMD_NR;
 151         }
 152
 153         return 0;
 154 }
 155
 156 static struct shrinker huge_zero_page_shrinker = {
 157         .count_objects = shrink_huge_zero_page_count,
 158         .scan_objects = shrink_huge_zero_page_scan,
 159         .seeks = DEFAULT_SEEKS,
 160 };
 161
 162 #ifdef CONFIG_SYSFS
 163 static ssize_t enabled_show(struct kobject *kobj,
 164                             struct kobj_attribute *attr, char *buf)
 165 {
 166         if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
 167                 return sprintf(buf, "[always] madvise never\n");
 168         else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
 169                 return sprintf(buf, "always [madvise] never\n");
 170         else
 171                 return sprintf(buf, "always madvise [never]\n");
 172 }
 173
 174 static ssize_t enabled_store(struct kobject *kobj,
 175                              struct kobj_attribute *attr,
 176                              const char *buf, size_t count)
 177 {
 178         ssize_t ret = count;
 179
 180         if (sysfs_streq(buf, "always")) {
 181                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 182                 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 183         } else if (sysfs_streq(buf, "madvise")) {
 184                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 185                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 186         } else if (sysfs_streq(buf, "never")) {
 187                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 188                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 189         } else
 190                 ret = -EINVAL;
 191
 192         if (ret > 0) {
 193                 int err = start_stop_khugepaged();
 194                 if (err)
 195                         ret = err;
 196         }
 197         return ret;
 198 }
 199 static struct kobj_attribute enabled_attr =
 200         __ATTR(enabled, 0644, enabled_show, enabled_store);
 201
 202 ssize_t single_hugepage_flag_show(struct kobject *kobj,
 203                                 struct kobj_attribute *attr, char *buf,
 204                                 enum transparent_hugepage_flag flag)
 205 {
 206         return sprintf(buf, "%d\n",
 207                        !!test_bit(flag, &transparent_hugepage_flags));
 208 }
 209
 210 ssize_t single_hugepage_flag_store(struct kobject *kobj,
 211                                  struct kobj_attribute *attr,
 212                                  const char *buf, size_t count,
 213                                  enum transparent_hugepage_flag flag)
 214 {
 215         unsigned long value;
 216         int ret;
 217
 218         ret = kstrtoul(buf, 10, &value);
 219         if (ret < 0)
 220                 return ret;
 221         if (value > 1)
 222                 return -EINVAL;
 223
 224         if (value)
 225                 set_bit(flag, &transparent_hugepage_flags);
 226         else
 227                 clear_bit(flag, &transparent_hugepage_flags);
 228
 229         return count;
 230 }
 231
 232 static ssize_t defrag_show(struct kobject *kobj,
 233                            struct kobj_attribute *attr, char *buf)
 234 {
 235         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 236                 return sprintf(buf, "[always] defer defer+madvise madvise never\n");
 237         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 238                 return sprintf(buf, "always [defer] defer+madvise madvise never\n");
 239         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 240                 return sprintf(buf, "always defer [defer+madvise] madvise never\n");
 241         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 242                 return sprintf(buf, "always defer defer+madvise [madvise] never\n");
 243         return sprintf(buf, "always defer defer+madvise madvise [never]\n");
 244 }
 245
 246 static ssize_t defrag_store(struct kobject *kobj,
 247                             struct kobj_attribute *attr,
 248                             const char *buf, size_t count)
 249 {
 250         if (sysfs_streq(buf, "always")) {
 251                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 252                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 253                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 254                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 255         } else if (sysfs_streq(buf, "defer+madvise")) {
 256                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 257                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 258                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 259                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 260         } else if (sysfs_streq(buf, "defer")) {
 261                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 262                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 263                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 264                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 265         } else if (sysfs_streq(buf, "madvise")) {
 266                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 267                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 268                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 269                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 270         } else if (sysfs_streq(buf, "never")) {
 271                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 272                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 273                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 274                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 275         } else
 276                 return -EINVAL;
 277
 278         return count;
 279 }
 280 static struct kobj_attribute defrag_attr =
 281         __ATTR(defrag, 0644, defrag_show, defrag_store);
 282
 283 static ssize_t use_zero_page_show(struct kobject *kobj,
 284                 struct kobj_attribute *attr, char *buf)
 285 {
 286         return single_hugepage_flag_show(kobj, attr, buf,
 287                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 288 }
 289 static ssize_t use_zero_page_store(struct kobject *kobj,
 290                 struct kobj_attribute *attr, const char *buf, size_t count)
 291 {
 292         return single_hugepage_flag_store(kobj, attr, buf, count,
 293                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 294 }
 295 static struct kobj_attribute use_zero_page_attr =
 296         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 297
 298 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
 299                 struct kobj_attribute *attr, char *buf)
 300 {
 301         return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE);
 302 }
 303 static struct kobj_attribute hpage_pmd_size_attr =
 304         __ATTR_RO(hpage_pmd_size);
 305
 306 #ifdef CONFIG_DEBUG_VM
 307 static ssize_t debug_cow_show(struct kobject *kobj,
 308                                 struct kobj_attribute *attr, char *buf)
 309 {
 310         return single_hugepage_flag_show(kobj, attr, buf,
 311                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 312 }
 313 static ssize_t debug_cow_store(struct kobject *kobj,
 314                                struct kobj_attribute *attr,
 315                                const char *buf, size_t count)
 316 {
 317         return single_hugepage_flag_store(kobj, attr, buf, count,
 318                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 319 }
 320 static struct kobj_attribute debug_cow_attr =
 321         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 322 #endif /* CONFIG_DEBUG_VM */
 323
 324 static struct attribute *hugepage_attr[] = {
 325         &enabled_attr.attr,
 326         &defrag_attr.attr,
 327         &use_zero_page_attr.attr,
 328         &hpage_pmd_size_attr.attr,
 329 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
 330         &shmem_enabled_attr.attr,
 331 #endif
 332 #ifdef CONFIG_DEBUG_VM
 333         &debug_cow_attr.attr,
 334 #endif
 335         NULL,
 336 };
 337
 338 static const struct attribute_group hugepage_attr_group = {
 339         .attrs = hugepage_attr,
 340 };
 341
 342 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 343 {
 344         int err;
 345
 346         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 347         if (unlikely(!*hugepage_kobj)) {
 348                 pr_err("failed to create transparent hugepage kobject\n");
 349                 return -ENOMEM;
 350         }
 351
 352         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 353         if (err) {
 354                 pr_err("failed to register transparent hugepage group\n");
 355                 goto delete_obj;
 356         }
 357
 358         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 359         if (err) {
 360                 pr_err("failed to register transparent hugepage group\n");
 361                 goto remove_hp_group;
 362         }
 363
 364         return 0;
 365
 366 remove_hp_group:
 367         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 368 delete_obj:
 369         kobject_put(*hugepage_kobj);
 370         return err;
 371 }
 372
 373 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 374 {
 375         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 376         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 377         kobject_put(hugepage_kobj);
 378 }
 379 #else
 380 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 381 {
 382         return 0;
 383 }
 384
 385 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 386 {
 387 }
 388 #endif /* CONFIG_SYSFS */
 389
 390 static int __init hugepage_init(void)
 391 {
 392         int err;
 393         struct kobject *hugepage_kobj;
 394
 395         if (!has_transparent_hugepage()) {
 396                 transparent_hugepage_flags = 0;
 397                 return -EINVAL;
 398         }
 399
 400         /*
 401          * hugepages can't be allocated by the buddy allocator
 402          */
 403         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
 404         /*
 405          * we use page->mapping and page->index in second tail page
 406          * as list_head: assuming THP order >= 2
 407          */
 408         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 409
 410         err = hugepage_init_sysfs(&hugepage_kobj);
 411         if (err)
 412                 goto err_sysfs;
 413
 414         err = khugepaged_init();
 415         if (err)
 416                 goto err_slab;
 417
 418         err = register_shrinker(&huge_zero_page_shrinker);
 419         if (err)
 420                 goto err_hzp_shrinker;
 421         err = register_shrinker(&deferred_split_shrinker);
 422         if (err)
 423                 goto err_split_shrinker;
 424
 425         /*
 426          * By default disable transparent hugepages on smaller systems,
 427          * where the extra memory used could hurt more than TLB overhead
 428          * is likely to save.  The admin can still enable it through /sys.
 429          */
 430         if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
 431                 transparent_hugepage_flags = 0;
 432                 return 0;
 433         }
 434
 435         err = start_stop_khugepaged();
 436         if (err)
 437                 goto err_khugepaged;
 438
 439         return 0;
 440 err_khugepaged:
 441         unregister_shrinker(&deferred_split_shrinker);
 442 err_split_shrinker:
 443         unregister_shrinker(&huge_zero_page_shrinker);
 444 err_hzp_shrinker:
 445         khugepaged_destroy();
 446 err_slab:
 447         hugepage_exit_sysfs(hugepage_kobj);
 448 err_sysfs:
 449         return err;
 450 }
 451 subsys_initcall(hugepage_init);
 452
 453 static int __init setup_transparent_hugepage(char *str)
 454 {
 455         int ret = 0;
 456         if (!str)
 457                 goto out;
 458         if (!strcmp(str, "always")) {
 459                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 460                         &transparent_hugepage_flags);
 461                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 462                           &transparent_hugepage_flags);
 463                 ret = 1;
 464         } else if (!strcmp(str, "madvise")) {
 465                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 466                           &transparent_hugepage_flags);
 467                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 468                         &transparent_hugepage_flags);
 469                 ret = 1;
 470         } else if (!strcmp(str, "never")) {
 471                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 472                           &transparent_hugepage_flags);
 473                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 474                           &transparent_hugepage_flags);
 475                 ret = 1;
 476         }
 477 out:
 478         if (!ret)
 479                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 480         return ret;
 481 }
 482 __setup("transparent_hugepage=", setup_transparent_hugepage);
 483
 484 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 485 {
 486         if (likely(vma->vm_flags & VM_WRITE))
 487                 pmd = pmd_mkwrite(pmd);
 488         return pmd;
 489 }
 490
 491 #ifdef CONFIG_MEMCG
 492 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
 493 {
 494         struct mem_cgroup *memcg = compound_head(page)->mem_cgroup;
 495         struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
 496
 497         if (memcg)
 498                 return &memcg->deferred_split_queue;
 499         else
 500                 return &pgdat->deferred_split_queue;
 501 }
 502 #else
 503 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
 504 {
 505         struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
 506
 507         return &pgdat->deferred_split_queue;
 508 }
 509 #endif
 510
 511 void prep_transhuge_page(struct page *page)
 512 {
 513         /*
 514          * we use page->mapping and page->indexlru in second tail page
 515          * as list_head: assuming THP order >= 2
 516          */
 517
 518         INIT_LIST_HEAD(page_deferred_list(page));
 519         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 520 }
 521
 522 bool is_transparent_hugepage(struct page *page)
 523 {
 524         if (!PageCompound(page))
 525                 return 0;
 526
 527         page = compound_head(page);
 528         return is_huge_zero_page(page) ||
 529                page[1].compound_dtor == TRANSHUGE_PAGE_DTOR;
 530 }
 531 EXPORT_SYMBOL_GPL(is_transparent_hugepage);
 532
 533 static unsigned long __thp_get_unmapped_area(struct file *filp,
 534                 unsigned long addr, unsigned long len,
 535                 loff_t off, unsigned long flags, unsigned long size)
 536 {
 537         loff_t off_end = off + len;
 538         loff_t off_align = round_up(off, size);
 539         unsigned long len_pad, ret;
 540
 541         if (off_end <= off_align || (off_end - off_align) < size)
 542                 return 0;
 543
 544         len_pad = len + size;
 545         if (len_pad < len || (off + len_pad) < off)
 546                 return 0;
 547
 548         ret = current->mm->get_unmapped_area(filp, addr, len_pad,
 549                                               off >> PAGE_SHIFT, flags);
 550
 551         /*
 552          * The failure might be due to length padding. The caller will retry
 553          * without the padding.
 554          */
 555         if (IS_ERR_VALUE(ret))
 556                 return 0;
 557
 558         /*
 559          * Do not try to align to THP boundary if allocation at the address
 560          * hint succeeds.
 561          */
 562         if (ret == addr)
 563                 return addr;
 564
 565         ret += (off - ret) & (size - 1);
 566         return ret;
 567 }
 568
 569 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 570                 unsigned long len, unsigned long pgoff, unsigned long flags)
 571 {
 572         unsigned long ret;
 573         loff_t off = (loff_t)pgoff << PAGE_SHIFT;
 574
 575         if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
 576                 goto out;
 577
 578         ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE);
 579         if (ret)
 580                 return ret;
 581 out:
 582         return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
 583 }
 584 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
 585
 586 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
 587                         struct page *page, gfp_t gfp)
 588 {
 589         struct vm_area_struct *vma = vmf->vma;
 590         struct mem_cgroup *memcg;
 591         pgtable_t pgtable;
 592         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 593         vm_fault_t ret = 0;
 594
 595         VM_BUG_ON_PAGE(!PageCompound(page), page);
 596
 597         if (mem_cgroup_try_charge_delay(page, vma->vm_mm, gfp, &memcg, true)) {
 598                 put_page(page);
 599                 count_vm_event(THP_FAULT_FALLBACK);
 600                 return VM_FAULT_FALLBACK;
 601         }
 602
 603         pgtable = pte_alloc_one(vma->vm_mm);
 604         if (unlikely(!pgtable)) {
 605                 ret = VM_FAULT_OOM;
 606                 goto release;
 607         }
 608
 609         clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
 610         /*
 611          * The memory barrier inside __SetPageUptodate makes sure that
 612          * clear_huge_page writes become visible before the set_pmd_at()
 613          * write.
 614          */
 615         __SetPageUptodate(page);
 616
 617         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 618         if (unlikely(!pmd_none(*vmf->pmd))) {
 619                 goto unlock_release;
 620         } else {
 621                 pmd_t entry;
 622
 623                 ret = check_stable_address_space(vma->vm_mm);
 624                 if (ret)
 625                         goto unlock_release;
 626
 627                 /* Deliver the page fault to userland */
 628                 if (userfaultfd_missing(vma)) {
 629                         vm_fault_t ret2;
 630
 631                         spin_unlock(vmf->ptl);
 632                         mem_cgroup_cancel_charge(page, memcg, true);
 633                         put_page(page);
 634                         pte_free(vma->vm_mm, pgtable);
 635                         ret2 = handle_userfault(vmf, VM_UFFD_MISSING);
 636                         VM_BUG_ON(ret2 & VM_FAULT_FALLBACK);
 637                         return ret2;
 638                 }
 639
 640                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 641                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 642                 page_add_new_anon_rmap(page, vma, haddr, true);
 643                 mem_cgroup_commit_charge(page, memcg, false, true);
 644                 lru_cache_add_active_or_unevictable(page, vma);
 645                 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
 646                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
 647                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 648                 mm_inc_nr_ptes(vma->vm_mm);
 649                 spin_unlock(vmf->ptl);
 650                 count_vm_event(THP_FAULT_ALLOC);
 651                 count_memcg_events(memcg, THP_FAULT_ALLOC, 1);
 652         }
 653
 654         return 0;
 655 unlock_release:
 656         spin_unlock(vmf->ptl);
 657 release:
 658         if (pgtable)
 659                 pte_free(vma->vm_mm, pgtable);
 660         mem_cgroup_cancel_charge(page, memcg, true);
 661         put_page(page);
 662         return ret;
 663
 664 }
 665
 666 /*
 667  * always: directly stall for all thp allocations
 668  * defer: wake kswapd and fail if not immediately available
 669  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
 670  *                fail if not immediately available
 671  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
 672  *          available
 673  * never: never stall for any thp allocation
 674  */
 675 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
 676 {
 677         const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
 678
 679         /* Always do synchronous compaction */
 680         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 681                 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
 682
 683         /* Kick kcompactd and fail quickly */
 684         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 685                 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
 686
 687         /* Synchronous compaction if madvised, otherwise kick kcompactd */
 688         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 689                 return GFP_TRANSHUGE_LIGHT |
 690                         (vma_madvised ? __GFP_DIRECT_RECLAIM :
 691                                         __GFP_KSWAPD_RECLAIM);
 692
 693         /* Only do synchronous compaction if madvised */
 694         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 695                 return GFP_TRANSHUGE_LIGHT |
 696                        (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
 697
 698         return GFP_TRANSHUGE_LIGHT;
 699 }
 700
 701 /* Caller must hold page table lock. */
 702 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 703                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 704                 struct page *zero_page)
 705 {
 706         pmd_t entry;
 707         if (!pmd_none(*pmd))
 708                 return false;
 709         entry = mk_pmd(zero_page, vma->vm_page_prot);
 710         entry = pmd_mkhuge(entry);
 711         if (pgtable)
 712                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 713         set_pmd_at(mm, haddr, pmd, entry);
 714         mm_inc_nr_ptes(mm);
 715         return true;
 716 }
 717
 718 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
 719 {
 720         struct vm_area_struct *vma = vmf->vma;
 721         gfp_t gfp;
 722         struct page *page;
 723         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 724
 725         if (!transhuge_vma_suitable(vma, haddr))
 726                 return VM_FAULT_FALLBACK;
 727         if (unlikely(anon_vma_prepare(vma)))
 728                 return VM_FAULT_OOM;
 729         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 730                 return VM_FAULT_OOM;
 731         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
 732                         !mm_forbids_zeropage(vma->vm_mm) &&
 733                         transparent_hugepage_use_zero_page()) {
 734                 pgtable_t pgtable;
 735                 struct page *zero_page;
 736                 bool set;
 737                 vm_fault_t ret;
 738                 pgtable = pte_alloc_one(vma->vm_mm);
 739                 if (unlikely(!pgtable))
 740                         return VM_FAULT_OOM;
 741                 zero_page = mm_get_huge_zero_page(vma->vm_mm);
 742                 if (unlikely(!zero_page)) {
 743                         pte_free(vma->vm_mm, pgtable);
 744                         count_vm_event(THP_FAULT_FALLBACK);
 745                         return VM_FAULT_FALLBACK;
 746                 }
 747                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 748                 ret = 0;
 749                 set = false;
 750                 if (pmd_none(*vmf->pmd)) {
 751                         ret = check_stable_address_space(vma->vm_mm);
 752                         if (ret) {
 753                                 spin_unlock(vmf->ptl);
 754                         } else if (userfaultfd_missing(vma)) {
 755                                 spin_unlock(vmf->ptl);
 756                                 ret = handle_userfault(vmf, VM_UFFD_MISSING);
 757                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 758                         } else {
 759                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
 760                                                    haddr, vmf->pmd, zero_page);
 761                                 spin_unlock(vmf->ptl);
 762                                 set = true;
 763                         }
 764                 } else
 765                         spin_unlock(vmf->ptl);
 766                 if (!set)
 767                         pte_free(vma->vm_mm, pgtable);
 768                 return ret;
 769         }
 770         gfp = alloc_hugepage_direct_gfpmask(vma);
 771         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 772         if (unlikely(!page)) {
 773                 count_vm_event(THP_FAULT_FALLBACK);
 774                 return VM_FAULT_FALLBACK;
 775         }
 776         prep_transhuge_page(page);
 777         return __do_huge_pmd_anonymous_page(vmf, page, gfp);
 778 }
 779
 780 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 781                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
 782                 pgtable_t pgtable)
 783 {
 784         struct mm_struct *mm = vma->vm_mm;
 785         pmd_t entry;
 786         spinlock_t *ptl;
 787
 788         ptl = pmd_lock(mm, pmd);
 789         if (!pmd_none(*pmd)) {
 790                 if (write) {
 791                         if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
 792                                 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
 793                                 goto out_unlock;
 794                         }
 795                         entry = pmd_mkyoung(*pmd);
 796                         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 797                         if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
 798                                 update_mmu_cache_pmd(vma, addr, pmd);
 799                 }
 800
 801                 goto out_unlock;
 802         }
 803
 804         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 805         if (pfn_t_devmap(pfn))
 806                 entry = pmd_mkdevmap(entry);
 807         if (write) {
 808                 entry = pmd_mkyoung(pmd_mkdirty(entry));
 809                 entry = maybe_pmd_mkwrite(entry, vma);
 810         }
 811
 812         if (pgtable) {
 813                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 814                 mm_inc_nr_ptes(mm);
 815                 pgtable = NULL;
 816         }
 817
 818         set_pmd_at(mm, addr, pmd, entry);
 819         update_mmu_cache_pmd(vma, addr, pmd);
 820
 821 out_unlock:
 822         spin_unlock(ptl);
 823         if (pgtable)
 824                 pte_free(mm, pgtable);
 825 }
 826
 827 vm_fault_t vmf_insert_pfn_pmd(struct vm_fault *vmf, pfn_t pfn, bool write)
 828 {
 829         unsigned long addr = vmf->address & PMD_MASK;
 830         struct vm_area_struct *vma = vmf->vma;
 831         pgprot_t pgprot = vma->vm_page_prot;
 832         pgtable_t pgtable = NULL;
 833
 834         /*
 835          * If we had pmd_special, we could avoid all these restrictions,
 836          * but we need to be consistent with PTEs and architectures that
 837          * can't support a 'special' bit.
 838          */
 839         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 840                         !pfn_t_devmap(pfn));
 841         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 842                                                 (VM_PFNMAP|VM_MIXEDMAP));
 843         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 844
 845         if (addr < vma->vm_start || addr >= vma->vm_end)
 846                 return VM_FAULT_SIGBUS;
 847
 848         if (arch_needs_pgtable_deposit()) {
 849                 pgtable = pte_alloc_one(vma->vm_mm);
 850                 if (!pgtable)
 851                         return VM_FAULT_OOM;
 852         }
 853
 854         track_pfn_insert(vma, &pgprot, pfn);
 855
 856         insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
 857         return VM_FAULT_NOPAGE;
 858 }
 859 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
 860
 861 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 862 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
 863 {
 864         if (likely(vma->vm_flags & VM_WRITE))
 865                 pud = pud_mkwrite(pud);
 866         return pud;
 867 }
 868
 869 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 870                 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
 871 {
 872         struct mm_struct *mm = vma->vm_mm;
 873         pud_t entry;
 874         spinlock_t *ptl;
 875
 876         ptl = pud_lock(mm, pud);
 877         if (!pud_none(*pud)) {
 878                 if (write) {
 879                         if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
 880                                 WARN_ON_ONCE(!is_huge_zero_pud(*pud));
 881                                 goto out_unlock;
 882                         }
 883                         entry = pud_mkyoung(*pud);
 884                         entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
 885                         if (pudp_set_access_flags(vma, addr, pud, entry, 1))
 886                                 update_mmu_cache_pud(vma, addr, pud);
 887                 }
 888                 goto out_unlock;
 889         }
 890
 891         entry = pud_mkhuge(pfn_t_pud(pfn, prot));
 892         if (pfn_t_devmap(pfn))
 893                 entry = pud_mkdevmap(entry);
 894         if (write) {
 895                 entry = pud_mkyoung(pud_mkdirty(entry));
 896                 entry = maybe_pud_mkwrite(entry, vma);
 897         }
 898         set_pud_at(mm, addr, pud, entry);
 899         update_mmu_cache_pud(vma, addr, pud);
 900
 901 out_unlock:
 902         spin_unlock(ptl);
 903 }
 904
 905 vm_fault_t vmf_insert_pfn_pud(struct vm_fault *vmf, pfn_t pfn, bool write)
 906 {
 907         unsigned long addr = vmf->address & PUD_MASK;
 908         struct vm_area_struct *vma = vmf->vma;
 909         pgprot_t pgprot = vma->vm_page_prot;
 910
 911         /*
 912          * If we had pud_special, we could avoid all these restrictions,
 913          * but we need to be consistent with PTEs and architectures that
 914          * can't support a 'special' bit.
 915          */
 916         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 917                         !pfn_t_devmap(pfn));
 918         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 919                                                 (VM_PFNMAP|VM_MIXEDMAP));
 920         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 921
 922         if (addr < vma->vm_start || addr >= vma->vm_end)
 923                 return VM_FAULT_SIGBUS;
 924
 925         track_pfn_insert(vma, &pgprot, pfn);
 926
 927         insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
 928         return VM_FAULT_NOPAGE;
 929 }
 930 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
 931 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
 932
 933 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
 934                 pmd_t *pmd, int flags)
 935 {
 936         pmd_t _pmd;
 937
 938         _pmd = pmd_mkyoung(*pmd);
 939         if (flags & FOLL_WRITE)
 940                 _pmd = pmd_mkdirty(_pmd);
 941         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
 942                                 pmd, _pmd, flags & FOLL_WRITE))
 943                 update_mmu_cache_pmd(vma, addr, pmd);
 944 }
 945
 946 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 947                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
 948 {
 949         unsigned long pfn = pmd_pfn(*pmd);
 950         struct mm_struct *mm = vma->vm_mm;
 951         struct page *page;
 952
 953         assert_spin_locked(pmd_lockptr(mm, pmd));
 954
 955         /*
 956          * When we COW a devmap PMD entry, we split it into PTEs, so we should
 957          * not be in this function with `flags & FOLL_COW` set.
 958          */
 959         WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
 960
 961         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
 962         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
 963                          (FOLL_PIN | FOLL_GET)))
 964                 return NULL;
 965
 966         if (flags & FOLL_WRITE && !pmd_write(*pmd))
 967                 return NULL;
 968
 969         if (pmd_present(*pmd) && pmd_devmap(*pmd))
 970                 /* pass */;
 971         else
 972                 return NULL;
 973
 974         if (flags & FOLL_TOUCH)
 975                 touch_pmd(vma, addr, pmd, flags);
 976
 977         /*
 978          * device mapped pages can only be returned if the
 979          * caller will manage the page reference count.
 980          */
 981         if (!(flags & (FOLL_GET | FOLL_PIN)))
 982                 return ERR_PTR(-EEXIST);
 983
 984         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
 985         *pgmap = get_dev_pagemap(pfn, *pgmap);
 986         if (!*pgmap)
 987                 return ERR_PTR(-EFAULT);
 988         page = pfn_to_page(pfn);
 989         if (!try_grab_page(page, flags))
 990                 page = ERR_PTR(-ENOMEM);
 991
 992         return page;
 993 }
 994
 995 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 996                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 997                   struct vm_area_struct *vma)
 998 {
 999         spinlock_t *dst_ptl, *src_ptl;
1000         struct page *src_page;
1001         pmd_t pmd;
1002         pgtable_t pgtable = NULL;
1003         int ret = -ENOMEM;
1004
1005         /* Skip if can be re-fill on fault */
1006         if (!vma_is_anonymous(vma))
1007                 return 0;
1008
1009         pgtable = pte_alloc_one(dst_mm);
1010         if (unlikely(!pgtable))
1011                 goto out;
1012
1013         dst_ptl = pmd_lock(dst_mm, dst_pmd);
1014         src_ptl = pmd_lockptr(src_mm, src_pmd);
1015         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1016
1017         ret = -EAGAIN;
1018         pmd = *src_pmd;
1019
1020 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1021         if (unlikely(is_swap_pmd(pmd))) {
1022                 swp_entry_t entry = pmd_to_swp_entry(pmd);
1023
1024                 VM_BUG_ON(!is_pmd_migration_entry(pmd));
1025                 if (is_write_migration_entry(entry)) {
1026                         make_migration_entry_read(&entry);
1027                         pmd = swp_entry_to_pmd(entry);
1028                         if (pmd_swp_soft_dirty(*src_pmd))
1029                                 pmd = pmd_swp_mksoft_dirty(pmd);
1030                         set_pmd_at(src_mm, addr, src_pmd, pmd);
1031                 }
1032                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1033                 mm_inc_nr_ptes(dst_mm);
1034                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1035                 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1036                 ret = 0;
1037                 goto out_unlock;
1038         }
1039 #endif
1040
1041         if (unlikely(!pmd_trans_huge(pmd))) {
1042                 pte_free(dst_mm, pgtable);
1043                 goto out_unlock;
1044         }
1045         /*
1046          * When page table lock is held, the huge zero pmd should not be
1047          * under splitting since we don't split the page itself, only pmd to
1048          * a page table.
1049          */
1050         if (is_huge_zero_pmd(pmd)) {
1051                 struct page *zero_page;
1052                 /*
1053                  * get_huge_zero_page() will never allocate a new page here,
1054                  * since we already have a zero page to copy. It just takes a
1055                  * reference.
1056                  */
1057                 zero_page = mm_get_huge_zero_page(dst_mm);
1058                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1059                                 zero_page);
1060                 ret = 0;
1061                 goto out_unlock;
1062         }
1063
1064         src_page = pmd_page(pmd);
1065         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1066         get_page(src_page);
1067         page_dup_rmap(src_page, true);
1068         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1069         mm_inc_nr_ptes(dst_mm);
1070         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1071
1072         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1073         pmd = pmd_mkold(pmd_wrprotect(pmd));
1074         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1075
1076         ret = 0;
1077 out_unlock:
1078         spin_unlock(src_ptl);
1079         spin_unlock(dst_ptl);
1080 out:
1081         return ret;
1082 }
1083
1084 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1085 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1086                 pud_t *pud, int flags)
1087 {
1088         pud_t _pud;
1089
1090         _pud = pud_mkyoung(*pud);
1091         if (flags & FOLL_WRITE)
1092                 _pud = pud_mkdirty(_pud);
1093         if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1094                                 pud, _pud, flags & FOLL_WRITE))
1095                 update_mmu_cache_pud(vma, addr, pud);
1096 }
1097
1098 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1099                 pud_t *pud, int flags, struct dev_pagemap **pgmap)
1100 {
1101         unsigned long pfn = pud_pfn(*pud);
1102         struct mm_struct *mm = vma->vm_mm;
1103         struct page *page;
1104
1105         assert_spin_locked(pud_lockptr(mm, pud));
1106
1107         if (flags & FOLL_WRITE && !pud_write(*pud))
1108                 return NULL;
1109
1110         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
1111         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
1112                          (FOLL_PIN | FOLL_GET)))
1113                 return NULL;
1114
1115         if (pud_present(*pud) && pud_devmap(*pud))
1116                 /* pass */;
1117         else
1118                 return NULL;
1119
1120         if (flags & FOLL_TOUCH)
1121                 touch_pud(vma, addr, pud, flags);
1122
1123         /*
1124          * device mapped pages can only be returned if the
1125          * caller will manage the page reference count.
1126          *
1127          * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here:
1128          */
1129         if (!(flags & (FOLL_GET | FOLL_PIN)))
1130                 return ERR_PTR(-EEXIST);
1131
1132         pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1133         *pgmap = get_dev_pagemap(pfn, *pgmap);
1134         if (!*pgmap)
1135                 return ERR_PTR(-EFAULT);
1136         page = pfn_to_page(pfn);
1137         if (!try_grab_page(page, flags))
1138                 page = ERR_PTR(-ENOMEM);
1139
1140         return page;
1141 }
1142
1143 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1144                   pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1145                   struct vm_area_struct *vma)
1146 {
1147         spinlock_t *dst_ptl, *src_ptl;
1148         pud_t pud;
1149         int ret;
1150
1151         dst_ptl = pud_lock(dst_mm, dst_pud);
1152         src_ptl = pud_lockptr(src_mm, src_pud);
1153         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1154
1155         ret = -EAGAIN;
1156         pud = *src_pud;
1157         if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1158                 goto out_unlock;
1159
1160         /*
1161          * When page table lock is held, the huge zero pud should not be
1162          * under splitting since we don't split the page itself, only pud to
1163          * a page table.
1164          */
1165         if (is_huge_zero_pud(pud)) {
1166                 /* No huge zero pud yet */
1167         }
1168
1169         pudp_set_wrprotect(src_mm, addr, src_pud);
1170         pud = pud_mkold(pud_wrprotect(pud));
1171         set_pud_at(dst_mm, addr, dst_pud, pud);
1172
1173         ret = 0;
1174 out_unlock:
1175         spin_unlock(src_ptl);
1176         spin_unlock(dst_ptl);
1177         return ret;
1178 }
1179
1180 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1181 {
1182         pud_t entry;
1183         unsigned long haddr;
1184         bool write = vmf->flags & FAULT_FLAG_WRITE;
1185
1186         vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1187         if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1188                 goto unlock;
1189
1190         entry = pud_mkyoung(orig_pud);
1191         if (write)
1192                 entry = pud_mkdirty(entry);
1193         haddr = vmf->address & HPAGE_PUD_MASK;
1194         if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1195                 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1196
1197 unlock:
1198         spin_unlock(vmf->ptl);
1199 }
1200 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1201
1202 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1203 {
1204         pmd_t entry;
1205         unsigned long haddr;
1206         bool write = vmf->flags & FAULT_FLAG_WRITE;
1207
1208         vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1209         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1210                 goto unlock;
1211
1212         entry = pmd_mkyoung(orig_pmd);
1213         if (write)
1214                 entry = pmd_mkdirty(entry);
1215         haddr = vmf->address & HPAGE_PMD_MASK;
1216         if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1217                 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1218
1219 unlock:
1220         spin_unlock(vmf->ptl);
1221 }
1222
1223 static vm_fault_t do_huge_pmd_wp_page_fallback(struct vm_fault *vmf,
1224                         pmd_t orig_pmd, struct page *page)
1225 {
1226         struct vm_area_struct *vma = vmf->vma;
1227         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1228         struct mem_cgroup *memcg;
1229         pgtable_t pgtable;
1230         pmd_t _pmd;
1231         int i;
1232         vm_fault_t ret = 0;
1233         struct page **pages;
1234         struct mmu_notifier_range range;
1235
1236         pages = kmalloc_array(HPAGE_PMD_NR, sizeof(struct page *),
1237                               GFP_KERNEL);
1238         if (unlikely(!pages)) {
1239                 ret |= VM_FAULT_OOM;
1240                 goto out;
1241         }
1242
1243         for (i = 0; i < HPAGE_PMD_NR; i++) {
1244                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
1245                                                vmf->address, page_to_nid(page));
1246                 if (unlikely(!pages[i] ||
1247                              mem_cgroup_try_charge_delay(pages[i], vma->vm_mm,
1248                                      GFP_KERNEL, &memcg, false))) {
1249                         if (pages[i])
1250                                 put_page(pages[i]);
1251                         while (--i >= 0) {
1252                                 memcg = (void *)page_private(pages[i]);
1253                                 set_page_private(pages[i], 0);
1254                                 mem_cgroup_cancel_charge(pages[i], memcg,
1255                                                 false);
1256                                 put_page(pages[i]);
1257                         }
1258                         kfree(pages);
1259                         ret |= VM_FAULT_OOM;
1260                         goto out;
1261                 }
1262                 set_page_private(pages[i], (unsigned long)memcg);
1263         }
1264
1265         for (i = 0; i < HPAGE_PMD_NR; i++) {
1266                 copy_user_highpage(pages[i], page + i,
1267                                    haddr + PAGE_SIZE * i, vma);
1268                 __SetPageUptodate(pages[i]);
1269                 cond_resched();
1270         }
1271
1272         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1273                                 haddr, haddr + HPAGE_PMD_SIZE);
1274         mmu_notifier_invalidate_range_start(&range);
1275
1276         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1277         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1278                 goto out_free_pages;
1279         VM_BUG_ON_PAGE(!PageHead(page), page);
1280
1281         /*
1282          * Leave pmd empty until pte is filled note we must notify here as
1283          * concurrent CPU thread might write to new page before the call to
1284          * mmu_notifier_invalidate_range_end() happens which can lead to a
1285          * device seeing memory write in different order than CPU.
1286          *
1287          * See Documentation/vm/mmu_notifier.rst
1288          */
1289         pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1290
1291         pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
1292         pmd_populate(vma->vm_mm, &_pmd, pgtable);
1293
1294         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1295                 pte_t entry;
1296                 entry = mk_pte(pages[i], vma->vm_page_prot);
1297                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1298                 memcg = (void *)page_private(pages[i]);
1299                 set_page_private(pages[i], 0);
1300                 page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
1301                 mem_cgroup_commit_charge(pages[i], memcg, false, false);
1302                 lru_cache_add_active_or_unevictable(pages[i], vma);
1303                 vmf->pte = pte_offset_map(&_pmd, haddr);
1304                 VM_BUG_ON(!pte_none(*vmf->pte));
1305                 set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
1306                 pte_unmap(vmf->pte);
1307         }
1308         kfree(pages);
1309
1310         smp_wmb(); /* make pte visible before pmd */
1311         pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
1312         page_remove_rmap(page, true);
1313         spin_unlock(vmf->ptl);
1314
1315         /*
1316          * No need to double call mmu_notifier->invalidate_range() callback as
1317          * the above pmdp_huge_clear_flush_notify() did already call it.
1318          */
1319         mmu_notifier_invalidate_range_only_end(&range);
1320
1321         ret |= VM_FAULT_WRITE;
1322         put_page(page);
1323
1324 out:
1325         return ret;
1326
1327 out_free_pages:
1328         spin_unlock(vmf->ptl);
1329         mmu_notifier_invalidate_range_end(&range);
1330         for (i = 0; i < HPAGE_PMD_NR; i++) {
1331                 memcg = (void *)page_private(pages[i]);
1332                 set_page_private(pages[i], 0);
1333                 mem_cgroup_cancel_charge(pages[i], memcg, false);
1334                 put_page(pages[i]);
1335         }
1336         kfree(pages);
1337         goto out;
1338 }
1339
1340 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1341 {
1342         struct vm_area_struct *vma = vmf->vma;
1343         struct page *page = NULL, *new_page;
1344         struct mem_cgroup *memcg;
1345         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1346         struct mmu_notifier_range range;
1347         gfp_t huge_gfp;                 /* for allocation and charge */
1348         vm_fault_t ret = 0;
1349
1350         vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1351         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1352         if (is_huge_zero_pmd(orig_pmd))
1353                 goto alloc;
1354         spin_lock(vmf->ptl);
1355         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1356                 goto out_unlock;
1357
1358         page = pmd_page(orig_pmd);
1359         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1360         /*
1361          * We can only reuse the page if nobody else maps the huge page or it's
1362          * part.
1363          */
1364         if (!trylock_page(page)) {
1365                 get_page(page);
1366                 spin_unlock(vmf->ptl);
1367                 lock_page(page);
1368                 spin_lock(vmf->ptl);
1369                 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1370                         unlock_page(page);
1371                         put_page(page);
1372                         goto out_unlock;
1373                 }
1374                 put_page(page);
1375         }
1376         if (reuse_swap_page(page, NULL)) {
1377                 pmd_t entry;
1378                 entry = pmd_mkyoung(orig_pmd);
1379                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1380                 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry,  1))
1381                         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1382                 ret |= VM_FAULT_WRITE;
1383                 unlock_page(page);
1384                 goto out_unlock;
1385         }
1386         unlock_page(page);
1387         get_page(page);
1388         spin_unlock(vmf->ptl);
1389 alloc:
1390         if (__transparent_hugepage_enabled(vma) &&
1391             !transparent_hugepage_debug_cow()) {
1392                 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1393                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1394         } else
1395                 new_page = NULL;
1396
1397         if (likely(new_page)) {
1398                 prep_transhuge_page(new_page);
1399         } else {
1400                 if (!page) {
1401                         split_huge_pmd(vma, vmf->pmd, vmf->address);
1402                         ret |= VM_FAULT_FALLBACK;
1403                 } else {
1404                         ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
1405                         if (ret & VM_FAULT_OOM) {
1406                                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1407                                 ret |= VM_FAULT_FALLBACK;
1408                         }
1409                         put_page(page);
1410                 }
1411                 count_vm_event(THP_FAULT_FALLBACK);
1412                 goto out;
1413         }
1414
1415         if (unlikely(mem_cgroup_try_charge_delay(new_page, vma->vm_mm,
1416                                         huge_gfp, &memcg, true))) {
1417                 put_page(new_page);
1418                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1419                 if (page)
1420                         put_page(page);
1421                 ret |= VM_FAULT_FALLBACK;
1422                 count_vm_event(THP_FAULT_FALLBACK);
1423                 goto out;
1424         }
1425
1426         count_vm_event(THP_FAULT_ALLOC);
1427         count_memcg_events(memcg, THP_FAULT_ALLOC, 1);
1428
1429         if (!page)
1430                 clear_huge_page(new_page, vmf->address, HPAGE_PMD_NR);
1431         else
1432                 copy_user_huge_page(new_page, page, vmf->address,
1433                                     vma, HPAGE_PMD_NR);
1434         __SetPageUptodate(new_page);
1435
1436         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1437                                 haddr, haddr + HPAGE_PMD_SIZE);
1438         mmu_notifier_invalidate_range_start(&range);
1439
1440         spin_lock(vmf->ptl);
1441         if (page)
1442                 put_page(page);
1443         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1444                 spin_unlock(vmf->ptl);
1445                 mem_cgroup_cancel_charge(new_page, memcg, true);
1446                 put_page(new_page);
1447                 goto out_mn;
1448         } else {
1449                 pmd_t entry;
1450                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1451                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1452                 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1453                 page_add_new_anon_rmap(new_page, vma, haddr, true);
1454                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1455                 lru_cache_add_active_or_unevictable(new_page, vma);
1456                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1457                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1458                 if (!page) {
1459                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1460                 } else {
1461                         VM_BUG_ON_PAGE(!PageHead(page), page);
1462                         page_remove_rmap(page, true);
1463                         put_page(page);
1464                 }
1465                 ret |= VM_FAULT_WRITE;
1466         }
1467         spin_unlock(vmf->ptl);
1468 out_mn:
1469         /*
1470          * No need to double call mmu_notifier->invalidate_range() callback as
1471          * the above pmdp_huge_clear_flush_notify() did already call it.
1472          */
1473         mmu_notifier_invalidate_range_only_end(&range);
1474 out:
1475         return ret;
1476 out_unlock:
1477         spin_unlock(vmf->ptl);
1478         return ret;
1479 }
1480
1481 /*
1482  * FOLL_FORCE can write to even unwritable pmd's, but only
1483  * after we've gone through a COW cycle and they are dirty.
1484  */
1485 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1486 {
1487         return pmd_write(pmd) ||
1488                ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1489 }
1490
1491 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1492                                    unsigned long addr,
1493                                    pmd_t *pmd,
1494                                    unsigned int flags)
1495 {
1496         struct mm_struct *mm = vma->vm_mm;
1497         struct page *page = NULL;
1498
1499         assert_spin_locked(pmd_lockptr(mm, pmd));
1500
1501         if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1502                 goto out;
1503
1504         /* Avoid dumping huge zero page */
1505         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1506                 return ERR_PTR(-EFAULT);
1507
1508         /* Full NUMA hinting faults to serialise migration in fault paths */
1509         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1510                 goto out;
1511
1512         page = pmd_page(*pmd);
1513         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1514
1515         if (!try_grab_page(page, flags))
1516                 return ERR_PTR(-ENOMEM);
1517
1518         if (flags & FOLL_TOUCH)
1519                 touch_pmd(vma, addr, pmd, flags);
1520
1521         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1522                 /*
1523                  * We don't mlock() pte-mapped THPs. This way we can avoid
1524                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1525                  *
1526                  * For anon THP:
1527                  *
1528                  * In most cases the pmd is the only mapping of the page as we
1529                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1530                  * writable private mappings in populate_vma_page_range().
1531                  *
1532                  * The only scenario when we have the page shared here is if we
1533                  * mlocking read-only mapping shared over fork(). We skip
1534                  * mlocking such pages.
1535                  *
1536                  * For file THP:
1537                  *
1538                  * We can expect PageDoubleMap() to be stable under page lock:
1539                  * for file pages we set it in page_add_file_rmap(), which
1540                  * requires page to be locked.
1541                  */
1542
1543                 if (PageAnon(page) && compound_mapcount(page) != 1)
1544                         goto skip_mlock;
1545                 if (PageDoubleMap(page) || !page->mapping)
1546                         goto skip_mlock;
1547                 if (!trylock_page(page))
1548                         goto skip_mlock;
1549                 lru_add_drain();
1550                 if (page->mapping && !PageDoubleMap(page))
1551                         mlock_vma_page(page);
1552                 unlock_page(page);
1553         }
1554 skip_mlock:
1555         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1556         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1557
1558 out:
1559         return page;
1560 }
1561
1562 /* NUMA hinting page fault entry point for trans huge pmds */
1563 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1564 {
1565         struct vm_area_struct *vma = vmf->vma;
1566         struct anon_vma *anon_vma = NULL;
1567         struct page *page;
1568         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1569         int page_nid = NUMA_NO_NODE, this_nid = numa_node_id();
1570         int target_nid, last_cpupid = -1;
1571         bool page_locked;
1572         bool migrated = false;
1573         bool was_writable;
1574         int flags = 0;
1575
1576         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1577         if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1578                 goto out_unlock;
1579
1580         /*
1581          * If there are potential migrations, wait for completion and retry
1582          * without disrupting NUMA hinting information. Do not relock and
1583          * check_same as the page may no longer be mapped.
1584          */
1585         if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1586                 page = pmd_page(*vmf->pmd);
1587                 if (!get_page_unless_zero(page))
1588                         goto out_unlock;
1589                 spin_unlock(vmf->ptl);
1590                 put_and_wait_on_page_locked(page);
1591                 goto out;
1592         }
1593
1594         page = pmd_page(pmd);
1595         BUG_ON(is_huge_zero_page(page));
1596         page_nid = page_to_nid(page);
1597         last_cpupid = page_cpupid_last(page);
1598         count_vm_numa_event(NUMA_HINT_FAULTS);
1599         if (page_nid == this_nid) {
1600                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1601                 flags |= TNF_FAULT_LOCAL;
1602         }
1603
1604         /* See similar comment in do_numa_page for explanation */
1605         if (!pmd_savedwrite(pmd))
1606                 flags |= TNF_NO_GROUP;
1607
1608         /*
1609          * Acquire the page lock to serialise THP migrations but avoid dropping
1610          * page_table_lock if at all possible
1611          */
1612         page_locked = trylock_page(page);
1613         target_nid = mpol_misplaced(page, vma, haddr);
1614         if (target_nid == NUMA_NO_NODE) {
1615                 /* If the page was locked, there are no parallel migrations */
1616                 if (page_locked)
1617                         goto clear_pmdnuma;
1618         }
1619
1620         /* Migration could have started since the pmd_trans_migrating check */
1621         if (!page_locked) {
1622                 page_nid = NUMA_NO_NODE;
1623                 if (!get_page_unless_zero(page))
1624                         goto out_unlock;
1625                 spin_unlock(vmf->ptl);
1626                 put_and_wait_on_page_locked(page);
1627                 goto out;
1628         }
1629
1630         /*
1631          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1632          * to serialises splits
1633          */
1634         get_page(page);
1635         spin_unlock(vmf->ptl);
1636         anon_vma = page_lock_anon_vma_read(page);
1637
1638         /* Confirm the PMD did not change while page_table_lock was released */
1639         spin_lock(vmf->ptl);
1640         if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1641                 unlock_page(page);
1642                 put_page(page);
1643                 page_nid = NUMA_NO_NODE;
1644                 goto out_unlock;
1645         }
1646
1647         /* Bail if we fail to protect against THP splits for any reason */
1648         if (unlikely(!anon_vma)) {
1649                 put_page(page);
1650                 page_nid = NUMA_NO_NODE;
1651                 goto clear_pmdnuma;
1652         }
1653
1654         /*
1655          * Since we took the NUMA fault, we must have observed the !accessible
1656          * bit. Make sure all other CPUs agree with that, to avoid them
1657          * modifying the page we're about to migrate.
1658          *
1659          * Must be done under PTL such that we'll observe the relevant
1660          * inc_tlb_flush_pending().
1661          *
1662          * We are not sure a pending tlb flush here is for a huge page
1663          * mapping or not. Hence use the tlb range variant
1664          */
1665         if (mm_tlb_flush_pending(vma->vm_mm)) {
1666                 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1667                 /*
1668                  * change_huge_pmd() released the pmd lock before
1669                  * invalidating the secondary MMUs sharing the primary
1670                  * MMU pagetables (with ->invalidate_range()). The
1671                  * mmu_notifier_invalidate_range_end() (which
1672                  * internally calls ->invalidate_range()) in
1673                  * change_pmd_range() will run after us, so we can't
1674                  * rely on it here and we need an explicit invalidate.
1675                  */
1676                 mmu_notifier_invalidate_range(vma->vm_mm, haddr,
1677                                               haddr + HPAGE_PMD_SIZE);
1678         }
1679
1680         /*
1681          * Migrate the THP to the requested node, returns with page unlocked
1682          * and access rights restored.
1683          */
1684         spin_unlock(vmf->ptl);
1685
1686         migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1687                                 vmf->pmd, pmd, vmf->address, page, target_nid);
1688         if (migrated) {
1689                 flags |= TNF_MIGRATED;
1690                 page_nid = target_nid;
1691         } else
1692                 flags |= TNF_MIGRATE_FAIL;
1693
1694         goto out;
1695 clear_pmdnuma:
1696         BUG_ON(!PageLocked(page));
1697         was_writable = pmd_savedwrite(pmd);
1698         pmd = pmd_modify(pmd, vma->vm_page_prot);
1699         pmd = pmd_mkyoung(pmd);
1700         if (was_writable)
1701                 pmd = pmd_mkwrite(pmd);
1702         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1703         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1704         unlock_page(page);
1705 out_unlock:
1706         spin_unlock(vmf->ptl);
1707
1708 out:
1709         if (anon_vma)
1710                 page_unlock_anon_vma_read(anon_vma);
1711
1712         if (page_nid != NUMA_NO_NODE)
1713                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1714                                 flags);
1715
1716         return 0;
1717 }
1718
1719 /*
1720  * Return true if we do MADV_FREE successfully on entire pmd page.
1721  * Otherwise, return false.
1722  */
1723 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1724                 pmd_t *pmd, unsigned long addr, unsigned long next)
1725 {
1726         spinlock_t *ptl;
1727         pmd_t orig_pmd;
1728         struct page *page;
1729         struct mm_struct *mm = tlb->mm;
1730         bool ret = false;
1731
1732         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1733
1734         ptl = pmd_trans_huge_lock(pmd, vma);
1735         if (!ptl)
1736                 goto out_unlocked;
1737
1738         orig_pmd = *pmd;
1739         if (is_huge_zero_pmd(orig_pmd))
1740                 goto out;
1741
1742         if (unlikely(!pmd_present(orig_pmd))) {
1743                 VM_BUG_ON(thp_migration_supported() &&
1744                                   !is_pmd_migration_entry(orig_pmd));
1745                 goto out;
1746         }
1747
1748         page = pmd_page(orig_pmd);
1749         /*
1750          * If other processes are mapping this page, we couldn't discard
1751          * the page unless they all do MADV_FREE so let's skip the page.
1752          */
1753         if (page_mapcount(page) != 1)
1754                 goto out;
1755
1756         if (!trylock_page(page))
1757                 goto out;
1758
1759         /*
1760          * If user want to discard part-pages of THP, split it so MADV_FREE
1761          * will deactivate only them.
1762          */
1763         if (next - addr != HPAGE_PMD_SIZE) {
1764                 get_page(page);
1765                 spin_unlock(ptl);
1766                 split_huge_page(page);
1767                 unlock_page(page);
1768                 put_page(page);
1769                 goto out_unlocked;
1770         }
1771
1772         if (PageDirty(page))
1773                 ClearPageDirty(page);
1774         unlock_page(page);
1775
1776         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1777                 pmdp_invalidate(vma, addr, pmd);
1778                 orig_pmd = pmd_mkold(orig_pmd);
1779                 orig_pmd = pmd_mkclean(orig_pmd);
1780
1781                 set_pmd_at(mm, addr, pmd, orig_pmd);
1782                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1783         }
1784
1785         mark_page_lazyfree(page);
1786         ret = true;
1787 out:
1788         spin_unlock(ptl);
1789 out_unlocked:
1790         return ret;
1791 }
1792
1793 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1794 {
1795         pgtable_t pgtable;
1796
1797         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1798         pte_free(mm, pgtable);
1799         mm_dec_nr_ptes(mm);
1800 }
1801
1802 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1803                  pmd_t *pmd, unsigned long addr)
1804 {
1805         pmd_t orig_pmd;
1806         spinlock_t *ptl;
1807
1808         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1809
1810         ptl = __pmd_trans_huge_lock(pmd, vma);
1811         if (!ptl)
1812                 return 0;
1813         /*
1814          * For architectures like ppc64 we look at deposited pgtable
1815          * when calling pmdp_huge_get_and_clear. So do the
1816          * pgtable_trans_huge_withdraw after finishing pmdp related
1817          * operations.
1818          */
1819         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1820                         tlb->fullmm);
1821         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1822         if (vma_is_dax(vma)) {
1823                 if (arch_needs_pgtable_deposit())
1824                         zap_deposited_table(tlb->mm, pmd);
1825                 spin_unlock(ptl);
1826                 if (is_huge_zero_pmd(orig_pmd))
1827                         tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1828         } else if (is_huge_zero_pmd(orig_pmd)) {
1829                 zap_deposited_table(tlb->mm, pmd);
1830                 spin_unlock(ptl);
1831                 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1832         } else {
1833                 struct page *page = NULL;
1834                 int flush_needed = 1;
1835
1836                 if (pmd_present(orig_pmd)) {
1837                         page = pmd_page(orig_pmd);
1838                         page_remove_rmap(page, true);
1839                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1840                         VM_BUG_ON_PAGE(!PageHead(page), page);
1841                 } else if (thp_migration_supported()) {
1842                         swp_entry_t entry;
1843
1844                         VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1845                         entry = pmd_to_swp_entry(orig_pmd);
1846                         page = pfn_to_page(swp_offset(entry));
1847                         flush_needed = 0;
1848                 } else
1849                         WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1850
1851                 if (PageAnon(page)) {
1852                         zap_deposited_table(tlb->mm, pmd);
1853                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1854                 } else {
1855                         if (arch_needs_pgtable_deposit())
1856                                 zap_deposited_table(tlb->mm, pmd);
1857                         add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1858                 }
1859
1860                 spin_unlock(ptl);
1861                 if (flush_needed)
1862                         tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1863         }
1864         return 1;
1865 }
1866
1867 #ifndef pmd_move_must_withdraw
1868 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1869                                          spinlock_t *old_pmd_ptl,
1870                                          struct vm_area_struct *vma)
1871 {
1872         /*
1873          * With split pmd lock we also need to move preallocated
1874          * PTE page table if new_pmd is on different PMD page table.
1875          *
1876          * We also don't deposit and withdraw tables for file pages.
1877          */
1878         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1879 }
1880 #endif
1881
1882 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1883 {
1884 #ifdef CONFIG_MEM_SOFT_DIRTY
1885         if (unlikely(is_pmd_migration_entry(pmd)))
1886                 pmd = pmd_swp_mksoft_dirty(pmd);
1887         else if (pmd_present(pmd))
1888                 pmd = pmd_mksoft_dirty(pmd);
1889 #endif
1890         return pmd;
1891 }
1892
1893 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1894                   unsigned long new_addr, unsigned long old_end,
1895                   pmd_t *old_pmd, pmd_t *new_pmd)
1896 {
1897         spinlock_t *old_ptl, *new_ptl;
1898         pmd_t pmd;
1899         struct mm_struct *mm = vma->vm_mm;
1900         bool force_flush = false;
1901
1902         if ((old_addr & ~HPAGE_PMD_MASK) ||
1903             (new_addr & ~HPAGE_PMD_MASK) ||
1904             old_end - old_addr < HPAGE_PMD_SIZE)
1905                 return false;
1906
1907         /*
1908          * The destination pmd shouldn't be established, free_pgtables()
1909          * should have release it.
1910          */
1911         if (WARN_ON(!pmd_none(*new_pmd))) {
1912                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1913                 return false;
1914         }
1915
1916         /*
1917          * We don't have to worry about the ordering of src and dst
1918          * ptlocks because exclusive mmap_sem prevents deadlock.
1919          */
1920         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1921         if (old_ptl) {
1922                 new_ptl = pmd_lockptr(mm, new_pmd);
1923                 if (new_ptl != old_ptl)
1924                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1925                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1926                 if (pmd_present(pmd))
1927                         force_flush = true;
1928                 VM_BUG_ON(!pmd_none(*new_pmd));
1929
1930                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1931                         pgtable_t pgtable;
1932                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1933                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1934                 }
1935                 pmd = move_soft_dirty_pmd(pmd);
1936                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1937                 if (force_flush)
1938                         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1939                 if (new_ptl != old_ptl)
1940                         spin_unlock(new_ptl);
1941                 spin_unlock(old_ptl);
1942                 return true;
1943         }
1944         return false;
1945 }
1946
1947 /*
1948  * Returns
1949  *  - 0 if PMD could not be locked
1950  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1951  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1952  */
1953 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1954                 unsigned long addr, pgprot_t newprot, int prot_numa)
1955 {
1956         struct mm_struct *mm = vma->vm_mm;
1957         spinlock_t *ptl;
1958         pmd_t entry;
1959         bool preserve_write;
1960         int ret;
1961
1962         ptl = __pmd_trans_huge_lock(pmd, vma);
1963         if (!ptl)
1964                 return 0;
1965
1966         preserve_write = prot_numa && pmd_write(*pmd);
1967         ret = 1;
1968
1969 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1970         if (is_swap_pmd(*pmd)) {
1971                 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1972
1973                 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1974                 if (is_write_migration_entry(entry)) {
1975                         pmd_t newpmd;
1976                         /*
1977                          * A protection check is difficult so
1978                          * just be safe and disable write
1979                          */
1980                         make_migration_entry_read(&entry);
1981                         newpmd = swp_entry_to_pmd(entry);
1982                         if (pmd_swp_soft_dirty(*pmd))
1983                                 newpmd = pmd_swp_mksoft_dirty(newpmd);
1984                         set_pmd_at(mm, addr, pmd, newpmd);
1985                 }
1986                 goto unlock;
1987         }
1988 #endif
1989
1990         /*
1991          * Avoid trapping faults against the zero page. The read-only
1992          * data is likely to be read-cached on the local CPU and
1993          * local/remote hits to the zero page are not interesting.
1994          */
1995         if (prot_numa && is_huge_zero_pmd(*pmd))
1996                 goto unlock;
1997
1998         if (prot_numa && pmd_protnone(*pmd))
1999                 goto unlock;
2000
2001         /*
2002          * In case prot_numa, we are under down_read(mmap_sem). It's critical
2003          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
2004          * which is also under down_read(mmap_sem):
2005          *
2006          *      CPU0:                           CPU1:
2007          *                              change_huge_pmd(prot_numa=1)
2008          *                               pmdp_huge_get_and_clear_notify()
2009          * madvise_dontneed()
2010          *  zap_pmd_range()
2011          *   pmd_trans_huge(*pmd) == 0 (without ptl)
2012          *   // skip the pmd
2013          *                               set_pmd_at();
2014          *                               // pmd is re-established
2015          *
2016          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
2017          * which may break userspace.
2018          *
2019          * pmdp_invalidate() is required to make sure we don't miss
2020          * dirty/young flags set by hardware.
2021          */
2022         entry = pmdp_invalidate(vma, addr, pmd);
2023
2024         entry = pmd_modify(entry, newprot);
2025         if (preserve_write)
2026                 entry = pmd_mk_savedwrite(entry);
2027         ret = HPAGE_PMD_NR;
2028         set_pmd_at(mm, addr, pmd, entry);
2029         BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
2030 unlock:
2031         spin_unlock(ptl);
2032         return ret;
2033 }
2034
2035 /*
2036  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
2037  *
2038  * Note that if it returns page table lock pointer, this routine returns without
2039  * unlocking page table lock. So callers must unlock it.
2040  */
2041 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
2042 {
2043         spinlock_t *ptl;
2044         ptl = pmd_lock(vma->vm_mm, pmd);
2045         if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
2046                         pmd_devmap(*pmd)))
2047                 return ptl;
2048         spin_unlock(ptl);
2049         return NULL;
2050 }
2051
2052 /*
2053  * Returns true if a given pud maps a thp, false otherwise.
2054  *
2055  * Note that if it returns true, this routine returns without unlocking page
2056  * table lock. So callers must unlock it.
2057  */
2058 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
2059 {
2060         spinlock_t *ptl;
2061
2062         ptl = pud_lock(vma->vm_mm, pud);
2063         if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
2064                 return ptl;
2065         spin_unlock(ptl);
2066         return NULL;
2067 }
2068
2069 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
2070 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
2071                  pud_t *pud, unsigned long addr)
2072 {
2073         spinlock_t *ptl;
2074
2075         ptl = __pud_trans_huge_lock(pud, vma);
2076         if (!ptl)
2077                 return 0;
2078         /*
2079          * For architectures like ppc64 we look at deposited pgtable
2080          * when calling pudp_huge_get_and_clear. So do the
2081          * pgtable_trans_huge_withdraw after finishing pudp related
2082          * operations.
2083          */
2084         pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
2085         tlb_remove_pud_tlb_entry(tlb, pud, addr);
2086         if (vma_is_dax(vma)) {
2087                 spin_unlock(ptl);
2088                 /* No zero page support yet */
2089         } else {
2090                 /* No support for anonymous PUD pages yet */
2091                 BUG();
2092         }
2093         return 1;
2094 }
2095
2096 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
2097                 unsigned long haddr)
2098 {
2099         VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
2100         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2101         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
2102         VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
2103
2104         count_vm_event(THP_SPLIT_PUD);
2105
2106         pudp_huge_clear_flush_notify(vma, haddr, pud);
2107 }
2108
2109 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2110                 unsigned long address)
2111 {
2112         spinlock_t *ptl;
2113         struct mmu_notifier_range range;
2114
2115         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2116                                 address & HPAGE_PUD_MASK,
2117                                 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
2118         mmu_notifier_invalidate_range_start(&range);
2119         ptl = pud_lock(vma->vm_mm, pud);
2120         if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2121                 goto out;
2122         __split_huge_pud_locked(vma, pud, range.start);
2123
2124 out:
2125         spin_unlock(ptl);
2126         /*
2127          * No need to double call mmu_notifier->invalidate_range() callback as
2128          * the above pudp_huge_clear_flush_notify() did already call it.
2129          */
2130         mmu_notifier_invalidate_range_only_end(&range);
2131 }
2132 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2133
2134 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2135                 unsigned long haddr, pmd_t *pmd)
2136 {
2137         struct mm_struct *mm = vma->vm_mm;
2138         pgtable_t pgtable;
2139         pmd_t _pmd;
2140         int i;
2141
2142         /*
2143          * Leave pmd empty until pte is filled note that it is fine to delay
2144          * notification until mmu_notifier_invalidate_range_end() as we are
2145          * replacing a zero pmd write protected page with a zero pte write
2146          * protected page.
2147          *
2148          * See Documentation/vm/mmu_notifier.rst
2149          */
2150         pmdp_huge_clear_flush(vma, haddr, pmd);
2151
2152         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2153         pmd_populate(mm, &_pmd, pgtable);
2154
2155         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2156                 pte_t *pte, entry;
2157                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2158                 entry = pte_mkspecial(entry);
2159                 pte = pte_offset_map(&_pmd, haddr);
2160                 VM_BUG_ON(!pte_none(*pte));
2161                 set_pte_at(mm, haddr, pte, entry);
2162                 pte_unmap(pte);
2163         }
2164         smp_wmb(); /* make pte visible before pmd */
2165         pmd_populate(mm, pmd, pgtable);
2166 }
2167
2168 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2169                 unsigned long haddr, bool freeze)
2170 {
2171         struct mm_struct *mm = vma->vm_mm;
2172         struct page *page;
2173         pgtable_t pgtable;
2174         pmd_t old_pmd, _pmd;
2175         bool young, write, soft_dirty, pmd_migration = false;
2176         unsigned long addr;
2177         int i;
2178
2179         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2180         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2181         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2182         VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2183                                 && !pmd_devmap(*pmd));
2184
2185         count_vm_event(THP_SPLIT_PMD);
2186
2187         if (!vma_is_anonymous(vma)) {
2188                 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2189                 /*
2190                  * We are going to unmap this huge page. So
2191                  * just go ahead and zap it
2192                  */
2193                 if (arch_needs_pgtable_deposit())
2194                         zap_deposited_table(mm, pmd);
2195                 if (vma_is_dax(vma))
2196                         return;
2197                 page = pmd_page(_pmd);
2198                 if (!PageDirty(page) && pmd_dirty(_pmd))
2199                         set_page_dirty(page);
2200                 if (!PageReferenced(page) && pmd_young(_pmd))
2201                         SetPageReferenced(page);
2202                 page_remove_rmap(page, true);
2203                 put_page(page);
2204                 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2205                 return;
2206         } else if (is_huge_zero_pmd(*pmd)) {
2207                 /*
2208                  * FIXME: Do we want to invalidate secondary mmu by calling
2209                  * mmu_notifier_invalidate_range() see comments below inside
2210                  * __split_huge_pmd() ?
2211                  *
2212                  * We are going from a zero huge page write protected to zero
2213                  * small page also write protected so it does not seems useful
2214                  * to invalidate secondary mmu at this time.
2215                  */
2216                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2217         }
2218
2219         /*
2220          * Up to this point the pmd is present and huge and userland has the
2221          * whole access to the hugepage during the split (which happens in
2222          * place). If we overwrite the pmd with the not-huge version pointing
2223          * to the pte here (which of course we could if all CPUs were bug
2224          * free), userland could trigger a small page size TLB miss on the
2225          * small sized TLB while the hugepage TLB entry is still established in
2226          * the huge TLB. Some CPU doesn't like that.
2227          * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2228          * 383 on page 93. Intel should be safe but is also warns that it's
2229          * only safe if the permission and cache attributes of the two entries
2230          * loaded in the two TLB is identical (which should be the case here).
2231          * But it is generally safer to never allow small and huge TLB entries
2232          * for the same virtual address to be loaded simultaneously. So instead
2233          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2234          * current pmd notpresent (atomically because here the pmd_trans_huge
2235          * must remain set at all times on the pmd until the split is complete
2236          * for this pmd), then we flush the SMP TLB and finally we write the
2237          * non-huge version of the pmd entry with pmd_populate.
2238          */
2239         old_pmd = pmdp_invalidate(vma, haddr, pmd);
2240
2241         pmd_migration = is_pmd_migration_entry(old_pmd);
2242         if (unlikely(pmd_migration)) {
2243                 swp_entry_t entry;
2244
2245                 entry = pmd_to_swp_entry(old_pmd);
2246                 page = pfn_to_page(swp_offset(entry));
2247                 write = is_write_migration_entry(entry);
2248                 young = false;
2249                 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2250         } else {
2251                 page = pmd_page(old_pmd);
2252                 if (pmd_dirty(old_pmd))
2253                         SetPageDirty(page);
2254                 write = pmd_write(old_pmd);
2255                 young = pmd_young(old_pmd);
2256                 soft_dirty = pmd_soft_dirty(old_pmd);
2257         }
2258         VM_BUG_ON_PAGE(!page_count(page), page);
2259         page_ref_add(page, HPAGE_PMD_NR - 1);
2260
2261         /*
2262          * Withdraw the table only after we mark the pmd entry invalid.
2263          * This's critical for some architectures (Power).
2264          */
2265         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2266         pmd_populate(mm, &_pmd, pgtable);
2267
2268         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2269                 pte_t entry, *pte;
2270                 /*
2271                  * Note that NUMA hinting access restrictions are not
2272                  * transferred to avoid any possibility of altering
2273                  * permissions across VMAs.
2274                  */
2275                 if (freeze || pmd_migration) {
2276                         swp_entry_t swp_entry;
2277                         swp_entry = make_migration_entry(page + i, write);
2278                         entry = swp_entry_to_pte(swp_entry);
2279                         if (soft_dirty)
2280                                 entry = pte_swp_mksoft_dirty(entry);
2281                 } else {
2282                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2283                         entry = maybe_mkwrite(entry, vma);
2284                         if (!write)
2285                                 entry = pte_wrprotect(entry);
2286                         if (!young)
2287                                 entry = pte_mkold(entry);
2288                         if (soft_dirty)
2289                                 entry = pte_mksoft_dirty(entry);
2290                 }
2291                 pte = pte_offset_map(&_pmd, addr);
2292                 BUG_ON(!pte_none(*pte));
2293                 set_pte_at(mm, addr, pte, entry);
2294                 atomic_inc(&page[i]._mapcount);
2295                 pte_unmap(pte);
2296         }
2297
2298         /*
2299          * Set PG_double_map before dropping compound_mapcount to avoid
2300          * false-negative page_mapped().
2301          */
2302         if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2303                 for (i = 0; i < HPAGE_PMD_NR; i++)
2304                         atomic_inc(&page[i]._mapcount);
2305         }
2306
2307         if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2308                 /* Last compound_mapcount is gone. */
2309                 __dec_node_page_state(page, NR_ANON_THPS);
2310                 if (TestClearPageDoubleMap(page)) {
2311                         /* No need in mapcount reference anymore */
2312                         for (i = 0; i < HPAGE_PMD_NR; i++)
2313                                 atomic_dec(&page[i]._mapcount);
2314                 }
2315         }
2316
2317         smp_wmb(); /* make pte visible before pmd */
2318         pmd_populate(mm, pmd, pgtable);
2319
2320         if (freeze) {
2321                 for (i = 0; i < HPAGE_PMD_NR; i++) {
2322                         page_remove_rmap(page + i, false);
2323                         put_page(page + i);
2324                 }
2325         }
2326 }
2327
2328 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2329                 unsigned long address, bool freeze, struct page *page)
2330 {
2331         spinlock_t *ptl;
2332         struct mmu_notifier_range range;
2333
2334         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2335                                 address & HPAGE_PMD_MASK,
2336                                 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2337         mmu_notifier_invalidate_range_start(&range);
2338         ptl = pmd_lock(vma->vm_mm, pmd);
2339
2340         /*
2341          * If caller asks to setup a migration entries, we need a page to check
2342          * pmd against. Otherwise we can end up replacing wrong page.
2343          */
2344         VM_BUG_ON(freeze && !page);
2345         if (page && page != pmd_page(*pmd))
2346                 goto out;
2347
2348         if (pmd_trans_huge(*pmd)) {
2349                 page = pmd_page(*pmd);
2350                 if (PageMlocked(page))
2351                         clear_page_mlock(page);
2352         } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2353                 goto out;
2354         __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2355 out:
2356         spin_unlock(ptl);
2357         /*
2358          * No need to double call mmu_notifier->invalidate_range() callback.
2359          * They are 3 cases to consider inside __split_huge_pmd_locked():
2360          *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2361          *  2) __split_huge_zero_page_pmd() read only zero page and any write
2362          *    fault will trigger a flush_notify before pointing to a new page
2363          *    (it is fine if the secondary mmu keeps pointing to the old zero
2364          *    page in the meantime)
2365          *  3) Split a huge pmd into pte pointing to the same page. No need
2366          *     to invalidate secondary tlb entry they are all still valid.
2367          *     any further changes to individual pte will notify. So no need
2368          *     to call mmu_notifier->invalidate_range()
2369          */
2370         mmu_notifier_invalidate_range_only_end(&range);
2371 }
2372
2373 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2374                 bool freeze, struct page *page)
2375 {
2376         pgd_t *pgd;
2377         p4d_t *p4d;
2378         pud_t *pud;
2379         pmd_t *pmd;
2380
2381         pgd = pgd_offset(vma->vm_mm, address);
2382         if (!pgd_present(*pgd))
2383                 return;
2384
2385         p4d = p4d_offset(pgd, address);
2386         if (!p4d_present(*p4d))
2387                 return;
2388
2389         pud = pud_offset(p4d, address);
2390         if (!pud_present(*pud))
2391                 return;
2392
2393         pmd = pmd_offset(pud, address);
2394
2395         __split_huge_pmd(vma, pmd, address, freeze, page);
2396 }
2397
2398 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2399                              unsigned long start,
2400                              unsigned long end,
2401                              long adjust_next)
2402 {
2403         /*
2404          * If the new start address isn't hpage aligned and it could
2405          * previously contain an hugepage: check if we need to split
2406          * an huge pmd.
2407          */
2408         if (start & ~HPAGE_PMD_MASK &&
2409             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2410             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2411                 split_huge_pmd_address(vma, start, false, NULL);
2412
2413         /*
2414          * If the new end address isn't hpage aligned and it could
2415          * previously contain an hugepage: check if we need to split
2416          * an huge pmd.
2417          */
2418         if (end & ~HPAGE_PMD_MASK &&
2419             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2420             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2421                 split_huge_pmd_address(vma, end, false, NULL);
2422
2423         /*
2424          * If we're also updating the vma->vm_next->vm_start, if the new
2425          * vm_next->vm_start isn't page aligned and it could previously
2426          * contain an hugepage: check if we need to split an huge pmd.
2427          */
2428         if (adjust_next > 0) {
2429                 struct vm_area_struct *next = vma->vm_next;
2430                 unsigned long nstart = next->vm_start;
2431                 nstart += adjust_next << PAGE_SHIFT;
2432                 if (nstart & ~HPAGE_PMD_MASK &&
2433                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2434                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2435                         split_huge_pmd_address(next, nstart, false, NULL);
2436         }
2437 }
2438
2439 static void unmap_page(struct page *page)
2440 {
2441         enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
2442                 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2443         bool unmap_success;
2444
2445         VM_BUG_ON_PAGE(!PageHead(page), page);
2446
2447         if (PageAnon(page))
2448                 ttu_flags |= TTU_SPLIT_FREEZE;
2449
2450         unmap_success = try_to_unmap(page, ttu_flags);
2451         VM_BUG_ON_PAGE(!unmap_success, page);
2452 }
2453
2454 static void remap_page(struct page *page)
2455 {
2456         int i;
2457         if (PageTransHuge(page)) {
2458                 remove_migration_ptes(page, page, true);
2459         } else {
2460                 for (i = 0; i < HPAGE_PMD_NR; i++)
2461                         remove_migration_ptes(page + i, page + i, true);
2462         }
2463 }
2464
2465 static void __split_huge_page_tail(struct page *head, int tail,
2466                 struct lruvec *lruvec, struct list_head *list)
2467 {
2468         struct page *page_tail = head + tail;
2469
2470         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2471
2472         /*
2473          * Clone page flags before unfreezing refcount.
2474          *
2475          * After successful get_page_unless_zero() might follow flags change,
2476          * for exmaple lock_page() which set PG_waiters.
2477          */
2478         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2479         page_tail->flags |= (head->flags &
2480                         ((1L << PG_referenced) |
2481                          (1L << PG_swapbacked) |
2482                          (1L << PG_swapcache) |
2483                          (1L << PG_mlocked) |
2484                          (1L << PG_uptodate) |
2485                          (1L << PG_active) |
2486                          (1L << PG_workingset) |
2487                          (1L << PG_locked) |
2488                          (1L << PG_unevictable) |
2489                          (1L << PG_dirty)));
2490
2491         /* ->mapping in first tail page is compound_mapcount */
2492         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2493                         page_tail);
2494         page_tail->mapping = head->mapping;
2495         page_tail->index = head->index + tail;
2496
2497         /* Page flags must be visible before we make the page non-compound. */
2498         smp_wmb();
2499
2500         /*
2501          * Clear PageTail before unfreezing page refcount.
2502          *
2503          * After successful get_page_unless_zero() might follow put_page()
2504          * which needs correct compound_head().
2505          */
2506         clear_compound_head(page_tail);
2507
2508         /* Finally unfreeze refcount. Additional reference from page cache. */
2509         page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2510                                           PageSwapCache(head)));
2511
2512         if (page_is_young(head))
2513                 set_page_young(page_tail);
2514         if (page_is_idle(head))
2515                 set_page_idle(page_tail);
2516
2517         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2518
2519         /*
2520          * always add to the tail because some iterators expect new
2521          * pages to show after the currently processed elements - e.g.
2522          * migrate_pages
2523          */
2524         lru_add_page_tail(head, page_tail, lruvec, list);
2525 }
2526
2527 static void __split_huge_page(struct page *page, struct list_head *list,
2528                 pgoff_t end, unsigned long flags)
2529 {
2530         struct page *head = compound_head(page);
2531         pg_data_t *pgdat = page_pgdat(head);
2532         struct lruvec *lruvec;
2533         struct address_space *swap_cache = NULL;
2534         unsigned long offset = 0;
2535         int i;
2536
2537         lruvec = mem_cgroup_page_lruvec(head, pgdat);
2538
2539         /* complete memcg works before add pages to LRU */
2540         mem_cgroup_split_huge_fixup(head);
2541
2542         if (PageAnon(head) && PageSwapCache(head)) {
2543                 swp_entry_t entry = { .val = page_private(head) };
2544
2545                 offset = swp_offset(entry);
2546                 swap_cache = swap_address_space(entry);
2547                 xa_lock(&swap_cache->i_pages);
2548         }
2549
2550         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
2551                 __split_huge_page_tail(head, i, lruvec, list);
2552                 /* Some pages can be beyond i_size: drop them from page cache */
2553                 if (head[i].index >= end) {
2554                         ClearPageDirty(head + i);
2555                         __delete_from_page_cache(head + i, NULL);
2556                         if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2557                                 shmem_uncharge(head->mapping->host, 1);
2558                         put_page(head + i);
2559                 } else if (!PageAnon(page)) {
2560                         __xa_store(&head->mapping->i_pages, head[i].index,
2561                                         head + i, 0);
2562                 } else if (swap_cache) {
2563                         __xa_store(&swap_cache->i_pages, offset + i,
2564                                         head + i, 0);
2565                 }
2566         }
2567
2568         ClearPageCompound(head);
2569
2570         split_page_owner(head, HPAGE_PMD_ORDER);
2571
2572         /* See comment in __split_huge_page_tail() */
2573         if (PageAnon(head)) {
2574                 /* Additional pin to swap cache */
2575                 if (PageSwapCache(head)) {
2576                         page_ref_add(head, 2);
2577                         xa_unlock(&swap_cache->i_pages);
2578                 } else {
2579                         page_ref_inc(head);
2580                 }
2581         } else {
2582                 /* Additional pin to page cache */
2583                 page_ref_add(head, 2);
2584                 xa_unlock(&head->mapping->i_pages);
2585         }
2586
2587         spin_unlock_irqrestore(&pgdat->lru_lock, flags);
2588
2589         remap_page(head);
2590
2591         for (i = 0; i < HPAGE_PMD_NR; i++) {
2592                 struct page *subpage = head + i;
2593                 if (subpage == page)
2594                         continue;
2595                 unlock_page(subpage);
2596
2597                 /*
2598                  * Subpages may be freed if there wasn't any mapping
2599                  * like if add_to_swap() is running on a lru page that
2600                  * had its mapping zapped. And freeing these pages
2601                  * requires taking the lru_lock so we do the put_page
2602                  * of the tail pages after the split is complete.
2603                  */
2604                 put_page(subpage);
2605         }
2606 }
2607
2608 int total_mapcount(struct page *page)
2609 {
2610         int i, compound, ret;
2611
2612         VM_BUG_ON_PAGE(PageTail(page), page);
2613
2614         if (likely(!PageCompound(page)))
2615                 return atomic_read(&page->_mapcount) + 1;
2616
2617         compound = compound_mapcount(page);
2618         if (PageHuge(page))
2619                 return compound;
2620         ret = compound;
2621         for (i = 0; i < HPAGE_PMD_NR; i++)
2622                 ret += atomic_read(&page[i]._mapcount) + 1;
2623         /* File pages has compound_mapcount included in _mapcount */
2624         if (!PageAnon(page))
2625                 return ret - compound * HPAGE_PMD_NR;
2626         if (PageDoubleMap(page))
2627                 ret -= HPAGE_PMD_NR;
2628         return ret;
2629 }
2630
2631 /*
2632  * This calculates accurately how many mappings a transparent hugepage
2633  * has (unlike page_mapcount() which isn't fully accurate). This full
2634  * accuracy is primarily needed to know if copy-on-write faults can
2635  * reuse the page and change the mapping to read-write instead of
2636  * copying them. At the same time this returns the total_mapcount too.
2637  *
2638  * The function returns the highest mapcount any one of the subpages
2639  * has. If the return value is one, even if different processes are
2640  * mapping different subpages of the transparent hugepage, they can
2641  * all reuse it, because each process is reusing a different subpage.
2642  *
2643  * The total_mapcount is instead counting all virtual mappings of the
2644  * subpages. If the total_mapcount is equal to "one", it tells the
2645  * caller all mappings belong to the same "mm" and in turn the
2646  * anon_vma of the transparent hugepage can become the vma->anon_vma
2647  * local one as no other process may be mapping any of the subpages.
2648  *
2649  * It would be more accurate to replace page_mapcount() with
2650  * page_trans_huge_mapcount(), however we only use
2651  * page_trans_huge_mapcount() in the copy-on-write faults where we
2652  * need full accuracy to avoid breaking page pinning, because
2653  * page_trans_huge_mapcount() is slower than page_mapcount().
2654  */
2655 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2656 {
2657         int i, ret, _total_mapcount, mapcount;
2658
2659         /* hugetlbfs shouldn't call it */
2660         VM_BUG_ON_PAGE(PageHuge(page), page);
2661
2662         if (likely(!PageTransCompound(page))) {
2663                 mapcount = atomic_read(&page->_mapcount) + 1;
2664                 if (total_mapcount)
2665                         *total_mapcount = mapcount;
2666                 return mapcount;
2667         }
2668
2669         page = compound_head(page);
2670
2671         _total_mapcount = ret = 0;
2672         for (i = 0; i < HPAGE_PMD_NR; i++) {
2673                 mapcount = atomic_read(&page[i]._mapcount) + 1;
2674                 ret = max(ret, mapcount);
2675                 _total_mapcount += mapcount;
2676         }
2677         if (PageDoubleMap(page)) {
2678                 ret -= 1;
2679                 _total_mapcount -= HPAGE_PMD_NR;
2680         }
2681         mapcount = compound_mapcount(page);
2682         ret += mapcount;
2683         _total_mapcount += mapcount;
2684         if (total_mapcount)
2685                 *total_mapcount = _total_mapcount;
2686         return ret;
2687 }
2688
2689 /* Racy check whether the huge page can be split */
2690 bool can_split_huge_page(struct page *page, int *pextra_pins)
2691 {
2692         int extra_pins;
2693
2694         /* Additional pins from page cache */
2695         if (PageAnon(page))
2696                 extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0;
2697         else
2698                 extra_pins = HPAGE_PMD_NR;
2699         if (pextra_pins)
2700                 *pextra_pins = extra_pins;
2701         return total_mapcount(page) == page_count(page) - extra_pins - 1;
2702 }
2703
2704 /*
2705  * This function splits huge page into normal pages. @page can point to any
2706  * subpage of huge page to split. Split doesn't change the position of @page.
2707  *
2708  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2709  * The huge page must be locked.
2710  *
2711  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2712  *
2713  * Both head page and tail pages will inherit mapping, flags, and so on from
2714  * the hugepage.
2715  *
2716  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2717  * they are not mapped.
2718  *
2719  * Returns 0 if the hugepage is split successfully.
2720  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2721  * us.
2722  */
2723 int split_huge_page_to_list(struct page *page, struct list_head *list)
2724 {
2725         struct page *head = compound_head(page);
2726         struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2727         struct deferred_split *ds_queue = get_deferred_split_queue(head);
2728         struct anon_vma *anon_vma = NULL;
2729         struct address_space *mapping = NULL;
2730         int count, mapcount, extra_pins, ret;
2731         bool mlocked;
2732         unsigned long flags;
2733         pgoff_t end;
2734
2735         VM_BUG_ON_PAGE(is_huge_zero_page(head), head);
2736         VM_BUG_ON_PAGE(!PageLocked(head), head);
2737         VM_BUG_ON_PAGE(!PageCompound(head), head);
2738
2739         if (PageWriteback(head))
2740                 return -EBUSY;
2741
2742         if (PageAnon(head)) {
2743                 /*
2744                  * The caller does not necessarily hold an mmap_sem that would
2745                  * prevent the anon_vma disappearing so we first we take a
2746                  * reference to it and then lock the anon_vma for write. This
2747                  * is similar to page_lock_anon_vma_read except the write lock
2748                  * is taken to serialise against parallel split or collapse
2749                  * operations.
2750                  */
2751                 anon_vma = page_get_anon_vma(head);
2752                 if (!anon_vma) {
2753                         ret = -EBUSY;
2754                         goto out;
2755                 }
2756                 end = -1;
2757                 mapping = NULL;
2758                 anon_vma_lock_write(anon_vma);
2759         } else {
2760                 mapping = head->mapping;
2761
2762                 /* Truncated ? */
2763                 if (!mapping) {
2764                         ret = -EBUSY;
2765                         goto out;
2766                 }
2767
2768                 anon_vma = NULL;
2769                 i_mmap_lock_read(mapping);
2770
2771                 /*
2772                  *__split_huge_page() may need to trim off pages beyond EOF:
2773                  * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2774                  * which cannot be nested inside the page tree lock. So note
2775                  * end now: i_size itself may be changed at any moment, but
2776                  * head page lock is good enough to serialize the trimming.
2777                  */
2778                 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2779         }
2780
2781         /*
2782          * Racy check if we can split the page, before unmap_page() will
2783          * split PMDs
2784          */
2785         if (!can_split_huge_page(head, &extra_pins)) {
2786                 ret = -EBUSY;
2787                 goto out_unlock;
2788         }
2789
2790         mlocked = PageMlocked(head);
2791         unmap_page(head);
2792         VM_BUG_ON_PAGE(compound_mapcount(head), head);
2793
2794         /* Make sure the page is not on per-CPU pagevec as it takes pin */
2795         if (mlocked)
2796                 lru_add_drain();
2797
2798         /* prevent PageLRU to go away from under us, and freeze lru stats */
2799         spin_lock_irqsave(&pgdata->lru_lock, flags);
2800
2801         if (mapping) {
2802                 XA_STATE(xas, &mapping->i_pages, page_index(head));
2803
2804                 /*
2805                  * Check if the head page is present in page cache.
2806                  * We assume all tail are present too, if head is there.
2807                  */
2808                 xa_lock(&mapping->i_pages);
2809                 if (xas_load(&xas) != head)
2810                         goto fail;
2811         }
2812
2813         /* Prevent deferred_split_scan() touching ->_refcount */
2814         spin_lock(&ds_queue->split_queue_lock);
2815         count = page_count(head);
2816         mapcount = total_mapcount(head);
2817         if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2818                 if (!list_empty(page_deferred_list(head))) {
2819                         ds_queue->split_queue_len--;
2820                         list_del(page_deferred_list(head));
2821                 }
2822                 spin_unlock(&ds_queue->split_queue_lock);
2823                 if (mapping) {
2824                         if (PageSwapBacked(head))
2825                                 __dec_node_page_state(head, NR_SHMEM_THPS);
2826                         else
2827                                 __dec_node_page_state(head, NR_FILE_THPS);
2828                 }
2829
2830                 __split_huge_page(page, list, end, flags);
2831                 if (PageSwapCache(head)) {
2832                         swp_entry_t entry = { .val = page_private(head) };
2833
2834                         ret = split_swap_cluster(entry);
2835                 } else
2836                         ret = 0;
2837         } else {
2838                 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2839                         pr_alert("total_mapcount: %u, page_count(): %u\n",
2840                                         mapcount, count);
2841                         if (PageTail(page))
2842                                 dump_page(head, NULL);
2843                         dump_page(page, "total_mapcount(head) > 0");
2844                         BUG();
2845                 }
2846                 spin_unlock(&ds_queue->split_queue_lock);
2847 fail:           if (mapping)
2848                         xa_unlock(&mapping->i_pages);
2849                 spin_unlock_irqrestore(&pgdata->lru_lock, flags);
2850                 remap_page(head);
2851                 ret = -EBUSY;
2852         }
2853
2854 out_unlock:
2855         if (anon_vma) {
2856                 anon_vma_unlock_write(anon_vma);
2857                 put_anon_vma(anon_vma);
2858         }
2859         if (mapping)
2860                 i_mmap_unlock_read(mapping);
2861 out:
2862         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2863         return ret;
2864 }
2865
2866 void free_transhuge_page(struct page *page)
2867 {
2868         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2869         unsigned long flags;
2870
2871         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2872         if (!list_empty(page_deferred_list(page))) {
2873                 ds_queue->split_queue_len--;
2874                 list_del(page_deferred_list(page));
2875         }
2876         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2877         free_compound_page(page);
2878 }
2879
2880 void deferred_split_huge_page(struct page *page)
2881 {
2882         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2883 #ifdef CONFIG_MEMCG
2884         struct mem_cgroup *memcg = compound_head(page)->mem_cgroup;
2885 #endif
2886         unsigned long flags;
2887
2888         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2889
2890         /*
2891          * The try_to_unmap() in page reclaim path might reach here too,
2892          * this may cause a race condition to corrupt deferred split queue.
2893          * And, if page reclaim is already handling the same page, it is
2894          * unnecessary to handle it again in shrinker.
2895          *
2896          * Check PageSwapCache to determine if the page is being
2897          * handled by page reclaim since THP swap would add the page into
2898          * swap cache before calling try_to_unmap().
2899          */
2900         if (PageSwapCache(page))
2901                 return;
2902
2903         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2904         if (list_empty(page_deferred_list(page))) {
2905                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2906                 list_add_tail(page_deferred_list(page), &ds_queue->split_queue);
2907                 ds_queue->split_queue_len++;
2908 #ifdef CONFIG_MEMCG
2909                 if (memcg)
2910                         memcg_set_shrinker_bit(memcg, page_to_nid(page),
2911                                                deferred_split_shrinker.id);
2912 #endif
2913         }
2914         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2915 }
2916
2917 static unsigned long deferred_split_count(struct shrinker *shrink,
2918                 struct shrink_control *sc)
2919 {
2920         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2921         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2922
2923 #ifdef CONFIG_MEMCG
2924         if (sc->memcg)
2925                 ds_queue = &sc->memcg->deferred_split_queue;
2926 #endif
2927         return READ_ONCE(ds_queue->split_queue_len);
2928 }
2929
2930 static unsigned long deferred_split_scan(struct shrinker *shrink,
2931                 struct shrink_control *sc)
2932 {
2933         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2934         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2935         unsigned long flags;
2936         LIST_HEAD(list), *pos, *next;
2937         struct page *page;
2938         int split = 0;
2939
2940 #ifdef CONFIG_MEMCG
2941         if (sc->memcg)
2942                 ds_queue = &sc->memcg->deferred_split_queue;
2943 #endif
2944
2945         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2946         /* Take pin on all head pages to avoid freeing them under us */
2947         list_for_each_safe(pos, next, &ds_queue->split_queue) {
2948                 page = list_entry((void *)pos, struct page, mapping);
2949                 page = compound_head(page);
2950                 if (get_page_unless_zero(page)) {
2951                         list_move(page_deferred_list(page), &list);
2952                 } else {
2953                         /* We lost race with put_compound_page() */
2954                         list_del_init(page_deferred_list(page));
2955                         ds_queue->split_queue_len--;
2956                 }
2957                 if (!--sc->nr_to_scan)
2958                         break;
2959         }
2960         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2961
2962         list_for_each_safe(pos, next, &list) {
2963                 page = list_entry((void *)pos, struct page, mapping);
2964                 if (!trylock_page(page))
2965                         goto next;
2966                 /* split_huge_page() removes page from list on success */
2967                 if (!split_huge_page(page))
2968                         split++;
2969                 unlock_page(page);
2970 next:
2971                 put_page(page);
2972         }
2973
2974         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2975         list_splice_tail(&list, &ds_queue->split_queue);
2976         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2977
2978         /*
2979          * Stop shrinker if we didn't split any page, but the queue is empty.
2980          * This can happen if pages were freed under us.
2981          */
2982         if (!split && list_empty(&ds_queue->split_queue))
2983                 return SHRINK_STOP;
2984         return split;
2985 }
2986
2987 static struct shrinker deferred_split_shrinker = {
2988         .count_objects = deferred_split_count,
2989         .scan_objects = deferred_split_scan,
2990         .seeks = DEFAULT_SEEKS,
2991         .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
2992                  SHRINKER_NONSLAB,
2993 };
2994
2995 #ifdef CONFIG_DEBUG_FS
2996 static int split_huge_pages_set(void *data, u64 val)
2997 {
2998         struct zone *zone;
2999         struct page *page;
3000         unsigned long pfn, max_zone_pfn;
3001         unsigned long total = 0, split = 0;
3002
3003         if (val != 1)
3004                 return -EINVAL;
3005
3006         for_each_populated_zone(zone) {
3007                 max_zone_pfn = zone_end_pfn(zone);
3008                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
3009                         if (!pfn_valid(pfn))
3010                                 continue;
3011
3012                         page = pfn_to_page(pfn);
3013                         if (!get_page_unless_zero(page))
3014                                 continue;
3015
3016                         if (zone != page_zone(page))
3017                                 goto next;
3018
3019                         if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
3020                                 goto next;
3021
3022                         total++;
3023                         lock_page(page);
3024                         if (!split_huge_page(page))
3025                                 split++;
3026                         unlock_page(page);
3027 next:
3028                         put_page(page);
3029                 }
3030         }
3031
3032         pr_info("%lu of %lu THP split\n", split, total);
3033
3034         return 0;
3035 }
3036 DEFINE_DEBUGFS_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
3037                 "%llu\n");
3038
3039 static int __init split_huge_pages_debugfs(void)
3040 {
3041         debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3042                             &split_huge_pages_fops);
3043         return 0;
3044 }
3045 late_initcall(split_huge_pages_debugfs);
3046 #endif
3047
3048 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
3049 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
3050                 struct page *page)
3051 {
3052         struct vm_area_struct *vma = pvmw->vma;
3053         struct mm_struct *mm = vma->vm_mm;
3054         unsigned long address = pvmw->address;
3055         pmd_t pmdval;
3056         swp_entry_t entry;
3057         pmd_t pmdswp;
3058
3059         if (!(pvmw->pmd && !pvmw->pte))
3060                 return;
3061
3062         flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
3063         pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
3064         if (pmd_dirty(pmdval))
3065                 set_page_dirty(page);
3066         entry = make_migration_entry(page, pmd_write(pmdval));
3067         pmdswp = swp_entry_to_pmd(entry);
3068         if (pmd_soft_dirty(pmdval))
3069                 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
3070         set_pmd_at(mm, address, pvmw->pmd, pmdswp);
3071         page_remove_rmap(page, true);
3072         put_page(page);
3073 }
3074
3075 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
3076 {
3077         struct vm_area_struct *vma = pvmw->vma;
3078         struct mm_struct *mm = vma->vm_mm;
3079         unsigned long address = pvmw->address;
3080         unsigned long mmun_start = address & HPAGE_PMD_MASK;
3081         pmd_t pmde;
3082         swp_entry_t entry;
3083
3084         if (!(pvmw->pmd && !pvmw->pte))
3085                 return;
3086
3087         entry = pmd_to_swp_entry(*pvmw->pmd);
3088         get_page(new);
3089         pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
3090         if (pmd_swp_soft_dirty(*pvmw->pmd))
3091                 pmde = pmd_mksoft_dirty(pmde);
3092         if (is_write_migration_entry(entry))
3093                 pmde = maybe_pmd_mkwrite(pmde, vma);
3094
3095         flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
3096         if (PageAnon(new))
3097                 page_add_anon_rmap(new, vma, mmun_start, true);
3098         else
3099                 page_add_file_rmap(new, true);
3100         set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
3101         if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new))
3102                 mlock_vma_page(new);
3103         update_mmu_cache_pmd(vma, address, pvmw->pmd);
3104 }
3105 #endif