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         const char *output;
 167
 168         if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
 169                 output = "[always] madvise never";
 170         else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 171                           &transparent_hugepage_flags))
 172                 output = "always [madvise] never";
 173         else
 174                 output = "always madvise [never]";
 175
 176         return sysfs_emit(buf, "%s\n", output);
 177 }
 178
 179 static ssize_t enabled_store(struct kobject *kobj,
 180                              struct kobj_attribute *attr,
 181                              const char *buf, size_t count)
 182 {
 183         ssize_t ret = count;
 184
 185         if (sysfs_streq(buf, "always")) {
 186                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 187                 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 188         } else if (sysfs_streq(buf, "madvise")) {
 189                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 190                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 191         } else if (sysfs_streq(buf, "never")) {
 192                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 193                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 194         } else
 195                 ret = -EINVAL;
 196
 197         if (ret > 0) {
 198                 int err = start_stop_khugepaged();
 199                 if (err)
 200                         ret = err;
 201         }
 202         return ret;
 203 }
 204 static struct kobj_attribute enabled_attr =
 205         __ATTR(enabled, 0644, enabled_show, enabled_store);
 206
 207 ssize_t single_hugepage_flag_show(struct kobject *kobj,
 208                                   struct kobj_attribute *attr, char *buf,
 209                                   enum transparent_hugepage_flag flag)
 210 {
 211         return sysfs_emit(buf, "%d\n",
 212                           !!test_bit(flag, &transparent_hugepage_flags));
 213 }
 214
 215 ssize_t single_hugepage_flag_store(struct kobject *kobj,
 216                                  struct kobj_attribute *attr,
 217                                  const char *buf, size_t count,
 218                                  enum transparent_hugepage_flag flag)
 219 {
 220         unsigned long value;
 221         int ret;
 222
 223         ret = kstrtoul(buf, 10, &value);
 224         if (ret < 0)
 225                 return ret;
 226         if (value > 1)
 227                 return -EINVAL;
 228
 229         if (value)
 230                 set_bit(flag, &transparent_hugepage_flags);
 231         else
 232                 clear_bit(flag, &transparent_hugepage_flags);
 233
 234         return count;
 235 }
 236
 237 static ssize_t defrag_show(struct kobject *kobj,
 238                            struct kobj_attribute *attr, char *buf)
 239 {
 240         const char *output;
 241
 242         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
 243                      &transparent_hugepage_flags))
 244                 output = "[always] defer defer+madvise madvise never";
 245         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
 246                           &transparent_hugepage_flags))
 247                 output = "always [defer] defer+madvise madvise never";
 248         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG,
 249                           &transparent_hugepage_flags))
 250                 output = "always defer [defer+madvise] madvise never";
 251         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
 252                           &transparent_hugepage_flags))
 253                 output = "always defer defer+madvise [madvise] never";
 254         else
 255                 output = "always defer defer+madvise madvise [never]";
 256
 257         return sysfs_emit(buf, "%s\n", output);
 258 }
 259
 260 static ssize_t defrag_store(struct kobject *kobj,
 261                             struct kobj_attribute *attr,
 262                             const char *buf, size_t count)
 263 {
 264         if (sysfs_streq(buf, "always")) {
 265                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 266                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 267                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 268                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 269         } else if (sysfs_streq(buf, "defer+madvise")) {
 270                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 271                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 272                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 273                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 274         } else if (sysfs_streq(buf, "defer")) {
 275                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 276                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 277                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 278                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 279         } else if (sysfs_streq(buf, "madvise")) {
 280                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 281                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 282                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 283                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 284         } else if (sysfs_streq(buf, "never")) {
 285                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 286                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 287                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 288                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 289         } else
 290                 return -EINVAL;
 291
 292         return count;
 293 }
 294 static struct kobj_attribute defrag_attr =
 295         __ATTR(defrag, 0644, defrag_show, defrag_store);
 296
 297 static ssize_t use_zero_page_show(struct kobject *kobj,
 298                                   struct kobj_attribute *attr, char *buf)
 299 {
 300         return single_hugepage_flag_show(kobj, attr, buf,
 301                                          TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 302 }
 303 static ssize_t use_zero_page_store(struct kobject *kobj,
 304                 struct kobj_attribute *attr, const char *buf, size_t count)
 305 {
 306         return single_hugepage_flag_store(kobj, attr, buf, count,
 307                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 308 }
 309 static struct kobj_attribute use_zero_page_attr =
 310         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 311
 312 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
 313                                    struct kobj_attribute *attr, char *buf)
 314 {
 315         return sysfs_emit(buf, "%lu\n", HPAGE_PMD_SIZE);
 316 }
 317 static struct kobj_attribute hpage_pmd_size_attr =
 318         __ATTR_RO(hpage_pmd_size);
 319
 320 static struct attribute *hugepage_attr[] = {
 321         &enabled_attr.attr,
 322         &defrag_attr.attr,
 323         &use_zero_page_attr.attr,
 324         &hpage_pmd_size_attr.attr,
 325 #ifdef CONFIG_SHMEM
 326         &shmem_enabled_attr.attr,
 327 #endif
 328         NULL,
 329 };
 330
 331 static const struct attribute_group hugepage_attr_group = {
 332         .attrs = hugepage_attr,
 333 };
 334
 335 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 336 {
 337         int err;
 338
 339         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 340         if (unlikely(!*hugepage_kobj)) {
 341                 pr_err("failed to create transparent hugepage kobject\n");
 342                 return -ENOMEM;
 343         }
 344
 345         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 346         if (err) {
 347                 pr_err("failed to register transparent hugepage group\n");
 348                 goto delete_obj;
 349         }
 350
 351         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 352         if (err) {
 353                 pr_err("failed to register transparent hugepage group\n");
 354                 goto remove_hp_group;
 355         }
 356
 357         return 0;
 358
 359 remove_hp_group:
 360         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 361 delete_obj:
 362         kobject_put(*hugepage_kobj);
 363         return err;
 364 }
 365
 366 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 367 {
 368         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 369         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 370         kobject_put(hugepage_kobj);
 371 }
 372 #else
 373 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 374 {
 375         return 0;
 376 }
 377
 378 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 379 {
 380 }
 381 #endif /* CONFIG_SYSFS */
 382
 383 static int __init hugepage_init(void)
 384 {
 385         int err;
 386         struct kobject *hugepage_kobj;
 387
 388         if (!has_transparent_hugepage()) {
 389                 /*
 390                  * Hardware doesn't support hugepages, hence disable
 391                  * DAX PMD support.
 392                  */
 393                 transparent_hugepage_flags = 1 << TRANSPARENT_HUGEPAGE_NEVER_DAX;
 394                 return -EINVAL;
 395         }
 396
 397         /*
 398          * hugepages can't be allocated by the buddy allocator
 399          */
 400         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
 401         /*
 402          * we use page->mapping and page->index in second tail page
 403          * as list_head: assuming THP order >= 2
 404          */
 405         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 406
 407         err = hugepage_init_sysfs(&hugepage_kobj);
 408         if (err)
 409                 goto err_sysfs;
 410
 411         err = khugepaged_init();
 412         if (err)
 413                 goto err_slab;
 414
 415         err = register_shrinker(&huge_zero_page_shrinker);
 416         if (err)
 417                 goto err_hzp_shrinker;
 418         err = register_shrinker(&deferred_split_shrinker);
 419         if (err)
 420                 goto err_split_shrinker;
 421
 422         /*
 423          * By default disable transparent hugepages on smaller systems,
 424          * where the extra memory used could hurt more than TLB overhead
 425          * is likely to save.  The admin can still enable it through /sys.
 426          */
 427         if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
 428                 transparent_hugepage_flags = 0;
 429                 return 0;
 430         }
 431
 432         err = start_stop_khugepaged();
 433         if (err)
 434                 goto err_khugepaged;
 435
 436         return 0;
 437 err_khugepaged:
 438         unregister_shrinker(&deferred_split_shrinker);
 439 err_split_shrinker:
 440         unregister_shrinker(&huge_zero_page_shrinker);
 441 err_hzp_shrinker:
 442         khugepaged_destroy();
 443 err_slab:
 444         hugepage_exit_sysfs(hugepage_kobj);
 445 err_sysfs:
 446         return err;
 447 }
 448 subsys_initcall(hugepage_init);
 449
 450 static int __init setup_transparent_hugepage(char *str)
 451 {
 452         int ret = 0;
 453         if (!str)
 454                 goto out;
 455         if (!strcmp(str, "always")) {
 456                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 457                         &transparent_hugepage_flags);
 458                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 459                           &transparent_hugepage_flags);
 460                 ret = 1;
 461         } else if (!strcmp(str, "madvise")) {
 462                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 463                           &transparent_hugepage_flags);
 464                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 465                         &transparent_hugepage_flags);
 466                 ret = 1;
 467         } else if (!strcmp(str, "never")) {
 468                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 469                           &transparent_hugepage_flags);
 470                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 471                           &transparent_hugepage_flags);
 472                 ret = 1;
 473         }
 474 out:
 475         if (!ret)
 476                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 477         return ret;
 478 }
 479 __setup("transparent_hugepage=", setup_transparent_hugepage);
 480
 481 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 482 {
 483         if (likely(vma->vm_flags & VM_WRITE))
 484                 pmd = pmd_mkwrite(pmd);
 485         return pmd;
 486 }
 487
 488 #ifdef CONFIG_MEMCG
 489 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
 490 {
 491         struct mem_cgroup *memcg = page_memcg(compound_head(page));
 492         struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
 493
 494         if (memcg)
 495                 return &memcg->deferred_split_queue;
 496         else
 497                 return &pgdat->deferred_split_queue;
 498 }
 499 #else
 500 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
 501 {
 502         struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
 503
 504         return &pgdat->deferred_split_queue;
 505 }
 506 #endif
 507
 508 void prep_transhuge_page(struct page *page)
 509 {
 510         /*
 511          * we use page->mapping and page->indexlru in second tail page
 512          * as list_head: assuming THP order >= 2
 513          */
 514
 515         INIT_LIST_HEAD(page_deferred_list(page));
 516         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 517 }
 518
 519 bool is_transparent_hugepage(struct page *page)
 520 {
 521         if (!PageCompound(page))
 522                 return false;
 523
 524         page = compound_head(page);
 525         return is_huge_zero_page(page) ||
 526                page[1].compound_dtor == TRANSHUGE_PAGE_DTOR;
 527 }
 528 EXPORT_SYMBOL_GPL(is_transparent_hugepage);
 529
 530 static unsigned long __thp_get_unmapped_area(struct file *filp,
 531                 unsigned long addr, unsigned long len,
 532                 loff_t off, unsigned long flags, unsigned long size)
 533 {
 534         loff_t off_end = off + len;
 535         loff_t off_align = round_up(off, size);
 536         unsigned long len_pad, ret;
 537
 538         if (off_end <= off_align || (off_end - off_align) < size)
 539                 return 0;
 540
 541         len_pad = len + size;
 542         if (len_pad < len || (off + len_pad) < off)
 543                 return 0;
 544
 545         ret = current->mm->get_unmapped_area(filp, addr, len_pad,
 546                                               off >> PAGE_SHIFT, flags);
 547
 548         /*
 549          * The failure might be due to length padding. The caller will retry
 550          * without the padding.
 551          */
 552         if (IS_ERR_VALUE(ret))
 553                 return 0;
 554
 555         /*
 556          * Do not try to align to THP boundary if allocation at the address
 557          * hint succeeds.
 558          */
 559         if (ret == addr)
 560                 return addr;
 561
 562         ret += (off - ret) & (size - 1);
 563         return ret;
 564 }
 565
 566 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 567                 unsigned long len, unsigned long pgoff, unsigned long flags)
 568 {
 569         unsigned long ret;
 570         loff_t off = (loff_t)pgoff << PAGE_SHIFT;
 571
 572         if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
 573                 goto out;
 574
 575         ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE);
 576         if (ret)
 577                 return ret;
 578 out:
 579         return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
 580 }
 581 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
 582
 583 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
 584                         struct page *page, gfp_t gfp)
 585 {
 586         struct vm_area_struct *vma = vmf->vma;
 587         pgtable_t pgtable;
 588         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 589         vm_fault_t ret = 0;
 590
 591         VM_BUG_ON_PAGE(!PageCompound(page), page);
 592
 593         if (mem_cgroup_charge(page, vma->vm_mm, gfp)) {
 594                 put_page(page);
 595                 count_vm_event(THP_FAULT_FALLBACK);
 596                 count_vm_event(THP_FAULT_FALLBACK_CHARGE);
 597                 return VM_FAULT_FALLBACK;
 598         }
 599         cgroup_throttle_swaprate(page, gfp);
 600
 601         pgtable = pte_alloc_one(vma->vm_mm);
 602         if (unlikely(!pgtable)) {
 603                 ret = VM_FAULT_OOM;
 604                 goto release;
 605         }
 606
 607         clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
 608         /*
 609          * The memory barrier inside __SetPageUptodate makes sure that
 610          * clear_huge_page writes become visible before the set_pmd_at()
 611          * write.
 612          */
 613         __SetPageUptodate(page);
 614
 615         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 616         if (unlikely(!pmd_none(*vmf->pmd))) {
 617                 goto unlock_release;
 618         } else {
 619                 pmd_t entry;
 620
 621                 ret = check_stable_address_space(vma->vm_mm);
 622                 if (ret)
 623                         goto unlock_release;
 624
 625                 /* Deliver the page fault to userland */
 626                 if (userfaultfd_missing(vma)) {
 627                         vm_fault_t ret2;
 628
 629                         spin_unlock(vmf->ptl);
 630                         put_page(page);
 631                         pte_free(vma->vm_mm, pgtable);
 632                         ret2 = handle_userfault(vmf, VM_UFFD_MISSING);
 633                         VM_BUG_ON(ret2 & VM_FAULT_FALLBACK);
 634                         return ret2;
 635                 }
 636
 637                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 638                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 639                 page_add_new_anon_rmap(page, vma, haddr, true);
 640                 lru_cache_add_inactive_or_unevictable(page, vma);
 641                 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
 642                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
 643                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
 644                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 645                 mm_inc_nr_ptes(vma->vm_mm);
 646                 spin_unlock(vmf->ptl);
 647                 count_vm_event(THP_FAULT_ALLOC);
 648                 count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC);
 649         }
 650
 651         return 0;
 652 unlock_release:
 653         spin_unlock(vmf->ptl);
 654 release:
 655         if (pgtable)
 656                 pte_free(vma->vm_mm, pgtable);
 657         put_page(page);
 658         return ret;
 659
 660 }
 661
 662 /*
 663  * always: directly stall for all thp allocations
 664  * defer: wake kswapd and fail if not immediately available
 665  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
 666  *                fail if not immediately available
 667  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
 668  *          available
 669  * never: never stall for any thp allocation
 670  */
 671 gfp_t vma_thp_gfp_mask(struct vm_area_struct *vma)
 672 {
 673         const bool vma_madvised = vma && (vma->vm_flags & VM_HUGEPAGE);
 674
 675         /* Always do synchronous compaction */
 676         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 677                 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
 678
 679         /* Kick kcompactd and fail quickly */
 680         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 681                 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
 682
 683         /* Synchronous compaction if madvised, otherwise kick kcompactd */
 684         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 685                 return GFP_TRANSHUGE_LIGHT |
 686                         (vma_madvised ? __GFP_DIRECT_RECLAIM :
 687                                         __GFP_KSWAPD_RECLAIM);
 688
 689         /* Only do synchronous compaction if madvised */
 690         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 691                 return GFP_TRANSHUGE_LIGHT |
 692                        (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
 693
 694         return GFP_TRANSHUGE_LIGHT;
 695 }
 696
 697 /* Caller must hold page table lock. */
 698 static void set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 699                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 700                 struct page *zero_page)
 701 {
 702         pmd_t entry;
 703         if (!pmd_none(*pmd))
 704                 return;
 705         entry = mk_pmd(zero_page, vma->vm_page_prot);
 706         entry = pmd_mkhuge(entry);
 707         if (pgtable)
 708                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 709         set_pmd_at(mm, haddr, pmd, entry);
 710         mm_inc_nr_ptes(mm);
 711 }
 712
 713 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
 714 {
 715         struct vm_area_struct *vma = vmf->vma;
 716         gfp_t gfp;
 717         struct page *page;
 718         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 719
 720         if (!transhuge_vma_suitable(vma, haddr))
 721                 return VM_FAULT_FALLBACK;
 722         if (unlikely(anon_vma_prepare(vma)))
 723                 return VM_FAULT_OOM;
 724         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 725                 return VM_FAULT_OOM;
 726         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
 727                         !mm_forbids_zeropage(vma->vm_mm) &&
 728                         transparent_hugepage_use_zero_page()) {
 729                 pgtable_t pgtable;
 730                 struct page *zero_page;
 731                 vm_fault_t ret;
 732                 pgtable = pte_alloc_one(vma->vm_mm);
 733                 if (unlikely(!pgtable))
 734                         return VM_FAULT_OOM;
 735                 zero_page = mm_get_huge_zero_page(vma->vm_mm);
 736                 if (unlikely(!zero_page)) {
 737                         pte_free(vma->vm_mm, pgtable);
 738                         count_vm_event(THP_FAULT_FALLBACK);
 739                         return VM_FAULT_FALLBACK;
 740                 }
 741                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 742                 ret = 0;
 743                 if (pmd_none(*vmf->pmd)) {
 744                         ret = check_stable_address_space(vma->vm_mm);
 745                         if (ret) {
 746                                 spin_unlock(vmf->ptl);
 747                                 pte_free(vma->vm_mm, pgtable);
 748                         } else if (userfaultfd_missing(vma)) {
 749                                 spin_unlock(vmf->ptl);
 750                                 pte_free(vma->vm_mm, pgtable);
 751                                 ret = handle_userfault(vmf, VM_UFFD_MISSING);
 752                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 753                         } else {
 754                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
 755                                                    haddr, vmf->pmd, zero_page);
 756                                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
 757                                 spin_unlock(vmf->ptl);
 758                         }
 759                 } else {
 760                         spin_unlock(vmf->ptl);
 761                         pte_free(vma->vm_mm, pgtable);
 762                 }
 763                 return ret;
 764         }
 765         gfp = vma_thp_gfp_mask(vma);
 766         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 767         if (unlikely(!page)) {
 768                 count_vm_event(THP_FAULT_FALLBACK);
 769                 return VM_FAULT_FALLBACK;
 770         }
 771         prep_transhuge_page(page);
 772         return __do_huge_pmd_anonymous_page(vmf, page, gfp);
 773 }
 774
 775 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 776                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
 777                 pgtable_t pgtable)
 778 {
 779         struct mm_struct *mm = vma->vm_mm;
 780         pmd_t entry;
 781         spinlock_t *ptl;
 782
 783         ptl = pmd_lock(mm, pmd);
 784         if (!pmd_none(*pmd)) {
 785                 if (write) {
 786                         if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
 787                                 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
 788                                 goto out_unlock;
 789                         }
 790                         entry = pmd_mkyoung(*pmd);
 791                         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 792                         if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
 793                                 update_mmu_cache_pmd(vma, addr, pmd);
 794                 }
 795
 796                 goto out_unlock;
 797         }
 798
 799         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 800         if (pfn_t_devmap(pfn))
 801                 entry = pmd_mkdevmap(entry);
 802         if (write) {
 803                 entry = pmd_mkyoung(pmd_mkdirty(entry));
 804                 entry = maybe_pmd_mkwrite(entry, vma);
 805         }
 806
 807         if (pgtable) {
 808                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 809                 mm_inc_nr_ptes(mm);
 810                 pgtable = NULL;
 811         }
 812
 813         set_pmd_at(mm, addr, pmd, entry);
 814         update_mmu_cache_pmd(vma, addr, pmd);
 815
 816 out_unlock:
 817         spin_unlock(ptl);
 818         if (pgtable)
 819                 pte_free(mm, pgtable);
 820 }
 821
 822 /**
 823  * vmf_insert_pfn_pmd_prot - insert a pmd size pfn
 824  * @vmf: Structure describing the fault
 825  * @pfn: pfn to insert
 826  * @pgprot: page protection to use
 827  * @write: whether it's a write fault
 828  *
 829  * Insert a pmd size pfn. See vmf_insert_pfn() for additional info and
 830  * also consult the vmf_insert_mixed_prot() documentation when
 831  * @pgprot != @vmf->vma->vm_page_prot.
 832  *
 833  * Return: vm_fault_t value.
 834  */
 835 vm_fault_t vmf_insert_pfn_pmd_prot(struct vm_fault *vmf, pfn_t pfn,
 836                                    pgprot_t pgprot, bool write)
 837 {
 838         unsigned long addr = vmf->address & PMD_MASK;
 839         struct vm_area_struct *vma = vmf->vma;
 840         pgtable_t pgtable = NULL;
 841
 842         /*
 843          * If we had pmd_special, we could avoid all these restrictions,
 844          * but we need to be consistent with PTEs and architectures that
 845          * can't support a 'special' bit.
 846          */
 847         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 848                         !pfn_t_devmap(pfn));
 849         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 850                                                 (VM_PFNMAP|VM_MIXEDMAP));
 851         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 852
 853         if (addr < vma->vm_start || addr >= vma->vm_end)
 854                 return VM_FAULT_SIGBUS;
 855
 856         if (arch_needs_pgtable_deposit()) {
 857                 pgtable = pte_alloc_one(vma->vm_mm);
 858                 if (!pgtable)
 859                         return VM_FAULT_OOM;
 860         }
 861
 862         track_pfn_insert(vma, &pgprot, pfn);
 863
 864         insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
 865         return VM_FAULT_NOPAGE;
 866 }
 867 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd_prot);
 868
 869 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 870 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
 871 {
 872         if (likely(vma->vm_flags & VM_WRITE))
 873                 pud = pud_mkwrite(pud);
 874         return pud;
 875 }
 876
 877 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 878                 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
 879 {
 880         struct mm_struct *mm = vma->vm_mm;
 881         pud_t entry;
 882         spinlock_t *ptl;
 883
 884         ptl = pud_lock(mm, pud);
 885         if (!pud_none(*pud)) {
 886                 if (write) {
 887                         if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
 888                                 WARN_ON_ONCE(!is_huge_zero_pud(*pud));
 889                                 goto out_unlock;
 890                         }
 891                         entry = pud_mkyoung(*pud);
 892                         entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
 893                         if (pudp_set_access_flags(vma, addr, pud, entry, 1))
 894                                 update_mmu_cache_pud(vma, addr, pud);
 895                 }
 896                 goto out_unlock;
 897         }
 898
 899         entry = pud_mkhuge(pfn_t_pud(pfn, prot));
 900         if (pfn_t_devmap(pfn))
 901                 entry = pud_mkdevmap(entry);
 902         if (write) {
 903                 entry = pud_mkyoung(pud_mkdirty(entry));
 904                 entry = maybe_pud_mkwrite(entry, vma);
 905         }
 906         set_pud_at(mm, addr, pud, entry);
 907         update_mmu_cache_pud(vma, addr, pud);
 908
 909 out_unlock:
 910         spin_unlock(ptl);
 911 }
 912
 913 /**
 914  * vmf_insert_pfn_pud_prot - insert a pud size pfn
 915  * @vmf: Structure describing the fault
 916  * @pfn: pfn to insert
 917  * @pgprot: page protection to use
 918  * @write: whether it's a write fault
 919  *
 920  * Insert a pud size pfn. See vmf_insert_pfn() for additional info and
 921  * also consult the vmf_insert_mixed_prot() documentation when
 922  * @pgprot != @vmf->vma->vm_page_prot.
 923  *
 924  * Return: vm_fault_t value.
 925  */
 926 vm_fault_t vmf_insert_pfn_pud_prot(struct vm_fault *vmf, pfn_t pfn,
 927                                    pgprot_t pgprot, bool write)
 928 {
 929         unsigned long addr = vmf->address & PUD_MASK;
 930         struct vm_area_struct *vma = vmf->vma;
 931
 932         /*
 933          * If we had pud_special, we could avoid all these restrictions,
 934          * but we need to be consistent with PTEs and architectures that
 935          * can't support a 'special' bit.
 936          */
 937         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 938                         !pfn_t_devmap(pfn));
 939         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 940                                                 (VM_PFNMAP|VM_MIXEDMAP));
 941         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 942
 943         if (addr < vma->vm_start || addr >= vma->vm_end)
 944                 return VM_FAULT_SIGBUS;
 945
 946         track_pfn_insert(vma, &pgprot, pfn);
 947
 948         insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
 949         return VM_FAULT_NOPAGE;
 950 }
 951 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud_prot);
 952 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
 953
 954 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
 955                 pmd_t *pmd, int flags)
 956 {
 957         pmd_t _pmd;
 958
 959         _pmd = pmd_mkyoung(*pmd);
 960         if (flags & FOLL_WRITE)
 961                 _pmd = pmd_mkdirty(_pmd);
 962         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
 963                                 pmd, _pmd, flags & FOLL_WRITE))
 964                 update_mmu_cache_pmd(vma, addr, pmd);
 965 }
 966
 967 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 968                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
 969 {
 970         unsigned long pfn = pmd_pfn(*pmd);
 971         struct mm_struct *mm = vma->vm_mm;
 972         struct page *page;
 973
 974         assert_spin_locked(pmd_lockptr(mm, pmd));
 975
 976         /*
 977          * When we COW a devmap PMD entry, we split it into PTEs, so we should
 978          * not be in this function with `flags & FOLL_COW` set.
 979          */
 980         WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
 981
 982         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
 983         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
 984                          (FOLL_PIN | FOLL_GET)))
 985                 return NULL;
 986
 987         if (flags & FOLL_WRITE && !pmd_write(*pmd))
 988                 return NULL;
 989
 990         if (pmd_present(*pmd) && pmd_devmap(*pmd))
 991                 /* pass */;
 992         else
 993                 return NULL;
 994
 995         if (flags & FOLL_TOUCH)
 996                 touch_pmd(vma, addr, pmd, flags);
 997
 998         /*
 999          * device mapped pages can only be returned if the
1000          * caller will manage the page reference count.
1001          */
1002         if (!(flags & (FOLL_GET | FOLL_PIN)))
1003                 return ERR_PTR(-EEXIST);
1004
1005         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1006         *pgmap = get_dev_pagemap(pfn, *pgmap);
1007         if (!*pgmap)
1008                 return ERR_PTR(-EFAULT);
1009         page = pfn_to_page(pfn);
1010         if (!try_grab_page(page, flags))
1011                 page = ERR_PTR(-ENOMEM);
1012
1013         return page;
1014 }
1015
1016 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1017                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1018                   struct vm_area_struct *vma)
1019 {
1020         spinlock_t *dst_ptl, *src_ptl;
1021         struct page *src_page;
1022         pmd_t pmd;
1023         pgtable_t pgtable = NULL;
1024         int ret = -ENOMEM;
1025
1026         /* Skip if can be re-fill on fault */
1027         if (!vma_is_anonymous(vma))
1028                 return 0;
1029
1030         pgtable = pte_alloc_one(dst_mm);
1031         if (unlikely(!pgtable))
1032                 goto out;
1033
1034         dst_ptl = pmd_lock(dst_mm, dst_pmd);
1035         src_ptl = pmd_lockptr(src_mm, src_pmd);
1036         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1037
1038         ret = -EAGAIN;
1039         pmd = *src_pmd;
1040
1041         /*
1042          * Make sure the _PAGE_UFFD_WP bit is cleared if the new VMA
1043          * does not have the VM_UFFD_WP, which means that the uffd
1044          * fork event is not enabled.
1045          */
1046         if (!(vma->vm_flags & VM_UFFD_WP))
1047                 pmd = pmd_clear_uffd_wp(pmd);
1048
1049 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1050         if (unlikely(is_swap_pmd(pmd))) {
1051                 swp_entry_t entry = pmd_to_swp_entry(pmd);
1052
1053                 VM_BUG_ON(!is_pmd_migration_entry(pmd));
1054                 if (is_write_migration_entry(entry)) {
1055                         make_migration_entry_read(&entry);
1056                         pmd = swp_entry_to_pmd(entry);
1057                         if (pmd_swp_soft_dirty(*src_pmd))
1058                                 pmd = pmd_swp_mksoft_dirty(pmd);
1059                         set_pmd_at(src_mm, addr, src_pmd, pmd);
1060                 }
1061                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1062                 mm_inc_nr_ptes(dst_mm);
1063                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1064                 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1065                 ret = 0;
1066                 goto out_unlock;
1067         }
1068 #endif
1069
1070         if (unlikely(!pmd_trans_huge(pmd))) {
1071                 pte_free(dst_mm, pgtable);
1072                 goto out_unlock;
1073         }
1074         /*
1075          * When page table lock is held, the huge zero pmd should not be
1076          * under splitting since we don't split the page itself, only pmd to
1077          * a page table.
1078          */
1079         if (is_huge_zero_pmd(pmd)) {
1080                 struct page *zero_page;
1081                 /*
1082                  * get_huge_zero_page() will never allocate a new page here,
1083                  * since we already have a zero page to copy. It just takes a
1084                  * reference.
1085                  */
1086                 zero_page = mm_get_huge_zero_page(dst_mm);
1087                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1088                                 zero_page);
1089                 ret = 0;
1090                 goto out_unlock;
1091         }
1092
1093         src_page = pmd_page(pmd);
1094         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1095
1096         /*
1097          * If this page is a potentially pinned page, split and retry the fault
1098          * with smaller page size.  Normally this should not happen because the
1099          * userspace should use MADV_DONTFORK upon pinned regions.  This is a
1100          * best effort that the pinned pages won't be replaced by another
1101          * random page during the coming copy-on-write.
1102          */
1103         if (unlikely(page_needs_cow_for_dma(vma, src_page))) {
1104                 pte_free(dst_mm, pgtable);
1105                 spin_unlock(src_ptl);
1106                 spin_unlock(dst_ptl);
1107                 __split_huge_pmd(vma, src_pmd, addr, false, NULL);
1108                 return -EAGAIN;
1109         }
1110
1111         get_page(src_page);
1112         page_dup_rmap(src_page, true);
1113         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1114         mm_inc_nr_ptes(dst_mm);
1115         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1116
1117         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1118         pmd = pmd_mkold(pmd_wrprotect(pmd));
1119         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1120
1121         ret = 0;
1122 out_unlock:
1123         spin_unlock(src_ptl);
1124         spin_unlock(dst_ptl);
1125 out:
1126         return ret;
1127 }
1128
1129 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1130 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1131                 pud_t *pud, int flags)
1132 {
1133         pud_t _pud;
1134
1135         _pud = pud_mkyoung(*pud);
1136         if (flags & FOLL_WRITE)
1137                 _pud = pud_mkdirty(_pud);
1138         if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1139                                 pud, _pud, flags & FOLL_WRITE))
1140                 update_mmu_cache_pud(vma, addr, pud);
1141 }
1142
1143 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1144                 pud_t *pud, int flags, struct dev_pagemap **pgmap)
1145 {
1146         unsigned long pfn = pud_pfn(*pud);
1147         struct mm_struct *mm = vma->vm_mm;
1148         struct page *page;
1149
1150         assert_spin_locked(pud_lockptr(mm, pud));
1151
1152         if (flags & FOLL_WRITE && !pud_write(*pud))
1153                 return NULL;
1154
1155         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
1156         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
1157                          (FOLL_PIN | FOLL_GET)))
1158                 return NULL;
1159
1160         if (pud_present(*pud) && pud_devmap(*pud))
1161                 /* pass */;
1162         else
1163                 return NULL;
1164
1165         if (flags & FOLL_TOUCH)
1166                 touch_pud(vma, addr, pud, flags);
1167
1168         /*
1169          * device mapped pages can only be returned if the
1170          * caller will manage the page reference count.
1171          *
1172          * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here:
1173          */
1174         if (!(flags & (FOLL_GET | FOLL_PIN)))
1175                 return ERR_PTR(-EEXIST);
1176
1177         pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1178         *pgmap = get_dev_pagemap(pfn, *pgmap);
1179         if (!*pgmap)
1180                 return ERR_PTR(-EFAULT);
1181         page = pfn_to_page(pfn);
1182         if (!try_grab_page(page, flags))
1183                 page = ERR_PTR(-ENOMEM);
1184
1185         return page;
1186 }
1187
1188 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1189                   pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1190                   struct vm_area_struct *vma)
1191 {
1192         spinlock_t *dst_ptl, *src_ptl;
1193         pud_t pud;
1194         int ret;
1195
1196         dst_ptl = pud_lock(dst_mm, dst_pud);
1197         src_ptl = pud_lockptr(src_mm, src_pud);
1198         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1199
1200         ret = -EAGAIN;
1201         pud = *src_pud;
1202         if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1203                 goto out_unlock;
1204
1205         /*
1206          * When page table lock is held, the huge zero pud should not be
1207          * under splitting since we don't split the page itself, only pud to
1208          * a page table.
1209          */
1210         if (is_huge_zero_pud(pud)) {
1211                 /* No huge zero pud yet */
1212         }
1213
1214         /* Please refer to comments in copy_huge_pmd() */
1215         if (unlikely(page_needs_cow_for_dma(vma, pud_page(pud)))) {
1216                 spin_unlock(src_ptl);
1217                 spin_unlock(dst_ptl);
1218                 __split_huge_pud(vma, src_pud, addr);
1219                 return -EAGAIN;
1220         }
1221
1222         pudp_set_wrprotect(src_mm, addr, src_pud);
1223         pud = pud_mkold(pud_wrprotect(pud));
1224         set_pud_at(dst_mm, addr, dst_pud, pud);
1225
1226         ret = 0;
1227 out_unlock:
1228         spin_unlock(src_ptl);
1229         spin_unlock(dst_ptl);
1230         return ret;
1231 }
1232
1233 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1234 {
1235         pud_t entry;
1236         unsigned long haddr;
1237         bool write = vmf->flags & FAULT_FLAG_WRITE;
1238
1239         vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1240         if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1241                 goto unlock;
1242
1243         entry = pud_mkyoung(orig_pud);
1244         if (write)
1245                 entry = pud_mkdirty(entry);
1246         haddr = vmf->address & HPAGE_PUD_MASK;
1247         if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1248                 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1249
1250 unlock:
1251         spin_unlock(vmf->ptl);
1252 }
1253 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1254
1255 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1256 {
1257         pmd_t entry;
1258         unsigned long haddr;
1259         bool write = vmf->flags & FAULT_FLAG_WRITE;
1260
1261         vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1262         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1263                 goto unlock;
1264
1265         entry = pmd_mkyoung(orig_pmd);
1266         if (write)
1267                 entry = pmd_mkdirty(entry);
1268         haddr = vmf->address & HPAGE_PMD_MASK;
1269         if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1270                 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1271
1272 unlock:
1273         spin_unlock(vmf->ptl);
1274 }
1275
1276 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1277 {
1278         struct vm_area_struct *vma = vmf->vma;
1279         struct page *page;
1280         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1281
1282         vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1283         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1284
1285         if (is_huge_zero_pmd(orig_pmd))
1286                 goto fallback;
1287
1288         spin_lock(vmf->ptl);
1289
1290         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1291                 spin_unlock(vmf->ptl);
1292                 return 0;
1293         }
1294
1295         page = pmd_page(orig_pmd);
1296         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1297
1298         /* Lock page for reuse_swap_page() */
1299         if (!trylock_page(page)) {
1300                 get_page(page);
1301                 spin_unlock(vmf->ptl);
1302                 lock_page(page);
1303                 spin_lock(vmf->ptl);
1304                 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1305                         spin_unlock(vmf->ptl);
1306                         unlock_page(page);
1307                         put_page(page);
1308                         return 0;
1309                 }
1310                 put_page(page);
1311         }
1312
1313         /*
1314          * We can only reuse the page if nobody else maps the huge page or it's
1315          * part.
1316          */
1317         if (reuse_swap_page(page, NULL)) {
1318                 pmd_t entry;
1319                 entry = pmd_mkyoung(orig_pmd);
1320                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1321                 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1322                         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1323                 unlock_page(page);
1324                 spin_unlock(vmf->ptl);
1325                 return VM_FAULT_WRITE;
1326         }
1327
1328         unlock_page(page);
1329         spin_unlock(vmf->ptl);
1330 fallback:
1331         __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL);
1332         return VM_FAULT_FALLBACK;
1333 }
1334
1335 /*
1336  * FOLL_FORCE can write to even unwritable pmd's, but only
1337  * after we've gone through a COW cycle and they are dirty.
1338  */
1339 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1340 {
1341         return pmd_write(pmd) ||
1342                ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1343 }
1344
1345 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1346                                    unsigned long addr,
1347                                    pmd_t *pmd,
1348                                    unsigned int flags)
1349 {
1350         struct mm_struct *mm = vma->vm_mm;
1351         struct page *page = NULL;
1352
1353         assert_spin_locked(pmd_lockptr(mm, pmd));
1354
1355         if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1356                 goto out;
1357
1358         /* Avoid dumping huge zero page */
1359         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1360                 return ERR_PTR(-EFAULT);
1361
1362         /* Full NUMA hinting faults to serialise migration in fault paths */
1363         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1364                 goto out;
1365
1366         page = pmd_page(*pmd);
1367         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1368
1369         if (!try_grab_page(page, flags))
1370                 return ERR_PTR(-ENOMEM);
1371
1372         if (flags & FOLL_TOUCH)
1373                 touch_pmd(vma, addr, pmd, flags);
1374
1375         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1376                 /*
1377                  * We don't mlock() pte-mapped THPs. This way we can avoid
1378                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1379                  *
1380                  * For anon THP:
1381                  *
1382                  * In most cases the pmd is the only mapping of the page as we
1383                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1384                  * writable private mappings in populate_vma_page_range().
1385                  *
1386                  * The only scenario when we have the page shared here is if we
1387                  * mlocking read-only mapping shared over fork(). We skip
1388                  * mlocking such pages.
1389                  *
1390                  * For file THP:
1391                  *
1392                  * We can expect PageDoubleMap() to be stable under page lock:
1393                  * for file pages we set it in page_add_file_rmap(), which
1394                  * requires page to be locked.
1395                  */
1396
1397                 if (PageAnon(page) && compound_mapcount(page) != 1)
1398                         goto skip_mlock;
1399                 if (PageDoubleMap(page) || !page->mapping)
1400                         goto skip_mlock;
1401                 if (!trylock_page(page))
1402                         goto skip_mlock;
1403                 if (page->mapping && !PageDoubleMap(page))
1404                         mlock_vma_page(page);
1405                 unlock_page(page);
1406         }
1407 skip_mlock:
1408         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1409         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1410
1411 out:
1412         return page;
1413 }
1414
1415 /* NUMA hinting page fault entry point for trans huge pmds */
1416 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1417 {
1418         struct vm_area_struct *vma = vmf->vma;
1419         struct anon_vma *anon_vma = NULL;
1420         struct page *page;
1421         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1422         int page_nid = NUMA_NO_NODE, this_nid = numa_node_id();
1423         int target_nid, last_cpupid = -1;
1424         bool page_locked;
1425         bool migrated = false;
1426         bool was_writable;
1427         int flags = 0;
1428
1429         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1430         if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1431                 goto out_unlock;
1432
1433         /*
1434          * If there are potential migrations, wait for completion and retry
1435          * without disrupting NUMA hinting information. Do not relock and
1436          * check_same as the page may no longer be mapped.
1437          */
1438         if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1439                 page = pmd_page(*vmf->pmd);
1440                 if (!get_page_unless_zero(page))
1441                         goto out_unlock;
1442                 spin_unlock(vmf->ptl);
1443                 put_and_wait_on_page_locked(page, TASK_UNINTERRUPTIBLE);
1444                 goto out;
1445         }
1446
1447         page = pmd_page(pmd);
1448         BUG_ON(is_huge_zero_page(page));
1449         page_nid = page_to_nid(page);
1450         last_cpupid = page_cpupid_last(page);
1451         count_vm_numa_event(NUMA_HINT_FAULTS);
1452         if (page_nid == this_nid) {
1453                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1454                 flags |= TNF_FAULT_LOCAL;
1455         }
1456
1457         /* See similar comment in do_numa_page for explanation */
1458         if (!pmd_savedwrite(pmd))
1459                 flags |= TNF_NO_GROUP;
1460
1461         /*
1462          * Acquire the page lock to serialise THP migrations but avoid dropping
1463          * page_table_lock if at all possible
1464          */
1465         page_locked = trylock_page(page);
1466         target_nid = mpol_misplaced(page, vma, haddr);
1467         if (target_nid == NUMA_NO_NODE) {
1468                 /* If the page was locked, there are no parallel migrations */
1469                 if (page_locked)
1470                         goto clear_pmdnuma;
1471         }
1472
1473         /* Migration could have started since the pmd_trans_migrating check */
1474         if (!page_locked) {
1475                 page_nid = NUMA_NO_NODE;
1476                 if (!get_page_unless_zero(page))
1477                         goto out_unlock;
1478                 spin_unlock(vmf->ptl);
1479                 put_and_wait_on_page_locked(page, TASK_UNINTERRUPTIBLE);
1480                 goto out;
1481         }
1482
1483         /*
1484          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1485          * to serialises splits
1486          */
1487         get_page(page);
1488         spin_unlock(vmf->ptl);
1489         anon_vma = page_lock_anon_vma_read(page);
1490
1491         /* Confirm the PMD did not change while page_table_lock was released */
1492         spin_lock(vmf->ptl);
1493         if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1494                 unlock_page(page);
1495                 put_page(page);
1496                 page_nid = NUMA_NO_NODE;
1497                 goto out_unlock;
1498         }
1499
1500         /* Bail if we fail to protect against THP splits for any reason */
1501         if (unlikely(!anon_vma)) {
1502                 put_page(page);
1503                 page_nid = NUMA_NO_NODE;
1504                 goto clear_pmdnuma;
1505         }
1506
1507         /*
1508          * Since we took the NUMA fault, we must have observed the !accessible
1509          * bit. Make sure all other CPUs agree with that, to avoid them
1510          * modifying the page we're about to migrate.
1511          *
1512          * Must be done under PTL such that we'll observe the relevant
1513          * inc_tlb_flush_pending().
1514          *
1515          * We are not sure a pending tlb flush here is for a huge page
1516          * mapping or not. Hence use the tlb range variant
1517          */
1518         if (mm_tlb_flush_pending(vma->vm_mm)) {
1519                 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1520                 /*
1521                  * change_huge_pmd() released the pmd lock before
1522                  * invalidating the secondary MMUs sharing the primary
1523                  * MMU pagetables (with ->invalidate_range()). The
1524                  * mmu_notifier_invalidate_range_end() (which
1525                  * internally calls ->invalidate_range()) in
1526                  * change_pmd_range() will run after us, so we can't
1527                  * rely on it here and we need an explicit invalidate.
1528                  */
1529                 mmu_notifier_invalidate_range(vma->vm_mm, haddr,
1530                                               haddr + HPAGE_PMD_SIZE);
1531         }
1532
1533         /*
1534          * Migrate the THP to the requested node, returns with page unlocked
1535          * and access rights restored.
1536          */
1537         spin_unlock(vmf->ptl);
1538
1539         migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1540                                 vmf->pmd, pmd, vmf->address, page, target_nid);
1541         if (migrated) {
1542                 flags |= TNF_MIGRATED;
1543                 page_nid = target_nid;
1544         } else
1545                 flags |= TNF_MIGRATE_FAIL;
1546
1547         goto out;
1548 clear_pmdnuma:
1549         BUG_ON(!PageLocked(page));
1550         was_writable = pmd_savedwrite(pmd);
1551         pmd = pmd_modify(pmd, vma->vm_page_prot);
1552         pmd = pmd_mkyoung(pmd);
1553         if (was_writable)
1554                 pmd = pmd_mkwrite(pmd);
1555         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1556         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1557         unlock_page(page);
1558 out_unlock:
1559         spin_unlock(vmf->ptl);
1560
1561 out:
1562         if (anon_vma)
1563                 page_unlock_anon_vma_read(anon_vma);
1564
1565         if (page_nid != NUMA_NO_NODE)
1566                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1567                                 flags);
1568
1569         return 0;
1570 }
1571
1572 /*
1573  * Return true if we do MADV_FREE successfully on entire pmd page.
1574  * Otherwise, return false.
1575  */
1576 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1577                 pmd_t *pmd, unsigned long addr, unsigned long next)
1578 {
1579         spinlock_t *ptl;
1580         pmd_t orig_pmd;
1581         struct page *page;
1582         struct mm_struct *mm = tlb->mm;
1583         bool ret = false;
1584
1585         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1586
1587         ptl = pmd_trans_huge_lock(pmd, vma);
1588         if (!ptl)
1589                 goto out_unlocked;
1590
1591         orig_pmd = *pmd;
1592         if (is_huge_zero_pmd(orig_pmd))
1593                 goto out;
1594
1595         if (unlikely(!pmd_present(orig_pmd))) {
1596                 VM_BUG_ON(thp_migration_supported() &&
1597                                   !is_pmd_migration_entry(orig_pmd));
1598                 goto out;
1599         }
1600
1601         page = pmd_page(orig_pmd);
1602         /*
1603          * If other processes are mapping this page, we couldn't discard
1604          * the page unless they all do MADV_FREE so let's skip the page.
1605          */
1606         if (page_mapcount(page) != 1)
1607                 goto out;
1608
1609         if (!trylock_page(page))
1610                 goto out;
1611
1612         /*
1613          * If user want to discard part-pages of THP, split it so MADV_FREE
1614          * will deactivate only them.
1615          */
1616         if (next - addr != HPAGE_PMD_SIZE) {
1617                 get_page(page);
1618                 spin_unlock(ptl);
1619                 split_huge_page(page);
1620                 unlock_page(page);
1621                 put_page(page);
1622                 goto out_unlocked;
1623         }
1624
1625         if (PageDirty(page))
1626                 ClearPageDirty(page);
1627         unlock_page(page);
1628
1629         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1630                 pmdp_invalidate(vma, addr, pmd);
1631                 orig_pmd = pmd_mkold(orig_pmd);
1632                 orig_pmd = pmd_mkclean(orig_pmd);
1633
1634                 set_pmd_at(mm, addr, pmd, orig_pmd);
1635                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1636         }
1637
1638         mark_page_lazyfree(page);
1639         ret = true;
1640 out:
1641         spin_unlock(ptl);
1642 out_unlocked:
1643         return ret;
1644 }
1645
1646 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1647 {
1648         pgtable_t pgtable;
1649
1650         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1651         pte_free(mm, pgtable);
1652         mm_dec_nr_ptes(mm);
1653 }
1654
1655 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1656                  pmd_t *pmd, unsigned long addr)
1657 {
1658         pmd_t orig_pmd;
1659         spinlock_t *ptl;
1660
1661         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1662
1663         ptl = __pmd_trans_huge_lock(pmd, vma);
1664         if (!ptl)
1665                 return 0;
1666         /*
1667          * For architectures like ppc64 we look at deposited pgtable
1668          * when calling pmdp_huge_get_and_clear. So do the
1669          * pgtable_trans_huge_withdraw after finishing pmdp related
1670          * operations.
1671          */
1672         orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd,
1673                                                 tlb->fullmm);
1674         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1675         if (vma_is_special_huge(vma)) {
1676                 if (arch_needs_pgtable_deposit())
1677                         zap_deposited_table(tlb->mm, pmd);
1678                 spin_unlock(ptl);
1679                 if (is_huge_zero_pmd(orig_pmd))
1680                         tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1681         } else if (is_huge_zero_pmd(orig_pmd)) {
1682                 zap_deposited_table(tlb->mm, pmd);
1683                 spin_unlock(ptl);
1684                 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1685         } else {
1686                 struct page *page = NULL;
1687                 int flush_needed = 1;
1688
1689                 if (pmd_present(orig_pmd)) {
1690                         page = pmd_page(orig_pmd);
1691                         page_remove_rmap(page, true);
1692                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1693                         VM_BUG_ON_PAGE(!PageHead(page), page);
1694                 } else if (thp_migration_supported()) {
1695                         swp_entry_t entry;
1696
1697                         VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1698                         entry = pmd_to_swp_entry(orig_pmd);
1699                         page = pfn_to_page(swp_offset(entry));
1700                         flush_needed = 0;
1701                 } else
1702                         WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1703
1704                 if (PageAnon(page)) {
1705                         zap_deposited_table(tlb->mm, pmd);
1706                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1707                 } else {
1708                         if (arch_needs_pgtable_deposit())
1709                                 zap_deposited_table(tlb->mm, pmd);
1710                         add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1711                 }
1712
1713                 spin_unlock(ptl);
1714                 if (flush_needed)
1715                         tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1716         }
1717         return 1;
1718 }
1719
1720 #ifndef pmd_move_must_withdraw
1721 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1722                                          spinlock_t *old_pmd_ptl,
1723                                          struct vm_area_struct *vma)
1724 {
1725         /*
1726          * With split pmd lock we also need to move preallocated
1727          * PTE page table if new_pmd is on different PMD page table.
1728          *
1729          * We also don't deposit and withdraw tables for file pages.
1730          */
1731         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1732 }
1733 #endif
1734
1735 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1736 {
1737 #ifdef CONFIG_MEM_SOFT_DIRTY
1738         if (unlikely(is_pmd_migration_entry(pmd)))
1739                 pmd = pmd_swp_mksoft_dirty(pmd);
1740         else if (pmd_present(pmd))
1741                 pmd = pmd_mksoft_dirty(pmd);
1742 #endif
1743         return pmd;
1744 }
1745
1746 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1747                   unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
1748 {
1749         spinlock_t *old_ptl, *new_ptl;
1750         pmd_t pmd;
1751         struct mm_struct *mm = vma->vm_mm;
1752         bool force_flush = false;
1753
1754         /*
1755          * The destination pmd shouldn't be established, free_pgtables()
1756          * should have release it.
1757          */
1758         if (WARN_ON(!pmd_none(*new_pmd))) {
1759                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1760                 return false;
1761         }
1762
1763         /*
1764          * We don't have to worry about the ordering of src and dst
1765          * ptlocks because exclusive mmap_lock prevents deadlock.
1766          */
1767         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1768         if (old_ptl) {
1769                 new_ptl = pmd_lockptr(mm, new_pmd);
1770                 if (new_ptl != old_ptl)
1771                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1772                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1773                 if (pmd_present(pmd))
1774                         force_flush = true;
1775                 VM_BUG_ON(!pmd_none(*new_pmd));
1776
1777                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1778                         pgtable_t pgtable;
1779                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1780                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1781                 }
1782                 pmd = move_soft_dirty_pmd(pmd);
1783                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1784                 if (force_flush)
1785                         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1786                 if (new_ptl != old_ptl)
1787                         spin_unlock(new_ptl);
1788                 spin_unlock(old_ptl);
1789                 return true;
1790         }
1791         return false;
1792 }
1793
1794 /*
1795  * Returns
1796  *  - 0 if PMD could not be locked
1797  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1798  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1799  */
1800 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1801                 unsigned long addr, pgprot_t newprot, unsigned long cp_flags)
1802 {
1803         struct mm_struct *mm = vma->vm_mm;
1804         spinlock_t *ptl;
1805         pmd_t entry;
1806         bool preserve_write;
1807         int ret;
1808         bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
1809         bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
1810         bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
1811
1812         ptl = __pmd_trans_huge_lock(pmd, vma);
1813         if (!ptl)
1814                 return 0;
1815
1816         preserve_write = prot_numa && pmd_write(*pmd);
1817         ret = 1;
1818
1819 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1820         if (is_swap_pmd(*pmd)) {
1821                 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1822
1823                 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1824                 if (is_write_migration_entry(entry)) {
1825                         pmd_t newpmd;
1826                         /*
1827                          * A protection check is difficult so
1828                          * just be safe and disable write
1829                          */
1830                         make_migration_entry_read(&entry);
1831                         newpmd = swp_entry_to_pmd(entry);
1832                         if (pmd_swp_soft_dirty(*pmd))
1833                                 newpmd = pmd_swp_mksoft_dirty(newpmd);
1834                         set_pmd_at(mm, addr, pmd, newpmd);
1835                 }
1836                 goto unlock;
1837         }
1838 #endif
1839
1840         /*
1841          * Avoid trapping faults against the zero page. The read-only
1842          * data is likely to be read-cached on the local CPU and
1843          * local/remote hits to the zero page are not interesting.
1844          */
1845         if (prot_numa && is_huge_zero_pmd(*pmd))
1846                 goto unlock;
1847
1848         if (prot_numa && pmd_protnone(*pmd))
1849                 goto unlock;
1850
1851         /*
1852          * In case prot_numa, we are under mmap_read_lock(mm). It's critical
1853          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1854          * which is also under mmap_read_lock(mm):
1855          *
1856          *      CPU0:                           CPU1:
1857          *                              change_huge_pmd(prot_numa=1)
1858          *                               pmdp_huge_get_and_clear_notify()
1859          * madvise_dontneed()
1860          *  zap_pmd_range()
1861          *   pmd_trans_huge(*pmd) == 0 (without ptl)
1862          *   // skip the pmd
1863          *                               set_pmd_at();
1864          *                               // pmd is re-established
1865          *
1866          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1867          * which may break userspace.
1868          *
1869          * pmdp_invalidate() is required to make sure we don't miss
1870          * dirty/young flags set by hardware.
1871          */
1872         entry = pmdp_invalidate(vma, addr, pmd);
1873
1874         entry = pmd_modify(entry, newprot);
1875         if (preserve_write)
1876                 entry = pmd_mk_savedwrite(entry);
1877         if (uffd_wp) {
1878                 entry = pmd_wrprotect(entry);
1879                 entry = pmd_mkuffd_wp(entry);
1880         } else if (uffd_wp_resolve) {
1881                 /*
1882                  * Leave the write bit to be handled by PF interrupt
1883                  * handler, then things like COW could be properly
1884                  * handled.
1885                  */
1886                 entry = pmd_clear_uffd_wp(entry);
1887         }
1888         ret = HPAGE_PMD_NR;
1889         set_pmd_at(mm, addr, pmd, entry);
1890         BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1891 unlock:
1892         spin_unlock(ptl);
1893         return ret;
1894 }
1895
1896 /*
1897  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1898  *
1899  * Note that if it returns page table lock pointer, this routine returns without
1900  * unlocking page table lock. So callers must unlock it.
1901  */
1902 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1903 {
1904         spinlock_t *ptl;
1905         ptl = pmd_lock(vma->vm_mm, pmd);
1906         if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1907                         pmd_devmap(*pmd)))
1908                 return ptl;
1909         spin_unlock(ptl);
1910         return NULL;
1911 }
1912
1913 /*
1914  * Returns true if a given pud maps a thp, false otherwise.
1915  *
1916  * Note that if it returns true, this routine returns without unlocking page
1917  * table lock. So callers must unlock it.
1918  */
1919 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1920 {
1921         spinlock_t *ptl;
1922
1923         ptl = pud_lock(vma->vm_mm, pud);
1924         if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1925                 return ptl;
1926         spin_unlock(ptl);
1927         return NULL;
1928 }
1929
1930 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1931 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1932                  pud_t *pud, unsigned long addr)
1933 {
1934         spinlock_t *ptl;
1935
1936         ptl = __pud_trans_huge_lock(pud, vma);
1937         if (!ptl)
1938                 return 0;
1939         /*
1940          * For architectures like ppc64 we look at deposited pgtable
1941          * when calling pudp_huge_get_and_clear. So do the
1942          * pgtable_trans_huge_withdraw after finishing pudp related
1943          * operations.
1944          */
1945         pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
1946         tlb_remove_pud_tlb_entry(tlb, pud, addr);
1947         if (vma_is_special_huge(vma)) {
1948                 spin_unlock(ptl);
1949                 /* No zero page support yet */
1950         } else {
1951                 /* No support for anonymous PUD pages yet */
1952                 BUG();
1953         }
1954         return 1;
1955 }
1956
1957 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1958                 unsigned long haddr)
1959 {
1960         VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1961         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1962         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
1963         VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
1964
1965         count_vm_event(THP_SPLIT_PUD);
1966
1967         pudp_huge_clear_flush_notify(vma, haddr, pud);
1968 }
1969
1970 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
1971                 unsigned long address)
1972 {
1973         spinlock_t *ptl;
1974         struct mmu_notifier_range range;
1975
1976         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1977                                 address & HPAGE_PUD_MASK,
1978                                 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
1979         mmu_notifier_invalidate_range_start(&range);
1980         ptl = pud_lock(vma->vm_mm, pud);
1981         if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
1982                 goto out;
1983         __split_huge_pud_locked(vma, pud, range.start);
1984
1985 out:
1986         spin_unlock(ptl);
1987         /*
1988          * No need to double call mmu_notifier->invalidate_range() callback as
1989          * the above pudp_huge_clear_flush_notify() did already call it.
1990          */
1991         mmu_notifier_invalidate_range_only_end(&range);
1992 }
1993 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1994
1995 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1996                 unsigned long haddr, pmd_t *pmd)
1997 {
1998         struct mm_struct *mm = vma->vm_mm;
1999         pgtable_t pgtable;
2000         pmd_t _pmd;
2001         int i;
2002
2003         /*
2004          * Leave pmd empty until pte is filled note that it is fine to delay
2005          * notification until mmu_notifier_invalidate_range_end() as we are
2006          * replacing a zero pmd write protected page with a zero pte write
2007          * protected page.
2008          *
2009          * See Documentation/vm/mmu_notifier.rst
2010          */
2011         pmdp_huge_clear_flush(vma, haddr, pmd);
2012
2013         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2014         pmd_populate(mm, &_pmd, pgtable);
2015
2016         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2017                 pte_t *pte, entry;
2018                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2019                 entry = pte_mkspecial(entry);
2020                 pte = pte_offset_map(&_pmd, haddr);
2021                 VM_BUG_ON(!pte_none(*pte));
2022                 set_pte_at(mm, haddr, pte, entry);
2023                 pte_unmap(pte);
2024         }
2025         smp_wmb(); /* make pte visible before pmd */
2026         pmd_populate(mm, pmd, pgtable);
2027 }
2028
2029 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2030                 unsigned long haddr, bool freeze)
2031 {
2032         struct mm_struct *mm = vma->vm_mm;
2033         struct page *page;
2034         pgtable_t pgtable;
2035         pmd_t old_pmd, _pmd;
2036         bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false;
2037         unsigned long addr;
2038         int i;
2039
2040         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2041         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2042         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2043         VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2044                                 && !pmd_devmap(*pmd));
2045
2046         count_vm_event(THP_SPLIT_PMD);
2047
2048         if (!vma_is_anonymous(vma)) {
2049                 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2050                 /*
2051                  * We are going to unmap this huge page. So
2052                  * just go ahead and zap it
2053                  */
2054                 if (arch_needs_pgtable_deposit())
2055                         zap_deposited_table(mm, pmd);
2056                 if (vma_is_special_huge(vma))
2057                         return;
2058                 page = pmd_page(_pmd);
2059                 if (!PageDirty(page) && pmd_dirty(_pmd))
2060                         set_page_dirty(page);
2061                 if (!PageReferenced(page) && pmd_young(_pmd))
2062                         SetPageReferenced(page);
2063                 page_remove_rmap(page, true);
2064                 put_page(page);
2065                 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2066                 return;
2067         } else if (pmd_trans_huge(*pmd) && is_huge_zero_pmd(*pmd)) {
2068                 /*
2069                  * FIXME: Do we want to invalidate secondary mmu by calling
2070                  * mmu_notifier_invalidate_range() see comments below inside
2071                  * __split_huge_pmd() ?
2072                  *
2073                  * We are going from a zero huge page write protected to zero
2074                  * small page also write protected so it does not seems useful
2075                  * to invalidate secondary mmu at this time.
2076                  */
2077                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2078         }
2079
2080         /*
2081          * Up to this point the pmd is present and huge and userland has the
2082          * whole access to the hugepage during the split (which happens in
2083          * place). If we overwrite the pmd with the not-huge version pointing
2084          * to the pte here (which of course we could if all CPUs were bug
2085          * free), userland could trigger a small page size TLB miss on the
2086          * small sized TLB while the hugepage TLB entry is still established in
2087          * the huge TLB. Some CPU doesn't like that.
2088          * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum
2089          * 383 on page 105. Intel should be safe but is also warns that it's
2090          * only safe if the permission and cache attributes of the two entries
2091          * loaded in the two TLB is identical (which should be the case here).
2092          * But it is generally safer to never allow small and huge TLB entries
2093          * for the same virtual address to be loaded simultaneously. So instead
2094          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2095          * current pmd notpresent (atomically because here the pmd_trans_huge
2096          * must remain set at all times on the pmd until the split is complete
2097          * for this pmd), then we flush the SMP TLB and finally we write the
2098          * non-huge version of the pmd entry with pmd_populate.
2099          */
2100         old_pmd = pmdp_invalidate(vma, haddr, pmd);
2101
2102         pmd_migration = is_pmd_migration_entry(old_pmd);
2103         if (unlikely(pmd_migration)) {
2104                 swp_entry_t entry;
2105
2106                 entry = pmd_to_swp_entry(old_pmd);
2107                 page = pfn_to_page(swp_offset(entry));
2108                 write = is_write_migration_entry(entry);
2109                 young = false;
2110                 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2111                 uffd_wp = pmd_swp_uffd_wp(old_pmd);
2112         } else {
2113                 page = pmd_page(old_pmd);
2114                 if (pmd_dirty(old_pmd))
2115                         SetPageDirty(page);
2116                 write = pmd_write(old_pmd);
2117                 young = pmd_young(old_pmd);
2118                 soft_dirty = pmd_soft_dirty(old_pmd);
2119                 uffd_wp = pmd_uffd_wp(old_pmd);
2120         }
2121         VM_BUG_ON_PAGE(!page_count(page), page);
2122         page_ref_add(page, HPAGE_PMD_NR - 1);
2123
2124         /*
2125          * Withdraw the table only after we mark the pmd entry invalid.
2126          * This's critical for some architectures (Power).
2127          */
2128         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2129         pmd_populate(mm, &_pmd, pgtable);
2130
2131         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2132                 pte_t entry, *pte;
2133                 /*
2134                  * Note that NUMA hinting access restrictions are not
2135                  * transferred to avoid any possibility of altering
2136                  * permissions across VMAs.
2137                  */
2138                 if (freeze || pmd_migration) {
2139                         swp_entry_t swp_entry;
2140                         swp_entry = make_migration_entry(page + i, write);
2141                         entry = swp_entry_to_pte(swp_entry);
2142                         if (soft_dirty)
2143                                 entry = pte_swp_mksoft_dirty(entry);
2144                         if (uffd_wp)
2145                                 entry = pte_swp_mkuffd_wp(entry);
2146                 } else {
2147                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2148                         entry = maybe_mkwrite(entry, vma);
2149                         if (!write)
2150                                 entry = pte_wrprotect(entry);
2151                         if (!young)
2152                                 entry = pte_mkold(entry);
2153                         if (soft_dirty)
2154                                 entry = pte_mksoft_dirty(entry);
2155                         if (uffd_wp)
2156                                 entry = pte_mkuffd_wp(entry);
2157                 }
2158                 pte = pte_offset_map(&_pmd, addr);
2159                 BUG_ON(!pte_none(*pte));
2160                 set_pte_at(mm, addr, pte, entry);
2161                 if (!pmd_migration)
2162                         atomic_inc(&page[i]._mapcount);
2163                 pte_unmap(pte);
2164         }
2165
2166         if (!pmd_migration) {
2167                 /*
2168                  * Set PG_double_map before dropping compound_mapcount to avoid
2169                  * false-negative page_mapped().
2170                  */
2171                 if (compound_mapcount(page) > 1 &&
2172                     !TestSetPageDoubleMap(page)) {
2173                         for (i = 0; i < HPAGE_PMD_NR; i++)
2174                                 atomic_inc(&page[i]._mapcount);
2175                 }
2176
2177                 lock_page_memcg(page);
2178                 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2179                         /* Last compound_mapcount is gone. */
2180                         __mod_lruvec_page_state(page, NR_ANON_THPS,
2181                                                 -HPAGE_PMD_NR);
2182                         if (TestClearPageDoubleMap(page)) {
2183                                 /* No need in mapcount reference anymore */
2184                                 for (i = 0; i < HPAGE_PMD_NR; i++)
2185                                         atomic_dec(&page[i]._mapcount);
2186                         }
2187                 }
2188                 unlock_page_memcg(page);
2189         }
2190
2191         smp_wmb(); /* make pte visible before pmd */
2192         pmd_populate(mm, pmd, pgtable);
2193
2194         if (freeze) {
2195                 for (i = 0; i < HPAGE_PMD_NR; i++) {
2196                         page_remove_rmap(page + i, false);
2197                         put_page(page + i);
2198                 }
2199         }
2200 }
2201
2202 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2203                 unsigned long address, bool freeze, struct page *page)
2204 {
2205         spinlock_t *ptl;
2206         struct mmu_notifier_range range;
2207         bool do_unlock_page = false;
2208         pmd_t _pmd;
2209
2210         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2211                                 address & HPAGE_PMD_MASK,
2212                                 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2213         mmu_notifier_invalidate_range_start(&range);
2214         ptl = pmd_lock(vma->vm_mm, pmd);
2215
2216         /*
2217          * If caller asks to setup a migration entries, we need a page to check
2218          * pmd against. Otherwise we can end up replacing wrong page.
2219          */
2220         VM_BUG_ON(freeze && !page);
2221         if (page) {
2222                 VM_WARN_ON_ONCE(!PageLocked(page));
2223                 if (page != pmd_page(*pmd))
2224                         goto out;
2225         }
2226
2227 repeat:
2228         if (pmd_trans_huge(*pmd)) {
2229                 if (!page) {
2230                         page = pmd_page(*pmd);
2231                         /*
2232                          * An anonymous page must be locked, to ensure that a
2233                          * concurrent reuse_swap_page() sees stable mapcount;
2234                          * but reuse_swap_page() is not used on shmem or file,
2235                          * and page lock must not be taken when zap_pmd_range()
2236                          * calls __split_huge_pmd() while i_mmap_lock is held.
2237                          */
2238                         if (PageAnon(page)) {
2239                                 if (unlikely(!trylock_page(page))) {
2240                                         get_page(page);
2241                                         _pmd = *pmd;
2242                                         spin_unlock(ptl);
2243                                         lock_page(page);
2244                                         spin_lock(ptl);
2245                                         if (unlikely(!pmd_same(*pmd, _pmd))) {
2246                                                 unlock_page(page);
2247                                                 put_page(page);
2248                                                 page = NULL;
2249                                                 goto repeat;
2250                                         }
2251                                         put_page(page);
2252                                 }
2253                                 do_unlock_page = true;
2254                         }
2255                 }
2256                 if (PageMlocked(page))
2257                         clear_page_mlock(page);
2258         } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2259                 goto out;
2260         __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2261 out:
2262         spin_unlock(ptl);
2263         if (do_unlock_page)
2264                 unlock_page(page);
2265         /*
2266          * No need to double call mmu_notifier->invalidate_range() callback.
2267          * They are 3 cases to consider inside __split_huge_pmd_locked():
2268          *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2269          *  2) __split_huge_zero_page_pmd() read only zero page and any write
2270          *    fault will trigger a flush_notify before pointing to a new page
2271          *    (it is fine if the secondary mmu keeps pointing to the old zero
2272          *    page in the meantime)
2273          *  3) Split a huge pmd into pte pointing to the same page. No need
2274          *     to invalidate secondary tlb entry they are all still valid.
2275          *     any further changes to individual pte will notify. So no need
2276          *     to call mmu_notifier->invalidate_range()
2277          */
2278         mmu_notifier_invalidate_range_only_end(&range);
2279 }
2280
2281 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2282                 bool freeze, struct page *page)
2283 {
2284         pgd_t *pgd;
2285         p4d_t *p4d;
2286         pud_t *pud;
2287         pmd_t *pmd;
2288
2289         pgd = pgd_offset(vma->vm_mm, address);
2290         if (!pgd_present(*pgd))
2291                 return;
2292
2293         p4d = p4d_offset(pgd, address);
2294         if (!p4d_present(*p4d))
2295                 return;
2296
2297         pud = pud_offset(p4d, address);
2298         if (!pud_present(*pud))
2299                 return;
2300
2301         pmd = pmd_offset(pud, address);
2302
2303         __split_huge_pmd(vma, pmd, address, freeze, page);
2304 }
2305
2306 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2307                              unsigned long start,
2308                              unsigned long end,
2309                              long adjust_next)
2310 {
2311         /*
2312          * If the new start address isn't hpage aligned and it could
2313          * previously contain an hugepage: check if we need to split
2314          * an huge pmd.
2315          */
2316         if (start & ~HPAGE_PMD_MASK &&
2317             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2318             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2319                 split_huge_pmd_address(vma, start, false, NULL);
2320
2321         /*
2322          * If the new end address isn't hpage aligned and it could
2323          * previously contain an hugepage: check if we need to split
2324          * an huge pmd.
2325          */
2326         if (end & ~HPAGE_PMD_MASK &&
2327             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2328             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2329                 split_huge_pmd_address(vma, end, false, NULL);
2330
2331         /*
2332          * If we're also updating the vma->vm_next->vm_start, if the new
2333          * vm_next->vm_start isn't hpage aligned and it could previously
2334          * contain an hugepage: check if we need to split an huge pmd.
2335          */
2336         if (adjust_next > 0) {
2337                 struct vm_area_struct *next = vma->vm_next;
2338                 unsigned long nstart = next->vm_start;
2339                 nstart += adjust_next;
2340                 if (nstart & ~HPAGE_PMD_MASK &&
2341                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2342                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2343                         split_huge_pmd_address(next, nstart, false, NULL);
2344         }
2345 }
2346
2347 static void unmap_page(struct page *page)
2348 {
2349         enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK |
2350                 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2351         bool unmap_success;
2352
2353         VM_BUG_ON_PAGE(!PageHead(page), page);
2354
2355         if (PageAnon(page))
2356                 ttu_flags |= TTU_SPLIT_FREEZE;
2357
2358         unmap_success = try_to_unmap(page, ttu_flags);
2359         VM_BUG_ON_PAGE(!unmap_success, page);
2360 }
2361
2362 static void remap_page(struct page *page, unsigned int nr)
2363 {
2364         int i;
2365         if (PageTransHuge(page)) {
2366                 remove_migration_ptes(page, page, true);
2367         } else {
2368                 for (i = 0; i < nr; i++)
2369                         remove_migration_ptes(page + i, page + i, true);
2370         }
2371 }
2372
2373 static void lru_add_page_tail(struct page *head, struct page *tail,
2374                 struct lruvec *lruvec, struct list_head *list)
2375 {
2376         VM_BUG_ON_PAGE(!PageHead(head), head);
2377         VM_BUG_ON_PAGE(PageCompound(tail), head);
2378         VM_BUG_ON_PAGE(PageLRU(tail), head);
2379         lockdep_assert_held(&lruvec->lru_lock);
2380
2381         if (list) {
2382                 /* page reclaim is reclaiming a huge page */
2383                 VM_WARN_ON(PageLRU(head));
2384                 get_page(tail);
2385                 list_add_tail(&tail->lru, list);
2386         } else {
2387                 /* head is still on lru (and we have it frozen) */
2388                 VM_WARN_ON(!PageLRU(head));
2389                 SetPageLRU(tail);
2390                 list_add_tail(&tail->lru, &head->lru);
2391         }
2392 }
2393
2394 static void __split_huge_page_tail(struct page *head, int tail,
2395                 struct lruvec *lruvec, struct list_head *list)
2396 {
2397         struct page *page_tail = head + tail;
2398
2399         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2400
2401         /*
2402          * Clone page flags before unfreezing refcount.
2403          *
2404          * After successful get_page_unless_zero() might follow flags change,
2405          * for example lock_page() which set PG_waiters.
2406          */
2407         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2408         page_tail->flags |= (head->flags &
2409                         ((1L << PG_referenced) |
2410                          (1L << PG_swapbacked) |
2411                          (1L << PG_swapcache) |
2412                          (1L << PG_mlocked) |
2413                          (1L << PG_uptodate) |
2414                          (1L << PG_active) |
2415                          (1L << PG_workingset) |
2416                          (1L << PG_locked) |
2417                          (1L << PG_unevictable) |
2418 #ifdef CONFIG_64BIT
2419                          (1L << PG_arch_2) |
2420 #endif
2421                          (1L << PG_dirty)));
2422
2423         /* ->mapping in first tail page is compound_mapcount */
2424         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2425                         page_tail);
2426         page_tail->mapping = head->mapping;
2427         page_tail->index = head->index + tail;
2428
2429         /* Page flags must be visible before we make the page non-compound. */
2430         smp_wmb();
2431
2432         /*
2433          * Clear PageTail before unfreezing page refcount.
2434          *
2435          * After successful get_page_unless_zero() might follow put_page()
2436          * which needs correct compound_head().
2437          */
2438         clear_compound_head(page_tail);
2439
2440         /* Finally unfreeze refcount. Additional reference from page cache. */
2441         page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2442                                           PageSwapCache(head)));
2443
2444         if (page_is_young(head))
2445                 set_page_young(page_tail);
2446         if (page_is_idle(head))
2447                 set_page_idle(page_tail);
2448
2449         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2450
2451         /*
2452          * always add to the tail because some iterators expect new
2453          * pages to show after the currently processed elements - e.g.
2454          * migrate_pages
2455          */
2456         lru_add_page_tail(head, page_tail, lruvec, list);
2457 }
2458
2459 static void __split_huge_page(struct page *page, struct list_head *list,
2460                 pgoff_t end)
2461 {
2462         struct page *head = compound_head(page);
2463         struct lruvec *lruvec;
2464         struct address_space *swap_cache = NULL;
2465         unsigned long offset = 0;
2466         unsigned int nr = thp_nr_pages(head);
2467         int i;
2468
2469         /* complete memcg works before add pages to LRU */
2470         split_page_memcg(head, nr);
2471
2472         if (PageAnon(head) && PageSwapCache(head)) {
2473                 swp_entry_t entry = { .val = page_private(head) };
2474
2475                 offset = swp_offset(entry);
2476                 swap_cache = swap_address_space(entry);
2477                 xa_lock(&swap_cache->i_pages);
2478         }
2479
2480         /* lock lru list/PageCompound, ref freezed by page_ref_freeze */
2481         lruvec = lock_page_lruvec(head);
2482
2483         for (i = nr - 1; i >= 1; i--) {
2484                 __split_huge_page_tail(head, i, lruvec, list);
2485                 /* Some pages can be beyond i_size: drop them from page cache */
2486                 if (head[i].index >= end) {
2487                         ClearPageDirty(head + i);
2488                         __delete_from_page_cache(head + i, NULL);
2489                         if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2490                                 shmem_uncharge(head->mapping->host, 1);
2491                         put_page(head + i);
2492                 } else if (!PageAnon(page)) {
2493                         __xa_store(&head->mapping->i_pages, head[i].index,
2494                                         head + i, 0);
2495                 } else if (swap_cache) {
2496                         __xa_store(&swap_cache->i_pages, offset + i,
2497                                         head + i, 0);
2498                 }
2499         }
2500
2501         ClearPageCompound(head);
2502         unlock_page_lruvec(lruvec);
2503         /* Caller disabled irqs, so they are still disabled here */
2504
2505         split_page_owner(head, nr);
2506
2507         /* See comment in __split_huge_page_tail() */
2508         if (PageAnon(head)) {
2509                 /* Additional pin to swap cache */
2510                 if (PageSwapCache(head)) {
2511                         page_ref_add(head, 2);
2512                         xa_unlock(&swap_cache->i_pages);
2513                 } else {
2514                         page_ref_inc(head);
2515                 }
2516         } else {
2517                 /* Additional pin to page cache */
2518                 page_ref_add(head, 2);
2519                 xa_unlock(&head->mapping->i_pages);
2520         }
2521         local_irq_enable();
2522
2523         remap_page(head, nr);
2524
2525         if (PageSwapCache(head)) {
2526                 swp_entry_t entry = { .val = page_private(head) };
2527
2528                 split_swap_cluster(entry);
2529         }
2530
2531         for (i = 0; i < nr; i++) {
2532                 struct page *subpage = head + i;
2533                 if (subpage == page)
2534                         continue;
2535                 unlock_page(subpage);
2536
2537                 /*
2538                  * Subpages may be freed if there wasn't any mapping
2539                  * like if add_to_swap() is running on a lru page that
2540                  * had its mapping zapped. And freeing these pages
2541                  * requires taking the lru_lock so we do the put_page
2542                  * of the tail pages after the split is complete.
2543                  */
2544                 put_page(subpage);
2545         }
2546 }
2547
2548 int total_mapcount(struct page *page)
2549 {
2550         int i, compound, nr, ret;
2551
2552         VM_BUG_ON_PAGE(PageTail(page), page);
2553
2554         if (likely(!PageCompound(page)))
2555                 return atomic_read(&page->_mapcount) + 1;
2556
2557         compound = compound_mapcount(page);
2558         nr = compound_nr(page);
2559         if (PageHuge(page))
2560                 return compound;
2561         ret = compound;
2562         for (i = 0; i < nr; i++)
2563                 ret += atomic_read(&page[i]._mapcount) + 1;
2564         /* File pages has compound_mapcount included in _mapcount */
2565         if (!PageAnon(page))
2566                 return ret - compound * nr;
2567         if (PageDoubleMap(page))
2568                 ret -= nr;
2569         return ret;
2570 }
2571
2572 /*
2573  * This calculates accurately how many mappings a transparent hugepage
2574  * has (unlike page_mapcount() which isn't fully accurate). This full
2575  * accuracy is primarily needed to know if copy-on-write faults can
2576  * reuse the page and change the mapping to read-write instead of
2577  * copying them. At the same time this returns the total_mapcount too.
2578  *
2579  * The function returns the highest mapcount any one of the subpages
2580  * has. If the return value is one, even if different processes are
2581  * mapping different subpages of the transparent hugepage, they can
2582  * all reuse it, because each process is reusing a different subpage.
2583  *
2584  * The total_mapcount is instead counting all virtual mappings of the
2585  * subpages. If the total_mapcount is equal to "one", it tells the
2586  * caller all mappings belong to the same "mm" and in turn the
2587  * anon_vma of the transparent hugepage can become the vma->anon_vma
2588  * local one as no other process may be mapping any of the subpages.
2589  *
2590  * It would be more accurate to replace page_mapcount() with
2591  * page_trans_huge_mapcount(), however we only use
2592  * page_trans_huge_mapcount() in the copy-on-write faults where we
2593  * need full accuracy to avoid breaking page pinning, because
2594  * page_trans_huge_mapcount() is slower than page_mapcount().
2595  */
2596 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2597 {
2598         int i, ret, _total_mapcount, mapcount;
2599
2600         /* hugetlbfs shouldn't call it */
2601         VM_BUG_ON_PAGE(PageHuge(page), page);
2602
2603         if (likely(!PageTransCompound(page))) {
2604                 mapcount = atomic_read(&page->_mapcount) + 1;
2605                 if (total_mapcount)
2606                         *total_mapcount = mapcount;
2607                 return mapcount;
2608         }
2609
2610         page = compound_head(page);
2611
2612         _total_mapcount = ret = 0;
2613         for (i = 0; i < thp_nr_pages(page); i++) {
2614                 mapcount = atomic_read(&page[i]._mapcount) + 1;
2615                 ret = max(ret, mapcount);
2616                 _total_mapcount += mapcount;
2617         }
2618         if (PageDoubleMap(page)) {
2619                 ret -= 1;
2620                 _total_mapcount -= thp_nr_pages(page);
2621         }
2622         mapcount = compound_mapcount(page);
2623         ret += mapcount;
2624         _total_mapcount += mapcount;
2625         if (total_mapcount)
2626                 *total_mapcount = _total_mapcount;
2627         return ret;
2628 }
2629
2630 /* Racy check whether the huge page can be split */
2631 bool can_split_huge_page(struct page *page, int *pextra_pins)
2632 {
2633         int extra_pins;
2634
2635         /* Additional pins from page cache */
2636         if (PageAnon(page))
2637                 extra_pins = PageSwapCache(page) ? thp_nr_pages(page) : 0;
2638         else
2639                 extra_pins = thp_nr_pages(page);
2640         if (pextra_pins)
2641                 *pextra_pins = extra_pins;
2642         return total_mapcount(page) == page_count(page) - extra_pins - 1;
2643 }
2644
2645 /*
2646  * This function splits huge page into normal pages. @page can point to any
2647  * subpage of huge page to split. Split doesn't change the position of @page.
2648  *
2649  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2650  * The huge page must be locked.
2651  *
2652  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2653  *
2654  * Both head page and tail pages will inherit mapping, flags, and so on from
2655  * the hugepage.
2656  *
2657  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2658  * they are not mapped.
2659  *
2660  * Returns 0 if the hugepage is split successfully.
2661  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2662  * us.
2663  */
2664 int split_huge_page_to_list(struct page *page, struct list_head *list)
2665 {
2666         struct page *head = compound_head(page);
2667         struct deferred_split *ds_queue = get_deferred_split_queue(head);
2668         struct anon_vma *anon_vma = NULL;
2669         struct address_space *mapping = NULL;
2670         int count, mapcount, extra_pins, ret;
2671         pgoff_t end;
2672
2673         VM_BUG_ON_PAGE(is_huge_zero_page(head), head);
2674         VM_BUG_ON_PAGE(!PageLocked(head), head);
2675         VM_BUG_ON_PAGE(!PageCompound(head), head);
2676
2677         if (PageWriteback(head))
2678                 return -EBUSY;
2679
2680         if (PageAnon(head)) {
2681                 /*
2682                  * The caller does not necessarily hold an mmap_lock that would
2683                  * prevent the anon_vma disappearing so we first we take a
2684                  * reference to it and then lock the anon_vma for write. This
2685                  * is similar to page_lock_anon_vma_read except the write lock
2686                  * is taken to serialise against parallel split or collapse
2687                  * operations.
2688                  */
2689                 anon_vma = page_get_anon_vma(head);
2690                 if (!anon_vma) {
2691                         ret = -EBUSY;
2692                         goto out;
2693                 }
2694                 end = -1;
2695                 mapping = NULL;
2696                 anon_vma_lock_write(anon_vma);
2697         } else {
2698                 mapping = head->mapping;
2699
2700                 /* Truncated ? */
2701                 if (!mapping) {
2702                         ret = -EBUSY;
2703                         goto out;
2704                 }
2705
2706                 anon_vma = NULL;
2707                 i_mmap_lock_read(mapping);
2708
2709                 /*
2710                  *__split_huge_page() may need to trim off pages beyond EOF:
2711                  * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2712                  * which cannot be nested inside the page tree lock. So note
2713                  * end now: i_size itself may be changed at any moment, but
2714                  * head page lock is good enough to serialize the trimming.
2715                  */
2716                 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2717         }
2718
2719         /*
2720          * Racy check if we can split the page, before unmap_page() will
2721          * split PMDs
2722          */
2723         if (!can_split_huge_page(head, &extra_pins)) {
2724                 ret = -EBUSY;
2725                 goto out_unlock;
2726         }
2727
2728         unmap_page(head);
2729         VM_BUG_ON_PAGE(compound_mapcount(head), head);
2730
2731         /* block interrupt reentry in xa_lock and spinlock */
2732         local_irq_disable();
2733         if (mapping) {
2734                 XA_STATE(xas, &mapping->i_pages, page_index(head));
2735
2736                 /*
2737                  * Check if the head page is present in page cache.
2738                  * We assume all tail are present too, if head is there.
2739                  */
2740                 xa_lock(&mapping->i_pages);
2741                 if (xas_load(&xas) != head)
2742                         goto fail;
2743         }
2744
2745         /* Prevent deferred_split_scan() touching ->_refcount */
2746         spin_lock(&ds_queue->split_queue_lock);
2747         count = page_count(head);
2748         mapcount = total_mapcount(head);
2749         if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2750                 if (!list_empty(page_deferred_list(head))) {
2751                         ds_queue->split_queue_len--;
2752                         list_del(page_deferred_list(head));
2753                 }
2754                 spin_unlock(&ds_queue->split_queue_lock);
2755                 if (mapping) {
2756                         int nr = thp_nr_pages(head);
2757
2758                         if (PageSwapBacked(head))
2759                                 __mod_lruvec_page_state(head, NR_SHMEM_THPS,
2760                                                         -nr);
2761                         else
2762                                 __mod_lruvec_page_state(head, NR_FILE_THPS,
2763                                                         -nr);
2764                 }
2765
2766                 __split_huge_page(page, list, end);
2767                 ret = 0;
2768         } else {
2769                 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2770                         pr_alert("total_mapcount: %u, page_count(): %u\n",
2771                                         mapcount, count);
2772                         if (PageTail(page))
2773                                 dump_page(head, NULL);
2774                         dump_page(page, "total_mapcount(head) > 0");
2775                         BUG();
2776                 }
2777                 spin_unlock(&ds_queue->split_queue_lock);
2778 fail:           if (mapping)
2779                         xa_unlock(&mapping->i_pages);
2780                 local_irq_enable();
2781                 remap_page(head, thp_nr_pages(head));
2782                 ret = -EBUSY;
2783         }
2784
2785 out_unlock:
2786         if (anon_vma) {
2787                 anon_vma_unlock_write(anon_vma);
2788                 put_anon_vma(anon_vma);
2789         }
2790         if (mapping)
2791                 i_mmap_unlock_read(mapping);
2792 out:
2793         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2794         return ret;
2795 }
2796
2797 void free_transhuge_page(struct page *page)
2798 {
2799         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2800         unsigned long flags;
2801
2802         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2803         if (!list_empty(page_deferred_list(page))) {
2804                 ds_queue->split_queue_len--;
2805                 list_del(page_deferred_list(page));
2806         }
2807         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2808         free_compound_page(page);
2809 }
2810
2811 void deferred_split_huge_page(struct page *page)
2812 {
2813         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2814 #ifdef CONFIG_MEMCG
2815         struct mem_cgroup *memcg = page_memcg(compound_head(page));
2816 #endif
2817         unsigned long flags;
2818
2819         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2820
2821         /*
2822          * The try_to_unmap() in page reclaim path might reach here too,
2823          * this may cause a race condition to corrupt deferred split queue.
2824          * And, if page reclaim is already handling the same page, it is
2825          * unnecessary to handle it again in shrinker.
2826          *
2827          * Check PageSwapCache to determine if the page is being
2828          * handled by page reclaim since THP swap would add the page into
2829          * swap cache before calling try_to_unmap().
2830          */
2831         if (PageSwapCache(page))
2832                 return;
2833
2834         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2835         if (list_empty(page_deferred_list(page))) {
2836                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2837                 list_add_tail(page_deferred_list(page), &ds_queue->split_queue);
2838                 ds_queue->split_queue_len++;
2839 #ifdef CONFIG_MEMCG
2840                 if (memcg)
2841                         memcg_set_shrinker_bit(memcg, page_to_nid(page),
2842                                                deferred_split_shrinker.id);
2843 #endif
2844         }
2845         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2846 }
2847
2848 static unsigned long deferred_split_count(struct shrinker *shrink,
2849                 struct shrink_control *sc)
2850 {
2851         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2852         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2853
2854 #ifdef CONFIG_MEMCG
2855         if (sc->memcg)
2856                 ds_queue = &sc->memcg->deferred_split_queue;
2857 #endif
2858         return READ_ONCE(ds_queue->split_queue_len);
2859 }
2860
2861 static unsigned long deferred_split_scan(struct shrinker *shrink,
2862                 struct shrink_control *sc)
2863 {
2864         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2865         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2866         unsigned long flags;
2867         LIST_HEAD(list), *pos, *next;
2868         struct page *page;
2869         int split = 0;
2870
2871 #ifdef CONFIG_MEMCG
2872         if (sc->memcg)
2873                 ds_queue = &sc->memcg->deferred_split_queue;
2874 #endif
2875
2876         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2877         /* Take pin on all head pages to avoid freeing them under us */
2878         list_for_each_safe(pos, next, &ds_queue->split_queue) {
2879                 page = list_entry((void *)pos, struct page, mapping);
2880                 page = compound_head(page);
2881                 if (get_page_unless_zero(page)) {
2882                         list_move(page_deferred_list(page), &list);
2883                 } else {
2884                         /* We lost race with put_compound_page() */
2885                         list_del_init(page_deferred_list(page));
2886                         ds_queue->split_queue_len--;
2887                 }
2888                 if (!--sc->nr_to_scan)
2889                         break;
2890         }
2891         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2892
2893         list_for_each_safe(pos, next, &list) {
2894                 page = list_entry((void *)pos, struct page, mapping);
2895                 if (!trylock_page(page))
2896                         goto next;
2897                 /* split_huge_page() removes page from list on success */
2898                 if (!split_huge_page(page))
2899                         split++;
2900                 unlock_page(page);
2901 next:
2902                 put_page(page);
2903         }
2904
2905         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2906         list_splice_tail(&list, &ds_queue->split_queue);
2907         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2908
2909         /*
2910          * Stop shrinker if we didn't split any page, but the queue is empty.
2911          * This can happen if pages were freed under us.
2912          */
2913         if (!split && list_empty(&ds_queue->split_queue))
2914                 return SHRINK_STOP;
2915         return split;
2916 }
2917
2918 static struct shrinker deferred_split_shrinker = {
2919         .count_objects = deferred_split_count,
2920         .scan_objects = deferred_split_scan,
2921         .seeks = DEFAULT_SEEKS,
2922         .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
2923                  SHRINKER_NONSLAB,
2924 };
2925
2926 #ifdef CONFIG_DEBUG_FS
2927 static int split_huge_pages_set(void *data, u64 val)
2928 {
2929         struct zone *zone;
2930         struct page *page;
2931         unsigned long pfn, max_zone_pfn;
2932         unsigned long total = 0, split = 0;
2933
2934         if (val != 1)
2935                 return -EINVAL;
2936
2937         for_each_populated_zone(zone) {
2938                 max_zone_pfn = zone_end_pfn(zone);
2939                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2940                         if (!pfn_valid(pfn))
2941                                 continue;
2942
2943                         page = pfn_to_page(pfn);
2944                         if (!get_page_unless_zero(page))
2945                                 continue;
2946
2947                         if (zone != page_zone(page))
2948                                 goto next;
2949
2950                         if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2951                                 goto next;
2952
2953                         total++;
2954                         lock_page(page);
2955                         if (!split_huge_page(page))
2956                                 split++;
2957                         unlock_page(page);
2958 next:
2959                         put_page(page);
2960                 }
2961         }
2962
2963         pr_info("%lu of %lu THP split\n", split, total);
2964
2965         return 0;
2966 }
2967 DEFINE_DEBUGFS_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2968                 "%llu\n");
2969
2970 static int __init split_huge_pages_debugfs(void)
2971 {
2972         debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2973                             &split_huge_pages_fops);
2974         return 0;
2975 }
2976 late_initcall(split_huge_pages_debugfs);
2977 #endif
2978
2979 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2980 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
2981                 struct page *page)
2982 {
2983         struct vm_area_struct *vma = pvmw->vma;
2984         struct mm_struct *mm = vma->vm_mm;
2985         unsigned long address = pvmw->address;
2986         pmd_t pmdval;
2987         swp_entry_t entry;
2988         pmd_t pmdswp;
2989
2990         if (!(pvmw->pmd && !pvmw->pte))
2991                 return;
2992
2993         flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
2994         pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
2995         if (pmd_dirty(pmdval))
2996                 set_page_dirty(page);
2997         entry = make_migration_entry(page, pmd_write(pmdval));
2998         pmdswp = swp_entry_to_pmd(entry);
2999         if (pmd_soft_dirty(pmdval))
3000                 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
3001         set_pmd_at(mm, address, pvmw->pmd, pmdswp);
3002         page_remove_rmap(page, true);
3003         put_page(page);
3004 }
3005
3006 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
3007 {
3008         struct vm_area_struct *vma = pvmw->vma;
3009         struct mm_struct *mm = vma->vm_mm;
3010         unsigned long address = pvmw->address;
3011         unsigned long mmun_start = address & HPAGE_PMD_MASK;
3012         pmd_t pmde;
3013         swp_entry_t entry;
3014
3015         if (!(pvmw->pmd && !pvmw->pte))
3016                 return;
3017
3018         entry = pmd_to_swp_entry(*pvmw->pmd);
3019         get_page(new);
3020         pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
3021         if (pmd_swp_soft_dirty(*pvmw->pmd))
3022                 pmde = pmd_mksoft_dirty(pmde);
3023         if (is_write_migration_entry(entry))
3024                 pmde = maybe_pmd_mkwrite(pmde, vma);
3025
3026         flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
3027         if (PageAnon(new))
3028                 page_add_anon_rmap(new, vma, mmun_start, true);
3029         else
3030                 page_add_file_rmap(new, true);
3031         set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
3032         if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new))
3033                 mlock_vma_page(new);
3034         update_mmu_cache_pmd(vma, address, pvmw->pmd);
3035 }
3036 #endif