mm/huge_memory.c

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