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