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                 count_memcg_events(memcg, THP_FAULT_ALLOC, 1);
 621         }
 622
 623         return 0;
 624 unlock_release:
 625         spin_unlock(vmf->ptl);
 626 release:
 627         if (pgtable)
 628                 pte_free(vma->vm_mm, pgtable);
 629         mem_cgroup_cancel_charge(page, memcg, true);
 630         put_page(page);
 631         return ret;
 632
 633 }
 634
 635 /*
 636  * always: directly stall for all thp allocations
 637  * defer: wake kswapd and fail if not immediately available
 638  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
 639  *                fail if not immediately available
 640  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
 641  *          available
 642  * never: never stall for any thp allocation
 643  */
 644 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
 645 {
 646         const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
 647
 648         /* Always do synchronous compaction */
 649         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 650                 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
 651
 652         /* Kick kcompactd and fail quickly */
 653         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 654                 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
 655
 656         /* Synchronous compaction if madvised, otherwise kick kcompactd */
 657         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 658                 return GFP_TRANSHUGE_LIGHT |
 659                         (vma_madvised ? __GFP_DIRECT_RECLAIM :
 660                                         __GFP_KSWAPD_RECLAIM);
 661
 662         /* Only do synchronous compaction if madvised */
 663         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 664                 return GFP_TRANSHUGE_LIGHT |
 665                        (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
 666
 667         return GFP_TRANSHUGE_LIGHT;
 668 }
 669
 670 /* Caller must hold page table lock. */
 671 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 672                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 673                 struct page *zero_page)
 674 {
 675         pmd_t entry;
 676         if (!pmd_none(*pmd))
 677                 return false;
 678         entry = mk_pmd(zero_page, vma->vm_page_prot);
 679         entry = pmd_mkhuge(entry);
 680         if (pgtable)
 681                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 682         set_pmd_at(mm, haddr, pmd, entry);
 683         mm_inc_nr_ptes(mm);
 684         return true;
 685 }
 686
 687 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
 688 {
 689         struct vm_area_struct *vma = vmf->vma;
 690         gfp_t gfp;
 691         struct page *page;
 692         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 693
 694         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 695                 return VM_FAULT_FALLBACK;
 696         if (unlikely(anon_vma_prepare(vma)))
 697                 return VM_FAULT_OOM;
 698         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 699                 return VM_FAULT_OOM;
 700         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
 701                         !mm_forbids_zeropage(vma->vm_mm) &&
 702                         transparent_hugepage_use_zero_page()) {
 703                 pgtable_t pgtable;
 704                 struct page *zero_page;
 705                 bool set;
 706                 vm_fault_t ret;
 707                 pgtable = pte_alloc_one(vma->vm_mm);
 708                 if (unlikely(!pgtable))
 709                         return VM_FAULT_OOM;
 710                 zero_page = mm_get_huge_zero_page(vma->vm_mm);
 711                 if (unlikely(!zero_page)) {
 712                         pte_free(vma->vm_mm, pgtable);
 713                         count_vm_event(THP_FAULT_FALLBACK);
 714                         return VM_FAULT_FALLBACK;
 715                 }
 716                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 717                 ret = 0;
 718                 set = false;
 719                 if (pmd_none(*vmf->pmd)) {
 720                         ret = check_stable_address_space(vma->vm_mm);
 721                         if (ret) {
 722                                 spin_unlock(vmf->ptl);
 723                         } else if (userfaultfd_missing(vma)) {
 724                                 spin_unlock(vmf->ptl);
 725                                 ret = handle_userfault(vmf, VM_UFFD_MISSING);
 726                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 727                         } else {
 728                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
 729                                                    haddr, vmf->pmd, zero_page);
 730                                 spin_unlock(vmf->ptl);
 731                                 set = true;
 732                         }
 733                 } else
 734                         spin_unlock(vmf->ptl);
 735                 if (!set)
 736                         pte_free(vma->vm_mm, pgtable);
 737                 return ret;
 738         }
 739         gfp = alloc_hugepage_direct_gfpmask(vma);
 740         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 741         if (unlikely(!page)) {
 742                 count_vm_event(THP_FAULT_FALLBACK);
 743                 return VM_FAULT_FALLBACK;
 744         }
 745         prep_transhuge_page(page);
 746         return __do_huge_pmd_anonymous_page(vmf, page, gfp);
 747 }
 748
 749 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 750                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
 751                 pgtable_t pgtable)
 752 {
 753         struct mm_struct *mm = vma->vm_mm;
 754         pmd_t entry;
 755         spinlock_t *ptl;
 756
 757         ptl = pmd_lock(mm, pmd);
 758         if (!pmd_none(*pmd)) {
 759                 if (write) {
 760                         if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
 761                                 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
 762                                 goto out_unlock;
 763                         }
 764                         entry = pmd_mkyoung(*pmd);
 765                         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 766                         if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
 767                                 update_mmu_cache_pmd(vma, addr, pmd);
 768                 }
 769
 770                 goto out_unlock;
 771         }
 772
 773         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 774         if (pfn_t_devmap(pfn))
 775                 entry = pmd_mkdevmap(entry);
 776         if (write) {
 777                 entry = pmd_mkyoung(pmd_mkdirty(entry));
 778                 entry = maybe_pmd_mkwrite(entry, vma);
 779         }
 780
 781         if (pgtable) {
 782                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 783                 mm_inc_nr_ptes(mm);
 784                 pgtable = NULL;
 785         }
 786
 787         set_pmd_at(mm, addr, pmd, entry);
 788         update_mmu_cache_pmd(vma, addr, pmd);
 789
 790 out_unlock:
 791         spin_unlock(ptl);
 792         if (pgtable)
 793                 pte_free(mm, pgtable);
 794 }
 795
 796 vm_fault_t vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 797                         pmd_t *pmd, pfn_t pfn, bool write)
 798 {
 799         pgprot_t pgprot = vma->vm_page_prot;
 800         pgtable_t pgtable = NULL;
 801         /*
 802          * If we had pmd_special, we could avoid all these restrictions,
 803          * but we need to be consistent with PTEs and architectures that
 804          * can't support a 'special' bit.
 805          */
 806         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 807                         !pfn_t_devmap(pfn));
 808         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 809                                                 (VM_PFNMAP|VM_MIXEDMAP));
 810         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 811
 812         if (addr < vma->vm_start || addr >= vma->vm_end)
 813                 return VM_FAULT_SIGBUS;
 814
 815         if (arch_needs_pgtable_deposit()) {
 816                 pgtable = pte_alloc_one(vma->vm_mm);
 817                 if (!pgtable)
 818                         return VM_FAULT_OOM;
 819         }
 820
 821         track_pfn_insert(vma, &pgprot, pfn);
 822
 823         insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write, pgtable);
 824         return VM_FAULT_NOPAGE;
 825 }
 826 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
 827
 828 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 829 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
 830 {
 831         if (likely(vma->vm_flags & VM_WRITE))
 832                 pud = pud_mkwrite(pud);
 833         return pud;
 834 }
 835
 836 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 837                 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
 838 {
 839         struct mm_struct *mm = vma->vm_mm;
 840         pud_t entry;
 841         spinlock_t *ptl;
 842
 843         ptl = pud_lock(mm, pud);
 844         if (!pud_none(*pud)) {
 845                 if (write) {
 846                         if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
 847                                 WARN_ON_ONCE(!is_huge_zero_pud(*pud));
 848                                 goto out_unlock;
 849                         }
 850                         entry = pud_mkyoung(*pud);
 851                         entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
 852                         if (pudp_set_access_flags(vma, addr, pud, entry, 1))
 853                                 update_mmu_cache_pud(vma, addr, pud);
 854                 }
 855                 goto out_unlock;
 856         }
 857
 858         entry = pud_mkhuge(pfn_t_pud(pfn, prot));
 859         if (pfn_t_devmap(pfn))
 860                 entry = pud_mkdevmap(entry);
 861         if (write) {
 862                 entry = pud_mkyoung(pud_mkdirty(entry));
 863                 entry = maybe_pud_mkwrite(entry, vma);
 864         }
 865         set_pud_at(mm, addr, pud, entry);
 866         update_mmu_cache_pud(vma, addr, pud);
 867
 868 out_unlock:
 869         spin_unlock(ptl);
 870 }
 871
 872 vm_fault_t vmf_insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 873                         pud_t *pud, pfn_t pfn, bool write)
 874 {
 875         pgprot_t pgprot = vma->vm_page_prot;
 876         /*
 877          * If we had pud_special, we could avoid all these restrictions,
 878          * but we need to be consistent with PTEs and architectures that
 879          * can't support a 'special' bit.
 880          */
 881         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 882                         !pfn_t_devmap(pfn));
 883         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 884                                                 (VM_PFNMAP|VM_MIXEDMAP));
 885         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 886
 887         if (addr < vma->vm_start || addr >= vma->vm_end)
 888                 return VM_FAULT_SIGBUS;
 889
 890         track_pfn_insert(vma, &pgprot, pfn);
 891
 892         insert_pfn_pud(vma, addr, pud, pfn, pgprot, write);
 893         return VM_FAULT_NOPAGE;
 894 }
 895 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
 896 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
 897
 898 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
 899                 pmd_t *pmd, int flags)
 900 {
 901         pmd_t _pmd;
 902
 903         _pmd = pmd_mkyoung(*pmd);
 904         if (flags & FOLL_WRITE)
 905                 _pmd = pmd_mkdirty(_pmd);
 906         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
 907                                 pmd, _pmd, flags & FOLL_WRITE))
 908                 update_mmu_cache_pmd(vma, addr, pmd);
 909 }
 910
 911 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 912                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
 913 {
 914         unsigned long pfn = pmd_pfn(*pmd);
 915         struct mm_struct *mm = vma->vm_mm;
 916         struct page *page;
 917
 918         assert_spin_locked(pmd_lockptr(mm, pmd));
 919
 920         /*
 921          * When we COW a devmap PMD entry, we split it into PTEs, so we should
 922          * not be in this function with `flags & FOLL_COW` set.
 923          */
 924         WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
 925
 926         if (flags & FOLL_WRITE && !pmd_write(*pmd))
 927                 return NULL;
 928
 929         if (pmd_present(*pmd) && pmd_devmap(*pmd))
 930                 /* pass */;
 931         else
 932                 return NULL;
 933
 934         if (flags & FOLL_TOUCH)
 935                 touch_pmd(vma, addr, pmd, flags);
 936
 937         /*
 938          * device mapped pages can only be returned if the
 939          * caller will manage the page reference count.
 940          */
 941         if (!(flags & FOLL_GET))
 942                 return ERR_PTR(-EEXIST);
 943
 944         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
 945         *pgmap = get_dev_pagemap(pfn, *pgmap);
 946         if (!*pgmap)
 947                 return ERR_PTR(-EFAULT);
 948         page = pfn_to_page(pfn);
 949         get_page(page);
 950
 951         return page;
 952 }
 953
 954 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 955                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 956                   struct vm_area_struct *vma)
 957 {
 958         spinlock_t *dst_ptl, *src_ptl;
 959         struct page *src_page;
 960         pmd_t pmd;
 961         pgtable_t pgtable = NULL;
 962         int ret = -ENOMEM;
 963
 964         /* Skip if can be re-fill on fault */
 965         if (!vma_is_anonymous(vma))
 966                 return 0;
 967
 968         pgtable = pte_alloc_one(dst_mm);
 969         if (unlikely(!pgtable))
 970                 goto out;
 971
 972         dst_ptl = pmd_lock(dst_mm, dst_pmd);
 973         src_ptl = pmd_lockptr(src_mm, src_pmd);
 974         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 975
 976         ret = -EAGAIN;
 977         pmd = *src_pmd;
 978
 979 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
 980         if (unlikely(is_swap_pmd(pmd))) {
 981                 swp_entry_t entry = pmd_to_swp_entry(pmd);
 982
 983                 VM_BUG_ON(!is_pmd_migration_entry(pmd));
 984                 if (is_write_migration_entry(entry)) {
 985                         make_migration_entry_read(&entry);
 986                         pmd = swp_entry_to_pmd(entry);
 987                         if (pmd_swp_soft_dirty(*src_pmd))
 988                                 pmd = pmd_swp_mksoft_dirty(pmd);
 989                         set_pmd_at(src_mm, addr, src_pmd, pmd);
 990                 }
 991                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 992                 mm_inc_nr_ptes(dst_mm);
 993                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
 994                 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 995                 ret = 0;
 996                 goto out_unlock;
 997         }
 998 #endif
 999
1000         if (unlikely(!pmd_trans_huge(pmd))) {
1001                 pte_free(dst_mm, pgtable);
1002                 goto out_unlock;
1003         }
1004         /*
1005          * When page table lock is held, the huge zero pmd should not be
1006          * under splitting since we don't split the page itself, only pmd to
1007          * a page table.
1008          */
1009         if (is_huge_zero_pmd(pmd)) {
1010                 struct page *zero_page;
1011                 /*
1012                  * get_huge_zero_page() will never allocate a new page here,
1013                  * since we already have a zero page to copy. It just takes a
1014                  * reference.
1015                  */
1016                 zero_page = mm_get_huge_zero_page(dst_mm);
1017                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1018                                 zero_page);
1019                 ret = 0;
1020                 goto out_unlock;
1021         }
1022
1023         src_page = pmd_page(pmd);
1024         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1025         get_page(src_page);
1026         page_dup_rmap(src_page, true);
1027         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1028         mm_inc_nr_ptes(dst_mm);
1029         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1030
1031         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1032         pmd = pmd_mkold(pmd_wrprotect(pmd));
1033         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1034
1035         ret = 0;
1036 out_unlock:
1037         spin_unlock(src_ptl);
1038         spin_unlock(dst_ptl);
1039 out:
1040         return ret;
1041 }
1042
1043 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1044 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1045                 pud_t *pud, int flags)
1046 {
1047         pud_t _pud;
1048
1049         _pud = pud_mkyoung(*pud);
1050         if (flags & FOLL_WRITE)
1051                 _pud = pud_mkdirty(_pud);
1052         if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1053                                 pud, _pud, flags & FOLL_WRITE))
1054                 update_mmu_cache_pud(vma, addr, pud);
1055 }
1056
1057 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1058                 pud_t *pud, int flags, struct dev_pagemap **pgmap)
1059 {
1060         unsigned long pfn = pud_pfn(*pud);
1061         struct mm_struct *mm = vma->vm_mm;
1062         struct page *page;
1063
1064         assert_spin_locked(pud_lockptr(mm, pud));
1065
1066         if (flags & FOLL_WRITE && !pud_write(*pud))
1067                 return NULL;
1068
1069         if (pud_present(*pud) && pud_devmap(*pud))
1070                 /* pass */;
1071         else
1072                 return NULL;
1073
1074         if (flags & FOLL_TOUCH)
1075                 touch_pud(vma, addr, pud, flags);
1076
1077         /*
1078          * device mapped pages can only be returned if the
1079          * caller will manage the page reference count.
1080          */
1081         if (!(flags & FOLL_GET))
1082                 return ERR_PTR(-EEXIST);
1083
1084         pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1085         *pgmap = get_dev_pagemap(pfn, *pgmap);
1086         if (!*pgmap)
1087                 return ERR_PTR(-EFAULT);
1088         page = pfn_to_page(pfn);
1089         get_page(page);
1090
1091         return page;
1092 }
1093
1094 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1095                   pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1096                   struct vm_area_struct *vma)
1097 {
1098         spinlock_t *dst_ptl, *src_ptl;
1099         pud_t pud;
1100         int ret;
1101
1102         dst_ptl = pud_lock(dst_mm, dst_pud);
1103         src_ptl = pud_lockptr(src_mm, src_pud);
1104         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1105
1106         ret = -EAGAIN;
1107         pud = *src_pud;
1108         if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1109                 goto out_unlock;
1110
1111         /*
1112          * When page table lock is held, the huge zero pud should not be
1113          * under splitting since we don't split the page itself, only pud to
1114          * a page table.
1115          */
1116         if (is_huge_zero_pud(pud)) {
1117                 /* No huge zero pud yet */
1118         }
1119
1120         pudp_set_wrprotect(src_mm, addr, src_pud);
1121         pud = pud_mkold(pud_wrprotect(pud));
1122         set_pud_at(dst_mm, addr, dst_pud, pud);
1123
1124         ret = 0;
1125 out_unlock:
1126         spin_unlock(src_ptl);
1127         spin_unlock(dst_ptl);
1128         return ret;
1129 }
1130
1131 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1132 {
1133         pud_t entry;
1134         unsigned long haddr;
1135         bool write = vmf->flags & FAULT_FLAG_WRITE;
1136
1137         vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1138         if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1139                 goto unlock;
1140
1141         entry = pud_mkyoung(orig_pud);
1142         if (write)
1143                 entry = pud_mkdirty(entry);
1144         haddr = vmf->address & HPAGE_PUD_MASK;
1145         if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1146                 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1147
1148 unlock:
1149         spin_unlock(vmf->ptl);
1150 }
1151 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1152
1153 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1154 {
1155         pmd_t entry;
1156         unsigned long haddr;
1157         bool write = vmf->flags & FAULT_FLAG_WRITE;
1158
1159         vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1160         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1161                 goto unlock;
1162
1163         entry = pmd_mkyoung(orig_pmd);
1164         if (write)
1165                 entry = pmd_mkdirty(entry);
1166         haddr = vmf->address & HPAGE_PMD_MASK;
1167         if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1168                 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1169
1170 unlock:
1171         spin_unlock(vmf->ptl);
1172 }
1173
1174 static vm_fault_t do_huge_pmd_wp_page_fallback(struct vm_fault *vmf,
1175                         pmd_t orig_pmd, struct page *page)
1176 {
1177         struct vm_area_struct *vma = vmf->vma;
1178         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1179         struct mem_cgroup *memcg;
1180         pgtable_t pgtable;
1181         pmd_t _pmd;
1182         int i;
1183         vm_fault_t ret = 0;
1184         struct page **pages;
1185         struct mmu_notifier_range range;
1186
1187         pages = kmalloc_array(HPAGE_PMD_NR, sizeof(struct page *),
1188                               GFP_KERNEL);
1189         if (unlikely(!pages)) {
1190                 ret |= VM_FAULT_OOM;
1191                 goto out;
1192         }
1193
1194         for (i = 0; i < HPAGE_PMD_NR; i++) {
1195                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
1196                                                vmf->address, page_to_nid(page));
1197                 if (unlikely(!pages[i] ||
1198                              mem_cgroup_try_charge_delay(pages[i], vma->vm_mm,
1199                                      GFP_KERNEL, &memcg, false))) {
1200                         if (pages[i])
1201                                 put_page(pages[i]);
1202                         while (--i >= 0) {
1203                                 memcg = (void *)page_private(pages[i]);
1204                                 set_page_private(pages[i], 0);
1205                                 mem_cgroup_cancel_charge(pages[i], memcg,
1206                                                 false);
1207                                 put_page(pages[i]);
1208                         }
1209                         kfree(pages);
1210                         ret |= VM_FAULT_OOM;
1211                         goto out;
1212                 }
1213                 set_page_private(pages[i], (unsigned long)memcg);
1214         }
1215
1216         for (i = 0; i < HPAGE_PMD_NR; i++) {
1217                 copy_user_highpage(pages[i], page + i,
1218                                    haddr + PAGE_SIZE * i, vma);
1219                 __SetPageUptodate(pages[i]);
1220                 cond_resched();
1221         }
1222
1223         mmu_notifier_range_init(&range, vma->vm_mm, haddr,
1224                                 haddr + HPAGE_PMD_SIZE);
1225         mmu_notifier_invalidate_range_start(&range);
1226
1227         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1228         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1229                 goto out_free_pages;
1230         VM_BUG_ON_PAGE(!PageHead(page), page);
1231
1232         /*
1233          * Leave pmd empty until pte is filled note we must notify here as
1234          * concurrent CPU thread might write to new page before the call to
1235          * mmu_notifier_invalidate_range_end() happens which can lead to a
1236          * device seeing memory write in different order than CPU.
1237          *
1238          * See Documentation/vm/mmu_notifier.rst
1239          */
1240         pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1241
1242         pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
1243         pmd_populate(vma->vm_mm, &_pmd, pgtable);
1244
1245         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1246                 pte_t entry;
1247                 entry = mk_pte(pages[i], vma->vm_page_prot);
1248                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1249                 memcg = (void *)page_private(pages[i]);
1250                 set_page_private(pages[i], 0);
1251                 page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
1252                 mem_cgroup_commit_charge(pages[i], memcg, false, false);
1253                 lru_cache_add_active_or_unevictable(pages[i], vma);
1254                 vmf->pte = pte_offset_map(&_pmd, haddr);
1255                 VM_BUG_ON(!pte_none(*vmf->pte));
1256                 set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
1257                 pte_unmap(vmf->pte);
1258         }
1259         kfree(pages);
1260
1261         smp_wmb(); /* make pte visible before pmd */
1262         pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
1263         page_remove_rmap(page, true);
1264         spin_unlock(vmf->ptl);
1265
1266         /*
1267          * No need to double call mmu_notifier->invalidate_range() callback as
1268          * the above pmdp_huge_clear_flush_notify() did already call it.
1269          */
1270         mmu_notifier_invalidate_range_only_end(&range);
1271
1272         ret |= VM_FAULT_WRITE;
1273         put_page(page);
1274
1275 out:
1276         return ret;
1277
1278 out_free_pages:
1279         spin_unlock(vmf->ptl);
1280         mmu_notifier_invalidate_range_end(&range);
1281         for (i = 0; i < HPAGE_PMD_NR; i++) {
1282                 memcg = (void *)page_private(pages[i]);
1283                 set_page_private(pages[i], 0);
1284                 mem_cgroup_cancel_charge(pages[i], memcg, false);
1285                 put_page(pages[i]);
1286         }
1287         kfree(pages);
1288         goto out;
1289 }
1290
1291 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1292 {
1293         struct vm_area_struct *vma = vmf->vma;
1294         struct page *page = NULL, *new_page;
1295         struct mem_cgroup *memcg;
1296         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1297         struct mmu_notifier_range range;
1298         gfp_t huge_gfp;                 /* for allocation and charge */
1299         vm_fault_t ret = 0;
1300
1301         vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1302         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1303         if (is_huge_zero_pmd(orig_pmd))
1304                 goto alloc;
1305         spin_lock(vmf->ptl);
1306         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1307                 goto out_unlock;
1308
1309         page = pmd_page(orig_pmd);
1310         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1311         /*
1312          * We can only reuse the page if nobody else maps the huge page or it's
1313          * part.
1314          */
1315         if (!trylock_page(page)) {
1316                 get_page(page);
1317                 spin_unlock(vmf->ptl);
1318                 lock_page(page);
1319                 spin_lock(vmf->ptl);
1320                 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1321                         unlock_page(page);
1322                         put_page(page);
1323                         goto out_unlock;
1324                 }
1325                 put_page(page);
1326         }
1327         if (reuse_swap_page(page, NULL)) {
1328                 pmd_t entry;
1329                 entry = pmd_mkyoung(orig_pmd);
1330                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1331                 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry,  1))
1332                         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1333                 ret |= VM_FAULT_WRITE;
1334                 unlock_page(page);
1335                 goto out_unlock;
1336         }
1337         unlock_page(page);
1338         get_page(page);
1339         spin_unlock(vmf->ptl);
1340 alloc:
1341         if (__transparent_hugepage_enabled(vma) &&
1342             !transparent_hugepage_debug_cow()) {
1343                 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1344                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1345         } else
1346                 new_page = NULL;
1347
1348         if (likely(new_page)) {
1349                 prep_transhuge_page(new_page);
1350         } else {
1351                 if (!page) {
1352                         split_huge_pmd(vma, vmf->pmd, vmf->address);
1353                         ret |= VM_FAULT_FALLBACK;
1354                 } else {
1355                         ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
1356                         if (ret & VM_FAULT_OOM) {
1357                                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1358                                 ret |= VM_FAULT_FALLBACK;
1359                         }
1360                         put_page(page);
1361                 }
1362                 count_vm_event(THP_FAULT_FALLBACK);
1363                 goto out;
1364         }
1365
1366         if (unlikely(mem_cgroup_try_charge_delay(new_page, vma->vm_mm,
1367                                         huge_gfp, &memcg, true))) {
1368                 put_page(new_page);
1369                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1370                 if (page)
1371                         put_page(page);
1372                 ret |= VM_FAULT_FALLBACK;
1373                 count_vm_event(THP_FAULT_FALLBACK);
1374                 goto out;
1375         }
1376
1377         count_vm_event(THP_FAULT_ALLOC);
1378         count_memcg_events(memcg, THP_FAULT_ALLOC, 1);
1379
1380         if (!page)
1381                 clear_huge_page(new_page, vmf->address, HPAGE_PMD_NR);
1382         else
1383                 copy_user_huge_page(new_page, page, vmf->address,
1384                                     vma, HPAGE_PMD_NR);
1385         __SetPageUptodate(new_page);
1386
1387         mmu_notifier_range_init(&range, vma->vm_mm, haddr,
1388                                 haddr + HPAGE_PMD_SIZE);
1389         mmu_notifier_invalidate_range_start(&range);
1390
1391         spin_lock(vmf->ptl);
1392         if (page)
1393                 put_page(page);
1394         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1395                 spin_unlock(vmf->ptl);
1396                 mem_cgroup_cancel_charge(new_page, memcg, true);
1397                 put_page(new_page);
1398                 goto out_mn;
1399         } else {
1400                 pmd_t entry;
1401                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1402                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1403                 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1404                 page_add_new_anon_rmap(new_page, vma, haddr, true);
1405                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1406                 lru_cache_add_active_or_unevictable(new_page, vma);
1407                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1408                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1409                 if (!page) {
1410                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1411                 } else {
1412                         VM_BUG_ON_PAGE(!PageHead(page), page);
1413                         page_remove_rmap(page, true);
1414                         put_page(page);
1415                 }
1416                 ret |= VM_FAULT_WRITE;
1417         }
1418         spin_unlock(vmf->ptl);
1419 out_mn:
1420         /*
1421          * No need to double call mmu_notifier->invalidate_range() callback as
1422          * the above pmdp_huge_clear_flush_notify() did already call it.
1423          */
1424         mmu_notifier_invalidate_range_only_end(&range);
1425 out:
1426         return ret;
1427 out_unlock:
1428         spin_unlock(vmf->ptl);
1429         return ret;
1430 }
1431
1432 /*
1433  * FOLL_FORCE can write to even unwritable pmd's, but only
1434  * after we've gone through a COW cycle and they are dirty.
1435  */
1436 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1437 {
1438         return pmd_write(pmd) ||
1439                ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1440 }
1441
1442 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1443                                    unsigned long addr,
1444                                    pmd_t *pmd,
1445                                    unsigned int flags)
1446 {
1447         struct mm_struct *mm = vma->vm_mm;
1448         struct page *page = NULL;
1449
1450         assert_spin_locked(pmd_lockptr(mm, pmd));
1451
1452         if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1453                 goto out;
1454
1455         /* Avoid dumping huge zero page */
1456         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1457                 return ERR_PTR(-EFAULT);
1458
1459         /* Full NUMA hinting faults to serialise migration in fault paths */
1460         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1461                 goto out;
1462
1463         page = pmd_page(*pmd);
1464         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1465         if (flags & FOLL_TOUCH)
1466                 touch_pmd(vma, addr, pmd, flags);
1467         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1468                 /*
1469                  * We don't mlock() pte-mapped THPs. This way we can avoid
1470                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1471                  *
1472                  * For anon THP:
1473                  *
1474                  * In most cases the pmd is the only mapping of the page as we
1475                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1476                  * writable private mappings in populate_vma_page_range().
1477                  *
1478                  * The only scenario when we have the page shared here is if we
1479                  * mlocking read-only mapping shared over fork(). We skip
1480                  * mlocking such pages.
1481                  *
1482                  * For file THP:
1483                  *
1484                  * We can expect PageDoubleMap() to be stable under page lock:
1485                  * for file pages we set it in page_add_file_rmap(), which
1486                  * requires page to be locked.
1487                  */
1488
1489                 if (PageAnon(page) && compound_mapcount(page) != 1)
1490                         goto skip_mlock;
1491                 if (PageDoubleMap(page) || !page->mapping)
1492                         goto skip_mlock;
1493                 if (!trylock_page(page))
1494                         goto skip_mlock;
1495                 lru_add_drain();
1496                 if (page->mapping && !PageDoubleMap(page))
1497                         mlock_vma_page(page);
1498                 unlock_page(page);
1499         }
1500 skip_mlock:
1501         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1502         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1503         if (flags & FOLL_GET)
1504                 get_page(page);
1505
1506 out:
1507         return page;
1508 }
1509
1510 /* NUMA hinting page fault entry point for trans huge pmds */
1511 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1512 {
1513         struct vm_area_struct *vma = vmf->vma;
1514         struct anon_vma *anon_vma = NULL;
1515         struct page *page;
1516         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1517         int page_nid = NUMA_NO_NODE, this_nid = numa_node_id();
1518         int target_nid, last_cpupid = -1;
1519         bool page_locked;
1520         bool migrated = false;
1521         bool was_writable;
1522         int flags = 0;
1523
1524         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1525         if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1526                 goto out_unlock;
1527
1528         /*
1529          * If there are potential migrations, wait for completion and retry
1530          * without disrupting NUMA hinting information. Do not relock and
1531          * check_same as the page may no longer be mapped.
1532          */
1533         if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1534                 page = pmd_page(*vmf->pmd);
1535                 if (!get_page_unless_zero(page))
1536                         goto out_unlock;
1537                 spin_unlock(vmf->ptl);
1538                 put_and_wait_on_page_locked(page);
1539                 goto out;
1540         }
1541
1542         page = pmd_page(pmd);
1543         BUG_ON(is_huge_zero_page(page));
1544         page_nid = page_to_nid(page);
1545         last_cpupid = page_cpupid_last(page);
1546         count_vm_numa_event(NUMA_HINT_FAULTS);
1547         if (page_nid == this_nid) {
1548                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1549                 flags |= TNF_FAULT_LOCAL;
1550         }
1551
1552         /* See similar comment in do_numa_page for explanation */
1553         if (!pmd_savedwrite(pmd))
1554                 flags |= TNF_NO_GROUP;
1555
1556         /*
1557          * Acquire the page lock to serialise THP migrations but avoid dropping
1558          * page_table_lock if at all possible
1559          */
1560         page_locked = trylock_page(page);
1561         target_nid = mpol_misplaced(page, vma, haddr);
1562         if (target_nid == NUMA_NO_NODE) {
1563                 /* If the page was locked, there are no parallel migrations */
1564                 if (page_locked)
1565                         goto clear_pmdnuma;
1566         }
1567
1568         /* Migration could have started since the pmd_trans_migrating check */
1569         if (!page_locked) {
1570                 page_nid = NUMA_NO_NODE;
1571                 if (!get_page_unless_zero(page))
1572                         goto out_unlock;
1573                 spin_unlock(vmf->ptl);
1574                 put_and_wait_on_page_locked(page);
1575                 goto out;
1576         }
1577
1578         /*
1579          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1580          * to serialises splits
1581          */
1582         get_page(page);
1583         spin_unlock(vmf->ptl);
1584         anon_vma = page_lock_anon_vma_read(page);
1585
1586         /* Confirm the PMD did not change while page_table_lock was released */
1587         spin_lock(vmf->ptl);
1588         if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1589                 unlock_page(page);
1590                 put_page(page);
1591                 page_nid = NUMA_NO_NODE;
1592                 goto out_unlock;
1593         }
1594
1595         /* Bail if we fail to protect against THP splits for any reason */
1596         if (unlikely(!anon_vma)) {
1597                 put_page(page);
1598                 page_nid = NUMA_NO_NODE;
1599                 goto clear_pmdnuma;
1600         }
1601
1602         /*
1603          * Since we took the NUMA fault, we must have observed the !accessible
1604          * bit. Make sure all other CPUs agree with that, to avoid them
1605          * modifying the page we're about to migrate.
1606          *
1607          * Must be done under PTL such that we'll observe the relevant
1608          * inc_tlb_flush_pending().
1609          *
1610          * We are not sure a pending tlb flush here is for a huge page
1611          * mapping or not. Hence use the tlb range variant
1612          */
1613         if (mm_tlb_flush_pending(vma->vm_mm)) {
1614                 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1615                 /*
1616                  * change_huge_pmd() released the pmd lock before
1617                  * invalidating the secondary MMUs sharing the primary
1618                  * MMU pagetables (with ->invalidate_range()). The
1619                  * mmu_notifier_invalidate_range_end() (which
1620                  * internally calls ->invalidate_range()) in
1621                  * change_pmd_range() will run after us, so we can't
1622                  * rely on it here and we need an explicit invalidate.
1623                  */
1624                 mmu_notifier_invalidate_range(vma->vm_mm, haddr,
1625                                               haddr + HPAGE_PMD_SIZE);
1626         }
1627
1628         /*
1629          * Migrate the THP to the requested node, returns with page unlocked
1630          * and access rights restored.
1631          */
1632         spin_unlock(vmf->ptl);
1633
1634         migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1635                                 vmf->pmd, pmd, vmf->address, page, target_nid);
1636         if (migrated) {
1637                 flags |= TNF_MIGRATED;
1638                 page_nid = target_nid;
1639         } else
1640                 flags |= TNF_MIGRATE_FAIL;
1641
1642         goto out;
1643 clear_pmdnuma:
1644         BUG_ON(!PageLocked(page));
1645         was_writable = pmd_savedwrite(pmd);
1646         pmd = pmd_modify(pmd, vma->vm_page_prot);
1647         pmd = pmd_mkyoung(pmd);
1648         if (was_writable)
1649                 pmd = pmd_mkwrite(pmd);
1650         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1651         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1652         unlock_page(page);
1653 out_unlock:
1654         spin_unlock(vmf->ptl);
1655
1656 out:
1657         if (anon_vma)
1658                 page_unlock_anon_vma_read(anon_vma);
1659
1660         if (page_nid != NUMA_NO_NODE)
1661                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1662                                 flags);
1663
1664         return 0;
1665 }
1666
1667 /*
1668  * Return true if we do MADV_FREE successfully on entire pmd page.
1669  * Otherwise, return false.
1670  */
1671 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1672                 pmd_t *pmd, unsigned long addr, unsigned long next)
1673 {
1674         spinlock_t *ptl;
1675         pmd_t orig_pmd;
1676         struct page *page;
1677         struct mm_struct *mm = tlb->mm;
1678         bool ret = false;
1679
1680         tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1681
1682         ptl = pmd_trans_huge_lock(pmd, vma);
1683         if (!ptl)
1684                 goto out_unlocked;
1685
1686         orig_pmd = *pmd;
1687         if (is_huge_zero_pmd(orig_pmd))
1688                 goto out;
1689
1690         if (unlikely(!pmd_present(orig_pmd))) {
1691                 VM_BUG_ON(thp_migration_supported() &&
1692                                   !is_pmd_migration_entry(orig_pmd));
1693                 goto out;
1694         }
1695
1696         page = pmd_page(orig_pmd);
1697         /*
1698          * If other processes are mapping this page, we couldn't discard
1699          * the page unless they all do MADV_FREE so let's skip the page.
1700          */
1701         if (page_mapcount(page) != 1)
1702                 goto out;
1703
1704         if (!trylock_page(page))
1705                 goto out;
1706
1707         /*
1708          * If user want to discard part-pages of THP, split it so MADV_FREE
1709          * will deactivate only them.
1710          */
1711         if (next - addr != HPAGE_PMD_SIZE) {
1712                 get_page(page);
1713                 spin_unlock(ptl);
1714                 split_huge_page(page);
1715                 unlock_page(page);
1716                 put_page(page);
1717                 goto out_unlocked;
1718         }
1719
1720         if (PageDirty(page))
1721                 ClearPageDirty(page);
1722         unlock_page(page);
1723
1724         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1725                 pmdp_invalidate(vma, addr, pmd);
1726                 orig_pmd = pmd_mkold(orig_pmd);
1727                 orig_pmd = pmd_mkclean(orig_pmd);
1728
1729                 set_pmd_at(mm, addr, pmd, orig_pmd);
1730                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1731         }
1732
1733         mark_page_lazyfree(page);
1734         ret = true;
1735 out:
1736         spin_unlock(ptl);
1737 out_unlocked:
1738         return ret;
1739 }
1740
1741 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1742 {
1743         pgtable_t pgtable;
1744
1745         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1746         pte_free(mm, pgtable);
1747         mm_dec_nr_ptes(mm);
1748 }
1749
1750 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1751                  pmd_t *pmd, unsigned long addr)
1752 {
1753         pmd_t orig_pmd;
1754         spinlock_t *ptl;
1755
1756         tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1757
1758         ptl = __pmd_trans_huge_lock(pmd, vma);
1759         if (!ptl)
1760                 return 0;
1761         /*
1762          * For architectures like ppc64 we look at deposited pgtable
1763          * when calling pmdp_huge_get_and_clear. So do the
1764          * pgtable_trans_huge_withdraw after finishing pmdp related
1765          * operations.
1766          */
1767         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1768                         tlb->fullmm);
1769         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1770         if (vma_is_dax(vma)) {
1771                 if (arch_needs_pgtable_deposit())
1772                         zap_deposited_table(tlb->mm, pmd);
1773                 spin_unlock(ptl);
1774                 if (is_huge_zero_pmd(orig_pmd))
1775                         tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1776         } else if (is_huge_zero_pmd(orig_pmd)) {
1777                 zap_deposited_table(tlb->mm, pmd);
1778                 spin_unlock(ptl);
1779                 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1780         } else {
1781                 struct page *page = NULL;
1782                 int flush_needed = 1;
1783
1784                 if (pmd_present(orig_pmd)) {
1785                         page = pmd_page(orig_pmd);
1786                         page_remove_rmap(page, true);
1787                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1788                         VM_BUG_ON_PAGE(!PageHead(page), page);
1789                 } else if (thp_migration_supported()) {
1790                         swp_entry_t entry;
1791
1792                         VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1793                         entry = pmd_to_swp_entry(orig_pmd);
1794                         page = pfn_to_page(swp_offset(entry));
1795                         flush_needed = 0;
1796                 } else
1797                         WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1798
1799                 if (PageAnon(page)) {
1800                         zap_deposited_table(tlb->mm, pmd);
1801                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1802                 } else {
1803                         if (arch_needs_pgtable_deposit())
1804                                 zap_deposited_table(tlb->mm, pmd);
1805                         add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1806                 }
1807
1808                 spin_unlock(ptl);
1809                 if (flush_needed)
1810                         tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1811         }
1812         return 1;
1813 }
1814
1815 #ifndef pmd_move_must_withdraw
1816 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1817                                          spinlock_t *old_pmd_ptl,
1818                                          struct vm_area_struct *vma)
1819 {
1820         /*
1821          * With split pmd lock we also need to move preallocated
1822          * PTE page table if new_pmd is on different PMD page table.
1823          *
1824          * We also don't deposit and withdraw tables for file pages.
1825          */
1826         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1827 }
1828 #endif
1829
1830 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1831 {
1832 #ifdef CONFIG_MEM_SOFT_DIRTY
1833         if (unlikely(is_pmd_migration_entry(pmd)))
1834                 pmd = pmd_swp_mksoft_dirty(pmd);
1835         else if (pmd_present(pmd))
1836                 pmd = pmd_mksoft_dirty(pmd);
1837 #endif
1838         return pmd;
1839 }
1840
1841 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1842                   unsigned long new_addr, unsigned long old_end,
1843                   pmd_t *old_pmd, pmd_t *new_pmd)
1844 {
1845         spinlock_t *old_ptl, *new_ptl;
1846         pmd_t pmd;
1847         struct mm_struct *mm = vma->vm_mm;
1848         bool force_flush = false;
1849
1850         if ((old_addr & ~HPAGE_PMD_MASK) ||
1851             (new_addr & ~HPAGE_PMD_MASK) ||
1852             old_end - old_addr < HPAGE_PMD_SIZE)
1853                 return false;
1854
1855         /*
1856          * The destination pmd shouldn't be established, free_pgtables()
1857          * should have release it.
1858          */
1859         if (WARN_ON(!pmd_none(*new_pmd))) {
1860                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1861                 return false;
1862         }
1863
1864         /*
1865          * We don't have to worry about the ordering of src and dst
1866          * ptlocks because exclusive mmap_sem prevents deadlock.
1867          */
1868         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1869         if (old_ptl) {
1870                 new_ptl = pmd_lockptr(mm, new_pmd);
1871                 if (new_ptl != old_ptl)
1872                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1873                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1874                 if (pmd_present(pmd))
1875                         force_flush = true;
1876                 VM_BUG_ON(!pmd_none(*new_pmd));
1877
1878                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1879                         pgtable_t pgtable;
1880                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1881                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1882                 }
1883                 pmd = move_soft_dirty_pmd(pmd);
1884                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1885                 if (force_flush)
1886                         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1887                 if (new_ptl != old_ptl)
1888                         spin_unlock(new_ptl);
1889                 spin_unlock(old_ptl);
1890                 return true;
1891         }
1892         return false;
1893 }
1894
1895 /*
1896  * Returns
1897  *  - 0 if PMD could not be locked
1898  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1899  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1900  */
1901 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1902                 unsigned long addr, pgprot_t newprot, int prot_numa)
1903 {
1904         struct mm_struct *mm = vma->vm_mm;
1905         spinlock_t *ptl;
1906         pmd_t entry;
1907         bool preserve_write;
1908         int ret;
1909
1910         ptl = __pmd_trans_huge_lock(pmd, vma);
1911         if (!ptl)
1912                 return 0;
1913
1914         preserve_write = prot_numa && pmd_write(*pmd);
1915         ret = 1;
1916
1917 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1918         if (is_swap_pmd(*pmd)) {
1919                 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1920
1921                 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1922                 if (is_write_migration_entry(entry)) {
1923                         pmd_t newpmd;
1924                         /*
1925                          * A protection check is difficult so
1926                          * just be safe and disable write
1927                          */
1928                         make_migration_entry_read(&entry);
1929                         newpmd = swp_entry_to_pmd(entry);
1930                         if (pmd_swp_soft_dirty(*pmd))
1931                                 newpmd = pmd_swp_mksoft_dirty(newpmd);
1932                         set_pmd_at(mm, addr, pmd, newpmd);
1933                 }
1934                 goto unlock;
1935         }
1936 #endif
1937
1938         /*
1939          * Avoid trapping faults against the zero page. The read-only
1940          * data is likely to be read-cached on the local CPU and
1941          * local/remote hits to the zero page are not interesting.
1942          */
1943         if (prot_numa && is_huge_zero_pmd(*pmd))
1944                 goto unlock;
1945
1946         if (prot_numa && pmd_protnone(*pmd))
1947                 goto unlock;
1948
1949         /*
1950          * In case prot_numa, we are under down_read(mmap_sem). It's critical
1951          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1952          * which is also under down_read(mmap_sem):
1953          *
1954          *      CPU0:                           CPU1:
1955          *                              change_huge_pmd(prot_numa=1)
1956          *                               pmdp_huge_get_and_clear_notify()
1957          * madvise_dontneed()
1958          *  zap_pmd_range()
1959          *   pmd_trans_huge(*pmd) == 0 (without ptl)
1960          *   // skip the pmd
1961          *                               set_pmd_at();
1962          *                               // pmd is re-established
1963          *
1964          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1965          * which may break userspace.
1966          *
1967          * pmdp_invalidate() is required to make sure we don't miss
1968          * dirty/young flags set by hardware.
1969          */
1970         entry = pmdp_invalidate(vma, addr, pmd);
1971
1972         entry = pmd_modify(entry, newprot);
1973         if (preserve_write)
1974                 entry = pmd_mk_savedwrite(entry);
1975         ret = HPAGE_PMD_NR;
1976         set_pmd_at(mm, addr, pmd, entry);
1977         BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1978 unlock:
1979         spin_unlock(ptl);
1980         return ret;
1981 }
1982
1983 /*
1984  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1985  *
1986  * Note that if it returns page table lock pointer, this routine returns without
1987  * unlocking page table lock. So callers must unlock it.
1988  */
1989 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1990 {
1991         spinlock_t *ptl;
1992         ptl = pmd_lock(vma->vm_mm, pmd);
1993         if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1994                         pmd_devmap(*pmd)))
1995                 return ptl;
1996         spin_unlock(ptl);
1997         return NULL;
1998 }
1999
2000 /*
2001  * Returns true if a given pud maps a thp, false otherwise.
2002  *
2003  * Note that if it returns true, this routine returns without unlocking page
2004  * table lock. So callers must unlock it.
2005  */
2006 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
2007 {
2008         spinlock_t *ptl;
2009
2010         ptl = pud_lock(vma->vm_mm, pud);
2011         if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
2012                 return ptl;
2013         spin_unlock(ptl);
2014         return NULL;
2015 }
2016
2017 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
2018 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
2019                  pud_t *pud, unsigned long addr)
2020 {
2021         spinlock_t *ptl;
2022
2023         ptl = __pud_trans_huge_lock(pud, vma);
2024         if (!ptl)
2025                 return 0;
2026         /*
2027          * For architectures like ppc64 we look at deposited pgtable
2028          * when calling pudp_huge_get_and_clear. So do the
2029          * pgtable_trans_huge_withdraw after finishing pudp related
2030          * operations.
2031          */
2032         pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
2033         tlb_remove_pud_tlb_entry(tlb, pud, addr);
2034         if (vma_is_dax(vma)) {
2035                 spin_unlock(ptl);
2036                 /* No zero page support yet */
2037         } else {
2038                 /* No support for anonymous PUD pages yet */
2039                 BUG();
2040         }
2041         return 1;
2042 }
2043
2044 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
2045                 unsigned long haddr)
2046 {
2047         VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
2048         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2049         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
2050         VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
2051
2052         count_vm_event(THP_SPLIT_PUD);
2053
2054         pudp_huge_clear_flush_notify(vma, haddr, pud);
2055 }
2056
2057 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2058                 unsigned long address)
2059 {
2060         spinlock_t *ptl;
2061         struct mmu_notifier_range range;
2062
2063         mmu_notifier_range_init(&range, vma->vm_mm, address & HPAGE_PUD_MASK,
2064                                 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
2065         mmu_notifier_invalidate_range_start(&range);
2066         ptl = pud_lock(vma->vm_mm, pud);
2067         if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2068                 goto out;
2069         __split_huge_pud_locked(vma, pud, range.start);
2070
2071 out:
2072         spin_unlock(ptl);
2073         /*
2074          * No need to double call mmu_notifier->invalidate_range() callback as
2075          * the above pudp_huge_clear_flush_notify() did already call it.
2076          */
2077         mmu_notifier_invalidate_range_only_end(&range);
2078 }
2079 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2080
2081 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2082                 unsigned long haddr, pmd_t *pmd)
2083 {
2084         struct mm_struct *mm = vma->vm_mm;
2085         pgtable_t pgtable;
2086         pmd_t _pmd;
2087         int i;
2088
2089         /*
2090          * Leave pmd empty until pte is filled note that it is fine to delay
2091          * notification until mmu_notifier_invalidate_range_end() as we are
2092          * replacing a zero pmd write protected page with a zero pte write
2093          * protected page.
2094          *
2095          * See Documentation/vm/mmu_notifier.rst
2096          */
2097         pmdp_huge_clear_flush(vma, haddr, pmd);
2098
2099         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2100         pmd_populate(mm, &_pmd, pgtable);
2101
2102         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2103                 pte_t *pte, entry;
2104                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2105                 entry = pte_mkspecial(entry);
2106                 pte = pte_offset_map(&_pmd, haddr);
2107                 VM_BUG_ON(!pte_none(*pte));
2108                 set_pte_at(mm, haddr, pte, entry);
2109                 pte_unmap(pte);
2110         }
2111         smp_wmb(); /* make pte visible before pmd */
2112         pmd_populate(mm, pmd, pgtable);
2113 }
2114
2115 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2116                 unsigned long haddr, bool freeze)
2117 {
2118         struct mm_struct *mm = vma->vm_mm;
2119         struct page *page;
2120         pgtable_t pgtable;
2121         pmd_t old_pmd, _pmd;
2122         bool young, write, soft_dirty, pmd_migration = false;
2123         unsigned long addr;
2124         int i;
2125
2126         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2127         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2128         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2129         VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2130                                 && !pmd_devmap(*pmd));
2131
2132         count_vm_event(THP_SPLIT_PMD);
2133
2134         if (!vma_is_anonymous(vma)) {
2135                 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2136                 /*
2137                  * We are going to unmap this huge page. So
2138                  * just go ahead and zap it
2139                  */
2140                 if (arch_needs_pgtable_deposit())
2141                         zap_deposited_table(mm, pmd);
2142                 if (vma_is_dax(vma))
2143                         return;
2144                 page = pmd_page(_pmd);
2145                 if (!PageDirty(page) && pmd_dirty(_pmd))
2146                         set_page_dirty(page);
2147                 if (!PageReferenced(page) && pmd_young(_pmd))
2148                         SetPageReferenced(page);
2149                 page_remove_rmap(page, true);
2150                 put_page(page);
2151                 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2152                 return;
2153         } else if (is_huge_zero_pmd(*pmd)) {
2154                 /*
2155                  * FIXME: Do we want to invalidate secondary mmu by calling
2156                  * mmu_notifier_invalidate_range() see comments below inside
2157                  * __split_huge_pmd() ?
2158                  *
2159                  * We are going from a zero huge page write protected to zero
2160                  * small page also write protected so it does not seems useful
2161                  * to invalidate secondary mmu at this time.
2162                  */
2163                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2164         }
2165
2166         /*
2167          * Up to this point the pmd is present and huge and userland has the
2168          * whole access to the hugepage during the split (which happens in
2169          * place). If we overwrite the pmd with the not-huge version pointing
2170          * to the pte here (which of course we could if all CPUs were bug
2171          * free), userland could trigger a small page size TLB miss on the
2172          * small sized TLB while the hugepage TLB entry is still established in
2173          * the huge TLB. Some CPU doesn't like that.
2174          * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2175          * 383 on page 93. Intel should be safe but is also warns that it's
2176          * only safe if the permission and cache attributes of the two entries
2177          * loaded in the two TLB is identical (which should be the case here).
2178          * But it is generally safer to never allow small and huge TLB entries
2179          * for the same virtual address to be loaded simultaneously. So instead
2180          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2181          * current pmd notpresent (atomically because here the pmd_trans_huge
2182          * must remain set at all times on the pmd until the split is complete
2183          * for this pmd), then we flush the SMP TLB and finally we write the
2184          * non-huge version of the pmd entry with pmd_populate.
2185          */
2186         old_pmd = pmdp_invalidate(vma, haddr, pmd);
2187
2188         pmd_migration = is_pmd_migration_entry(old_pmd);
2189         if (unlikely(pmd_migration)) {
2190                 swp_entry_t entry;
2191
2192                 entry = pmd_to_swp_entry(old_pmd);
2193                 page = pfn_to_page(swp_offset(entry));
2194                 write = is_write_migration_entry(entry);
2195                 young = false;
2196                 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2197         } else {
2198                 page = pmd_page(old_pmd);
2199                 if (pmd_dirty(old_pmd))
2200                         SetPageDirty(page);
2201                 write = pmd_write(old_pmd);
2202                 young = pmd_young(old_pmd);
2203                 soft_dirty = pmd_soft_dirty(old_pmd);
2204         }
2205         VM_BUG_ON_PAGE(!page_count(page), page);
2206         page_ref_add(page, HPAGE_PMD_NR - 1);
2207
2208         /*
2209          * Withdraw the table only after we mark the pmd entry invalid.
2210          * This's critical for some architectures (Power).
2211          */
2212         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2213         pmd_populate(mm, &_pmd, pgtable);
2214
2215         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2216                 pte_t entry, *pte;
2217                 /*
2218                  * Note that NUMA hinting access restrictions are not
2219                  * transferred to avoid any possibility of altering
2220                  * permissions across VMAs.
2221                  */
2222                 if (freeze || pmd_migration) {
2223                         swp_entry_t swp_entry;
2224                         swp_entry = make_migration_entry(page + i, write);
2225                         entry = swp_entry_to_pte(swp_entry);
2226                         if (soft_dirty)
2227                                 entry = pte_swp_mksoft_dirty(entry);
2228                 } else {
2229                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2230                         entry = maybe_mkwrite(entry, vma);
2231                         if (!write)
2232                                 entry = pte_wrprotect(entry);
2233                         if (!young)
2234                                 entry = pte_mkold(entry);
2235                         if (soft_dirty)
2236                                 entry = pte_mksoft_dirty(entry);
2237                 }
2238                 pte = pte_offset_map(&_pmd, addr);
2239                 BUG_ON(!pte_none(*pte));
2240                 set_pte_at(mm, addr, pte, entry);
2241                 atomic_inc(&page[i]._mapcount);
2242                 pte_unmap(pte);
2243         }
2244
2245         /*
2246          * Set PG_double_map before dropping compound_mapcount to avoid
2247          * false-negative page_mapped().
2248          */
2249         if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2250                 for (i = 0; i < HPAGE_PMD_NR; i++)
2251                         atomic_inc(&page[i]._mapcount);
2252         }
2253
2254         if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2255                 /* Last compound_mapcount is gone. */
2256                 __dec_node_page_state(page, NR_ANON_THPS);
2257                 if (TestClearPageDoubleMap(page)) {
2258                         /* No need in mapcount reference anymore */
2259                         for (i = 0; i < HPAGE_PMD_NR; i++)
2260                                 atomic_dec(&page[i]._mapcount);
2261                 }
2262         }
2263
2264         smp_wmb(); /* make pte visible before pmd */
2265         pmd_populate(mm, pmd, pgtable);
2266
2267         if (freeze) {
2268                 for (i = 0; i < HPAGE_PMD_NR; i++) {
2269                         page_remove_rmap(page + i, false);
2270                         put_page(page + i);
2271                 }
2272         }
2273 }
2274
2275 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2276                 unsigned long address, bool freeze, struct page *page)
2277 {
2278         spinlock_t *ptl;
2279         struct mmu_notifier_range range;
2280
2281         mmu_notifier_range_init(&range, vma->vm_mm, address & HPAGE_PMD_MASK,
2282                                 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2283         mmu_notifier_invalidate_range_start(&range);
2284         ptl = pmd_lock(vma->vm_mm, pmd);
2285
2286         /*
2287          * If caller asks to setup a migration entries, we need a page to check
2288          * pmd against. Otherwise we can end up replacing wrong page.
2289          */
2290         VM_BUG_ON(freeze && !page);
2291         if (page && page != pmd_page(*pmd))
2292                 goto out;
2293
2294         if (pmd_trans_huge(*pmd)) {
2295                 page = pmd_page(*pmd);
2296                 if (PageMlocked(page))
2297                         clear_page_mlock(page);
2298         } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2299                 goto out;
2300         __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2301 out:
2302         spin_unlock(ptl);
2303         /*
2304          * No need to double call mmu_notifier->invalidate_range() callback.
2305          * They are 3 cases to consider inside __split_huge_pmd_locked():
2306          *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2307          *  2) __split_huge_zero_page_pmd() read only zero page and any write
2308          *    fault will trigger a flush_notify before pointing to a new page
2309          *    (it is fine if the secondary mmu keeps pointing to the old zero
2310          *    page in the meantime)
2311          *  3) Split a huge pmd into pte pointing to the same page. No need
2312          *     to invalidate secondary tlb entry they are all still valid.
2313          *     any further changes to individual pte will notify. So no need
2314          *     to call mmu_notifier->invalidate_range()
2315          */
2316         mmu_notifier_invalidate_range_only_end(&range);
2317 }
2318
2319 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2320                 bool freeze, struct page *page)
2321 {
2322         pgd_t *pgd;
2323         p4d_t *p4d;
2324         pud_t *pud;
2325         pmd_t *pmd;
2326
2327         pgd = pgd_offset(vma->vm_mm, address);
2328         if (!pgd_present(*pgd))
2329                 return;
2330
2331         p4d = p4d_offset(pgd, address);
2332         if (!p4d_present(*p4d))
2333                 return;
2334
2335         pud = pud_offset(p4d, address);
2336         if (!pud_present(*pud))
2337                 return;
2338
2339         pmd = pmd_offset(pud, address);
2340
2341         __split_huge_pmd(vma, pmd, address, freeze, page);
2342 }
2343
2344 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2345                              unsigned long start,
2346                              unsigned long end,
2347                              long adjust_next)
2348 {
2349         /*
2350          * If the new start address isn't hpage aligned and it could
2351          * previously contain an hugepage: check if we need to split
2352          * an huge pmd.
2353          */
2354         if (start & ~HPAGE_PMD_MASK &&
2355             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2356             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2357                 split_huge_pmd_address(vma, start, false, NULL);
2358
2359         /*
2360          * If the new end address isn't hpage aligned and it could
2361          * previously contain an hugepage: check if we need to split
2362          * an huge pmd.
2363          */
2364         if (end & ~HPAGE_PMD_MASK &&
2365             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2366             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2367                 split_huge_pmd_address(vma, end, false, NULL);
2368
2369         /*
2370          * If we're also updating the vma->vm_next->vm_start, if the new
2371          * vm_next->vm_start isn't page aligned and it could previously
2372          * contain an hugepage: check if we need to split an huge pmd.
2373          */
2374         if (adjust_next > 0) {
2375                 struct vm_area_struct *next = vma->vm_next;
2376                 unsigned long nstart = next->vm_start;
2377                 nstart += adjust_next << PAGE_SHIFT;
2378                 if (nstart & ~HPAGE_PMD_MASK &&
2379                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2380                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2381                         split_huge_pmd_address(next, nstart, false, NULL);
2382         }
2383 }
2384
2385 static void unmap_page(struct page *page)
2386 {
2387         enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
2388                 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2389         bool unmap_success;
2390
2391         VM_BUG_ON_PAGE(!PageHead(page), page);
2392
2393         if (PageAnon(page))
2394                 ttu_flags |= TTU_SPLIT_FREEZE;
2395
2396         unmap_success = try_to_unmap(page, ttu_flags);
2397         VM_BUG_ON_PAGE(!unmap_success, page);
2398 }
2399
2400 static void remap_page(struct page *page)
2401 {
2402         int i;
2403         if (PageTransHuge(page)) {
2404                 remove_migration_ptes(page, page, true);
2405         } else {
2406                 for (i = 0; i < HPAGE_PMD_NR; i++)
2407                         remove_migration_ptes(page + i, page + i, true);
2408         }
2409 }
2410
2411 static void __split_huge_page_tail(struct page *head, int tail,
2412                 struct lruvec *lruvec, struct list_head *list)
2413 {
2414         struct page *page_tail = head + tail;
2415
2416         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2417
2418         /*
2419          * Clone page flags before unfreezing refcount.
2420          *
2421          * After successful get_page_unless_zero() might follow flags change,
2422          * for exmaple lock_page() which set PG_waiters.
2423          */
2424         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2425         page_tail->flags |= (head->flags &
2426                         ((1L << PG_referenced) |
2427                          (1L << PG_swapbacked) |
2428                          (1L << PG_swapcache) |
2429                          (1L << PG_mlocked) |
2430                          (1L << PG_uptodate) |
2431                          (1L << PG_active) |
2432                          (1L << PG_workingset) |
2433                          (1L << PG_locked) |
2434                          (1L << PG_unevictable) |
2435                          (1L << PG_dirty)));
2436
2437         /* ->mapping in first tail page is compound_mapcount */
2438         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2439                         page_tail);
2440         page_tail->mapping = head->mapping;
2441         page_tail->index = head->index + tail;
2442
2443         /* Page flags must be visible before we make the page non-compound. */
2444         smp_wmb();
2445
2446         /*
2447          * Clear PageTail before unfreezing page refcount.
2448          *
2449          * After successful get_page_unless_zero() might follow put_page()
2450          * which needs correct compound_head().
2451          */
2452         clear_compound_head(page_tail);
2453
2454         /* Finally unfreeze refcount. Additional reference from page cache. */
2455         page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2456                                           PageSwapCache(head)));
2457
2458         if (page_is_young(head))
2459                 set_page_young(page_tail);
2460         if (page_is_idle(head))
2461                 set_page_idle(page_tail);
2462
2463         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2464
2465         /*
2466          * always add to the tail because some iterators expect new
2467          * pages to show after the currently processed elements - e.g.
2468          * migrate_pages
2469          */
2470         lru_add_page_tail(head, page_tail, lruvec, list);
2471 }
2472
2473 static void __split_huge_page(struct page *page, struct list_head *list,
2474                 pgoff_t end, unsigned long flags)
2475 {
2476         struct page *head = compound_head(page);
2477         pg_data_t *pgdat = page_pgdat(head);
2478         struct lruvec *lruvec;
2479         int i;
2480
2481         lruvec = mem_cgroup_page_lruvec(head, pgdat);
2482
2483         /* complete memcg works before add pages to LRU */
2484         mem_cgroup_split_huge_fixup(head);
2485
2486         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
2487                 __split_huge_page_tail(head, i, lruvec, list);
2488                 /* Some pages can be beyond i_size: drop them from page cache */
2489                 if (head[i].index >= end) {
2490                         ClearPageDirty(head + i);
2491                         __delete_from_page_cache(head + i, NULL);
2492                         if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2493                                 shmem_uncharge(head->mapping->host, 1);
2494                         put_page(head + i);
2495                 }
2496         }
2497
2498         ClearPageCompound(head);
2499         /* See comment in __split_huge_page_tail() */
2500         if (PageAnon(head)) {
2501                 /* Additional pin to swap cache */
2502                 if (PageSwapCache(head))
2503                         page_ref_add(head, 2);
2504                 else
2505                         page_ref_inc(head);
2506         } else {
2507                 /* Additional pin to page cache */
2508                 page_ref_add(head, 2);
2509                 xa_unlock(&head->mapping->i_pages);
2510         }
2511
2512         spin_unlock_irqrestore(&pgdat->lru_lock, flags);
2513
2514         remap_page(head);
2515
2516         for (i = 0; i < HPAGE_PMD_NR; i++) {
2517                 struct page *subpage = head + i;
2518                 if (subpage == page)
2519                         continue;
2520                 unlock_page(subpage);
2521
2522                 /*
2523                  * Subpages may be freed if there wasn't any mapping
2524                  * like if add_to_swap() is running on a lru page that
2525                  * had its mapping zapped. And freeing these pages
2526                  * requires taking the lru_lock so we do the put_page
2527                  * of the tail pages after the split is complete.
2528                  */
2529                 put_page(subpage);
2530         }
2531 }
2532
2533 int total_mapcount(struct page *page)
2534 {
2535         int i, compound, ret;
2536
2537         VM_BUG_ON_PAGE(PageTail(page), page);
2538
2539         if (likely(!PageCompound(page)))
2540                 return atomic_read(&page->_mapcount) + 1;
2541
2542         compound = compound_mapcount(page);
2543         if (PageHuge(page))
2544                 return compound;
2545         ret = compound;
2546         for (i = 0; i < HPAGE_PMD_NR; i++)
2547                 ret += atomic_read(&page[i]._mapcount) + 1;
2548         /* File pages has compound_mapcount included in _mapcount */
2549         if (!PageAnon(page))
2550                 return ret - compound * HPAGE_PMD_NR;
2551         if (PageDoubleMap(page))
2552                 ret -= HPAGE_PMD_NR;
2553         return ret;
2554 }
2555
2556 /*
2557  * This calculates accurately how many mappings a transparent hugepage
2558  * has (unlike page_mapcount() which isn't fully accurate). This full
2559  * accuracy is primarily needed to know if copy-on-write faults can
2560  * reuse the page and change the mapping to read-write instead of
2561  * copying them. At the same time this returns the total_mapcount too.
2562  *
2563  * The function returns the highest mapcount any one of the subpages
2564  * has. If the return value is one, even if different processes are
2565  * mapping different subpages of the transparent hugepage, they can
2566  * all reuse it, because each process is reusing a different subpage.
2567  *
2568  * The total_mapcount is instead counting all virtual mappings of the
2569  * subpages. If the total_mapcount is equal to "one", it tells the
2570  * caller all mappings belong to the same "mm" and in turn the
2571  * anon_vma of the transparent hugepage can become the vma->anon_vma
2572  * local one as no other process may be mapping any of the subpages.
2573  *
2574  * It would be more accurate to replace page_mapcount() with
2575  * page_trans_huge_mapcount(), however we only use
2576  * page_trans_huge_mapcount() in the copy-on-write faults where we
2577  * need full accuracy to avoid breaking page pinning, because
2578  * page_trans_huge_mapcount() is slower than page_mapcount().
2579  */
2580 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2581 {
2582         int i, ret, _total_mapcount, mapcount;
2583
2584         /* hugetlbfs shouldn't call it */
2585         VM_BUG_ON_PAGE(PageHuge(page), page);
2586
2587         if (likely(!PageTransCompound(page))) {
2588                 mapcount = atomic_read(&page->_mapcount) + 1;
2589                 if (total_mapcount)
2590                         *total_mapcount = mapcount;
2591                 return mapcount;
2592         }
2593
2594         page = compound_head(page);
2595
2596         _total_mapcount = ret = 0;
2597         for (i = 0; i < HPAGE_PMD_NR; i++) {
2598                 mapcount = atomic_read(&page[i]._mapcount) + 1;
2599                 ret = max(ret, mapcount);
2600                 _total_mapcount += mapcount;
2601         }
2602         if (PageDoubleMap(page)) {
2603                 ret -= 1;
2604                 _total_mapcount -= HPAGE_PMD_NR;
2605         }
2606         mapcount = compound_mapcount(page);
2607         ret += mapcount;
2608         _total_mapcount += mapcount;
2609         if (total_mapcount)
2610                 *total_mapcount = _total_mapcount;
2611         return ret;
2612 }
2613
2614 /* Racy check whether the huge page can be split */
2615 bool can_split_huge_page(struct page *page, int *pextra_pins)
2616 {
2617         int extra_pins;
2618
2619         /* Additional pins from page cache */
2620         if (PageAnon(page))
2621                 extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0;
2622         else
2623                 extra_pins = HPAGE_PMD_NR;
2624         if (pextra_pins)
2625                 *pextra_pins = extra_pins;
2626         return total_mapcount(page) == page_count(page) - extra_pins - 1;
2627 }
2628
2629 /*
2630  * This function splits huge page into normal pages. @page can point to any
2631  * subpage of huge page to split. Split doesn't change the position of @page.
2632  *
2633  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2634  * The huge page must be locked.
2635  *
2636  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2637  *
2638  * Both head page and tail pages will inherit mapping, flags, and so on from
2639  * the hugepage.
2640  *
2641  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2642  * they are not mapped.
2643  *
2644  * Returns 0 if the hugepage is split successfully.
2645  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2646  * us.
2647  */
2648 int split_huge_page_to_list(struct page *page, struct list_head *list)
2649 {
2650         struct page *head = compound_head(page);
2651         struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2652         struct anon_vma *anon_vma = NULL;
2653         struct address_space *mapping = NULL;
2654         int count, mapcount, extra_pins, ret;
2655         bool mlocked;
2656         unsigned long flags;
2657         pgoff_t end;
2658
2659         VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
2660         VM_BUG_ON_PAGE(!PageLocked(page), page);
2661         VM_BUG_ON_PAGE(!PageCompound(page), page);
2662
2663         if (PageWriteback(page))
2664                 return -EBUSY;
2665
2666         if (PageAnon(head)) {
2667                 /*
2668                  * The caller does not necessarily hold an mmap_sem that would
2669                  * prevent the anon_vma disappearing so we first we take a
2670                  * reference to it and then lock the anon_vma for write. This
2671                  * is similar to page_lock_anon_vma_read except the write lock
2672                  * is taken to serialise against parallel split or collapse
2673                  * operations.
2674                  */
2675                 anon_vma = page_get_anon_vma(head);
2676                 if (!anon_vma) {
2677                         ret = -EBUSY;
2678                         goto out;
2679                 }
2680                 end = -1;
2681                 mapping = NULL;
2682                 anon_vma_lock_write(anon_vma);
2683         } else {
2684                 mapping = head->mapping;
2685
2686                 /* Truncated ? */
2687                 if (!mapping) {
2688                         ret = -EBUSY;
2689                         goto out;
2690                 }
2691
2692                 anon_vma = NULL;
2693                 i_mmap_lock_read(mapping);
2694
2695                 /*
2696                  *__split_huge_page() may need to trim off pages beyond EOF:
2697                  * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2698                  * which cannot be nested inside the page tree lock. So note
2699                  * end now: i_size itself may be changed at any moment, but
2700                  * head page lock is good enough to serialize the trimming.
2701                  */
2702                 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2703         }
2704
2705         /*
2706          * Racy check if we can split the page, before unmap_page() will
2707          * split PMDs
2708          */
2709         if (!can_split_huge_page(head, &extra_pins)) {
2710                 ret = -EBUSY;
2711                 goto out_unlock;
2712         }
2713
2714         mlocked = PageMlocked(page);
2715         unmap_page(head);
2716         VM_BUG_ON_PAGE(compound_mapcount(head), head);
2717
2718         /* Make sure the page is not on per-CPU pagevec as it takes pin */
2719         if (mlocked)
2720                 lru_add_drain();
2721
2722         /* prevent PageLRU to go away from under us, and freeze lru stats */
2723         spin_lock_irqsave(&pgdata->lru_lock, flags);
2724
2725         if (mapping) {
2726                 XA_STATE(xas, &mapping->i_pages, page_index(head));
2727
2728                 /*
2729                  * Check if the head page is present in page cache.
2730                  * We assume all tail are present too, if head is there.
2731                  */
2732                 xa_lock(&mapping->i_pages);
2733                 if (xas_load(&xas) != head)
2734                         goto fail;
2735         }
2736
2737         /* Prevent deferred_split_scan() touching ->_refcount */
2738         spin_lock(&pgdata->split_queue_lock);
2739         count = page_count(head);
2740         mapcount = total_mapcount(head);
2741         if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2742                 if (!list_empty(page_deferred_list(head))) {
2743                         pgdata->split_queue_len--;
2744                         list_del(page_deferred_list(head));
2745                 }
2746                 if (mapping)
2747                         __dec_node_page_state(page, NR_SHMEM_THPS);
2748                 spin_unlock(&pgdata->split_queue_lock);
2749                 __split_huge_page(page, list, end, flags);
2750                 if (PageSwapCache(head)) {
2751                         swp_entry_t entry = { .val = page_private(head) };
2752
2753                         ret = split_swap_cluster(entry);
2754                 } else
2755                         ret = 0;
2756         } else {
2757                 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2758                         pr_alert("total_mapcount: %u, page_count(): %u\n",
2759                                         mapcount, count);
2760                         if (PageTail(page))
2761                                 dump_page(head, NULL);
2762                         dump_page(page, "total_mapcount(head) > 0");
2763                         BUG();
2764                 }
2765                 spin_unlock(&pgdata->split_queue_lock);
2766 fail:           if (mapping)
2767                         xa_unlock(&mapping->i_pages);
2768                 spin_unlock_irqrestore(&pgdata->lru_lock, flags);
2769                 remap_page(head);
2770                 ret = -EBUSY;
2771         }
2772
2773 out_unlock:
2774         if (anon_vma) {
2775                 anon_vma_unlock_write(anon_vma);
2776                 put_anon_vma(anon_vma);
2777         }
2778         if (mapping)
2779                 i_mmap_unlock_read(mapping);
2780 out:
2781         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2782         return ret;
2783 }
2784
2785 void free_transhuge_page(struct page *page)
2786 {
2787         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2788         unsigned long flags;
2789
2790         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2791         if (!list_empty(page_deferred_list(page))) {
2792                 pgdata->split_queue_len--;
2793                 list_del(page_deferred_list(page));
2794         }
2795         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2796         free_compound_page(page);
2797 }
2798
2799 void deferred_split_huge_page(struct page *page)
2800 {
2801         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2802         unsigned long flags;
2803
2804         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2805
2806         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2807         if (list_empty(page_deferred_list(page))) {
2808                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2809                 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2810                 pgdata->split_queue_len++;
2811         }
2812         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2813 }
2814
2815 static unsigned long deferred_split_count(struct shrinker *shrink,
2816                 struct shrink_control *sc)
2817 {
2818         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2819         return READ_ONCE(pgdata->split_queue_len);
2820 }
2821
2822 static unsigned long deferred_split_scan(struct shrinker *shrink,
2823                 struct shrink_control *sc)
2824 {
2825         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2826         unsigned long flags;
2827         LIST_HEAD(list), *pos, *next;
2828         struct page *page;
2829         int split = 0;
2830
2831         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2832         /* Take pin on all head pages to avoid freeing them under us */
2833         list_for_each_safe(pos, next, &pgdata->split_queue) {
2834                 page = list_entry((void *)pos, struct page, mapping);
2835                 page = compound_head(page);
2836                 if (get_page_unless_zero(page)) {
2837                         list_move(page_deferred_list(page), &list);
2838                 } else {
2839                         /* We lost race with put_compound_page() */
2840                         list_del_init(page_deferred_list(page));
2841                         pgdata->split_queue_len--;
2842                 }
2843                 if (!--sc->nr_to_scan)
2844                         break;
2845         }
2846         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2847
2848         list_for_each_safe(pos, next, &list) {
2849                 page = list_entry((void *)pos, struct page, mapping);
2850                 if (!trylock_page(page))
2851                         goto next;
2852                 /* split_huge_page() removes page from list on success */
2853                 if (!split_huge_page(page))
2854                         split++;
2855                 unlock_page(page);
2856 next:
2857                 put_page(page);
2858         }
2859
2860         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2861         list_splice_tail(&list, &pgdata->split_queue);
2862         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2863
2864         /*
2865          * Stop shrinker if we didn't split any page, but the queue is empty.
2866          * This can happen if pages were freed under us.
2867          */
2868         if (!split && list_empty(&pgdata->split_queue))
2869                 return SHRINK_STOP;
2870         return split;
2871 }
2872
2873 static struct shrinker deferred_split_shrinker = {
2874         .count_objects = deferred_split_count,
2875         .scan_objects = deferred_split_scan,
2876         .seeks = DEFAULT_SEEKS,
2877         .flags = SHRINKER_NUMA_AWARE,
2878 };
2879
2880 #ifdef CONFIG_DEBUG_FS
2881 static int split_huge_pages_set(void *data, u64 val)
2882 {
2883         struct zone *zone;
2884         struct page *page;
2885         unsigned long pfn, max_zone_pfn;
2886         unsigned long total = 0, split = 0;
2887
2888         if (val != 1)
2889                 return -EINVAL;
2890
2891         for_each_populated_zone(zone) {
2892                 max_zone_pfn = zone_end_pfn(zone);
2893                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2894                         if (!pfn_valid(pfn))
2895                                 continue;
2896
2897                         page = pfn_to_page(pfn);
2898                         if (!get_page_unless_zero(page))
2899                                 continue;
2900
2901                         if (zone != page_zone(page))
2902                                 goto next;
2903
2904                         if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2905                                 goto next;
2906
2907                         total++;
2908                         lock_page(page);
2909                         if (!split_huge_page(page))
2910                                 split++;
2911                         unlock_page(page);
2912 next:
2913                         put_page(page);
2914                 }
2915         }
2916
2917         pr_info("%lu of %lu THP split\n", split, total);
2918
2919         return 0;
2920 }
2921 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2922                 "%llu\n");
2923
2924 static int __init split_huge_pages_debugfs(void)
2925 {
2926         debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2927                             &split_huge_pages_fops);
2928         return 0;
2929 }
2930 late_initcall(split_huge_pages_debugfs);
2931 #endif
2932
2933 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2934 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
2935                 struct page *page)
2936 {
2937         struct vm_area_struct *vma = pvmw->vma;
2938         struct mm_struct *mm = vma->vm_mm;
2939         unsigned long address = pvmw->address;
2940         pmd_t pmdval;
2941         swp_entry_t entry;
2942         pmd_t pmdswp;
2943
2944         if (!(pvmw->pmd && !pvmw->pte))
2945                 return;
2946
2947         flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
2948         pmdval = *pvmw->pmd;
2949         pmdp_invalidate(vma, address, pvmw->pmd);
2950         if (pmd_dirty(pmdval))
2951                 set_page_dirty(page);
2952         entry = make_migration_entry(page, pmd_write(pmdval));
2953         pmdswp = swp_entry_to_pmd(entry);
2954         if (pmd_soft_dirty(pmdval))
2955                 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
2956         set_pmd_at(mm, address, pvmw->pmd, pmdswp);
2957         page_remove_rmap(page, true);
2958         put_page(page);
2959 }
2960
2961 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
2962 {
2963         struct vm_area_struct *vma = pvmw->vma;
2964         struct mm_struct *mm = vma->vm_mm;
2965         unsigned long address = pvmw->address;
2966         unsigned long mmun_start = address & HPAGE_PMD_MASK;
2967         pmd_t pmde;
2968         swp_entry_t entry;
2969
2970         if (!(pvmw->pmd && !pvmw->pte))
2971                 return;
2972
2973         entry = pmd_to_swp_entry(*pvmw->pmd);
2974         get_page(new);
2975         pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
2976         if (pmd_swp_soft_dirty(*pvmw->pmd))
2977                 pmde = pmd_mksoft_dirty(pmde);
2978         if (is_write_migration_entry(entry))
2979                 pmde = maybe_pmd_mkwrite(pmde, vma);
2980
2981         flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
2982         if (PageAnon(new))
2983                 page_add_anon_rmap(new, vma, mmun_start, true);
2984         else
2985                 page_add_file_rmap(new, true);
2986         set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
2987         if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new))
2988                 mlock_vma_page(new);
2989         update_mmu_cache_pmd(vma, address, pvmw->pmd);
2990 }
2991 #endif