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
  37 #include <asm/tlb.h>
  38 #include <asm/pgalloc.h>
  39 #include "internal.h"
  40
  41 /*
  42  * By default, transparent hugepage support is disabled in order to avoid
  43  * risking an increased memory footprint for applications that are not
  44  * guaranteed to benefit from it. When transparent hugepage support is
  45  * enabled, it is for all mappings, and khugepaged scans all mappings.
  46  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  47  * for all hugepage allocations.
  48  */
  49 unsigned long transparent_hugepage_flags __read_mostly =
  50 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  51         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  52 #endif
  53 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  54         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  55 #endif
  56         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
  57         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  58         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  59
  60 static struct shrinker deferred_split_shrinker;
  61
  62 static atomic_t huge_zero_refcount;
  63 struct page *huge_zero_page __read_mostly;
  64
  65 static struct page *get_huge_zero_page(void)
  66 {
  67         struct page *zero_page;
  68 retry:
  69         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
  70                 return READ_ONCE(huge_zero_page);
  71
  72         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
  73                         HPAGE_PMD_ORDER);
  74         if (!zero_page) {
  75                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
  76                 return NULL;
  77         }
  78         count_vm_event(THP_ZERO_PAGE_ALLOC);
  79         preempt_disable();
  80         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
  81                 preempt_enable();
  82                 __free_pages(zero_page, compound_order(zero_page));
  83                 goto retry;
  84         }
  85
  86         /* We take additional reference here. It will be put back by shrinker */
  87         atomic_set(&huge_zero_refcount, 2);
  88         preempt_enable();
  89         return READ_ONCE(huge_zero_page);
  90 }
  91
  92 static void put_huge_zero_page(void)
  93 {
  94         /*
  95          * Counter should never go to zero here. Only shrinker can put
  96          * last reference.
  97          */
  98         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
  99 }
 100
 101 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
 102 {
 103         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 104                 return READ_ONCE(huge_zero_page);
 105
 106         if (!get_huge_zero_page())
 107                 return NULL;
 108
 109         if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 110                 put_huge_zero_page();
 111
 112         return READ_ONCE(huge_zero_page);
 113 }
 114
 115 void mm_put_huge_zero_page(struct mm_struct *mm)
 116 {
 117         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 118                 put_huge_zero_page();
 119 }
 120
 121 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 122                                         struct shrink_control *sc)
 123 {
 124         /* we can free zero page only if last reference remains */
 125         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 126 }
 127
 128 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 129                                        struct shrink_control *sc)
 130 {
 131         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 132                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 133                 BUG_ON(zero_page == NULL);
 134                 __free_pages(zero_page, compound_order(zero_page));
 135                 return HPAGE_PMD_NR;
 136         }
 137
 138         return 0;
 139 }
 140
 141 static struct shrinker huge_zero_page_shrinker = {
 142         .count_objects = shrink_huge_zero_page_count,
 143         .scan_objects = shrink_huge_zero_page_scan,
 144         .seeks = DEFAULT_SEEKS,
 145 };
 146
 147 #ifdef CONFIG_SYSFS
 148 static ssize_t enabled_show(struct kobject *kobj,
 149                             struct kobj_attribute *attr, char *buf)
 150 {
 151         if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
 152                 return sprintf(buf, "[always] madvise never\n");
 153         else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
 154                 return sprintf(buf, "always [madvise] never\n");
 155         else
 156                 return sprintf(buf, "always madvise [never]\n");
 157 }
 158
 159 static ssize_t enabled_store(struct kobject *kobj,
 160                              struct kobj_attribute *attr,
 161                              const char *buf, size_t count)
 162 {
 163         ssize_t ret = count;
 164
 165         if (!memcmp("always", buf,
 166                     min(sizeof("always")-1, count))) {
 167                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 168                 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 169         } else if (!memcmp("madvise", buf,
 170                            min(sizeof("madvise")-1, count))) {
 171                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 172                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 173         } else if (!memcmp("never", buf,
 174                            min(sizeof("never")-1, count))) {
 175                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 176                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 177         } else
 178                 ret = -EINVAL;
 179
 180         if (ret > 0) {
 181                 int err = start_stop_khugepaged();
 182                 if (err)
 183                         ret = err;
 184         }
 185         return ret;
 186 }
 187 static struct kobj_attribute enabled_attr =
 188         __ATTR(enabled, 0644, enabled_show, enabled_store);
 189
 190 ssize_t single_hugepage_flag_show(struct kobject *kobj,
 191                                 struct kobj_attribute *attr, char *buf,
 192                                 enum transparent_hugepage_flag flag)
 193 {
 194         return sprintf(buf, "%d\n",
 195                        !!test_bit(flag, &transparent_hugepage_flags));
 196 }
 197
 198 ssize_t single_hugepage_flag_store(struct kobject *kobj,
 199                                  struct kobj_attribute *attr,
 200                                  const char *buf, size_t count,
 201                                  enum transparent_hugepage_flag flag)
 202 {
 203         unsigned long value;
 204         int ret;
 205
 206         ret = kstrtoul(buf, 10, &value);
 207         if (ret < 0)
 208                 return ret;
 209         if (value > 1)
 210                 return -EINVAL;
 211
 212         if (value)
 213                 set_bit(flag, &transparent_hugepage_flags);
 214         else
 215                 clear_bit(flag, &transparent_hugepage_flags);
 216
 217         return count;
 218 }
 219
 220 static ssize_t defrag_show(struct kobject *kobj,
 221                            struct kobj_attribute *attr, char *buf)
 222 {
 223         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 224                 return sprintf(buf, "[always] defer defer+madvise madvise never\n");
 225         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 226                 return sprintf(buf, "always [defer] defer+madvise madvise never\n");
 227         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 228                 return sprintf(buf, "always defer [defer+madvise] madvise never\n");
 229         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 230                 return sprintf(buf, "always defer defer+madvise [madvise] never\n");
 231         return sprintf(buf, "always defer defer+madvise madvise [never]\n");
 232 }
 233
 234 static ssize_t defrag_store(struct kobject *kobj,
 235                             struct kobj_attribute *attr,
 236                             const char *buf, size_t count)
 237 {
 238         if (!memcmp("always", buf,
 239                     min(sizeof("always")-1, count))) {
 240                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 241                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 242                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 243                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 244         } else if (!memcmp("defer+madvise", buf,
 245                     min(sizeof("defer+madvise")-1, count))) {
 246                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 247                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 248                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 249                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 250         } else if (!memcmp("defer", buf,
 251                     min(sizeof("defer")-1, count))) {
 252                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 253                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 254                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 255                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 256         } else if (!memcmp("madvise", buf,
 257                            min(sizeof("madvise")-1, count))) {
 258                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 259                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 260                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 261                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 262         } else if (!memcmp("never", buf,
 263                            min(sizeof("never")-1, count))) {
 264                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 265                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 266                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 267                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 268         } else
 269                 return -EINVAL;
 270
 271         return count;
 272 }
 273 static struct kobj_attribute defrag_attr =
 274         __ATTR(defrag, 0644, defrag_show, defrag_store);
 275
 276 static ssize_t use_zero_page_show(struct kobject *kobj,
 277                 struct kobj_attribute *attr, char *buf)
 278 {
 279         return single_hugepage_flag_show(kobj, attr, buf,
 280                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 281 }
 282 static ssize_t use_zero_page_store(struct kobject *kobj,
 283                 struct kobj_attribute *attr, const char *buf, size_t count)
 284 {
 285         return single_hugepage_flag_store(kobj, attr, buf, count,
 286                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 287 }
 288 static struct kobj_attribute use_zero_page_attr =
 289         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 290
 291 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
 292                 struct kobj_attribute *attr, char *buf)
 293 {
 294         return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE);
 295 }
 296 static struct kobj_attribute hpage_pmd_size_attr =
 297         __ATTR_RO(hpage_pmd_size);
 298
 299 #ifdef CONFIG_DEBUG_VM
 300 static ssize_t debug_cow_show(struct kobject *kobj,
 301                                 struct kobj_attribute *attr, char *buf)
 302 {
 303         return single_hugepage_flag_show(kobj, attr, buf,
 304                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 305 }
 306 static ssize_t debug_cow_store(struct kobject *kobj,
 307                                struct kobj_attribute *attr,
 308                                const char *buf, size_t count)
 309 {
 310         return single_hugepage_flag_store(kobj, attr, buf, count,
 311                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 312 }
 313 static struct kobj_attribute debug_cow_attr =
 314         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 315 #endif /* CONFIG_DEBUG_VM */
 316
 317 static struct attribute *hugepage_attr[] = {
 318         &enabled_attr.attr,
 319         &defrag_attr.attr,
 320         &use_zero_page_attr.attr,
 321         &hpage_pmd_size_attr.attr,
 322 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
 323         &shmem_enabled_attr.attr,
 324 #endif
 325 #ifdef CONFIG_DEBUG_VM
 326         &debug_cow_attr.attr,
 327 #endif
 328         NULL,
 329 };
 330
 331 static const struct attribute_group hugepage_attr_group = {
 332         .attrs = hugepage_attr,
 333 };
 334
 335 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 336 {
 337         int err;
 338
 339         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 340         if (unlikely(!*hugepage_kobj)) {
 341                 pr_err("failed to create transparent hugepage kobject\n");
 342                 return -ENOMEM;
 343         }
 344
 345         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 346         if (err) {
 347                 pr_err("failed to register transparent hugepage group\n");
 348                 goto delete_obj;
 349         }
 350
 351         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 352         if (err) {
 353                 pr_err("failed to register transparent hugepage group\n");
 354                 goto remove_hp_group;
 355         }
 356
 357         return 0;
 358
 359 remove_hp_group:
 360         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 361 delete_obj:
 362         kobject_put(*hugepage_kobj);
 363         return err;
 364 }
 365
 366 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 367 {
 368         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 369         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 370         kobject_put(hugepage_kobj);
 371 }
 372 #else
 373 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 374 {
 375         return 0;
 376 }
 377
 378 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 379 {
 380 }
 381 #endif /* CONFIG_SYSFS */
 382
 383 static int __init hugepage_init(void)
 384 {
 385         int err;
 386         struct kobject *hugepage_kobj;
 387
 388         if (!has_transparent_hugepage()) {
 389                 transparent_hugepage_flags = 0;
 390                 return -EINVAL;
 391         }
 392
 393         /*
 394          * hugepages can't be allocated by the buddy allocator
 395          */
 396         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
 397         /*
 398          * we use page->mapping and page->index in second tail page
 399          * as list_head: assuming THP order >= 2
 400          */
 401         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 402
 403         err = hugepage_init_sysfs(&hugepage_kobj);
 404         if (err)
 405                 goto err_sysfs;
 406
 407         err = khugepaged_init();
 408         if (err)
 409                 goto err_slab;
 410
 411         err = register_shrinker(&huge_zero_page_shrinker);
 412         if (err)
 413                 goto err_hzp_shrinker;
 414         err = register_shrinker(&deferred_split_shrinker);
 415         if (err)
 416                 goto err_split_shrinker;
 417
 418         /*
 419          * By default disable transparent hugepages on smaller systems,
 420          * where the extra memory used could hurt more than TLB overhead
 421          * is likely to save.  The admin can still enable it through /sys.
 422          */
 423         if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
 424                 transparent_hugepage_flags = 0;
 425                 return 0;
 426         }
 427
 428         err = start_stop_khugepaged();
 429         if (err)
 430                 goto err_khugepaged;
 431
 432         return 0;
 433 err_khugepaged:
 434         unregister_shrinker(&deferred_split_shrinker);
 435 err_split_shrinker:
 436         unregister_shrinker(&huge_zero_page_shrinker);
 437 err_hzp_shrinker:
 438         khugepaged_destroy();
 439 err_slab:
 440         hugepage_exit_sysfs(hugepage_kobj);
 441 err_sysfs:
 442         return err;
 443 }
 444 subsys_initcall(hugepage_init);
 445
 446 static int __init setup_transparent_hugepage(char *str)
 447 {
 448         int ret = 0;
 449         if (!str)
 450                 goto out;
 451         if (!strcmp(str, "always")) {
 452                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 453                         &transparent_hugepage_flags);
 454                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 455                           &transparent_hugepage_flags);
 456                 ret = 1;
 457         } else if (!strcmp(str, "madvise")) {
 458                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 459                           &transparent_hugepage_flags);
 460                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 461                         &transparent_hugepage_flags);
 462                 ret = 1;
 463         } else if (!strcmp(str, "never")) {
 464                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 465                           &transparent_hugepage_flags);
 466                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 467                           &transparent_hugepage_flags);
 468                 ret = 1;
 469         }
 470 out:
 471         if (!ret)
 472                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 473         return ret;
 474 }
 475 __setup("transparent_hugepage=", setup_transparent_hugepage);
 476
 477 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma, bool dirty)
 478 {
 479         if (likely(vma->vm_flags & VM_WRITE)) {
 480                 pmd = pmd_mkwrite(pmd);
 481                 if (dirty)
 482                         pmd = pmd_mkdirty(pmd);
 483         }
 484         return pmd;
 485 }
 486
 487 static inline struct list_head *page_deferred_list(struct page *page)
 488 {
 489         /*
 490          * ->lru in the tail pages is occupied by compound_head.
 491          * Let's use ->mapping + ->index in the second tail page as list_head.
 492          */
 493         return (struct list_head *)&page[2].mapping;
 494 }
 495
 496 void prep_transhuge_page(struct page *page)
 497 {
 498         /*
 499          * we use page->mapping and page->indexlru in second tail page
 500          * as list_head: assuming THP order >= 2
 501          */
 502
 503         INIT_LIST_HEAD(page_deferred_list(page));
 504         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 505 }
 506
 507 unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len,
 508                 loff_t off, unsigned long flags, unsigned long size)
 509 {
 510         unsigned long addr;
 511         loff_t off_end = off + len;
 512         loff_t off_align = round_up(off, size);
 513         unsigned long len_pad;
 514
 515         if (off_end <= off_align || (off_end - off_align) < size)
 516                 return 0;
 517
 518         len_pad = len + size;
 519         if (len_pad < len || (off + len_pad) < off)
 520                 return 0;
 521
 522         addr = current->mm->get_unmapped_area(filp, 0, len_pad,
 523                                               off >> PAGE_SHIFT, flags);
 524         if (IS_ERR_VALUE(addr))
 525                 return 0;
 526
 527         addr += (off - addr) & (size - 1);
 528         return addr;
 529 }
 530
 531 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 532                 unsigned long len, unsigned long pgoff, unsigned long flags)
 533 {
 534         loff_t off = (loff_t)pgoff << PAGE_SHIFT;
 535
 536         if (addr)
 537                 goto out;
 538         if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
 539                 goto out;
 540
 541         addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE);
 542         if (addr)
 543                 return addr;
 544
 545  out:
 546         return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
 547 }
 548 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
 549
 550 static int __do_huge_pmd_anonymous_page(struct vm_fault *vmf, struct page *page,
 551                 gfp_t gfp)
 552 {
 553         struct vm_area_struct *vma = vmf->vma;
 554         struct mem_cgroup *memcg;
 555         pgtable_t pgtable;
 556         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 557         int ret = 0;
 558
 559         VM_BUG_ON_PAGE(!PageCompound(page), page);
 560
 561         if (mem_cgroup_try_charge(page, vma->vm_mm, gfp, &memcg, true)) {
 562                 put_page(page);
 563                 count_vm_event(THP_FAULT_FALLBACK);
 564                 return VM_FAULT_FALLBACK;
 565         }
 566
 567         pgtable = pte_alloc_one(vma->vm_mm, haddr);
 568         if (unlikely(!pgtable)) {
 569                 ret = VM_FAULT_OOM;
 570                 goto release;
 571         }
 572
 573         clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
 574         /*
 575          * The memory barrier inside __SetPageUptodate makes sure that
 576          * clear_huge_page writes become visible before the set_pmd_at()
 577          * write.
 578          */
 579         __SetPageUptodate(page);
 580
 581         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 582         if (unlikely(!pmd_none(*vmf->pmd))) {
 583                 goto unlock_release;
 584         } else {
 585                 pmd_t entry;
 586
 587                 ret = check_stable_address_space(vma->vm_mm);
 588                 if (ret)
 589                         goto unlock_release;
 590
 591                 /* Deliver the page fault to userland */
 592                 if (userfaultfd_missing(vma)) {
 593                         int ret;
 594
 595                         spin_unlock(vmf->ptl);
 596                         mem_cgroup_cancel_charge(page, memcg, true);
 597                         put_page(page);
 598                         pte_free(vma->vm_mm, pgtable);
 599                         ret = handle_userfault(vmf, VM_UFFD_MISSING);
 600                         VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 601                         return ret;
 602                 }
 603
 604                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 605                 entry = maybe_pmd_mkwrite(entry, vma, true);
 606                 page_add_new_anon_rmap(page, vma, haddr, true);
 607                 mem_cgroup_commit_charge(page, memcg, false, true);
 608                 lru_cache_add_active_or_unevictable(page, vma);
 609                 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
 610                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
 611                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 612                 mm_inc_nr_ptes(vma->vm_mm);
 613                 spin_unlock(vmf->ptl);
 614                 count_vm_event(THP_FAULT_ALLOC);
 615         }
 616
 617         return 0;
 618 unlock_release:
 619         spin_unlock(vmf->ptl);
 620 release:
 621         if (pgtable)
 622                 pte_free(vma->vm_mm, pgtable);
 623         mem_cgroup_cancel_charge(page, memcg, true);
 624         put_page(page);
 625         return ret;
 626
 627 }
 628
 629 /*
 630  * always: directly stall for all thp allocations
 631  * defer: wake kswapd and fail if not immediately available
 632  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
 633  *                fail if not immediately available
 634  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
 635  *          available
 636  * never: never stall for any thp allocation
 637  */
 638 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
 639 {
 640         const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
 641
 642         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 643                 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
 644         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 645                 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
 646         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 647                 return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM :
 648                                                              __GFP_KSWAPD_RECLAIM);
 649         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 650                 return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM :
 651                                                              0);
 652         return GFP_TRANSHUGE_LIGHT;
 653 }
 654
 655 /* Caller must hold page table lock. */
 656 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 657                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 658                 struct page *zero_page)
 659 {
 660         pmd_t entry;
 661         if (!pmd_none(*pmd))
 662                 return false;
 663         entry = mk_pmd(zero_page, vma->vm_page_prot);
 664         entry = pmd_mkhuge(entry);
 665         if (pgtable)
 666                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 667         set_pmd_at(mm, haddr, pmd, entry);
 668         mm_inc_nr_ptes(mm);
 669         return true;
 670 }
 671
 672 int do_huge_pmd_anonymous_page(struct vm_fault *vmf)
 673 {
 674         struct vm_area_struct *vma = vmf->vma;
 675         gfp_t gfp;
 676         struct page *page;
 677         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 678
 679         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 680                 return VM_FAULT_FALLBACK;
 681         if (unlikely(anon_vma_prepare(vma)))
 682                 return VM_FAULT_OOM;
 683         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 684                 return VM_FAULT_OOM;
 685         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
 686                         !mm_forbids_zeropage(vma->vm_mm) &&
 687                         transparent_hugepage_use_zero_page()) {
 688                 pgtable_t pgtable;
 689                 struct page *zero_page;
 690                 bool set;
 691                 int ret;
 692                 pgtable = pte_alloc_one(vma->vm_mm, haddr);
 693                 if (unlikely(!pgtable))
 694                         return VM_FAULT_OOM;
 695                 zero_page = mm_get_huge_zero_page(vma->vm_mm);
 696                 if (unlikely(!zero_page)) {
 697                         pte_free(vma->vm_mm, pgtable);
 698                         count_vm_event(THP_FAULT_FALLBACK);
 699                         return VM_FAULT_FALLBACK;
 700                 }
 701                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 702                 ret = 0;
 703                 set = false;
 704                 if (pmd_none(*vmf->pmd)) {
 705                         ret = check_stable_address_space(vma->vm_mm);
 706                         if (ret) {
 707                                 spin_unlock(vmf->ptl);
 708                         } else if (userfaultfd_missing(vma)) {
 709                                 spin_unlock(vmf->ptl);
 710                                 ret = handle_userfault(vmf, VM_UFFD_MISSING);
 711                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 712                         } else {
 713                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
 714                                                    haddr, vmf->pmd, zero_page);
 715                                 spin_unlock(vmf->ptl);
 716                                 set = true;
 717                         }
 718                 } else
 719                         spin_unlock(vmf->ptl);
 720                 if (!set)
 721                         pte_free(vma->vm_mm, pgtable);
 722                 return ret;
 723         }
 724         gfp = alloc_hugepage_direct_gfpmask(vma);
 725         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 726         if (unlikely(!page)) {
 727                 count_vm_event(THP_FAULT_FALLBACK);
 728                 return VM_FAULT_FALLBACK;
 729         }
 730         prep_transhuge_page(page);
 731         return __do_huge_pmd_anonymous_page(vmf, page, gfp);
 732 }
 733
 734 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 735                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
 736                 pgtable_t pgtable)
 737 {
 738         struct mm_struct *mm = vma->vm_mm;
 739         pmd_t entry;
 740         spinlock_t *ptl;
 741
 742         ptl = pmd_lock(mm, pmd);
 743         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 744         if (pfn_t_devmap(pfn))
 745                 entry = pmd_mkdevmap(entry);
 746         if (write) {
 747                 entry = pmd_mkyoung(entry);
 748                 entry = maybe_pmd_mkwrite(entry, vma, true);
 749         }
 750
 751         if (pgtable) {
 752                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 753                 mm_inc_nr_ptes(mm);
 754         }
 755
 756         set_pmd_at(mm, addr, pmd, entry);
 757         update_mmu_cache_pmd(vma, addr, pmd);
 758         spin_unlock(ptl);
 759 }
 760
 761 int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 762                         pmd_t *pmd, pfn_t pfn, bool write)
 763 {
 764         pgprot_t pgprot = vma->vm_page_prot;
 765         pgtable_t pgtable = NULL;
 766         /*
 767          * If we had pmd_special, we could avoid all these restrictions,
 768          * but we need to be consistent with PTEs and architectures that
 769          * can't support a 'special' bit.
 770          */
 771         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
 772         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 773                                                 (VM_PFNMAP|VM_MIXEDMAP));
 774         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 775         BUG_ON(!pfn_t_devmap(pfn));
 776
 777         if (addr < vma->vm_start || addr >= vma->vm_end)
 778                 return VM_FAULT_SIGBUS;
 779
 780         if (arch_needs_pgtable_deposit()) {
 781                 pgtable = pte_alloc_one(vma->vm_mm, addr);
 782                 if (!pgtable)
 783                         return VM_FAULT_OOM;
 784         }
 785
 786         track_pfn_insert(vma, &pgprot, pfn);
 787
 788         insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write, pgtable);
 789         return VM_FAULT_NOPAGE;
 790 }
 791 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
 792
 793 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 794 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma,
 795                 bool dirty)
 796 {
 797         if (likely(vma->vm_flags & VM_WRITE)) {
 798                 pud = pud_mkwrite(pud);
 799                 if (dirty)
 800                         pud = pud_mkdirty(pud);
 801         }
 802         return pud;
 803 }
 804
 805 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 806                 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
 807 {
 808         struct mm_struct *mm = vma->vm_mm;
 809         pud_t entry;
 810         spinlock_t *ptl;
 811
 812         ptl = pud_lock(mm, pud);
 813         entry = pud_mkhuge(pfn_t_pud(pfn, prot));
 814         if (pfn_t_devmap(pfn))
 815                 entry = pud_mkdevmap(entry);
 816         if (write) {
 817                 entry = pud_mkyoung(entry);
 818                 entry = maybe_pud_mkwrite(entry, vma, true);
 819         }
 820         set_pud_at(mm, addr, pud, entry);
 821         update_mmu_cache_pud(vma, addr, pud);
 822         spin_unlock(ptl);
 823 }
 824
 825 int vmf_insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 826                         pud_t *pud, pfn_t pfn, bool write)
 827 {
 828         pgprot_t pgprot = vma->vm_page_prot;
 829         /*
 830          * If we had pud_special, we could avoid all these restrictions,
 831          * but we need to be consistent with PTEs and architectures that
 832          * can't support a 'special' bit.
 833          */
 834         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
 835         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 836                                                 (VM_PFNMAP|VM_MIXEDMAP));
 837         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 838         BUG_ON(!pfn_t_devmap(pfn));
 839
 840         if (addr < vma->vm_start || addr >= vma->vm_end)
 841                 return VM_FAULT_SIGBUS;
 842
 843         track_pfn_insert(vma, &pgprot, pfn);
 844
 845         insert_pfn_pud(vma, addr, pud, pfn, pgprot, write);
 846         return VM_FAULT_NOPAGE;
 847 }
 848 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
 849 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
 850
 851 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
 852                 pmd_t *pmd, int flags)
 853 {
 854         pmd_t _pmd;
 855
 856         _pmd = pmd_mkyoung(*pmd);
 857         if (flags & FOLL_WRITE)
 858                 _pmd = pmd_mkdirty(_pmd);
 859         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
 860                                 pmd, _pmd, flags & FOLL_WRITE))
 861                 update_mmu_cache_pmd(vma, addr, pmd);
 862 }
 863
 864 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 865                 pmd_t *pmd, int flags)
 866 {
 867         unsigned long pfn = pmd_pfn(*pmd);
 868         struct mm_struct *mm = vma->vm_mm;
 869         struct dev_pagemap *pgmap;
 870         struct page *page;
 871
 872         assert_spin_locked(pmd_lockptr(mm, pmd));
 873
 874         /*
 875          * When we COW a devmap PMD entry, we split it into PTEs, so we should
 876          * not be in this function with `flags & FOLL_COW` set.
 877          */
 878         WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
 879
 880         if (flags & FOLL_WRITE && !pmd_write(*pmd))
 881                 return NULL;
 882
 883         if (pmd_present(*pmd) && pmd_devmap(*pmd))
 884                 /* pass */;
 885         else
 886                 return NULL;
 887
 888         if (flags & FOLL_TOUCH)
 889                 touch_pmd(vma, addr, pmd, flags);
 890
 891         /*
 892          * device mapped pages can only be returned if the
 893          * caller will manage the page reference count.
 894          */
 895         if (!(flags & FOLL_GET))
 896                 return ERR_PTR(-EEXIST);
 897
 898         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
 899         pgmap = get_dev_pagemap(pfn, NULL);
 900         if (!pgmap)
 901                 return ERR_PTR(-EFAULT);
 902         page = pfn_to_page(pfn);
 903         get_page(page);
 904         put_dev_pagemap(pgmap);
 905
 906         return page;
 907 }
 908
 909 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 910                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 911                   struct vm_area_struct *vma)
 912 {
 913         spinlock_t *dst_ptl, *src_ptl;
 914         struct page *src_page;
 915         pmd_t pmd;
 916         pgtable_t pgtable = NULL;
 917         int ret = -ENOMEM;
 918
 919         /* Skip if can be re-fill on fault */
 920         if (!vma_is_anonymous(vma))
 921                 return 0;
 922
 923         pgtable = pte_alloc_one(dst_mm, addr);
 924         if (unlikely(!pgtable))
 925                 goto out;
 926
 927         dst_ptl = pmd_lock(dst_mm, dst_pmd);
 928         src_ptl = pmd_lockptr(src_mm, src_pmd);
 929         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 930
 931         ret = -EAGAIN;
 932         pmd = *src_pmd;
 933
 934 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
 935         if (unlikely(is_swap_pmd(pmd))) {
 936                 swp_entry_t entry = pmd_to_swp_entry(pmd);
 937
 938                 VM_BUG_ON(!is_pmd_migration_entry(pmd));
 939                 if (is_write_migration_entry(entry)) {
 940                         make_migration_entry_read(&entry);
 941                         pmd = swp_entry_to_pmd(entry);
 942                         if (pmd_swp_soft_dirty(*src_pmd))
 943                                 pmd = pmd_swp_mksoft_dirty(pmd);
 944                         set_pmd_at(src_mm, addr, src_pmd, pmd);
 945                 }
 946                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 947                 mm_inc_nr_ptes(dst_mm);
 948                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
 949                 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 950                 ret = 0;
 951                 goto out_unlock;
 952         }
 953 #endif
 954
 955         if (unlikely(!pmd_trans_huge(pmd))) {
 956                 pte_free(dst_mm, pgtable);
 957                 goto out_unlock;
 958         }
 959         /*
 960          * When page table lock is held, the huge zero pmd should not be
 961          * under splitting since we don't split the page itself, only pmd to
 962          * a page table.
 963          */
 964         if (is_huge_zero_pmd(pmd)) {
 965                 struct page *zero_page;
 966                 /*
 967                  * get_huge_zero_page() will never allocate a new page here,
 968                  * since we already have a zero page to copy. It just takes a
 969                  * reference.
 970                  */
 971                 zero_page = mm_get_huge_zero_page(dst_mm);
 972                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
 973                                 zero_page);
 974                 ret = 0;
 975                 goto out_unlock;
 976         }
 977
 978         src_page = pmd_page(pmd);
 979         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
 980         get_page(src_page);
 981         page_dup_rmap(src_page, true);
 982         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 983         mm_inc_nr_ptes(dst_mm);
 984         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
 985
 986         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 987         pmd = pmd_mkold(pmd_wrprotect(pmd));
 988         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 989
 990         ret = 0;
 991 out_unlock:
 992         spin_unlock(src_ptl);
 993         spin_unlock(dst_ptl);
 994 out:
 995         return ret;
 996 }
 997
 998 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 999 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1000                 pud_t *pud, int flags)
1001 {
1002         pud_t _pud;
1003
1004         _pud = pud_mkyoung(*pud);
1005         if (flags & FOLL_WRITE)
1006                 _pud = pud_mkdirty(_pud);
1007         if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1008                                 pud, _pud, flags & FOLL_WRITE))
1009                 update_mmu_cache_pud(vma, addr, pud);
1010 }
1011
1012 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1013                 pud_t *pud, int flags)
1014 {
1015         unsigned long pfn = pud_pfn(*pud);
1016         struct mm_struct *mm = vma->vm_mm;
1017         struct dev_pagemap *pgmap;
1018         struct page *page;
1019
1020         assert_spin_locked(pud_lockptr(mm, pud));
1021
1022         if (!pud_access_permitted(*pud, flags & FOLL_WRITE))
1023                 return NULL;
1024
1025         if (pud_present(*pud) && pud_devmap(*pud))
1026                 /* pass */;
1027         else
1028                 return NULL;
1029
1030         if (flags & FOLL_TOUCH)
1031                 touch_pud(vma, addr, pud, flags);
1032
1033         /*
1034          * device mapped pages can only be returned if the
1035          * caller will manage the page reference count.
1036          */
1037         if (!(flags & FOLL_GET))
1038                 return ERR_PTR(-EEXIST);
1039
1040         pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1041         pgmap = get_dev_pagemap(pfn, NULL);
1042         if (!pgmap)
1043                 return ERR_PTR(-EFAULT);
1044         page = pfn_to_page(pfn);
1045         get_page(page);
1046         put_dev_pagemap(pgmap);
1047
1048         return page;
1049 }
1050
1051 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1052                   pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1053                   struct vm_area_struct *vma)
1054 {
1055         spinlock_t *dst_ptl, *src_ptl;
1056         pud_t pud;
1057         int ret;
1058
1059         dst_ptl = pud_lock(dst_mm, dst_pud);
1060         src_ptl = pud_lockptr(src_mm, src_pud);
1061         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1062
1063         ret = -EAGAIN;
1064         pud = *src_pud;
1065         if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1066                 goto out_unlock;
1067
1068         /*
1069          * When page table lock is held, the huge zero pud should not be
1070          * under splitting since we don't split the page itself, only pud to
1071          * a page table.
1072          */
1073         if (is_huge_zero_pud(pud)) {
1074                 /* No huge zero pud yet */
1075         }
1076
1077         pudp_set_wrprotect(src_mm, addr, src_pud);
1078         pud = pud_mkold(pud_wrprotect(pud));
1079         set_pud_at(dst_mm, addr, dst_pud, pud);
1080
1081         ret = 0;
1082 out_unlock:
1083         spin_unlock(src_ptl);
1084         spin_unlock(dst_ptl);
1085         return ret;
1086 }
1087
1088 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1089 {
1090         pud_t entry;
1091         unsigned long haddr;
1092         bool write = vmf->flags & FAULT_FLAG_WRITE;
1093
1094         vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1095         if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1096                 goto unlock;
1097
1098         entry = pud_mkyoung(orig_pud);
1099         if (write)
1100                 entry = pud_mkdirty(entry);
1101         haddr = vmf->address & HPAGE_PUD_MASK;
1102         if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1103                 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1104
1105 unlock:
1106         spin_unlock(vmf->ptl);
1107 }
1108 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1109
1110 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1111 {
1112         pmd_t entry;
1113         unsigned long haddr;
1114         bool write = vmf->flags & FAULT_FLAG_WRITE;
1115
1116         vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1117         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1118                 goto unlock;
1119
1120         entry = pmd_mkyoung(orig_pmd);
1121         if (write)
1122                 entry = pmd_mkdirty(entry);
1123         haddr = vmf->address & HPAGE_PMD_MASK;
1124         if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1125                 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1126
1127 unlock:
1128         spin_unlock(vmf->ptl);
1129 }
1130
1131 static int do_huge_pmd_wp_page_fallback(struct vm_fault *vmf, pmd_t orig_pmd,
1132                 struct page *page)
1133 {
1134         struct vm_area_struct *vma = vmf->vma;
1135         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1136         struct mem_cgroup *memcg;
1137         pgtable_t pgtable;
1138         pmd_t _pmd;
1139         int ret = 0, i;
1140         struct page **pages;
1141         unsigned long mmun_start;       /* For mmu_notifiers */
1142         unsigned long mmun_end;         /* For mmu_notifiers */
1143
1144         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1145                         GFP_KERNEL);
1146         if (unlikely(!pages)) {
1147                 ret |= VM_FAULT_OOM;
1148                 goto out;
1149         }
1150
1151         for (i = 0; i < HPAGE_PMD_NR; i++) {
1152                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
1153                                                vmf->address, page_to_nid(page));
1154                 if (unlikely(!pages[i] ||
1155                              mem_cgroup_try_charge(pages[i], vma->vm_mm,
1156                                      GFP_KERNEL, &memcg, false))) {
1157                         if (pages[i])
1158                                 put_page(pages[i]);
1159                         while (--i >= 0) {
1160                                 memcg = (void *)page_private(pages[i]);
1161                                 set_page_private(pages[i], 0);
1162                                 mem_cgroup_cancel_charge(pages[i], memcg,
1163                                                 false);
1164                                 put_page(pages[i]);
1165                         }
1166                         kfree(pages);
1167                         ret |= VM_FAULT_OOM;
1168                         goto out;
1169                 }
1170                 set_page_private(pages[i], (unsigned long)memcg);
1171         }
1172
1173         for (i = 0; i < HPAGE_PMD_NR; i++) {
1174                 copy_user_highpage(pages[i], page + i,
1175                                    haddr + PAGE_SIZE * i, vma);
1176                 __SetPageUptodate(pages[i]);
1177                 cond_resched();
1178         }
1179
1180         mmun_start = haddr;
1181         mmun_end   = haddr + HPAGE_PMD_SIZE;
1182         mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
1183
1184         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1185         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1186                 goto out_free_pages;
1187         VM_BUG_ON_PAGE(!PageHead(page), page);
1188
1189         /*
1190          * Leave pmd empty until pte is filled note we must notify here as
1191          * concurrent CPU thread might write to new page before the call to
1192          * mmu_notifier_invalidate_range_end() happens which can lead to a
1193          * device seeing memory write in different order than CPU.
1194          *
1195          * See Documentation/vm/mmu_notifier.txt
1196          */
1197         pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1198
1199         pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
1200         pmd_populate(vma->vm_mm, &_pmd, pgtable);
1201
1202         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1203                 pte_t entry;
1204                 entry = mk_pte(pages[i], vma->vm_page_prot);
1205                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1206                 memcg = (void *)page_private(pages[i]);
1207                 set_page_private(pages[i], 0);
1208                 page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
1209                 mem_cgroup_commit_charge(pages[i], memcg, false, false);
1210                 lru_cache_add_active_or_unevictable(pages[i], vma);
1211                 vmf->pte = pte_offset_map(&_pmd, haddr);
1212                 VM_BUG_ON(!pte_none(*vmf->pte));
1213                 set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
1214                 pte_unmap(vmf->pte);
1215         }
1216         kfree(pages);
1217
1218         smp_wmb(); /* make pte visible before pmd */
1219         pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
1220         page_remove_rmap(page, true);
1221         spin_unlock(vmf->ptl);
1222
1223         /*
1224          * No need to double call mmu_notifier->invalidate_range() callback as
1225          * the above pmdp_huge_clear_flush_notify() did already call it.
1226          */
1227         mmu_notifier_invalidate_range_only_end(vma->vm_mm, mmun_start,
1228                                                 mmun_end);
1229
1230         ret |= VM_FAULT_WRITE;
1231         put_page(page);
1232
1233 out:
1234         return ret;
1235
1236 out_free_pages:
1237         spin_unlock(vmf->ptl);
1238         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1239         for (i = 0; i < HPAGE_PMD_NR; i++) {
1240                 memcg = (void *)page_private(pages[i]);
1241                 set_page_private(pages[i], 0);
1242                 mem_cgroup_cancel_charge(pages[i], memcg, false);
1243                 put_page(pages[i]);
1244         }
1245         kfree(pages);
1246         goto out;
1247 }
1248
1249 int do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1250 {
1251         struct vm_area_struct *vma = vmf->vma;
1252         struct page *page = NULL, *new_page;
1253         struct mem_cgroup *memcg;
1254         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1255         unsigned long mmun_start;       /* For mmu_notifiers */
1256         unsigned long mmun_end;         /* For mmu_notifiers */
1257         gfp_t huge_gfp;                 /* for allocation and charge */
1258         int ret = 0;
1259
1260         vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1261         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1262         if (is_huge_zero_pmd(orig_pmd))
1263                 goto alloc;
1264         spin_lock(vmf->ptl);
1265         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1266                 goto out_unlock;
1267
1268         page = pmd_page(orig_pmd);
1269         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1270         /*
1271          * We can only reuse the page if nobody else maps the huge page or it's
1272          * part.
1273          */
1274         if (!trylock_page(page)) {
1275                 get_page(page);
1276                 spin_unlock(vmf->ptl);
1277                 lock_page(page);
1278                 spin_lock(vmf->ptl);
1279                 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1280                         unlock_page(page);
1281                         put_page(page);
1282                         goto out_unlock;
1283                 }
1284                 put_page(page);
1285         }
1286         if (reuse_swap_page(page, NULL)) {
1287                 pmd_t entry;
1288                 entry = pmd_mkyoung(orig_pmd);
1289                 entry = maybe_pmd_mkwrite(entry, vma, true);
1290                 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry,  1))
1291                         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1292                 ret |= VM_FAULT_WRITE;
1293                 unlock_page(page);
1294                 goto out_unlock;
1295         }
1296         unlock_page(page);
1297         get_page(page);
1298         spin_unlock(vmf->ptl);
1299 alloc:
1300         if (transparent_hugepage_enabled(vma) &&
1301             !transparent_hugepage_debug_cow()) {
1302                 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1303                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1304         } else
1305                 new_page = NULL;
1306
1307         if (likely(new_page)) {
1308                 prep_transhuge_page(new_page);
1309         } else {
1310                 if (!page) {
1311                         split_huge_pmd(vma, vmf->pmd, vmf->address);
1312                         ret |= VM_FAULT_FALLBACK;
1313                 } else {
1314                         ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
1315                         if (ret & VM_FAULT_OOM) {
1316                                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1317                                 ret |= VM_FAULT_FALLBACK;
1318                         }
1319                         put_page(page);
1320                 }
1321                 count_vm_event(THP_FAULT_FALLBACK);
1322                 goto out;
1323         }
1324
1325         if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm,
1326                                         huge_gfp, &memcg, true))) {
1327                 put_page(new_page);
1328                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1329                 if (page)
1330                         put_page(page);
1331                 ret |= VM_FAULT_FALLBACK;
1332                 count_vm_event(THP_FAULT_FALLBACK);
1333                 goto out;
1334         }
1335
1336         count_vm_event(THP_FAULT_ALLOC);
1337
1338         if (!page)
1339                 clear_huge_page(new_page, vmf->address, HPAGE_PMD_NR);
1340         else
1341                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1342         __SetPageUptodate(new_page);
1343
1344         mmun_start = haddr;
1345         mmun_end   = haddr + HPAGE_PMD_SIZE;
1346         mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
1347
1348         spin_lock(vmf->ptl);
1349         if (page)
1350                 put_page(page);
1351         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1352                 spin_unlock(vmf->ptl);
1353                 mem_cgroup_cancel_charge(new_page, memcg, true);
1354                 put_page(new_page);
1355                 goto out_mn;
1356         } else {
1357                 pmd_t entry;
1358                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1359                 entry = maybe_pmd_mkwrite(entry, vma, true);
1360                 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1361                 page_add_new_anon_rmap(new_page, vma, haddr, true);
1362                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1363                 lru_cache_add_active_or_unevictable(new_page, vma);
1364                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1365                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1366                 if (!page) {
1367                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1368                 } else {
1369                         VM_BUG_ON_PAGE(!PageHead(page), page);
1370                         page_remove_rmap(page, true);
1371                         put_page(page);
1372                 }
1373                 ret |= VM_FAULT_WRITE;
1374         }
1375         spin_unlock(vmf->ptl);
1376 out_mn:
1377         /*
1378          * No need to double call mmu_notifier->invalidate_range() callback as
1379          * the above pmdp_huge_clear_flush_notify() did already call it.
1380          */
1381         mmu_notifier_invalidate_range_only_end(vma->vm_mm, mmun_start,
1382                                                mmun_end);
1383 out:
1384         return ret;
1385 out_unlock:
1386         spin_unlock(vmf->ptl);
1387         return ret;
1388 }
1389
1390 /*
1391  * FOLL_FORCE can write to even unwritable pmd's, but only
1392  * after we've gone through a COW cycle and they are dirty.
1393  */
1394 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1395 {
1396         return pmd_write(pmd) ||
1397                ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1398 }
1399
1400 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1401                                    unsigned long addr,
1402                                    pmd_t *pmd,
1403                                    unsigned int flags)
1404 {
1405         struct mm_struct *mm = vma->vm_mm;
1406         struct page *page = NULL;
1407
1408         assert_spin_locked(pmd_lockptr(mm, pmd));
1409
1410         if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1411                 goto out;
1412
1413         /* Avoid dumping huge zero page */
1414         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1415                 return ERR_PTR(-EFAULT);
1416
1417         /* Full NUMA hinting faults to serialise migration in fault paths */
1418         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1419                 goto out;
1420
1421         page = pmd_page(*pmd);
1422         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1423         if (flags & FOLL_TOUCH)
1424                 touch_pmd(vma, addr, pmd, flags);
1425         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1426                 /*
1427                  * We don't mlock() pte-mapped THPs. This way we can avoid
1428                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1429                  *
1430                  * For anon THP:
1431                  *
1432                  * In most cases the pmd is the only mapping of the page as we
1433                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1434                  * writable private mappings in populate_vma_page_range().
1435                  *
1436                  * The only scenario when we have the page shared here is if we
1437                  * mlocking read-only mapping shared over fork(). We skip
1438                  * mlocking such pages.
1439                  *
1440                  * For file THP:
1441                  *
1442                  * We can expect PageDoubleMap() to be stable under page lock:
1443                  * for file pages we set it in page_add_file_rmap(), which
1444                  * requires page to be locked.
1445                  */
1446
1447                 if (PageAnon(page) && compound_mapcount(page) != 1)
1448                         goto skip_mlock;
1449                 if (PageDoubleMap(page) || !page->mapping)
1450                         goto skip_mlock;
1451                 if (!trylock_page(page))
1452                         goto skip_mlock;
1453                 lru_add_drain();
1454                 if (page->mapping && !PageDoubleMap(page))
1455                         mlock_vma_page(page);
1456                 unlock_page(page);
1457         }
1458 skip_mlock:
1459         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1460         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1461         if (flags & FOLL_GET)
1462                 get_page(page);
1463
1464 out:
1465         return page;
1466 }
1467
1468 /* NUMA hinting page fault entry point for trans huge pmds */
1469 int do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1470 {
1471         struct vm_area_struct *vma = vmf->vma;
1472         struct anon_vma *anon_vma = NULL;
1473         struct page *page;
1474         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1475         int page_nid = -1, this_nid = numa_node_id();
1476         int target_nid, last_cpupid = -1;
1477         bool page_locked;
1478         bool migrated = false;
1479         bool was_writable;
1480         int flags = 0;
1481
1482         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1483         if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1484                 goto out_unlock;
1485
1486         /*
1487          * If there are potential migrations, wait for completion and retry
1488          * without disrupting NUMA hinting information. Do not relock and
1489          * check_same as the page may no longer be mapped.
1490          */
1491         if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1492                 page = pmd_page(*vmf->pmd);
1493                 if (!get_page_unless_zero(page))
1494                         goto out_unlock;
1495                 spin_unlock(vmf->ptl);
1496                 wait_on_page_locked(page);
1497                 put_page(page);
1498                 goto out;
1499         }
1500
1501         page = pmd_page(pmd);
1502         BUG_ON(is_huge_zero_page(page));
1503         page_nid = page_to_nid(page);
1504         last_cpupid = page_cpupid_last(page);
1505         count_vm_numa_event(NUMA_HINT_FAULTS);
1506         if (page_nid == this_nid) {
1507                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1508                 flags |= TNF_FAULT_LOCAL;
1509         }
1510
1511         /* See similar comment in do_numa_page for explanation */
1512         if (!pmd_savedwrite(pmd))
1513                 flags |= TNF_NO_GROUP;
1514
1515         /*
1516          * Acquire the page lock to serialise THP migrations but avoid dropping
1517          * page_table_lock if at all possible
1518          */
1519         page_locked = trylock_page(page);
1520         target_nid = mpol_misplaced(page, vma, haddr);
1521         if (target_nid == -1) {
1522                 /* If the page was locked, there are no parallel migrations */
1523                 if (page_locked)
1524                         goto clear_pmdnuma;
1525         }
1526
1527         /* Migration could have started since the pmd_trans_migrating check */
1528         if (!page_locked) {
1529                 page_nid = -1;
1530                 if (!get_page_unless_zero(page))
1531                         goto out_unlock;
1532                 spin_unlock(vmf->ptl);
1533                 wait_on_page_locked(page);
1534                 put_page(page);
1535                 goto out;
1536         }
1537
1538         /*
1539          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1540          * to serialises splits
1541          */
1542         get_page(page);
1543         spin_unlock(vmf->ptl);
1544         anon_vma = page_lock_anon_vma_read(page);
1545
1546         /* Confirm the PMD did not change while page_table_lock was released */
1547         spin_lock(vmf->ptl);
1548         if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1549                 unlock_page(page);
1550                 put_page(page);
1551                 page_nid = -1;
1552                 goto out_unlock;
1553         }
1554
1555         /* Bail if we fail to protect against THP splits for any reason */
1556         if (unlikely(!anon_vma)) {
1557                 put_page(page);
1558                 page_nid = -1;
1559                 goto clear_pmdnuma;
1560         }
1561
1562         /*
1563          * Since we took the NUMA fault, we must have observed the !accessible
1564          * bit. Make sure all other CPUs agree with that, to avoid them
1565          * modifying the page we're about to migrate.
1566          *
1567          * Must be done under PTL such that we'll observe the relevant
1568          * inc_tlb_flush_pending().
1569          *
1570          * We are not sure a pending tlb flush here is for a huge page
1571          * mapping or not. Hence use the tlb range variant
1572          */
1573         if (mm_tlb_flush_pending(vma->vm_mm))
1574                 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1575
1576         /*
1577          * Migrate the THP to the requested node, returns with page unlocked
1578          * and access rights restored.
1579          */
1580         spin_unlock(vmf->ptl);
1581
1582         migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1583                                 vmf->pmd, pmd, vmf->address, page, target_nid);
1584         if (migrated) {
1585                 flags |= TNF_MIGRATED;
1586                 page_nid = target_nid;
1587         } else
1588                 flags |= TNF_MIGRATE_FAIL;
1589
1590         goto out;
1591 clear_pmdnuma:
1592         BUG_ON(!PageLocked(page));
1593         was_writable = pmd_savedwrite(pmd);
1594         pmd = pmd_modify(pmd, vma->vm_page_prot);
1595         pmd = pmd_mkyoung(pmd);
1596         if (was_writable)
1597                 pmd = pmd_mkwrite(pmd);
1598         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1599         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1600         unlock_page(page);
1601 out_unlock:
1602         spin_unlock(vmf->ptl);
1603
1604 out:
1605         if (anon_vma)
1606                 page_unlock_anon_vma_read(anon_vma);
1607
1608         if (page_nid != -1)
1609                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1610                                 flags);
1611
1612         return 0;
1613 }
1614
1615 /*
1616  * Return true if we do MADV_FREE successfully on entire pmd page.
1617  * Otherwise, return false.
1618  */
1619 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1620                 pmd_t *pmd, unsigned long addr, unsigned long next)
1621 {
1622         spinlock_t *ptl;
1623         pmd_t orig_pmd;
1624         struct page *page;
1625         struct mm_struct *mm = tlb->mm;
1626         bool ret = false;
1627
1628         tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1629
1630         ptl = pmd_trans_huge_lock(pmd, vma);
1631         if (!ptl)
1632                 goto out_unlocked;
1633
1634         orig_pmd = *pmd;
1635         if (is_huge_zero_pmd(orig_pmd))
1636                 goto out;
1637
1638         if (unlikely(!pmd_present(orig_pmd))) {
1639                 VM_BUG_ON(thp_migration_supported() &&
1640                                   !is_pmd_migration_entry(orig_pmd));
1641                 goto out;
1642         }
1643
1644         page = pmd_page(orig_pmd);
1645         /*
1646          * If other processes are mapping this page, we couldn't discard
1647          * the page unless they all do MADV_FREE so let's skip the page.
1648          */
1649         if (page_mapcount(page) != 1)
1650                 goto out;
1651
1652         if (!trylock_page(page))
1653                 goto out;
1654
1655         /*
1656          * If user want to discard part-pages of THP, split it so MADV_FREE
1657          * will deactivate only them.
1658          */
1659         if (next - addr != HPAGE_PMD_SIZE) {
1660                 get_page(page);
1661                 spin_unlock(ptl);
1662                 split_huge_page(page);
1663                 unlock_page(page);
1664                 put_page(page);
1665                 goto out_unlocked;
1666         }
1667
1668         if (PageDirty(page))
1669                 ClearPageDirty(page);
1670         unlock_page(page);
1671
1672         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1673                 pmdp_invalidate(vma, addr, pmd);
1674                 orig_pmd = pmd_mkold(orig_pmd);
1675                 orig_pmd = pmd_mkclean(orig_pmd);
1676
1677                 set_pmd_at(mm, addr, pmd, orig_pmd);
1678                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1679         }
1680
1681         mark_page_lazyfree(page);
1682         ret = true;
1683 out:
1684         spin_unlock(ptl);
1685 out_unlocked:
1686         return ret;
1687 }
1688
1689 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1690 {
1691         pgtable_t pgtable;
1692
1693         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1694         pte_free(mm, pgtable);
1695         mm_dec_nr_ptes(mm);
1696 }
1697
1698 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1699                  pmd_t *pmd, unsigned long addr)
1700 {
1701         pmd_t orig_pmd;
1702         spinlock_t *ptl;
1703
1704         tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1705
1706         ptl = __pmd_trans_huge_lock(pmd, vma);
1707         if (!ptl)
1708                 return 0;
1709         /*
1710          * For architectures like ppc64 we look at deposited pgtable
1711          * when calling pmdp_huge_get_and_clear. So do the
1712          * pgtable_trans_huge_withdraw after finishing pmdp related
1713          * operations.
1714          */
1715         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1716                         tlb->fullmm);
1717         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1718         if (vma_is_dax(vma)) {
1719                 if (arch_needs_pgtable_deposit())
1720                         zap_deposited_table(tlb->mm, pmd);
1721                 spin_unlock(ptl);
1722                 if (is_huge_zero_pmd(orig_pmd))
1723                         tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1724         } else if (is_huge_zero_pmd(orig_pmd)) {
1725                 zap_deposited_table(tlb->mm, pmd);
1726                 spin_unlock(ptl);
1727                 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1728         } else {
1729                 struct page *page = NULL;
1730                 int flush_needed = 1;
1731
1732                 if (pmd_present(orig_pmd)) {
1733                         page = pmd_page(orig_pmd);
1734                         page_remove_rmap(page, true);
1735                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1736                         VM_BUG_ON_PAGE(!PageHead(page), page);
1737                 } else if (thp_migration_supported()) {
1738                         swp_entry_t entry;
1739
1740                         VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1741                         entry = pmd_to_swp_entry(orig_pmd);
1742                         page = pfn_to_page(swp_offset(entry));
1743                         flush_needed = 0;
1744                 } else
1745                         WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1746
1747                 if (PageAnon(page)) {
1748                         zap_deposited_table(tlb->mm, pmd);
1749                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1750                 } else {
1751                         if (arch_needs_pgtable_deposit())
1752                                 zap_deposited_table(tlb->mm, pmd);
1753                         add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1754                 }
1755
1756                 spin_unlock(ptl);
1757                 if (flush_needed)
1758                         tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1759         }
1760         return 1;
1761 }
1762
1763 #ifndef pmd_move_must_withdraw
1764 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1765                                          spinlock_t *old_pmd_ptl,
1766                                          struct vm_area_struct *vma)
1767 {
1768         /*
1769          * With split pmd lock we also need to move preallocated
1770          * PTE page table if new_pmd is on different PMD page table.
1771          *
1772          * We also don't deposit and withdraw tables for file pages.
1773          */
1774         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1775 }
1776 #endif
1777
1778 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1779 {
1780 #ifdef CONFIG_MEM_SOFT_DIRTY
1781         if (unlikely(is_pmd_migration_entry(pmd)))
1782                 pmd = pmd_swp_mksoft_dirty(pmd);
1783         else if (pmd_present(pmd))
1784                 pmd = pmd_mksoft_dirty(pmd);
1785 #endif
1786         return pmd;
1787 }
1788
1789 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1790                   unsigned long new_addr, unsigned long old_end,
1791                   pmd_t *old_pmd, pmd_t *new_pmd, bool *need_flush)
1792 {
1793         spinlock_t *old_ptl, *new_ptl;
1794         pmd_t pmd;
1795         struct mm_struct *mm = vma->vm_mm;
1796         bool force_flush = false;
1797
1798         if ((old_addr & ~HPAGE_PMD_MASK) ||
1799             (new_addr & ~HPAGE_PMD_MASK) ||
1800             old_end - old_addr < HPAGE_PMD_SIZE)
1801                 return false;
1802
1803         /*
1804          * The destination pmd shouldn't be established, free_pgtables()
1805          * should have release it.
1806          */
1807         if (WARN_ON(!pmd_none(*new_pmd))) {
1808                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1809                 return false;
1810         }
1811
1812         /*
1813          * We don't have to worry about the ordering of src and dst
1814          * ptlocks because exclusive mmap_sem prevents deadlock.
1815          */
1816         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1817         if (old_ptl) {
1818                 new_ptl = pmd_lockptr(mm, new_pmd);
1819                 if (new_ptl != old_ptl)
1820                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1821                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1822                 if (pmd_present(pmd) && pmd_dirty(pmd))
1823                         force_flush = true;
1824                 VM_BUG_ON(!pmd_none(*new_pmd));
1825
1826                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1827                         pgtable_t pgtable;
1828                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1829                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1830                 }
1831                 pmd = move_soft_dirty_pmd(pmd);
1832                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1833                 if (new_ptl != old_ptl)
1834                         spin_unlock(new_ptl);
1835                 if (force_flush)
1836                         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1837                 else
1838                         *need_flush = true;
1839                 spin_unlock(old_ptl);
1840                 return true;
1841         }
1842         return false;
1843 }
1844
1845 /*
1846  * Returns
1847  *  - 0 if PMD could not be locked
1848  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1849  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1850  */
1851 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1852                 unsigned long addr, pgprot_t newprot, int prot_numa)
1853 {
1854         struct mm_struct *mm = vma->vm_mm;
1855         spinlock_t *ptl;
1856         pmd_t entry;
1857         bool preserve_write;
1858         int ret;
1859
1860         ptl = __pmd_trans_huge_lock(pmd, vma);
1861         if (!ptl)
1862                 return 0;
1863
1864         preserve_write = prot_numa && pmd_write(*pmd);
1865         ret = 1;
1866
1867 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1868         if (is_swap_pmd(*pmd)) {
1869                 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1870
1871                 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1872                 if (is_write_migration_entry(entry)) {
1873                         pmd_t newpmd;
1874                         /*
1875                          * A protection check is difficult so
1876                          * just be safe and disable write
1877                          */
1878                         make_migration_entry_read(&entry);
1879                         newpmd = swp_entry_to_pmd(entry);
1880                         if (pmd_swp_soft_dirty(*pmd))
1881                                 newpmd = pmd_swp_mksoft_dirty(newpmd);
1882                         set_pmd_at(mm, addr, pmd, newpmd);
1883                 }
1884                 goto unlock;
1885         }
1886 #endif
1887
1888         /*
1889          * Avoid trapping faults against the zero page. The read-only
1890          * data is likely to be read-cached on the local CPU and
1891          * local/remote hits to the zero page are not interesting.
1892          */
1893         if (prot_numa && is_huge_zero_pmd(*pmd))
1894                 goto unlock;
1895
1896         if (prot_numa && pmd_protnone(*pmd))
1897                 goto unlock;
1898
1899         /*
1900          * In case prot_numa, we are under down_read(mmap_sem). It's critical
1901          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1902          * which is also under down_read(mmap_sem):
1903          *
1904          *      CPU0:                           CPU1:
1905          *                              change_huge_pmd(prot_numa=1)
1906          *                               pmdp_huge_get_and_clear_notify()
1907          * madvise_dontneed()
1908          *  zap_pmd_range()
1909          *   pmd_trans_huge(*pmd) == 0 (without ptl)
1910          *   // skip the pmd
1911          *                               set_pmd_at();
1912          *                               // pmd is re-established
1913          *
1914          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1915          * which may break userspace.
1916          *
1917          * pmdp_invalidate() is required to make sure we don't miss
1918          * dirty/young flags set by hardware.
1919          */
1920         entry = *pmd;
1921         pmdp_invalidate(vma, addr, pmd);
1922
1923         /*
1924          * Recover dirty/young flags.  It relies on pmdp_invalidate to not
1925          * corrupt them.
1926          */
1927         if (pmd_dirty(*pmd))
1928                 entry = pmd_mkdirty(entry);
1929         if (pmd_young(*pmd))
1930                 entry = pmd_mkyoung(entry);
1931
1932         entry = pmd_modify(entry, newprot);
1933         if (preserve_write)
1934                 entry = pmd_mk_savedwrite(entry);
1935         ret = HPAGE_PMD_NR;
1936         set_pmd_at(mm, addr, pmd, entry);
1937         BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1938 unlock:
1939         spin_unlock(ptl);
1940         return ret;
1941 }
1942
1943 /*
1944  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1945  *
1946  * Note that if it returns page table lock pointer, this routine returns without
1947  * unlocking page table lock. So callers must unlock it.
1948  */
1949 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1950 {
1951         spinlock_t *ptl;
1952         ptl = pmd_lock(vma->vm_mm, pmd);
1953         if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1954                         pmd_devmap(*pmd)))
1955                 return ptl;
1956         spin_unlock(ptl);
1957         return NULL;
1958 }
1959
1960 /*
1961  * Returns true if a given pud maps a thp, false otherwise.
1962  *
1963  * Note that if it returns true, this routine returns without unlocking page
1964  * table lock. So callers must unlock it.
1965  */
1966 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1967 {
1968         spinlock_t *ptl;
1969
1970         ptl = pud_lock(vma->vm_mm, pud);
1971         if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1972                 return ptl;
1973         spin_unlock(ptl);
1974         return NULL;
1975 }
1976
1977 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1978 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1979                  pud_t *pud, unsigned long addr)
1980 {
1981         pud_t orig_pud;
1982         spinlock_t *ptl;
1983
1984         ptl = __pud_trans_huge_lock(pud, vma);
1985         if (!ptl)
1986                 return 0;
1987         /*
1988          * For architectures like ppc64 we look at deposited pgtable
1989          * when calling pudp_huge_get_and_clear. So do the
1990          * pgtable_trans_huge_withdraw after finishing pudp related
1991          * operations.
1992          */
1993         orig_pud = pudp_huge_get_and_clear_full(tlb->mm, addr, pud,
1994                         tlb->fullmm);
1995         tlb_remove_pud_tlb_entry(tlb, pud, addr);
1996         if (vma_is_dax(vma)) {
1997                 spin_unlock(ptl);
1998                 /* No zero page support yet */
1999         } else {
2000                 /* No support for anonymous PUD pages yet */
2001                 BUG();
2002         }
2003         return 1;
2004 }
2005
2006 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
2007                 unsigned long haddr)
2008 {
2009         VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
2010         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2011         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
2012         VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
2013
2014         count_vm_event(THP_SPLIT_PUD);
2015
2016         pudp_huge_clear_flush_notify(vma, haddr, pud);
2017 }
2018
2019 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2020                 unsigned long address)
2021 {
2022         spinlock_t *ptl;
2023         struct mm_struct *mm = vma->vm_mm;
2024         unsigned long haddr = address & HPAGE_PUD_MASK;
2025
2026         mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PUD_SIZE);
2027         ptl = pud_lock(mm, pud);
2028         if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2029                 goto out;
2030         __split_huge_pud_locked(vma, pud, haddr);
2031
2032 out:
2033         spin_unlock(ptl);
2034         /*
2035          * No need to double call mmu_notifier->invalidate_range() callback as
2036          * the above pudp_huge_clear_flush_notify() did already call it.
2037          */
2038         mmu_notifier_invalidate_range_only_end(mm, haddr, haddr +
2039                                                HPAGE_PUD_SIZE);
2040 }
2041 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2042
2043 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2044                 unsigned long haddr, pmd_t *pmd)
2045 {
2046         struct mm_struct *mm = vma->vm_mm;
2047         pgtable_t pgtable;
2048         pmd_t _pmd;
2049         int i;
2050
2051         /*
2052          * Leave pmd empty until pte is filled note that it is fine to delay
2053          * notification until mmu_notifier_invalidate_range_end() as we are
2054          * replacing a zero pmd write protected page with a zero pte write
2055          * protected page.
2056          *
2057          * See Documentation/vm/mmu_notifier.txt
2058          */
2059         pmdp_huge_clear_flush(vma, haddr, pmd);
2060
2061         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2062         pmd_populate(mm, &_pmd, pgtable);
2063
2064         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2065                 pte_t *pte, entry;
2066                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2067                 entry = pte_mkspecial(entry);
2068                 pte = pte_offset_map(&_pmd, haddr);
2069                 VM_BUG_ON(!pte_none(*pte));
2070                 set_pte_at(mm, haddr, pte, entry);
2071                 pte_unmap(pte);
2072         }
2073         smp_wmb(); /* make pte visible before pmd */
2074         pmd_populate(mm, pmd, pgtable);
2075 }
2076
2077 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2078                 unsigned long haddr, bool freeze)
2079 {
2080         struct mm_struct *mm = vma->vm_mm;
2081         struct page *page;
2082         pgtable_t pgtable;
2083         pmd_t _pmd;
2084         bool young, write, dirty, soft_dirty, pmd_migration = false;
2085         unsigned long addr;
2086         int i;
2087
2088         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2089         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2090         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2091         VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2092                                 && !pmd_devmap(*pmd));
2093
2094         count_vm_event(THP_SPLIT_PMD);
2095
2096         if (!vma_is_anonymous(vma)) {
2097                 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2098                 /*
2099                  * We are going to unmap this huge page. So
2100                  * just go ahead and zap it
2101                  */
2102                 if (arch_needs_pgtable_deposit())
2103                         zap_deposited_table(mm, pmd);
2104                 if (vma_is_dax(vma))
2105                         return;
2106                 page = pmd_page(_pmd);
2107                 if (!PageReferenced(page) && pmd_young(_pmd))
2108                         SetPageReferenced(page);
2109                 page_remove_rmap(page, true);
2110                 put_page(page);
2111                 add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR);
2112                 return;
2113         } else if (is_huge_zero_pmd(*pmd)) {
2114                 /*
2115                  * FIXME: Do we want to invalidate secondary mmu by calling
2116                  * mmu_notifier_invalidate_range() see comments below inside
2117                  * __split_huge_pmd() ?
2118                  *
2119                  * We are going from a zero huge page write protected to zero
2120                  * small page also write protected so it does not seems useful
2121                  * to invalidate secondary mmu at this time.
2122                  */
2123                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2124         }
2125
2126 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2127         pmd_migration = is_pmd_migration_entry(*pmd);
2128         if (pmd_migration) {
2129                 swp_entry_t entry;
2130
2131                 entry = pmd_to_swp_entry(*pmd);
2132                 page = pfn_to_page(swp_offset(entry));
2133         } else
2134 #endif
2135                 page = pmd_page(*pmd);
2136         VM_BUG_ON_PAGE(!page_count(page), page);
2137         page_ref_add(page, HPAGE_PMD_NR - 1);
2138         write = pmd_write(*pmd);
2139         young = pmd_young(*pmd);
2140         dirty = pmd_dirty(*pmd);
2141         soft_dirty = pmd_soft_dirty(*pmd);
2142
2143         pmdp_huge_split_prepare(vma, haddr, pmd);
2144         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2145         pmd_populate(mm, &_pmd, pgtable);
2146
2147         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2148                 pte_t entry, *pte;
2149                 /*
2150                  * Note that NUMA hinting access restrictions are not
2151                  * transferred to avoid any possibility of altering
2152                  * permissions across VMAs.
2153                  */
2154                 if (freeze || pmd_migration) {
2155                         swp_entry_t swp_entry;
2156                         swp_entry = make_migration_entry(page + i, write);
2157                         entry = swp_entry_to_pte(swp_entry);
2158                         if (soft_dirty)
2159                                 entry = pte_swp_mksoft_dirty(entry);
2160                 } else {
2161                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2162                         entry = maybe_mkwrite(entry, vma);
2163                         if (!write)
2164                                 entry = pte_wrprotect(entry);
2165                         if (!young)
2166                                 entry = pte_mkold(entry);
2167                         if (soft_dirty)
2168                                 entry = pte_mksoft_dirty(entry);
2169                 }
2170                 if (dirty)
2171                         SetPageDirty(page + i);
2172                 pte = pte_offset_map(&_pmd, addr);
2173                 BUG_ON(!pte_none(*pte));
2174                 set_pte_at(mm, addr, pte, entry);
2175                 atomic_inc(&page[i]._mapcount);
2176                 pte_unmap(pte);
2177         }
2178
2179         /*
2180          * Set PG_double_map before dropping compound_mapcount to avoid
2181          * false-negative page_mapped().
2182          */
2183         if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2184                 for (i = 0; i < HPAGE_PMD_NR; i++)
2185                         atomic_inc(&page[i]._mapcount);
2186         }
2187
2188         if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2189                 /* Last compound_mapcount is gone. */
2190                 __dec_node_page_state(page, NR_ANON_THPS);
2191                 if (TestClearPageDoubleMap(page)) {
2192                         /* No need in mapcount reference anymore */
2193                         for (i = 0; i < HPAGE_PMD_NR; i++)
2194                                 atomic_dec(&page[i]._mapcount);
2195                 }
2196         }
2197
2198         smp_wmb(); /* make pte visible before pmd */
2199         /*
2200          * Up to this point the pmd is present and huge and userland has the
2201          * whole access to the hugepage during the split (which happens in
2202          * place). If we overwrite the pmd with the not-huge version pointing
2203          * to the pte here (which of course we could if all CPUs were bug
2204          * free), userland could trigger a small page size TLB miss on the
2205          * small sized TLB while the hugepage TLB entry is still established in
2206          * the huge TLB. Some CPU doesn't like that.
2207          * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2208          * 383 on page 93. Intel should be safe but is also warns that it's
2209          * only safe if the permission and cache attributes of the two entries
2210          * loaded in the two TLB is identical (which should be the case here).
2211          * But it is generally safer to never allow small and huge TLB entries
2212          * for the same virtual address to be loaded simultaneously. So instead
2213          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2214          * current pmd notpresent (atomically because here the pmd_trans_huge
2215          * and pmd_trans_splitting must remain set at all times on the pmd
2216          * until the split is complete for this pmd), then we flush the SMP TLB
2217          * and finally we write the non-huge version of the pmd entry with
2218          * pmd_populate.
2219          */
2220         pmdp_invalidate(vma, haddr, pmd);
2221         pmd_populate(mm, pmd, pgtable);
2222
2223         if (freeze) {
2224                 for (i = 0; i < HPAGE_PMD_NR; i++) {
2225                         page_remove_rmap(page + i, false);
2226                         put_page(page + i);
2227                 }
2228         }
2229 }
2230
2231 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2232                 unsigned long address, bool freeze, struct page *page)
2233 {
2234         spinlock_t *ptl;
2235         struct mm_struct *mm = vma->vm_mm;
2236         unsigned long haddr = address & HPAGE_PMD_MASK;
2237
2238         mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
2239         ptl = pmd_lock(mm, pmd);
2240
2241         /*
2242          * If caller asks to setup a migration entries, we need a page to check
2243          * pmd against. Otherwise we can end up replacing wrong page.
2244          */
2245         VM_BUG_ON(freeze && !page);
2246         if (page && page != pmd_page(*pmd))
2247                 goto out;
2248
2249         if (pmd_trans_huge(*pmd)) {
2250                 page = pmd_page(*pmd);
2251                 if (PageMlocked(page))
2252                         clear_page_mlock(page);
2253         } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2254                 goto out;
2255         __split_huge_pmd_locked(vma, pmd, haddr, freeze);
2256 out:
2257         spin_unlock(ptl);
2258         /*
2259          * No need to double call mmu_notifier->invalidate_range() callback.
2260          * They are 3 cases to consider inside __split_huge_pmd_locked():
2261          *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2262          *  2) __split_huge_zero_page_pmd() read only zero page and any write
2263          *    fault will trigger a flush_notify before pointing to a new page
2264          *    (it is fine if the secondary mmu keeps pointing to the old zero
2265          *    page in the meantime)
2266          *  3) Split a huge pmd into pte pointing to the same page. No need
2267          *     to invalidate secondary tlb entry they are all still valid.
2268          *     any further changes to individual pte will notify. So no need
2269          *     to call mmu_notifier->invalidate_range()
2270          */
2271         mmu_notifier_invalidate_range_only_end(mm, haddr, haddr +
2272                                                HPAGE_PMD_SIZE);
2273 }
2274
2275 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2276                 bool freeze, struct page *page)
2277 {
2278         pgd_t *pgd;
2279         p4d_t *p4d;
2280         pud_t *pud;
2281         pmd_t *pmd;
2282
2283         pgd = pgd_offset(vma->vm_mm, address);
2284         if (!pgd_present(*pgd))
2285                 return;
2286
2287         p4d = p4d_offset(pgd, address);
2288         if (!p4d_present(*p4d))
2289                 return;
2290
2291         pud = pud_offset(p4d, address);
2292         if (!pud_present(*pud))
2293                 return;
2294
2295         pmd = pmd_offset(pud, address);
2296
2297         __split_huge_pmd(vma, pmd, address, freeze, page);
2298 }
2299
2300 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2301                              unsigned long start,
2302                              unsigned long end,
2303                              long adjust_next)
2304 {
2305         /*
2306          * If the new start address isn't hpage aligned and it could
2307          * previously contain an hugepage: check if we need to split
2308          * an huge pmd.
2309          */
2310         if (start & ~HPAGE_PMD_MASK &&
2311             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2312             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2313                 split_huge_pmd_address(vma, start, false, NULL);
2314
2315         /*
2316          * If the new end address isn't hpage aligned and it could
2317          * previously contain an hugepage: check if we need to split
2318          * an huge pmd.
2319          */
2320         if (end & ~HPAGE_PMD_MASK &&
2321             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2322             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2323                 split_huge_pmd_address(vma, end, false, NULL);
2324
2325         /*
2326          * If we're also updating the vma->vm_next->vm_start, if the new
2327          * vm_next->vm_start isn't page aligned and it could previously
2328          * contain an hugepage: check if we need to split an huge pmd.
2329          */
2330         if (adjust_next > 0) {
2331                 struct vm_area_struct *next = vma->vm_next;
2332                 unsigned long nstart = next->vm_start;
2333                 nstart += adjust_next << PAGE_SHIFT;
2334                 if (nstart & ~HPAGE_PMD_MASK &&
2335                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2336                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2337                         split_huge_pmd_address(next, nstart, false, NULL);
2338         }
2339 }
2340
2341 static void freeze_page(struct page *page)
2342 {
2343         enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
2344                 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2345         bool unmap_success;
2346
2347         VM_BUG_ON_PAGE(!PageHead(page), page);
2348
2349         if (PageAnon(page))
2350                 ttu_flags |= TTU_SPLIT_FREEZE;
2351
2352         unmap_success = try_to_unmap(page, ttu_flags);
2353         VM_BUG_ON_PAGE(!unmap_success, page);
2354 }
2355
2356 static void unfreeze_page(struct page *page)
2357 {
2358         int i;
2359         if (PageTransHuge(page)) {
2360                 remove_migration_ptes(page, page, true);
2361         } else {
2362                 for (i = 0; i < HPAGE_PMD_NR; i++)
2363                         remove_migration_ptes(page + i, page + i, true);
2364         }
2365 }
2366
2367 static void __split_huge_page_tail(struct page *head, int tail,
2368                 struct lruvec *lruvec, struct list_head *list)
2369 {
2370         struct page *page_tail = head + tail;
2371
2372         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2373         VM_BUG_ON_PAGE(page_ref_count(page_tail) != 0, page_tail);
2374
2375         /*
2376          * tail_page->_refcount is zero and not changing from under us. But
2377          * get_page_unless_zero() may be running from under us on the
2378          * tail_page. If we used atomic_set() below instead of atomic_inc() or
2379          * atomic_add(), we would then run atomic_set() concurrently with
2380          * get_page_unless_zero(), and atomic_set() is implemented in C not
2381          * using locked ops. spin_unlock on x86 sometime uses locked ops
2382          * because of PPro errata 66, 92, so unless somebody can guarantee
2383          * atomic_set() here would be safe on all archs (and not only on x86),
2384          * it's safer to use atomic_inc()/atomic_add().
2385          */
2386         if (PageAnon(head) && !PageSwapCache(head)) {
2387                 page_ref_inc(page_tail);
2388         } else {
2389                 /* Additional pin to radix tree */
2390                 page_ref_add(page_tail, 2);
2391         }
2392
2393         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2394         page_tail->flags |= (head->flags &
2395                         ((1L << PG_referenced) |
2396                          (1L << PG_swapbacked) |
2397                          (1L << PG_swapcache) |
2398                          (1L << PG_mlocked) |
2399                          (1L << PG_uptodate) |
2400                          (1L << PG_active) |
2401                          (1L << PG_locked) |
2402                          (1L << PG_unevictable) |
2403                          (1L << PG_dirty)));
2404
2405         /*
2406          * After clearing PageTail the gup refcount can be released.
2407          * Page flags also must be visible before we make the page non-compound.
2408          */
2409         smp_wmb();
2410
2411         clear_compound_head(page_tail);
2412
2413         if (page_is_young(head))
2414                 set_page_young(page_tail);
2415         if (page_is_idle(head))
2416                 set_page_idle(page_tail);
2417
2418         /* ->mapping in first tail page is compound_mapcount */
2419         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2420                         page_tail);
2421         page_tail->mapping = head->mapping;
2422
2423         page_tail->index = head->index + tail;
2424         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2425         lru_add_page_tail(head, page_tail, lruvec, list);
2426 }
2427
2428 static void __split_huge_page(struct page *page, struct list_head *list,
2429                 unsigned long flags)
2430 {
2431         struct page *head = compound_head(page);
2432         struct zone *zone = page_zone(head);
2433         struct lruvec *lruvec;
2434         pgoff_t end = -1;
2435         int i;
2436
2437         lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat);
2438
2439         /* complete memcg works before add pages to LRU */
2440         mem_cgroup_split_huge_fixup(head);
2441
2442         if (!PageAnon(page))
2443                 end = DIV_ROUND_UP(i_size_read(head->mapping->host), PAGE_SIZE);
2444
2445         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
2446                 __split_huge_page_tail(head, i, lruvec, list);
2447                 /* Some pages can be beyond i_size: drop them from page cache */
2448                 if (head[i].index >= end) {
2449                         __ClearPageDirty(head + i);
2450                         __delete_from_page_cache(head + i, NULL);
2451                         if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2452                                 shmem_uncharge(head->mapping->host, 1);
2453                         put_page(head + i);
2454                 }
2455         }
2456
2457         ClearPageCompound(head);
2458         /* See comment in __split_huge_page_tail() */
2459         if (PageAnon(head)) {
2460                 /* Additional pin to radix tree of swap cache */
2461                 if (PageSwapCache(head))
2462                         page_ref_add(head, 2);
2463                 else
2464                         page_ref_inc(head);
2465         } else {
2466                 /* Additional pin to radix tree */
2467                 page_ref_add(head, 2);
2468                 spin_unlock(&head->mapping->tree_lock);
2469         }
2470
2471         spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
2472
2473         unfreeze_page(head);
2474
2475         for (i = 0; i < HPAGE_PMD_NR; i++) {
2476                 struct page *subpage = head + i;
2477                 if (subpage == page)
2478                         continue;
2479                 unlock_page(subpage);
2480
2481                 /*
2482                  * Subpages may be freed if there wasn't any mapping
2483                  * like if add_to_swap() is running on a lru page that
2484                  * had its mapping zapped. And freeing these pages
2485                  * requires taking the lru_lock so we do the put_page
2486                  * of the tail pages after the split is complete.
2487                  */
2488                 put_page(subpage);
2489         }
2490 }
2491
2492 int total_mapcount(struct page *page)
2493 {
2494         int i, compound, ret;
2495
2496         VM_BUG_ON_PAGE(PageTail(page), page);
2497
2498         if (likely(!PageCompound(page)))
2499                 return atomic_read(&page->_mapcount) + 1;
2500
2501         compound = compound_mapcount(page);
2502         if (PageHuge(page))
2503                 return compound;
2504         ret = compound;
2505         for (i = 0; i < HPAGE_PMD_NR; i++)
2506                 ret += atomic_read(&page[i]._mapcount) + 1;
2507         /* File pages has compound_mapcount included in _mapcount */
2508         if (!PageAnon(page))
2509                 return ret - compound * HPAGE_PMD_NR;
2510         if (PageDoubleMap(page))
2511                 ret -= HPAGE_PMD_NR;
2512         return ret;
2513 }
2514
2515 /*
2516  * This calculates accurately how many mappings a transparent hugepage
2517  * has (unlike page_mapcount() which isn't fully accurate). This full
2518  * accuracy is primarily needed to know if copy-on-write faults can
2519  * reuse the page and change the mapping to read-write instead of
2520  * copying them. At the same time this returns the total_mapcount too.
2521  *
2522  * The function returns the highest mapcount any one of the subpages
2523  * has. If the return value is one, even if different processes are
2524  * mapping different subpages of the transparent hugepage, they can
2525  * all reuse it, because each process is reusing a different subpage.
2526  *
2527  * The total_mapcount is instead counting all virtual mappings of the
2528  * subpages. If the total_mapcount is equal to "one", it tells the
2529  * caller all mappings belong to the same "mm" and in turn the
2530  * anon_vma of the transparent hugepage can become the vma->anon_vma
2531  * local one as no other process may be mapping any of the subpages.
2532  *
2533  * It would be more accurate to replace page_mapcount() with
2534  * page_trans_huge_mapcount(), however we only use
2535  * page_trans_huge_mapcount() in the copy-on-write faults where we
2536  * need full accuracy to avoid breaking page pinning, because
2537  * page_trans_huge_mapcount() is slower than page_mapcount().
2538  */
2539 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2540 {
2541         int i, ret, _total_mapcount, mapcount;
2542
2543         /* hugetlbfs shouldn't call it */
2544         VM_BUG_ON_PAGE(PageHuge(page), page);
2545
2546         if (likely(!PageTransCompound(page))) {
2547                 mapcount = atomic_read(&page->_mapcount) + 1;
2548                 if (total_mapcount)
2549                         *total_mapcount = mapcount;
2550                 return mapcount;
2551         }
2552
2553         page = compound_head(page);
2554
2555         _total_mapcount = ret = 0;
2556         for (i = 0; i < HPAGE_PMD_NR; i++) {
2557                 mapcount = atomic_read(&page[i]._mapcount) + 1;
2558                 ret = max(ret, mapcount);
2559                 _total_mapcount += mapcount;
2560         }
2561         if (PageDoubleMap(page)) {
2562                 ret -= 1;
2563                 _total_mapcount -= HPAGE_PMD_NR;
2564         }
2565         mapcount = compound_mapcount(page);
2566         ret += mapcount;
2567         _total_mapcount += mapcount;
2568         if (total_mapcount)
2569                 *total_mapcount = _total_mapcount;
2570         return ret;
2571 }
2572
2573 /* Racy check whether the huge page can be split */
2574 bool can_split_huge_page(struct page *page, int *pextra_pins)
2575 {
2576         int extra_pins;
2577
2578         /* Additional pins from radix tree */
2579         if (PageAnon(page))
2580                 extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0;
2581         else
2582                 extra_pins = HPAGE_PMD_NR;
2583         if (pextra_pins)
2584                 *pextra_pins = extra_pins;
2585         return total_mapcount(page) == page_count(page) - extra_pins - 1;
2586 }
2587
2588 /*
2589  * This function splits huge page into normal pages. @page can point to any
2590  * subpage of huge page to split. Split doesn't change the position of @page.
2591  *
2592  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2593  * The huge page must be locked.
2594  *
2595  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2596  *
2597  * Both head page and tail pages will inherit mapping, flags, and so on from
2598  * the hugepage.
2599  *
2600  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2601  * they are not mapped.
2602  *
2603  * Returns 0 if the hugepage is split successfully.
2604  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2605  * us.
2606  */
2607 int split_huge_page_to_list(struct page *page, struct list_head *list)
2608 {
2609         struct page *head = compound_head(page);
2610         struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2611         struct anon_vma *anon_vma = NULL;
2612         struct address_space *mapping = NULL;
2613         int count, mapcount, extra_pins, ret;
2614         bool mlocked;
2615         unsigned long flags;
2616
2617         VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
2618         VM_BUG_ON_PAGE(!PageLocked(page), page);
2619         VM_BUG_ON_PAGE(!PageCompound(page), page);
2620
2621         if (PageWriteback(page))
2622                 return -EBUSY;
2623
2624         if (PageAnon(head)) {
2625                 /*
2626                  * The caller does not necessarily hold an mmap_sem that would
2627                  * prevent the anon_vma disappearing so we first we take a
2628                  * reference to it and then lock the anon_vma for write. This
2629                  * is similar to page_lock_anon_vma_read except the write lock
2630                  * is taken to serialise against parallel split or collapse
2631                  * operations.
2632                  */
2633                 anon_vma = page_get_anon_vma(head);
2634                 if (!anon_vma) {
2635                         ret = -EBUSY;
2636                         goto out;
2637                 }
2638                 mapping = NULL;
2639                 anon_vma_lock_write(anon_vma);
2640         } else {
2641                 mapping = head->mapping;
2642
2643                 /* Truncated ? */
2644                 if (!mapping) {
2645                         ret = -EBUSY;
2646                         goto out;
2647                 }
2648
2649                 anon_vma = NULL;
2650                 i_mmap_lock_read(mapping);
2651         }
2652
2653         /*
2654          * Racy check if we can split the page, before freeze_page() will
2655          * split PMDs
2656          */
2657         if (!can_split_huge_page(head, &extra_pins)) {
2658                 ret = -EBUSY;
2659                 goto out_unlock;
2660         }
2661
2662         mlocked = PageMlocked(page);
2663         freeze_page(head);
2664         VM_BUG_ON_PAGE(compound_mapcount(head), head);
2665
2666         /* Make sure the page is not on per-CPU pagevec as it takes pin */
2667         if (mlocked)
2668                 lru_add_drain();
2669
2670         /* prevent PageLRU to go away from under us, and freeze lru stats */
2671         spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags);
2672
2673         if (mapping) {
2674                 void **pslot;
2675
2676                 spin_lock(&mapping->tree_lock);
2677                 pslot = radix_tree_lookup_slot(&mapping->page_tree,
2678                                 page_index(head));
2679                 /*
2680                  * Check if the head page is present in radix tree.
2681                  * We assume all tail are present too, if head is there.
2682                  */
2683                 if (radix_tree_deref_slot_protected(pslot,
2684                                         &mapping->tree_lock) != head)
2685                         goto fail;
2686         }
2687
2688         /* Prevent deferred_split_scan() touching ->_refcount */
2689         spin_lock(&pgdata->split_queue_lock);
2690         count = page_count(head);
2691         mapcount = total_mapcount(head);
2692         if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2693                 if (!list_empty(page_deferred_list(head))) {
2694                         pgdata->split_queue_len--;
2695                         list_del(page_deferred_list(head));
2696                 }
2697                 if (mapping)
2698                         __dec_node_page_state(page, NR_SHMEM_THPS);
2699                 spin_unlock(&pgdata->split_queue_lock);
2700                 __split_huge_page(page, list, flags);
2701                 if (PageSwapCache(head)) {
2702                         swp_entry_t entry = { .val = page_private(head) };
2703
2704                         ret = split_swap_cluster(entry);
2705                 } else
2706                         ret = 0;
2707         } else {
2708                 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2709                         pr_alert("total_mapcount: %u, page_count(): %u\n",
2710                                         mapcount, count);
2711                         if (PageTail(page))
2712                                 dump_page(head, NULL);
2713                         dump_page(page, "total_mapcount(head) > 0");
2714                         BUG();
2715                 }
2716                 spin_unlock(&pgdata->split_queue_lock);
2717 fail:           if (mapping)
2718                         spin_unlock(&mapping->tree_lock);
2719                 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
2720                 unfreeze_page(head);
2721                 ret = -EBUSY;
2722         }
2723
2724 out_unlock:
2725         if (anon_vma) {
2726                 anon_vma_unlock_write(anon_vma);
2727                 put_anon_vma(anon_vma);
2728         }
2729         if (mapping)
2730                 i_mmap_unlock_read(mapping);
2731 out:
2732         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2733         return ret;
2734 }
2735
2736 void free_transhuge_page(struct page *page)
2737 {
2738         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2739         unsigned long flags;
2740
2741         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2742         if (!list_empty(page_deferred_list(page))) {
2743                 pgdata->split_queue_len--;
2744                 list_del(page_deferred_list(page));
2745         }
2746         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2747         free_compound_page(page);
2748 }
2749
2750 void deferred_split_huge_page(struct page *page)
2751 {
2752         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2753         unsigned long flags;
2754
2755         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2756
2757         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2758         if (list_empty(page_deferred_list(page))) {
2759                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2760                 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2761                 pgdata->split_queue_len++;
2762         }
2763         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2764 }
2765
2766 static unsigned long deferred_split_count(struct shrinker *shrink,
2767                 struct shrink_control *sc)
2768 {
2769         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2770         return READ_ONCE(pgdata->split_queue_len);
2771 }
2772
2773 static unsigned long deferred_split_scan(struct shrinker *shrink,
2774                 struct shrink_control *sc)
2775 {
2776         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2777         unsigned long flags;
2778         LIST_HEAD(list), *pos, *next;
2779         struct page *page;
2780         int split = 0;
2781
2782         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2783         /* Take pin on all head pages to avoid freeing them under us */
2784         list_for_each_safe(pos, next, &pgdata->split_queue) {
2785                 page = list_entry((void *)pos, struct page, mapping);
2786                 page = compound_head(page);
2787                 if (get_page_unless_zero(page)) {
2788                         list_move(page_deferred_list(page), &list);
2789                 } else {
2790                         /* We lost race with put_compound_page() */
2791                         list_del_init(page_deferred_list(page));
2792                         pgdata->split_queue_len--;
2793                 }
2794                 if (!--sc->nr_to_scan)
2795                         break;
2796         }
2797         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2798
2799         list_for_each_safe(pos, next, &list) {
2800                 page = list_entry((void *)pos, struct page, mapping);
2801                 lock_page(page);
2802                 /* split_huge_page() removes page from list on success */
2803                 if (!split_huge_page(page))
2804                         split++;
2805                 unlock_page(page);
2806                 put_page(page);
2807         }
2808
2809         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2810         list_splice_tail(&list, &pgdata->split_queue);
2811         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2812
2813         /*
2814          * Stop shrinker if we didn't split any page, but the queue is empty.
2815          * This can happen if pages were freed under us.
2816          */
2817         if (!split && list_empty(&pgdata->split_queue))
2818                 return SHRINK_STOP;
2819         return split;
2820 }
2821
2822 static struct shrinker deferred_split_shrinker = {
2823         .count_objects = deferred_split_count,
2824         .scan_objects = deferred_split_scan,
2825         .seeks = DEFAULT_SEEKS,
2826         .flags = SHRINKER_NUMA_AWARE,
2827 };
2828
2829 #ifdef CONFIG_DEBUG_FS
2830 static int split_huge_pages_set(void *data, u64 val)
2831 {
2832         struct zone *zone;
2833         struct page *page;
2834         unsigned long pfn, max_zone_pfn;
2835         unsigned long total = 0, split = 0;
2836
2837         if (val != 1)
2838                 return -EINVAL;
2839
2840         for_each_populated_zone(zone) {
2841                 max_zone_pfn = zone_end_pfn(zone);
2842                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2843                         if (!pfn_valid(pfn))
2844                                 continue;
2845
2846                         page = pfn_to_page(pfn);
2847                         if (!get_page_unless_zero(page))
2848                                 continue;
2849
2850                         if (zone != page_zone(page))
2851                                 goto next;
2852
2853                         if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2854                                 goto next;
2855
2856                         total++;
2857                         lock_page(page);
2858                         if (!split_huge_page(page))
2859                                 split++;
2860                         unlock_page(page);
2861 next:
2862                         put_page(page);
2863                 }
2864         }
2865
2866         pr_info("%lu of %lu THP split\n", split, total);
2867
2868         return 0;
2869 }
2870 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2871                 "%llu\n");
2872
2873 static int __init split_huge_pages_debugfs(void)
2874 {
2875         void *ret;
2876
2877         ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2878                         &split_huge_pages_fops);
2879         if (!ret)
2880                 pr_warn("Failed to create split_huge_pages in debugfs");
2881         return 0;
2882 }
2883 late_initcall(split_huge_pages_debugfs);
2884 #endif
2885
2886 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2887 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
2888                 struct page *page)
2889 {
2890         struct vm_area_struct *vma = pvmw->vma;
2891         struct mm_struct *mm = vma->vm_mm;
2892         unsigned long address = pvmw->address;
2893         pmd_t pmdval;
2894         swp_entry_t entry;
2895         pmd_t pmdswp;
2896
2897         if (!(pvmw->pmd && !pvmw->pte))
2898                 return;
2899
2900         mmu_notifier_invalidate_range_start(mm, address,
2901                         address + HPAGE_PMD_SIZE);
2902
2903         flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
2904         pmdval = *pvmw->pmd;
2905         pmdp_invalidate(vma, address, pvmw->pmd);
2906         if (pmd_dirty(pmdval))
2907                 set_page_dirty(page);
2908         entry = make_migration_entry(page, pmd_write(pmdval));
2909         pmdswp = swp_entry_to_pmd(entry);
2910         if (pmd_soft_dirty(pmdval))
2911                 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
2912         set_pmd_at(mm, address, pvmw->pmd, pmdswp);
2913         page_remove_rmap(page, true);
2914         put_page(page);
2915
2916         mmu_notifier_invalidate_range_end(mm, address,
2917                         address + HPAGE_PMD_SIZE);
2918 }
2919
2920 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
2921 {
2922         struct vm_area_struct *vma = pvmw->vma;
2923         struct mm_struct *mm = vma->vm_mm;
2924         unsigned long address = pvmw->address;
2925         unsigned long mmun_start = address & HPAGE_PMD_MASK;
2926         pmd_t pmde;
2927         swp_entry_t entry;
2928
2929         if (!(pvmw->pmd && !pvmw->pte))
2930                 return;
2931
2932         entry = pmd_to_swp_entry(*pvmw->pmd);
2933         get_page(new);
2934         pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
2935         if (pmd_swp_soft_dirty(*pvmw->pmd))
2936                 pmde = pmd_mksoft_dirty(pmde);
2937         if (is_write_migration_entry(entry))
2938                 pmde = maybe_pmd_mkwrite(pmde, vma, false);
2939
2940         flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
2941         page_add_anon_rmap(new, vma, mmun_start, true);
2942         set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
2943         if (vma->vm_flags & VM_LOCKED)
2944                 mlock_vma_page(new);
2945         update_mmu_cache_pmd(vma, address, pvmw->pmd);
2946 }
2947 #endif