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