mm/huge_memory.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *  Copyright (C) 2009  Red Hat, Inc.
   4  */
   5
   6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   7
   8 #include <linux/mm.h>
   9 #include <linux/sched.h>
  10 #include <linux/sched/coredump.h>
  11 #include <linux/sched/numa_balancing.h>
  12 #include <linux/highmem.h>
  13 #include <linux/hugetlb.h>
  14 #include <linux/mmu_notifier.h>
  15 #include <linux/rmap.h>
  16 #include <linux/swap.h>
  17 #include <linux/shrinker.h>
  18 #include <linux/mm_inline.h>
  19 #include <linux/swapops.h>
  20 #include <linux/dax.h>
  21 #include <linux/khugepaged.h>
  22 #include <linux/freezer.h>
  23 #include <linux/pfn_t.h>
  24 #include <linux/mman.h>
  25 #include <linux/memremap.h>
  26 #include <linux/pagemap.h>
  27 #include <linux/debugfs.h>
  28 #include <linux/migrate.h>
  29 #include <linux/hashtable.h>
  30 #include <linux/userfaultfd_k.h>
  31 #include <linux/page_idle.h>
  32 #include <linux/shmem_fs.h>
  33 #include <linux/oom.h>
  34 #include <linux/numa.h>
  35 #include <linux/page_owner.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 bool transparent_hugepage_enabled(struct vm_area_struct *vma)
  66 {
  67         /* The addr is used to check if the vma size fits */
  68         unsigned long addr = (vma->vm_end & HPAGE_PMD_MASK) - HPAGE_PMD_SIZE;
  69
  70         if (!transhuge_vma_suitable(vma, addr))
  71                 return false;
  72         if (vma_is_anonymous(vma))
  73                 return __transparent_hugepage_enabled(vma);
  74         if (vma_is_shmem(vma))
  75                 return shmem_huge_enabled(vma);
  76
  77         return false;
  78 }
  79
  80 static struct page *get_huge_zero_page(void)
  81 {
  82         struct page *zero_page;
  83 retry:
  84         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
  85                 return READ_ONCE(huge_zero_page);
  86
  87         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
  88                         HPAGE_PMD_ORDER);
  89         if (!zero_page) {
  90                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
  91                 return NULL;
  92         }
  93         count_vm_event(THP_ZERO_PAGE_ALLOC);
  94         preempt_disable();
  95         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
  96                 preempt_enable();
  97                 __free_pages(zero_page, compound_order(zero_page));
  98                 goto retry;
  99         }
 100
 101         /* We take additional reference here. It will be put back by shrinker */
 102         atomic_set(&huge_zero_refcount, 2);
 103         preempt_enable();
 104         return READ_ONCE(huge_zero_page);
 105 }
 106
 107 static void put_huge_zero_page(void)
 108 {
 109         /*
 110          * Counter should never go to zero here. Only shrinker can put
 111          * last reference.
 112          */
 113         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 114 }
 115
 116 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
 117 {
 118         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 119                 return READ_ONCE(huge_zero_page);
 120
 121         if (!get_huge_zero_page())
 122                 return NULL;
 123
 124         if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 125                 put_huge_zero_page();
 126
 127         return READ_ONCE(huge_zero_page);
 128 }
 129
 130 void mm_put_huge_zero_page(struct mm_struct *mm)
 131 {
 132         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 133                 put_huge_zero_page();
 134 }
 135
 136 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 137                                         struct shrink_control *sc)
 138 {
 139         /* we can free zero page only if last reference remains */
 140         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 141 }
 142
 143 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 144                                        struct shrink_control *sc)
 145 {
 146         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 147                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 148                 BUG_ON(zero_page == NULL);
 149                 __free_pages(zero_page, compound_order(zero_page));
 150                 return HPAGE_PMD_NR;
 151         }
 152
 153         return 0;
 154 }
 155
 156 static struct shrinker huge_zero_page_shrinker = {
 157         .count_objects = shrink_huge_zero_page_count,
 158         .scan_objects = shrink_huge_zero_page_scan,
 159         .seeks = DEFAULT_SEEKS,
 160 };
 161
 162 #ifdef CONFIG_SYSFS
 163 static ssize_t enabled_show(struct kobject *kobj,
 164                             struct kobj_attribute *attr, char *buf)
 165 {
 166         const char *output;
 167
 168         if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
 169                 output = "[always] madvise never";
 170         else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 171                           &transparent_hugepage_flags))
 172                 output = "always [madvise] never";
 173         else
 174                 output = "always madvise [never]";
 175
 176         return sysfs_emit(buf, "%s\n", output);
 177 }
 178
 179 static ssize_t enabled_store(struct kobject *kobj,
 180                              struct kobj_attribute *attr,
 181                              const char *buf, size_t count)
 182 {
 183         ssize_t ret = count;
 184
 185         if (sysfs_streq(buf, "always")) {
 186                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 187                 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 188         } else if (sysfs_streq(buf, "madvise")) {
 189                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 190                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 191         } else if (sysfs_streq(buf, "never")) {
 192                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 193                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 194         } else
 195                 ret = -EINVAL;
 196
 197         if (ret > 0) {
 198                 int err = start_stop_khugepaged();
 199                 if (err)
 200                         ret = err;
 201         }
 202         return ret;
 203 }
 204 static struct kobj_attribute enabled_attr =
 205         __ATTR(enabled, 0644, enabled_show, enabled_store);
 206
 207 ssize_t single_hugepage_flag_show(struct kobject *kobj,
 208                                   struct kobj_attribute *attr, char *buf,
 209                                   enum transparent_hugepage_flag flag)
 210 {
 211         return sysfs_emit(buf, "%d\n",
 212                           !!test_bit(flag, &transparent_hugepage_flags));
 213 }
 214
 215 ssize_t single_hugepage_flag_store(struct kobject *kobj,
 216                                  struct kobj_attribute *attr,
 217                                  const char *buf, size_t count,
 218                                  enum transparent_hugepage_flag flag)
 219 {
 220         unsigned long value;
 221         int ret;
 222
 223         ret = kstrtoul(buf, 10, &value);
 224         if (ret < 0)
 225                 return ret;
 226         if (value > 1)
 227                 return -EINVAL;
 228
 229         if (value)
 230                 set_bit(flag, &transparent_hugepage_flags);
 231         else
 232                 clear_bit(flag, &transparent_hugepage_flags);
 233
 234         return count;
 235 }
 236
 237 static ssize_t defrag_show(struct kobject *kobj,
 238                            struct kobj_attribute *attr, char *buf)
 239 {
 240         const char *output;
 241
 242         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
 243                      &transparent_hugepage_flags))
 244                 output = "[always] defer defer+madvise madvise never";
 245         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
 246                           &transparent_hugepage_flags))
 247                 output = "always [defer] defer+madvise madvise never";
 248         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG,
 249                           &transparent_hugepage_flags))
 250                 output = "always defer [defer+madvise] madvise never";
 251         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
 252                           &transparent_hugepage_flags))
 253                 output = "always defer defer+madvise [madvise] never";
 254         else
 255                 output = "always defer defer+madvise madvise [never]";
 256
 257         return sysfs_emit(buf, "%s\n", output);
 258 }
 259
 260 static ssize_t defrag_store(struct kobject *kobj,
 261                             struct kobj_attribute *attr,
 262                             const char *buf, size_t count)
 263 {
 264         if (sysfs_streq(buf, "always")) {
 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                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 269         } else if (sysfs_streq(buf, "defer+madvise")) {
 270                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 271                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 272                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 273                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 274         } else if (sysfs_streq(buf, "defer")) {
 275                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 276                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 277                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 278                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 279         } else if (sysfs_streq(buf, "madvise")) {
 280                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 281                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 282                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 283                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 284         } else if (sysfs_streq(buf, "never")) {
 285                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 286                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 287                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 288                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 289         } else
 290                 return -EINVAL;
 291
 292         return count;
 293 }
 294 static struct kobj_attribute defrag_attr =
 295         __ATTR(defrag, 0644, defrag_show, defrag_store);
 296
 297 static ssize_t use_zero_page_show(struct kobject *kobj,
 298                                   struct kobj_attribute *attr, char *buf)
 299 {
 300         return single_hugepage_flag_show(kobj, attr, buf,
 301                                          TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 302 }
 303 static ssize_t use_zero_page_store(struct kobject *kobj,
 304                 struct kobj_attribute *attr, const char *buf, size_t count)
 305 {
 306         return single_hugepage_flag_store(kobj, attr, buf, count,
 307                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 308 }
 309 static struct kobj_attribute use_zero_page_attr =
 310         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 311
 312 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
 313                                    struct kobj_attribute *attr, char *buf)
 314 {
 315         return sysfs_emit(buf, "%lu\n", HPAGE_PMD_SIZE);
 316 }
 317 static struct kobj_attribute hpage_pmd_size_attr =
 318         __ATTR_RO(hpage_pmd_size);
 319
 320 static struct attribute *hugepage_attr[] = {
 321         &enabled_attr.attr,
 322         &defrag_attr.attr,
 323         &use_zero_page_attr.attr,
 324         &hpage_pmd_size_attr.attr,
 325 #ifdef CONFIG_SHMEM
 326         &shmem_enabled_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                 /*
 390                  * Hardware doesn't support hugepages, hence disable
 391                  * DAX PMD support.
 392                  */
 393                 transparent_hugepage_flags = 1 << TRANSPARENT_HUGEPAGE_NEVER_DAX;
 394                 return -EINVAL;
 395         }
 396
 397         /*
 398          * hugepages can't be allocated by the buddy allocator
 399          */
 400         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
 401         /*
 402          * we use page->mapping and page->index in second tail page
 403          * as list_head: assuming THP order >= 2
 404          */
 405         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 406
 407         err = hugepage_init_sysfs(&hugepage_kobj);
 408         if (err)
 409                 goto err_sysfs;
 410
 411         err = khugepaged_init();
 412         if (err)
 413                 goto err_slab;
 414
 415         err = register_shrinker(&huge_zero_page_shrinker);
 416         if (err)
 417                 goto err_hzp_shrinker;
 418         err = register_shrinker(&deferred_split_shrinker);
 419         if (err)
 420                 goto err_split_shrinker;
 421
 422         /*
 423          * By default disable transparent hugepages on smaller systems,
 424          * where the extra memory used could hurt more than TLB overhead
 425          * is likely to save.  The admin can still enable it through /sys.
 426          */
 427         if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
 428                 transparent_hugepage_flags = 0;
 429                 return 0;
 430         }
 431
 432         err = start_stop_khugepaged();
 433         if (err)
 434                 goto err_khugepaged;
 435
 436         return 0;
 437 err_khugepaged:
 438         unregister_shrinker(&deferred_split_shrinker);
 439 err_split_shrinker:
 440         unregister_shrinker(&huge_zero_page_shrinker);
 441 err_hzp_shrinker:
 442         khugepaged_destroy();
 443 err_slab:
 444         hugepage_exit_sysfs(hugepage_kobj);
 445 err_sysfs:
 446         return err;
 447 }
 448 subsys_initcall(hugepage_init);
 449
 450 static int __init setup_transparent_hugepage(char *str)
 451 {
 452         int ret = 0;
 453         if (!str)
 454                 goto out;
 455         if (!strcmp(str, "always")) {
 456                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 457                         &transparent_hugepage_flags);
 458                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 459                           &transparent_hugepage_flags);
 460                 ret = 1;
 461         } else if (!strcmp(str, "madvise")) {
 462                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 463                           &transparent_hugepage_flags);
 464                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 465                         &transparent_hugepage_flags);
 466                 ret = 1;
 467         } else if (!strcmp(str, "never")) {
 468                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 469                           &transparent_hugepage_flags);
 470                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 471                           &transparent_hugepage_flags);
 472                 ret = 1;
 473         }
 474 out:
 475         if (!ret)
 476                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 477         return ret;
 478 }
 479 __setup("transparent_hugepage=", setup_transparent_hugepage);
 480
 481 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 482 {
 483         if (likely(vma->vm_flags & VM_WRITE))
 484                 pmd = pmd_mkwrite(pmd);
 485         return pmd;
 486 }
 487
 488 #ifdef CONFIG_MEMCG
 489 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
 490 {
 491         struct mem_cgroup *memcg = page_memcg(compound_head(page));
 492         struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
 493
 494         if (memcg)
 495                 return &memcg->deferred_split_queue;
 496         else
 497                 return &pgdat->deferred_split_queue;
 498 }
 499 #else
 500 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
 501 {
 502         struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
 503
 504         return &pgdat->deferred_split_queue;
 505 }
 506 #endif
 507
 508 void prep_transhuge_page(struct page *page)
 509 {
 510         /*
 511          * we use page->mapping and page->indexlru in second tail page
 512          * as list_head: assuming THP order >= 2
 513          */
 514
 515         INIT_LIST_HEAD(page_deferred_list(page));
 516         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 517 }
 518
 519 bool is_transparent_hugepage(struct page *page)
 520 {
 521         if (!PageCompound(page))
 522                 return false;
 523
 524         page = compound_head(page);
 525         return is_huge_zero_page(page) ||
 526                page[1].compound_dtor == TRANSHUGE_PAGE_DTOR;
 527 }
 528 EXPORT_SYMBOL_GPL(is_transparent_hugepage);
 529
 530 static unsigned long __thp_get_unmapped_area(struct file *filp,
 531                 unsigned long addr, unsigned long len,
 532                 loff_t off, unsigned long flags, unsigned long size)
 533 {
 534         loff_t off_end = off + len;
 535         loff_t off_align = round_up(off, size);
 536         unsigned long len_pad, ret;
 537
 538         if (off_end <= off_align || (off_end - off_align) < size)
 539                 return 0;
 540
 541         len_pad = len + size;
 542         if (len_pad < len || (off + len_pad) < off)
 543                 return 0;
 544
 545         ret = current->mm->get_unmapped_area(filp, addr, len_pad,
 546                                               off >> PAGE_SHIFT, flags);
 547
 548         /*
 549          * The failure might be due to length padding. The caller will retry
 550          * without the padding.
 551          */
 552         if (IS_ERR_VALUE(ret))
 553                 return 0;
 554
 555         /*
 556          * Do not try to align to THP boundary if allocation at the address
 557          * hint succeeds.
 558          */
 559         if (ret == addr)
 560                 return addr;
 561
 562         ret += (off - ret) & (size - 1);
 563         return ret;
 564 }
 565
 566 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 567                 unsigned long len, unsigned long pgoff, unsigned long flags)
 568 {
 569         unsigned long ret;
 570         loff_t off = (loff_t)pgoff << PAGE_SHIFT;
 571
 572         if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
 573                 goto out;
 574
 575         ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE);
 576         if (ret)
 577                 return ret;
 578 out:
 579         return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
 580 }
 581 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
 582
 583 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
 584                         struct page *page, gfp_t gfp)
 585 {
 586         struct vm_area_struct *vma = vmf->vma;
 587         pgtable_t pgtable;
 588         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 589         vm_fault_t ret = 0;
 590
 591         VM_BUG_ON_PAGE(!PageCompound(page), page);
 592
 593         if (mem_cgroup_charge(page, vma->vm_mm, gfp)) {
 594                 put_page(page);
 595                 count_vm_event(THP_FAULT_FALLBACK);
 596                 count_vm_event(THP_FAULT_FALLBACK_CHARGE);
 597                 return VM_FAULT_FALLBACK;
 598         }
 599         cgroup_throttle_swaprate(page, gfp);
 600
 601         pgtable = pte_alloc_one(vma->vm_mm);
 602         if (unlikely(!pgtable)) {
 603                 ret = VM_FAULT_OOM;
 604                 goto release;
 605         }
 606
 607         clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
 608         /*
 609          * The memory barrier inside __SetPageUptodate makes sure that
 610          * clear_huge_page writes become visible before the set_pmd_at()
 611          * write.
 612          */
 613         __SetPageUptodate(page);
 614
 615         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 616         if (unlikely(!pmd_none(*vmf->pmd))) {
 617                 goto unlock_release;
 618         } else {
 619                 pmd_t entry;
 620
 621                 ret = check_stable_address_space(vma->vm_mm);
 622                 if (ret)
 623                         goto unlock_release;
 624
 625                 /* Deliver the page fault to userland */
 626                 if (userfaultfd_missing(vma)) {
 627                         vm_fault_t ret2;
 628
 629                         spin_unlock(vmf->ptl);
 630                         put_page(page);
 631                         pte_free(vma->vm_mm, pgtable);
 632                         ret2 = handle_userfault(vmf, VM_UFFD_MISSING);
 633                         VM_BUG_ON(ret2 & VM_FAULT_FALLBACK);
 634                         return ret2;
 635                 }
 636
 637                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 638                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 639                 page_add_new_anon_rmap(page, vma, haddr, true);
 640                 lru_cache_add_inactive_or_unevictable(page, vma);
 641                 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
 642                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
 643                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
 644                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 645                 mm_inc_nr_ptes(vma->vm_mm);
 646                 spin_unlock(vmf->ptl);
 647                 count_vm_event(THP_FAULT_ALLOC);
 648                 count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC);
 649         }
 650
 651         return 0;
 652 unlock_release:
 653         spin_unlock(vmf->ptl);
 654 release:
 655         if (pgtable)
 656                 pte_free(vma->vm_mm, pgtable);
 657         put_page(page);
 658         return ret;
 659
 660 }
 661
 662 /*
 663  * always: directly stall for all thp allocations
 664  * defer: wake kswapd and fail if not immediately available
 665  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
 666  *                fail if not immediately available
 667  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
 668  *          available
 669  * never: never stall for any thp allocation
 670  */
 671 gfp_t vma_thp_gfp_mask(struct vm_area_struct *vma)
 672 {
 673         const bool vma_madvised = vma && (vma->vm_flags & VM_HUGEPAGE);
 674
 675         /* Always do synchronous compaction */
 676         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 677                 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
 678
 679         /* Kick kcompactd and fail quickly */
 680         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 681                 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
 682
 683         /* Synchronous compaction if madvised, otherwise kick kcompactd */
 684         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 685                 return GFP_TRANSHUGE_LIGHT |
 686                         (vma_madvised ? __GFP_DIRECT_RECLAIM :
 687                                         __GFP_KSWAPD_RECLAIM);
 688
 689         /* Only do synchronous compaction if madvised */
 690         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 691                 return GFP_TRANSHUGE_LIGHT |
 692                        (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
 693
 694         return GFP_TRANSHUGE_LIGHT;
 695 }
 696
 697 /* Caller must hold page table lock. */
 698 static void set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 699                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 700                 struct page *zero_page)
 701 {
 702         pmd_t entry;
 703         if (!pmd_none(*pmd))
 704                 return;
 705         entry = mk_pmd(zero_page, vma->vm_page_prot);
 706         entry = pmd_mkhuge(entry);
 707         if (pgtable)
 708                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 709         set_pmd_at(mm, haddr, pmd, entry);
 710         mm_inc_nr_ptes(mm);
 711 }
 712
 713 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
 714 {
 715         struct vm_area_struct *vma = vmf->vma;
 716         gfp_t gfp;
 717         struct page *page;
 718         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 719
 720         if (!transhuge_vma_suitable(vma, haddr))
 721                 return VM_FAULT_FALLBACK;
 722         if (unlikely(anon_vma_prepare(vma)))
 723                 return VM_FAULT_OOM;
 724         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 725                 return VM_FAULT_OOM;
 726         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
 727                         !mm_forbids_zeropage(vma->vm_mm) &&
 728                         transparent_hugepage_use_zero_page()) {
 729                 pgtable_t pgtable;
 730                 struct page *zero_page;
 731                 vm_fault_t ret;
 732                 pgtable = pte_alloc_one(vma->vm_mm);
 733                 if (unlikely(!pgtable))
 734                         return VM_FAULT_OOM;
 735                 zero_page = mm_get_huge_zero_page(vma->vm_mm);
 736                 if (unlikely(!zero_page)) {
 737                         pte_free(vma->vm_mm, pgtable);
 738                         count_vm_event(THP_FAULT_FALLBACK);
 739                         return VM_FAULT_FALLBACK;
 740                 }
 741                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 742                 ret = 0;
 743                 if (pmd_none(*vmf->pmd)) {
 744                         ret = check_stable_address_space(vma->vm_mm);
 745                         if (ret) {
 746                                 spin_unlock(vmf->ptl);
 747                                 pte_free(vma->vm_mm, pgtable);
 748                         } else if (userfaultfd_missing(vma)) {
 749                                 spin_unlock(vmf->ptl);
 750                                 pte_free(vma->vm_mm, pgtable);
 751                                 ret = handle_userfault(vmf, VM_UFFD_MISSING);
 752                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 753                         } else {
 754                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
 755                                                    haddr, vmf->pmd, zero_page);
 756                                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
 757                                 spin_unlock(vmf->ptl);
 758                         }
 759                 } else {
 760                         spin_unlock(vmf->ptl);
 761                         pte_free(vma->vm_mm, pgtable);
 762                 }
 763                 return ret;
 764         }
 765         gfp = vma_thp_gfp_mask(vma);
 766         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 767         if (unlikely(!page)) {
 768                 count_vm_event(THP_FAULT_FALLBACK);
 769                 return VM_FAULT_FALLBACK;
 770         }
 771         prep_transhuge_page(page);
 772         return __do_huge_pmd_anonymous_page(vmf, page, gfp);
 773 }
 774
 775 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 776                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
 777                 pgtable_t pgtable)
 778 {
 779         struct mm_struct *mm = vma->vm_mm;
 780         pmd_t entry;
 781         spinlock_t *ptl;
 782
 783         ptl = pmd_lock(mm, pmd);
 784         if (!pmd_none(*pmd)) {
 785                 if (write) {
 786                         if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
 787                                 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
 788                                 goto out_unlock;
 789                         }
 790                         entry = pmd_mkyoung(*pmd);
 791                         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 792                         if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
 793                                 update_mmu_cache_pmd(vma, addr, pmd);
 794                 }
 795
 796                 goto out_unlock;
 797         }
 798
 799         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 800         if (pfn_t_devmap(pfn))
 801                 entry = pmd_mkdevmap(entry);
 802         if (write) {
 803                 entry = pmd_mkyoung(pmd_mkdirty(entry));
 804                 entry = maybe_pmd_mkwrite(entry, vma);
 805         }
 806
 807         if (pgtable) {
 808                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 809                 mm_inc_nr_ptes(mm);
 810                 pgtable = NULL;
 811         }
 812
 813         set_pmd_at(mm, addr, pmd, entry);
 814         update_mmu_cache_pmd(vma, addr, pmd);
 815
 816 out_unlock:
 817         spin_unlock(ptl);
 818         if (pgtable)
 819                 pte_free(mm, pgtable);
 820 }
 821
 822 /**
 823  * vmf_insert_pfn_pmd_prot - insert a pmd size pfn
 824  * @vmf: Structure describing the fault
 825  * @pfn: pfn to insert
 826  * @pgprot: page protection to use
 827  * @write: whether it's a write fault
 828  *
 829  * Insert a pmd size pfn. See vmf_insert_pfn() for additional info and
 830  * also consult the vmf_insert_mixed_prot() documentation when
 831  * @pgprot != @vmf->vma->vm_page_prot.
 832  *
 833  * Return: vm_fault_t value.
 834  */
 835 vm_fault_t vmf_insert_pfn_pmd_prot(struct vm_fault *vmf, pfn_t pfn,
 836                                    pgprot_t pgprot, bool write)
 837 {
 838         unsigned long addr = vmf->address & PMD_MASK;
 839         struct vm_area_struct *vma = vmf->vma;
 840         pgtable_t pgtable = NULL;
 841
 842         /*
 843          * If we had pmd_special, we could avoid all these restrictions,
 844          * but we need to be consistent with PTEs and architectures that
 845          * can't support a 'special' bit.
 846          */
 847         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 848                         !pfn_t_devmap(pfn));
 849         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 850                                                 (VM_PFNMAP|VM_MIXEDMAP));
 851         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 852
 853         if (addr < vma->vm_start || addr >= vma->vm_end)
 854                 return VM_FAULT_SIGBUS;
 855
 856         if (arch_needs_pgtable_deposit()) {
 857                 pgtable = pte_alloc_one(vma->vm_mm);
 858                 if (!pgtable)
 859                         return VM_FAULT_OOM;
 860         }
 861
 862         track_pfn_insert(vma, &pgprot, pfn);
 863
 864         insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
 865         return VM_FAULT_NOPAGE;
 866 }
 867 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd_prot);
 868
 869 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 870 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
 871 {
 872         if (likely(vma->vm_flags & VM_WRITE))
 873                 pud = pud_mkwrite(pud);
 874         return pud;
 875 }
 876
 877 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 878                 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
 879 {
 880         struct mm_struct *mm = vma->vm_mm;
 881         pud_t entry;
 882         spinlock_t *ptl;
 883
 884         ptl = pud_lock(mm, pud);
 885         if (!pud_none(*pud)) {
 886                 if (write) {
 887                         if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
 888                                 WARN_ON_ONCE(!is_huge_zero_pud(*pud));
 889                                 goto out_unlock;
 890                         }
 891                         entry = pud_mkyoung(*pud);
 892                         entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
 893                         if (pudp_set_access_flags(vma, addr, pud, entry, 1))
 894                                 update_mmu_cache_pud(vma, addr, pud);
 895                 }
 896                 goto out_unlock;
 897         }
 898
 899         entry = pud_mkhuge(pfn_t_pud(pfn, prot));
 900         if (pfn_t_devmap(pfn))
 901                 entry = pud_mkdevmap(entry);
 902         if (write) {
 903                 entry = pud_mkyoung(pud_mkdirty(entry));
 904                 entry = maybe_pud_mkwrite(entry, vma);
 905         }
 906         set_pud_at(mm, addr, pud, entry);
 907         update_mmu_cache_pud(vma, addr, pud);
 908
 909 out_unlock:
 910         spin_unlock(ptl);
 911 }
 912
 913 /**
 914  * vmf_insert_pfn_pud_prot - insert a pud size pfn
 915  * @vmf: Structure describing the fault
 916  * @pfn: pfn to insert
 917  * @pgprot: page protection to use
 918  * @write: whether it's a write fault
 919  *
 920  * Insert a pud size pfn. See vmf_insert_pfn() for additional info and
 921  * also consult the vmf_insert_mixed_prot() documentation when
 922  * @pgprot != @vmf->vma->vm_page_prot.
 923  *
 924  * Return: vm_fault_t value.
 925  */
 926 vm_fault_t vmf_insert_pfn_pud_prot(struct vm_fault *vmf, pfn_t pfn,
 927                                    pgprot_t pgprot, bool write)
 928 {
 929         unsigned long addr = vmf->address & PUD_MASK;
 930         struct vm_area_struct *vma = vmf->vma;
 931
 932         /*
 933          * If we had pud_special, we could avoid all these restrictions,
 934          * but we need to be consistent with PTEs and architectures that
 935          * can't support a 'special' bit.
 936          */
 937         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 938                         !pfn_t_devmap(pfn));
 939         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 940                                                 (VM_PFNMAP|VM_MIXEDMAP));
 941         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 942
 943         if (addr < vma->vm_start || addr >= vma->vm_end)
 944                 return VM_FAULT_SIGBUS;
 945
 946         track_pfn_insert(vma, &pgprot, pfn);
 947
 948         insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
 949         return VM_FAULT_NOPAGE;
 950 }
 951 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud_prot);
 952 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
 953
 954 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
 955                 pmd_t *pmd, int flags)
 956 {
 957         pmd_t _pmd;
 958
 959         _pmd = pmd_mkyoung(*pmd);
 960         if (flags & FOLL_WRITE)
 961                 _pmd = pmd_mkdirty(_pmd);
 962         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
 963                                 pmd, _pmd, flags & FOLL_WRITE))
 964                 update_mmu_cache_pmd(vma, addr, pmd);
 965 }
 966
 967 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 968                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
 969 {
 970         unsigned long pfn = pmd_pfn(*pmd);
 971         struct mm_struct *mm = vma->vm_mm;
 972         struct page *page;
 973
 974         assert_spin_locked(pmd_lockptr(mm, pmd));
 975
 976         /*
 977          * When we COW a devmap PMD entry, we split it into PTEs, so we should
 978          * not be in this function with `flags & FOLL_COW` set.
 979          */
 980         WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
 981
 982         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
 983         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
 984                          (FOLL_PIN | FOLL_GET)))
 985                 return NULL;
 986
 987         if (flags & FOLL_WRITE && !pmd_write(*pmd))
 988                 return NULL;
 989
 990         if (pmd_present(*pmd) && pmd_devmap(*pmd))
 991                 /* pass */;
 992         else
 993                 return NULL;
 994
 995         if (flags & FOLL_TOUCH)
 996                 touch_pmd(vma, addr, pmd, flags);
 997
 998         /*
 999          * device mapped pages can only be returned if the
1000          * caller will manage the page reference count.
1001          */
1002         if (!(flags & (FOLL_GET | FOLL_PIN)))
1003                 return ERR_PTR(-EEXIST);
1004
1005         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1006         *pgmap = get_dev_pagemap(pfn, *pgmap);
1007         if (!*pgmap)
1008                 return ERR_PTR(-EFAULT);
1009         page = pfn_to_page(pfn);
1010         if (!try_grab_page(page, flags))
1011                 page = ERR_PTR(-ENOMEM);
1012
1013         return page;
1014 }
1015
1016 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1017                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1018                   struct vm_area_struct *vma)
1019 {
1020         spinlock_t *dst_ptl, *src_ptl;
1021         struct page *src_page;
1022         pmd_t pmd;
1023         pgtable_t pgtable = NULL;
1024         int ret = -ENOMEM;
1025
1026         /* Skip if can be re-fill on fault */
1027         if (!vma_is_anonymous(vma))
1028                 return 0;
1029
1030         pgtable = pte_alloc_one(dst_mm);
1031         if (unlikely(!pgtable))
1032                 goto out;
1033
1034         dst_ptl = pmd_lock(dst_mm, dst_pmd);
1035         src_ptl = pmd_lockptr(src_mm, src_pmd);
1036         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1037
1038         ret = -EAGAIN;
1039         pmd = *src_pmd;
1040
1041         /*
1042          * Make sure the _PAGE_UFFD_WP bit is cleared if the new VMA
1043          * does not have the VM_UFFD_WP, which means that the uffd
1044          * fork event is not enabled.
1045          */
1046         if (!(vma->vm_flags & VM_UFFD_WP))
1047                 pmd = pmd_clear_uffd_wp(pmd);
1048
1049 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1050         if (unlikely(is_swap_pmd(pmd))) {
1051                 swp_entry_t entry = pmd_to_swp_entry(pmd);
1052
1053                 VM_BUG_ON(!is_pmd_migration_entry(pmd));
1054                 if (is_write_migration_entry(entry)) {
1055                         make_migration_entry_read(&entry);
1056                         pmd = swp_entry_to_pmd(entry);
1057                         if (pmd_swp_soft_dirty(*src_pmd))
1058                                 pmd = pmd_swp_mksoft_dirty(pmd);
1059                         set_pmd_at(src_mm, addr, src_pmd, pmd);
1060                 }
1061                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1062                 mm_inc_nr_ptes(dst_mm);
1063                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1064                 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1065                 ret = 0;
1066                 goto out_unlock;
1067         }
1068 #endif
1069
1070         if (unlikely(!pmd_trans_huge(pmd))) {
1071                 pte_free(dst_mm, pgtable);
1072                 goto out_unlock;
1073         }
1074         /*
1075          * When page table lock is held, the huge zero pmd should not be
1076          * under splitting since we don't split the page itself, only pmd to
1077          * a page table.
1078          */
1079         if (is_huge_zero_pmd(pmd)) {
1080                 struct page *zero_page;
1081                 /*
1082                  * get_huge_zero_page() will never allocate a new page here,
1083                  * since we already have a zero page to copy. It just takes a
1084                  * reference.
1085                  */
1086                 zero_page = mm_get_huge_zero_page(dst_mm);
1087                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1088                                 zero_page);
1089                 ret = 0;
1090                 goto out_unlock;
1091         }
1092
1093         src_page = pmd_page(pmd);
1094         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1095
1096         /*
1097          * If this page is a potentially pinned page, split and retry the fault
1098          * with smaller page size.  Normally this should not happen because the
1099          * userspace should use MADV_DONTFORK upon pinned regions.  This is a
1100          * best effort that the pinned pages won't be replaced by another
1101          * random page during the coming copy-on-write.
1102          */
1103         if (unlikely(is_cow_mapping(vma->vm_flags) &&
1104                      atomic_read(&src_mm->has_pinned) &&
1105                      page_maybe_dma_pinned(src_page))) {
1106                 pte_free(dst_mm, pgtable);
1107                 spin_unlock(src_ptl);
1108                 spin_unlock(dst_ptl);
1109                 __split_huge_pmd(vma, src_pmd, addr, false, NULL);
1110                 return -EAGAIN;
1111         }
1112
1113         get_page(src_page);
1114         page_dup_rmap(src_page, true);
1115         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1116         mm_inc_nr_ptes(dst_mm);
1117         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1118
1119         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1120         pmd = pmd_mkold(pmd_wrprotect(pmd));
1121         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1122
1123         ret = 0;
1124 out_unlock:
1125         spin_unlock(src_ptl);
1126         spin_unlock(dst_ptl);
1127 out:
1128         return ret;
1129 }
1130
1131 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1132 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1133                 pud_t *pud, int flags)
1134 {
1135         pud_t _pud;
1136
1137         _pud = pud_mkyoung(*pud);
1138         if (flags & FOLL_WRITE)
1139                 _pud = pud_mkdirty(_pud);
1140         if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1141                                 pud, _pud, flags & FOLL_WRITE))
1142                 update_mmu_cache_pud(vma, addr, pud);
1143 }
1144
1145 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1146                 pud_t *pud, int flags, struct dev_pagemap **pgmap)
1147 {
1148         unsigned long pfn = pud_pfn(*pud);
1149         struct mm_struct *mm = vma->vm_mm;
1150         struct page *page;
1151
1152         assert_spin_locked(pud_lockptr(mm, pud));
1153
1154         if (flags & FOLL_WRITE && !pud_write(*pud))
1155                 return NULL;
1156
1157         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
1158         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
1159                          (FOLL_PIN | FOLL_GET)))
1160                 return NULL;
1161
1162         if (pud_present(*pud) && pud_devmap(*pud))
1163                 /* pass */;
1164         else
1165                 return NULL;
1166
1167         if (flags & FOLL_TOUCH)
1168                 touch_pud(vma, addr, pud, flags);
1169
1170         /*
1171          * device mapped pages can only be returned if the
1172          * caller will manage the page reference count.
1173          *
1174          * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here:
1175          */
1176         if (!(flags & (FOLL_GET | FOLL_PIN)))
1177                 return ERR_PTR(-EEXIST);
1178
1179         pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1180         *pgmap = get_dev_pagemap(pfn, *pgmap);
1181         if (!*pgmap)
1182                 return ERR_PTR(-EFAULT);
1183         page = pfn_to_page(pfn);
1184         if (!try_grab_page(page, flags))
1185                 page = ERR_PTR(-ENOMEM);
1186
1187         return page;
1188 }
1189
1190 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1191                   pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1192                   struct vm_area_struct *vma)
1193 {
1194         spinlock_t *dst_ptl, *src_ptl;
1195         pud_t pud;
1196         int ret;
1197
1198         dst_ptl = pud_lock(dst_mm, dst_pud);
1199         src_ptl = pud_lockptr(src_mm, src_pud);
1200         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1201
1202         ret = -EAGAIN;
1203         pud = *src_pud;
1204         if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1205                 goto out_unlock;
1206
1207         /*
1208          * When page table lock is held, the huge zero pud should not be
1209          * under splitting since we don't split the page itself, only pud to
1210          * a page table.
1211          */
1212         if (is_huge_zero_pud(pud)) {
1213                 /* No huge zero pud yet */
1214         }
1215
1216         /* Please refer to comments in copy_huge_pmd() */
1217         if (unlikely(is_cow_mapping(vma->vm_flags) &&
1218                      atomic_read(&src_mm->has_pinned) &&
1219                      page_maybe_dma_pinned(pud_page(pud)))) {
1220                 spin_unlock(src_ptl);
1221                 spin_unlock(dst_ptl);
1222                 __split_huge_pud(vma, src_pud, addr);
1223                 return -EAGAIN;
1224         }
1225
1226         pudp_set_wrprotect(src_mm, addr, src_pud);
1227         pud = pud_mkold(pud_wrprotect(pud));
1228         set_pud_at(dst_mm, addr, dst_pud, pud);
1229
1230         ret = 0;
1231 out_unlock:
1232         spin_unlock(src_ptl);
1233         spin_unlock(dst_ptl);
1234         return ret;
1235 }
1236
1237 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1238 {
1239         pud_t entry;
1240         unsigned long haddr;
1241         bool write = vmf->flags & FAULT_FLAG_WRITE;
1242
1243         vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1244         if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1245                 goto unlock;
1246
1247         entry = pud_mkyoung(orig_pud);
1248         if (write)
1249                 entry = pud_mkdirty(entry);
1250         haddr = vmf->address & HPAGE_PUD_MASK;
1251         if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1252                 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1253
1254 unlock:
1255         spin_unlock(vmf->ptl);
1256 }
1257 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1258
1259 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1260 {
1261         pmd_t entry;
1262         unsigned long haddr;
1263         bool write = vmf->flags & FAULT_FLAG_WRITE;
1264
1265         vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1266         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1267                 goto unlock;
1268
1269         entry = pmd_mkyoung(orig_pmd);
1270         if (write)
1271                 entry = pmd_mkdirty(entry);
1272         haddr = vmf->address & HPAGE_PMD_MASK;
1273         if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1274                 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1275
1276 unlock:
1277         spin_unlock(vmf->ptl);
1278 }
1279
1280 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1281 {
1282         struct vm_area_struct *vma = vmf->vma;
1283         struct page *page;
1284         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1285
1286         vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1287         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1288
1289         if (is_huge_zero_pmd(orig_pmd))
1290                 goto fallback;
1291
1292         spin_lock(vmf->ptl);
1293
1294         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1295                 spin_unlock(vmf->ptl);
1296                 return 0;
1297         }
1298
1299         page = pmd_page(orig_pmd);
1300         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1301
1302         /* Lock page for reuse_swap_page() */
1303         if (!trylock_page(page)) {
1304                 get_page(page);
1305                 spin_unlock(vmf->ptl);
1306                 lock_page(page);
1307                 spin_lock(vmf->ptl);
1308                 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1309                         spin_unlock(vmf->ptl);
1310                         unlock_page(page);
1311                         put_page(page);
1312                         return 0;
1313                 }
1314                 put_page(page);
1315         }
1316
1317         /*
1318          * We can only reuse the page if nobody else maps the huge page or it's
1319          * part.
1320          */
1321         if (reuse_swap_page(page, NULL)) {
1322                 pmd_t entry;
1323                 entry = pmd_mkyoung(orig_pmd);
1324                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1325                 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1326                         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1327                 unlock_page(page);
1328                 spin_unlock(vmf->ptl);
1329                 return VM_FAULT_WRITE;
1330         }
1331
1332         unlock_page(page);
1333         spin_unlock(vmf->ptl);
1334 fallback:
1335         __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL);
1336         return VM_FAULT_FALLBACK;
1337 }
1338
1339 /*
1340  * FOLL_FORCE can write to even unwritable pmd's, but only
1341  * after we've gone through a COW cycle and they are dirty.
1342  */
1343 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1344 {
1345         return pmd_write(pmd) ||
1346                ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1347 }
1348
1349 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1350                                    unsigned long addr,
1351                                    pmd_t *pmd,
1352                                    unsigned int flags)
1353 {
1354         struct mm_struct *mm = vma->vm_mm;
1355         struct page *page = NULL;
1356
1357         assert_spin_locked(pmd_lockptr(mm, pmd));
1358
1359         if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1360                 goto out;
1361
1362         /* Avoid dumping huge zero page */
1363         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1364                 return ERR_PTR(-EFAULT);
1365
1366         /* Full NUMA hinting faults to serialise migration in fault paths */
1367         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1368                 goto out;
1369
1370         page = pmd_page(*pmd);
1371         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1372
1373         if (!try_grab_page(page, flags))
1374                 return ERR_PTR(-ENOMEM);
1375
1376         if (flags & FOLL_TOUCH)
1377                 touch_pmd(vma, addr, pmd, flags);
1378
1379         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1380                 /*
1381                  * We don't mlock() pte-mapped THPs. This way we can avoid
1382                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1383                  *
1384                  * For anon THP:
1385                  *
1386                  * In most cases the pmd is the only mapping of the page as we
1387                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1388                  * writable private mappings in populate_vma_page_range().
1389                  *
1390                  * The only scenario when we have the page shared here is if we
1391                  * mlocking read-only mapping shared over fork(). We skip
1392                  * mlocking such pages.
1393                  *
1394                  * For file THP:
1395                  *
1396                  * We can expect PageDoubleMap() to be stable under page lock:
1397                  * for file pages we set it in page_add_file_rmap(), which
1398                  * requires page to be locked.
1399                  */
1400
1401                 if (PageAnon(page) && compound_mapcount(page) != 1)
1402                         goto skip_mlock;
1403                 if (PageDoubleMap(page) || !page->mapping)
1404                         goto skip_mlock;
1405                 if (!trylock_page(page))
1406                         goto skip_mlock;
1407                 if (page->mapping && !PageDoubleMap(page))
1408                         mlock_vma_page(page);
1409                 unlock_page(page);
1410         }
1411 skip_mlock:
1412         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1413         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1414
1415 out:
1416         return page;
1417 }
1418
1419 /* NUMA hinting page fault entry point for trans huge pmds */
1420 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1421 {
1422         struct vm_area_struct *vma = vmf->vma;
1423         struct anon_vma *anon_vma = NULL;
1424         struct page *page;
1425         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1426         int page_nid = NUMA_NO_NODE, this_nid = numa_node_id();
1427         int target_nid, last_cpupid = -1;
1428         bool page_locked;
1429         bool migrated = false;
1430         bool was_writable;
1431         int flags = 0;
1432
1433         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1434         if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1435                 goto out_unlock;
1436
1437         /*
1438          * If there are potential migrations, wait for completion and retry
1439          * without disrupting NUMA hinting information. Do not relock and
1440          * check_same as the page may no longer be mapped.
1441          */
1442         if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1443                 page = pmd_page(*vmf->pmd);
1444                 if (!get_page_unless_zero(page))
1445                         goto out_unlock;
1446                 spin_unlock(vmf->ptl);
1447                 put_and_wait_on_page_locked(page, TASK_UNINTERRUPTIBLE);
1448                 goto out;
1449         }
1450
1451         page = pmd_page(pmd);
1452         BUG_ON(is_huge_zero_page(page));
1453         page_nid = page_to_nid(page);
1454         last_cpupid = page_cpupid_last(page);
1455         count_vm_numa_event(NUMA_HINT_FAULTS);
1456         if (page_nid == this_nid) {
1457                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1458                 flags |= TNF_FAULT_LOCAL;
1459         }
1460
1461         /* See similar comment in do_numa_page for explanation */
1462         if (!pmd_savedwrite(pmd))
1463                 flags |= TNF_NO_GROUP;
1464
1465         /*
1466          * Acquire the page lock to serialise THP migrations but avoid dropping
1467          * page_table_lock if at all possible
1468          */
1469         page_locked = trylock_page(page);
1470         target_nid = mpol_misplaced(page, vma, haddr);
1471         if (target_nid == NUMA_NO_NODE) {
1472                 /* If the page was locked, there are no parallel migrations */
1473                 if (page_locked)
1474                         goto clear_pmdnuma;
1475         }
1476
1477         /* Migration could have started since the pmd_trans_migrating check */
1478         if (!page_locked) {
1479                 page_nid = NUMA_NO_NODE;
1480                 if (!get_page_unless_zero(page))
1481                         goto out_unlock;
1482                 spin_unlock(vmf->ptl);
1483                 put_and_wait_on_page_locked(page, TASK_UNINTERRUPTIBLE);
1484                 goto out;
1485         }
1486
1487         /*
1488          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1489          * to serialises splits
1490          */
1491         get_page(page);
1492         spin_unlock(vmf->ptl);
1493         anon_vma = page_lock_anon_vma_read(page);
1494
1495         /* Confirm the PMD did not change while page_table_lock was released */
1496         spin_lock(vmf->ptl);
1497         if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1498                 unlock_page(page);
1499                 put_page(page);
1500                 page_nid = NUMA_NO_NODE;
1501                 goto out_unlock;
1502         }
1503
1504         /* Bail if we fail to protect against THP splits for any reason */
1505         if (unlikely(!anon_vma)) {
1506                 put_page(page);
1507                 page_nid = NUMA_NO_NODE;
1508                 goto clear_pmdnuma;
1509         }
1510
1511         /*
1512          * Since we took the NUMA fault, we must have observed the !accessible
1513          * bit. Make sure all other CPUs agree with that, to avoid them
1514          * modifying the page we're about to migrate.
1515          *
1516          * Must be done under PTL such that we'll observe the relevant
1517          * inc_tlb_flush_pending().
1518          *
1519          * We are not sure a pending tlb flush here is for a huge page
1520          * mapping or not. Hence use the tlb range variant
1521          */
1522         if (mm_tlb_flush_pending(vma->vm_mm)) {
1523                 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1524                 /*
1525                  * change_huge_pmd() released the pmd lock before
1526                  * invalidating the secondary MMUs sharing the primary
1527                  * MMU pagetables (with ->invalidate_range()). The
1528                  * mmu_notifier_invalidate_range_end() (which
1529                  * internally calls ->invalidate_range()) in
1530                  * change_pmd_range() will run after us, so we can't
1531                  * rely on it here and we need an explicit invalidate.
1532                  */
1533                 mmu_notifier_invalidate_range(vma->vm_mm, haddr,
1534                                               haddr + HPAGE_PMD_SIZE);
1535         }
1536
1537         /*
1538          * Migrate the THP to the requested node, returns with page unlocked
1539          * and access rights restored.
1540          */
1541         spin_unlock(vmf->ptl);
1542
1543         migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1544                                 vmf->pmd, pmd, vmf->address, page, target_nid);
1545         if (migrated) {
1546                 flags |= TNF_MIGRATED;
1547                 page_nid = target_nid;
1548         } else
1549                 flags |= TNF_MIGRATE_FAIL;
1550
1551         goto out;
1552 clear_pmdnuma:
1553         BUG_ON(!PageLocked(page));
1554         was_writable = pmd_savedwrite(pmd);
1555         pmd = pmd_modify(pmd, vma->vm_page_prot);
1556         pmd = pmd_mkyoung(pmd);
1557         if (was_writable)
1558                 pmd = pmd_mkwrite(pmd);
1559         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1560         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1561         unlock_page(page);
1562 out_unlock:
1563         spin_unlock(vmf->ptl);
1564
1565 out:
1566         if (anon_vma)
1567                 page_unlock_anon_vma_read(anon_vma);
1568
1569         if (page_nid != NUMA_NO_NODE)
1570                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1571                                 flags);
1572
1573         return 0;
1574 }
1575
1576 /*
1577  * Return true if we do MADV_FREE successfully on entire pmd page.
1578  * Otherwise, return false.
1579  */
1580 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1581                 pmd_t *pmd, unsigned long addr, unsigned long next)
1582 {
1583         spinlock_t *ptl;
1584         pmd_t orig_pmd;
1585         struct page *page;
1586         struct mm_struct *mm = tlb->mm;
1587         bool ret = false;
1588
1589         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1590
1591         ptl = pmd_trans_huge_lock(pmd, vma);
1592         if (!ptl)
1593                 goto out_unlocked;
1594
1595         orig_pmd = *pmd;
1596         if (is_huge_zero_pmd(orig_pmd))
1597                 goto out;
1598
1599         if (unlikely(!pmd_present(orig_pmd))) {
1600                 VM_BUG_ON(thp_migration_supported() &&
1601                                   !is_pmd_migration_entry(orig_pmd));
1602                 goto out;
1603         }
1604
1605         page = pmd_page(orig_pmd);
1606         /*
1607          * If other processes are mapping this page, we couldn't discard
1608          * the page unless they all do MADV_FREE so let's skip the page.
1609          */
1610         if (page_mapcount(page) != 1)
1611                 goto out;
1612
1613         if (!trylock_page(page))
1614                 goto out;
1615
1616         /*
1617          * If user want to discard part-pages of THP, split it so MADV_FREE
1618          * will deactivate only them.
1619          */
1620         if (next - addr != HPAGE_PMD_SIZE) {
1621                 get_page(page);
1622                 spin_unlock(ptl);
1623                 split_huge_page(page);
1624                 unlock_page(page);
1625                 put_page(page);
1626                 goto out_unlocked;
1627         }
1628
1629         if (PageDirty(page))
1630                 ClearPageDirty(page);
1631         unlock_page(page);
1632
1633         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1634                 pmdp_invalidate(vma, addr, pmd);
1635                 orig_pmd = pmd_mkold(orig_pmd);
1636                 orig_pmd = pmd_mkclean(orig_pmd);
1637
1638                 set_pmd_at(mm, addr, pmd, orig_pmd);
1639                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1640         }
1641
1642         mark_page_lazyfree(page);
1643         ret = true;
1644 out:
1645         spin_unlock(ptl);
1646 out_unlocked:
1647         return ret;
1648 }
1649
1650 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1651 {
1652         pgtable_t pgtable;
1653
1654         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1655         pte_free(mm, pgtable);
1656         mm_dec_nr_ptes(mm);
1657 }
1658
1659 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1660                  pmd_t *pmd, unsigned long addr)
1661 {
1662         pmd_t orig_pmd;
1663         spinlock_t *ptl;
1664
1665         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1666
1667         ptl = __pmd_trans_huge_lock(pmd, vma);
1668         if (!ptl)
1669                 return 0;
1670         /*
1671          * For architectures like ppc64 we look at deposited pgtable
1672          * when calling pmdp_huge_get_and_clear. So do the
1673          * pgtable_trans_huge_withdraw after finishing pmdp related
1674          * operations.
1675          */
1676         orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd,
1677                                                 tlb->fullmm);
1678         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1679         if (vma_is_special_huge(vma)) {
1680                 if (arch_needs_pgtable_deposit())
1681                         zap_deposited_table(tlb->mm, pmd);
1682                 spin_unlock(ptl);
1683                 if (is_huge_zero_pmd(orig_pmd))
1684                         tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1685         } else if (is_huge_zero_pmd(orig_pmd)) {
1686                 zap_deposited_table(tlb->mm, pmd);
1687                 spin_unlock(ptl);
1688                 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1689         } else {
1690                 struct page *page = NULL;
1691                 int flush_needed = 1;
1692
1693                 if (pmd_present(orig_pmd)) {
1694                         page = pmd_page(orig_pmd);
1695                         page_remove_rmap(page, true);
1696                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1697                         VM_BUG_ON_PAGE(!PageHead(page), page);
1698                 } else if (thp_migration_supported()) {
1699                         swp_entry_t entry;
1700
1701                         VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1702                         entry = pmd_to_swp_entry(orig_pmd);
1703                         page = pfn_to_page(swp_offset(entry));
1704                         flush_needed = 0;
1705                 } else
1706                         WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1707
1708                 if (PageAnon(page)) {
1709                         zap_deposited_table(tlb->mm, pmd);
1710                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1711                 } else {
1712                         if (arch_needs_pgtable_deposit())
1713                                 zap_deposited_table(tlb->mm, pmd);
1714                         add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1715                 }
1716
1717                 spin_unlock(ptl);
1718                 if (flush_needed)
1719                         tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1720         }
1721         return 1;
1722 }
1723
1724 #ifndef pmd_move_must_withdraw
1725 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1726                                          spinlock_t *old_pmd_ptl,
1727                                          struct vm_area_struct *vma)
1728 {
1729         /*
1730          * With split pmd lock we also need to move preallocated
1731          * PTE page table if new_pmd is on different PMD page table.
1732          *
1733          * We also don't deposit and withdraw tables for file pages.
1734          */
1735         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1736 }
1737 #endif
1738
1739 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1740 {
1741 #ifdef CONFIG_MEM_SOFT_DIRTY
1742         if (unlikely(is_pmd_migration_entry(pmd)))
1743                 pmd = pmd_swp_mksoft_dirty(pmd);
1744         else if (pmd_present(pmd))
1745                 pmd = pmd_mksoft_dirty(pmd);
1746 #endif
1747         return pmd;
1748 }
1749
1750 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1751                   unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
1752 {
1753         spinlock_t *old_ptl, *new_ptl;
1754         pmd_t pmd;
1755         struct mm_struct *mm = vma->vm_mm;
1756         bool force_flush = false;
1757
1758         /*
1759          * The destination pmd shouldn't be established, free_pgtables()
1760          * should have release it.
1761          */
1762         if (WARN_ON(!pmd_none(*new_pmd))) {
1763                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1764                 return false;
1765         }
1766
1767         /*
1768          * We don't have to worry about the ordering of src and dst
1769          * ptlocks because exclusive mmap_lock prevents deadlock.
1770          */
1771         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1772         if (old_ptl) {
1773                 new_ptl = pmd_lockptr(mm, new_pmd);
1774                 if (new_ptl != old_ptl)
1775                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1776                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1777                 if (pmd_present(pmd))
1778                         force_flush = true;
1779                 VM_BUG_ON(!pmd_none(*new_pmd));
1780
1781                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1782                         pgtable_t pgtable;
1783                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1784                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1785                 }
1786                 pmd = move_soft_dirty_pmd(pmd);
1787                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1788                 if (force_flush)
1789                         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1790                 if (new_ptl != old_ptl)
1791                         spin_unlock(new_ptl);
1792                 spin_unlock(old_ptl);
1793                 return true;
1794         }
1795         return false;
1796 }
1797
1798 /*
1799  * Returns
1800  *  - 0 if PMD could not be locked
1801  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1802  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1803  */
1804 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1805                 unsigned long addr, pgprot_t newprot, unsigned long cp_flags)
1806 {
1807         struct mm_struct *mm = vma->vm_mm;
1808         spinlock_t *ptl;
1809         pmd_t entry;
1810         bool preserve_write;
1811         int ret;
1812         bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
1813         bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
1814         bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
1815
1816         ptl = __pmd_trans_huge_lock(pmd, vma);
1817         if (!ptl)
1818                 return 0;
1819
1820         preserve_write = prot_numa && pmd_write(*pmd);
1821         ret = 1;
1822
1823 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1824         if (is_swap_pmd(*pmd)) {
1825                 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1826
1827                 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1828                 if (is_write_migration_entry(entry)) {
1829                         pmd_t newpmd;
1830                         /*
1831                          * A protection check is difficult so
1832                          * just be safe and disable write
1833                          */
1834                         make_migration_entry_read(&entry);
1835                         newpmd = swp_entry_to_pmd(entry);
1836                         if (pmd_swp_soft_dirty(*pmd))
1837                                 newpmd = pmd_swp_mksoft_dirty(newpmd);
1838                         set_pmd_at(mm, addr, pmd, newpmd);
1839                 }
1840                 goto unlock;
1841         }
1842 #endif
1843
1844         /*
1845          * Avoid trapping faults against the zero page. The read-only
1846          * data is likely to be read-cached on the local CPU and
1847          * local/remote hits to the zero page are not interesting.
1848          */
1849         if (prot_numa && is_huge_zero_pmd(*pmd))
1850                 goto unlock;
1851
1852         if (prot_numa && pmd_protnone(*pmd))
1853                 goto unlock;
1854
1855         /*
1856          * In case prot_numa, we are under mmap_read_lock(mm). It's critical
1857          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1858          * which is also under mmap_read_lock(mm):
1859          *
1860          *      CPU0:                           CPU1:
1861          *                              change_huge_pmd(prot_numa=1)
1862          *                               pmdp_huge_get_and_clear_notify()
1863          * madvise_dontneed()
1864          *  zap_pmd_range()
1865          *   pmd_trans_huge(*pmd) == 0 (without ptl)
1866          *   // skip the pmd
1867          *                               set_pmd_at();
1868          *                               // pmd is re-established
1869          *
1870          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1871          * which may break userspace.
1872          *
1873          * pmdp_invalidate() is required to make sure we don't miss
1874          * dirty/young flags set by hardware.
1875          */
1876         entry = pmdp_invalidate(vma, addr, pmd);
1877
1878         entry = pmd_modify(entry, newprot);
1879         if (preserve_write)
1880                 entry = pmd_mk_savedwrite(entry);
1881         if (uffd_wp) {
1882                 entry = pmd_wrprotect(entry);
1883                 entry = pmd_mkuffd_wp(entry);
1884         } else if (uffd_wp_resolve) {
1885                 /*
1886                  * Leave the write bit to be handled by PF interrupt
1887                  * handler, then things like COW could be properly
1888                  * handled.
1889                  */
1890                 entry = pmd_clear_uffd_wp(entry);
1891         }
1892         ret = HPAGE_PMD_NR;
1893         set_pmd_at(mm, addr, pmd, entry);
1894         BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1895 unlock:
1896         spin_unlock(ptl);
1897         return ret;
1898 }
1899
1900 /*
1901  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1902  *
1903  * Note that if it returns page table lock pointer, this routine returns without
1904  * unlocking page table lock. So callers must unlock it.
1905  */
1906 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1907 {
1908         spinlock_t *ptl;
1909         ptl = pmd_lock(vma->vm_mm, pmd);
1910         if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1911                         pmd_devmap(*pmd)))
1912                 return ptl;
1913         spin_unlock(ptl);
1914         return NULL;
1915 }
1916
1917 /*
1918  * Returns true if a given pud maps a thp, false otherwise.
1919  *
1920  * Note that if it returns true, this routine returns without unlocking page
1921  * table lock. So callers must unlock it.
1922  */
1923 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1924 {
1925         spinlock_t *ptl;
1926
1927         ptl = pud_lock(vma->vm_mm, pud);
1928         if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1929                 return ptl;
1930         spin_unlock(ptl);
1931         return NULL;
1932 }
1933
1934 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1935 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1936                  pud_t *pud, unsigned long addr)
1937 {
1938         spinlock_t *ptl;
1939
1940         ptl = __pud_trans_huge_lock(pud, vma);
1941         if (!ptl)
1942                 return 0;
1943         /*
1944          * For architectures like ppc64 we look at deposited pgtable
1945          * when calling pudp_huge_get_and_clear. So do the
1946          * pgtable_trans_huge_withdraw after finishing pudp related
1947          * operations.
1948          */
1949         pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
1950         tlb_remove_pud_tlb_entry(tlb, pud, addr);
1951         if (vma_is_special_huge(vma)) {
1952                 spin_unlock(ptl);
1953                 /* No zero page support yet */
1954         } else {
1955                 /* No support for anonymous PUD pages yet */
1956                 BUG();
1957         }
1958         return 1;
1959 }
1960
1961 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1962                 unsigned long haddr)
1963 {
1964         VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1965         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1966         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
1967         VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
1968
1969         count_vm_event(THP_SPLIT_PUD);
1970
1971         pudp_huge_clear_flush_notify(vma, haddr, pud);
1972 }
1973
1974 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
1975                 unsigned long address)
1976 {
1977         spinlock_t *ptl;
1978         struct mmu_notifier_range range;
1979
1980         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1981                                 address & HPAGE_PUD_MASK,
1982                                 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
1983         mmu_notifier_invalidate_range_start(&range);
1984         ptl = pud_lock(vma->vm_mm, pud);
1985         if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
1986                 goto out;
1987         __split_huge_pud_locked(vma, pud, range.start);
1988
1989 out:
1990         spin_unlock(ptl);
1991         /*
1992          * No need to double call mmu_notifier->invalidate_range() callback as
1993          * the above pudp_huge_clear_flush_notify() did already call it.
1994          */
1995         mmu_notifier_invalidate_range_only_end(&range);
1996 }
1997 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1998
1999 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2000                 unsigned long haddr, pmd_t *pmd)
2001 {
2002         struct mm_struct *mm = vma->vm_mm;
2003         pgtable_t pgtable;
2004         pmd_t _pmd;
2005         int i;
2006
2007         /*
2008          * Leave pmd empty until pte is filled note that it is fine to delay
2009          * notification until mmu_notifier_invalidate_range_end() as we are
2010          * replacing a zero pmd write protected page with a zero pte write
2011          * protected page.
2012          *
2013          * See Documentation/vm/mmu_notifier.rst
2014          */
2015         pmdp_huge_clear_flush(vma, haddr, pmd);
2016
2017         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2018         pmd_populate(mm, &_pmd, pgtable);
2019
2020         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2021                 pte_t *pte, entry;
2022                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2023                 entry = pte_mkspecial(entry);
2024                 pte = pte_offset_map(&_pmd, haddr);
2025                 VM_BUG_ON(!pte_none(*pte));
2026                 set_pte_at(mm, haddr, pte, entry);
2027                 pte_unmap(pte);
2028         }
2029         smp_wmb(); /* make pte visible before pmd */
2030         pmd_populate(mm, pmd, pgtable);
2031 }
2032
2033 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2034                 unsigned long haddr, bool freeze)
2035 {
2036         struct mm_struct *mm = vma->vm_mm;
2037         struct page *page;
2038         pgtable_t pgtable;
2039         pmd_t old_pmd, _pmd;
2040         bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false;
2041         unsigned long addr;
2042         int i;
2043
2044         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2045         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2046         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2047         VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2048                                 && !pmd_devmap(*pmd));
2049
2050         count_vm_event(THP_SPLIT_PMD);
2051
2052         if (!vma_is_anonymous(vma)) {
2053                 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2054                 /*
2055                  * We are going to unmap this huge page. So
2056                  * just go ahead and zap it
2057                  */
2058                 if (arch_needs_pgtable_deposit())
2059                         zap_deposited_table(mm, pmd);
2060                 if (vma_is_special_huge(vma))
2061                         return;
2062                 page = pmd_page(_pmd);
2063                 if (!PageDirty(page) && pmd_dirty(_pmd))
2064                         set_page_dirty(page);
2065                 if (!PageReferenced(page) && pmd_young(_pmd))
2066                         SetPageReferenced(page);
2067                 page_remove_rmap(page, true);
2068                 put_page(page);
2069                 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2070                 return;
2071         } else if (pmd_trans_huge(*pmd) && is_huge_zero_pmd(*pmd)) {
2072                 /*
2073                  * FIXME: Do we want to invalidate secondary mmu by calling
2074                  * mmu_notifier_invalidate_range() see comments below inside
2075                  * __split_huge_pmd() ?
2076                  *
2077                  * We are going from a zero huge page write protected to zero
2078                  * small page also write protected so it does not seems useful
2079                  * to invalidate secondary mmu at this time.
2080                  */
2081                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2082         }
2083
2084         /*
2085          * Up to this point the pmd is present and huge and userland has the
2086          * whole access to the hugepage during the split (which happens in
2087          * place). If we overwrite the pmd with the not-huge version pointing
2088          * to the pte here (which of course we could if all CPUs were bug
2089          * free), userland could trigger a small page size TLB miss on the
2090          * small sized TLB while the hugepage TLB entry is still established in
2091          * the huge TLB. Some CPU doesn't like that.
2092          * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum
2093          * 383 on page 105. Intel should be safe but is also warns that it's
2094          * only safe if the permission and cache attributes of the two entries
2095          * loaded in the two TLB is identical (which should be the case here).
2096          * But it is generally safer to never allow small and huge TLB entries
2097          * for the same virtual address to be loaded simultaneously. So instead
2098          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2099          * current pmd notpresent (atomically because here the pmd_trans_huge
2100          * must remain set at all times on the pmd until the split is complete
2101          * for this pmd), then we flush the SMP TLB and finally we write the
2102          * non-huge version of the pmd entry with pmd_populate.
2103          */
2104         old_pmd = pmdp_invalidate(vma, haddr, pmd);
2105
2106         pmd_migration = is_pmd_migration_entry(old_pmd);
2107         if (unlikely(pmd_migration)) {
2108                 swp_entry_t entry;
2109
2110                 entry = pmd_to_swp_entry(old_pmd);
2111                 page = pfn_to_page(swp_offset(entry));
2112                 write = is_write_migration_entry(entry);
2113                 young = false;
2114                 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2115                 uffd_wp = pmd_swp_uffd_wp(old_pmd);
2116         } else {
2117                 page = pmd_page(old_pmd);
2118                 if (pmd_dirty(old_pmd))
2119                         SetPageDirty(page);
2120                 write = pmd_write(old_pmd);
2121                 young = pmd_young(old_pmd);
2122                 soft_dirty = pmd_soft_dirty(old_pmd);
2123                 uffd_wp = pmd_uffd_wp(old_pmd);
2124         }
2125         VM_BUG_ON_PAGE(!page_count(page), page);
2126         page_ref_add(page, HPAGE_PMD_NR - 1);
2127
2128         /*
2129          * Withdraw the table only after we mark the pmd entry invalid.
2130          * This's critical for some architectures (Power).
2131          */
2132         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2133         pmd_populate(mm, &_pmd, pgtable);
2134
2135         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2136                 pte_t entry, *pte;
2137                 /*
2138                  * Note that NUMA hinting access restrictions are not
2139                  * transferred to avoid any possibility of altering
2140                  * permissions across VMAs.
2141                  */
2142                 if (freeze || pmd_migration) {
2143                         swp_entry_t swp_entry;
2144                         swp_entry = make_migration_entry(page + i, write);
2145                         entry = swp_entry_to_pte(swp_entry);
2146                         if (soft_dirty)
2147                                 entry = pte_swp_mksoft_dirty(entry);
2148                         if (uffd_wp)
2149                                 entry = pte_swp_mkuffd_wp(entry);
2150                 } else {
2151                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2152                         entry = maybe_mkwrite(entry, vma);
2153                         if (!write)
2154                                 entry = pte_wrprotect(entry);
2155                         if (!young)
2156                                 entry = pte_mkold(entry);
2157                         if (soft_dirty)
2158                                 entry = pte_mksoft_dirty(entry);
2159                         if (uffd_wp)
2160                                 entry = pte_mkuffd_wp(entry);
2161                 }
2162                 pte = pte_offset_map(&_pmd, addr);
2163                 BUG_ON(!pte_none(*pte));
2164                 set_pte_at(mm, addr, pte, entry);
2165                 if (!pmd_migration)
2166                         atomic_inc(&page[i]._mapcount);
2167                 pte_unmap(pte);
2168         }
2169
2170         if (!pmd_migration) {
2171                 /*
2172                  * Set PG_double_map before dropping compound_mapcount to avoid
2173                  * false-negative page_mapped().
2174                  */
2175                 if (compound_mapcount(page) > 1 &&
2176                     !TestSetPageDoubleMap(page)) {
2177                         for (i = 0; i < HPAGE_PMD_NR; i++)
2178                                 atomic_inc(&page[i]._mapcount);
2179                 }
2180
2181                 lock_page_memcg(page);
2182                 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2183                         /* Last compound_mapcount is gone. */
2184                         __mod_lruvec_page_state(page, NR_ANON_THPS,
2185                                                 -HPAGE_PMD_NR);
2186                         if (TestClearPageDoubleMap(page)) {
2187                                 /* No need in mapcount reference anymore */
2188                                 for (i = 0; i < HPAGE_PMD_NR; i++)
2189                                         atomic_dec(&page[i]._mapcount);
2190                         }
2191                 }
2192                 unlock_page_memcg(page);
2193         }
2194
2195         smp_wmb(); /* make pte visible before pmd */
2196         pmd_populate(mm, pmd, pgtable);
2197
2198         if (freeze) {
2199                 for (i = 0; i < HPAGE_PMD_NR; i++) {
2200                         page_remove_rmap(page + i, false);
2201                         put_page(page + i);
2202                 }
2203         }
2204 }
2205
2206 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2207                 unsigned long address, bool freeze, struct page *page)
2208 {
2209         spinlock_t *ptl;
2210         struct mmu_notifier_range range;
2211         bool do_unlock_page = false;
2212         pmd_t _pmd;
2213
2214         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2215                                 address & HPAGE_PMD_MASK,
2216                                 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2217         mmu_notifier_invalidate_range_start(&range);
2218         ptl = pmd_lock(vma->vm_mm, pmd);
2219
2220         /*
2221          * If caller asks to setup a migration entries, we need a page to check
2222          * pmd against. Otherwise we can end up replacing wrong page.
2223          */
2224         VM_BUG_ON(freeze && !page);
2225         if (page) {
2226                 VM_WARN_ON_ONCE(!PageLocked(page));
2227                 if (page != pmd_page(*pmd))
2228                         goto out;
2229         }
2230
2231 repeat:
2232         if (pmd_trans_huge(*pmd)) {
2233                 if (!page) {
2234                         page = pmd_page(*pmd);
2235                         /*
2236                          * An anonymous page must be locked, to ensure that a
2237                          * concurrent reuse_swap_page() sees stable mapcount;
2238                          * but reuse_swap_page() is not used on shmem or file,
2239                          * and page lock must not be taken when zap_pmd_range()
2240                          * calls __split_huge_pmd() while i_mmap_lock is held.
2241                          */
2242                         if (PageAnon(page)) {
2243                                 if (unlikely(!trylock_page(page))) {
2244                                         get_page(page);
2245                                         _pmd = *pmd;
2246                                         spin_unlock(ptl);
2247                                         lock_page(page);
2248                                         spin_lock(ptl);
2249                                         if (unlikely(!pmd_same(*pmd, _pmd))) {
2250                                                 unlock_page(page);
2251                                                 put_page(page);
2252                                                 page = NULL;
2253                                                 goto repeat;
2254                                         }
2255                                         put_page(page);
2256                                 }
2257                                 do_unlock_page = true;
2258                         }
2259                 }
2260                 if (PageMlocked(page))
2261                         clear_page_mlock(page);
2262         } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2263                 goto out;
2264         __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2265 out:
2266         spin_unlock(ptl);
2267         if (do_unlock_page)
2268                 unlock_page(page);
2269         /*
2270          * No need to double call mmu_notifier->invalidate_range() callback.
2271          * They are 3 cases to consider inside __split_huge_pmd_locked():
2272          *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2273          *  2) __split_huge_zero_page_pmd() read only zero page and any write
2274          *    fault will trigger a flush_notify before pointing to a new page
2275          *    (it is fine if the secondary mmu keeps pointing to the old zero
2276          *    page in the meantime)
2277          *  3) Split a huge pmd into pte pointing to the same page. No need
2278          *     to invalidate secondary tlb entry they are all still valid.
2279          *     any further changes to individual pte will notify. So no need
2280          *     to call mmu_notifier->invalidate_range()
2281          */
2282         mmu_notifier_invalidate_range_only_end(&range);
2283 }
2284
2285 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2286                 bool freeze, struct page *page)
2287 {
2288         pgd_t *pgd;
2289         p4d_t *p4d;
2290         pud_t *pud;
2291         pmd_t *pmd;
2292
2293         pgd = pgd_offset(vma->vm_mm, address);
2294         if (!pgd_present(*pgd))
2295                 return;
2296
2297         p4d = p4d_offset(pgd, address);
2298         if (!p4d_present(*p4d))
2299                 return;
2300
2301         pud = pud_offset(p4d, address);
2302         if (!pud_present(*pud))
2303                 return;
2304
2305         pmd = pmd_offset(pud, address);
2306
2307         __split_huge_pmd(vma, pmd, address, freeze, page);
2308 }
2309
2310 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2311                              unsigned long start,
2312                              unsigned long end,
2313                              long adjust_next)
2314 {
2315         /*
2316          * If the new start address isn't hpage aligned and it could
2317          * previously contain an hugepage: check if we need to split
2318          * an huge pmd.
2319          */
2320         if (start & ~HPAGE_PMD_MASK &&
2321             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2322             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2323                 split_huge_pmd_address(vma, start, false, NULL);
2324
2325         /*
2326          * If the new end address isn't hpage aligned and it could
2327          * previously contain an hugepage: check if we need to split
2328          * an huge pmd.
2329          */
2330         if (end & ~HPAGE_PMD_MASK &&
2331             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2332             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2333                 split_huge_pmd_address(vma, end, false, NULL);
2334
2335         /*
2336          * If we're also updating the vma->vm_next->vm_start, if the new
2337          * vm_next->vm_start isn't hpage aligned and it could previously
2338          * contain an hugepage: check if we need to split an huge pmd.
2339          */
2340         if (adjust_next > 0) {
2341                 struct vm_area_struct *next = vma->vm_next;
2342                 unsigned long nstart = next->vm_start;
2343                 nstart += adjust_next;
2344                 if (nstart & ~HPAGE_PMD_MASK &&
2345                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2346                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2347                         split_huge_pmd_address(next, nstart, false, NULL);
2348         }
2349 }
2350
2351 static void unmap_page(struct page *page)
2352 {
2353         enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK |
2354                 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2355         bool unmap_success;
2356
2357         VM_BUG_ON_PAGE(!PageHead(page), page);
2358
2359         if (PageAnon(page))
2360                 ttu_flags |= TTU_SPLIT_FREEZE;
2361
2362         unmap_success = try_to_unmap(page, ttu_flags);
2363         VM_BUG_ON_PAGE(!unmap_success, page);
2364 }
2365
2366 static void remap_page(struct page *page, unsigned int nr)
2367 {
2368         int i;
2369         if (PageTransHuge(page)) {
2370                 remove_migration_ptes(page, page, true);
2371         } else {
2372                 for (i = 0; i < nr; i++)
2373                         remove_migration_ptes(page + i, page + i, true);
2374         }
2375 }
2376
2377 static void lru_add_page_tail(struct page *head, struct page *tail,
2378                 struct lruvec *lruvec, struct list_head *list)
2379 {
2380         VM_BUG_ON_PAGE(!PageHead(head), head);
2381         VM_BUG_ON_PAGE(PageCompound(tail), head);
2382         VM_BUG_ON_PAGE(PageLRU(tail), head);
2383         lockdep_assert_held(&lruvec->lru_lock);
2384
2385         if (list) {
2386                 /* page reclaim is reclaiming a huge page */
2387                 VM_WARN_ON(PageLRU(head));
2388                 get_page(tail);
2389                 list_add_tail(&tail->lru, list);
2390         } else {
2391                 /* head is still on lru (and we have it frozen) */
2392                 VM_WARN_ON(!PageLRU(head));
2393                 SetPageLRU(tail);
2394                 list_add_tail(&tail->lru, &head->lru);
2395         }
2396 }
2397
2398 static void __split_huge_page_tail(struct page *head, int tail,
2399                 struct lruvec *lruvec, struct list_head *list)
2400 {
2401         struct page *page_tail = head + tail;
2402
2403         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2404
2405         /*
2406          * Clone page flags before unfreezing refcount.
2407          *
2408          * After successful get_page_unless_zero() might follow flags change,
2409          * for example lock_page() which set PG_waiters.
2410          */
2411         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2412         page_tail->flags |= (head->flags &
2413                         ((1L << PG_referenced) |
2414                          (1L << PG_swapbacked) |
2415                          (1L << PG_swapcache) |
2416                          (1L << PG_mlocked) |
2417                          (1L << PG_uptodate) |
2418                          (1L << PG_active) |
2419                          (1L << PG_workingset) |
2420                          (1L << PG_locked) |
2421                          (1L << PG_unevictable) |
2422 #ifdef CONFIG_64BIT
2423                          (1L << PG_arch_2) |
2424 #endif
2425                          (1L << PG_dirty)));
2426
2427         /* ->mapping in first tail page is compound_mapcount */
2428         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2429                         page_tail);
2430         page_tail->mapping = head->mapping;
2431         page_tail->index = head->index + tail;
2432
2433         /* Page flags must be visible before we make the page non-compound. */
2434         smp_wmb();
2435
2436         /*
2437          * Clear PageTail before unfreezing page refcount.
2438          *
2439          * After successful get_page_unless_zero() might follow put_page()
2440          * which needs correct compound_head().
2441          */
2442         clear_compound_head(page_tail);
2443
2444         /* Finally unfreeze refcount. Additional reference from page cache. */
2445         page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2446                                           PageSwapCache(head)));
2447
2448         if (page_is_young(head))
2449                 set_page_young(page_tail);
2450         if (page_is_idle(head))
2451                 set_page_idle(page_tail);
2452
2453         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2454
2455         /*
2456          * always add to the tail because some iterators expect new
2457          * pages to show after the currently processed elements - e.g.
2458          * migrate_pages
2459          */
2460         lru_add_page_tail(head, page_tail, lruvec, list);
2461 }
2462
2463 static void __split_huge_page(struct page *page, struct list_head *list,
2464                 pgoff_t end)
2465 {
2466         struct page *head = compound_head(page);
2467         struct lruvec *lruvec;
2468         struct address_space *swap_cache = NULL;
2469         unsigned long offset = 0;
2470         unsigned int nr = thp_nr_pages(head);
2471         int i;
2472
2473         /* complete memcg works before add pages to LRU */
2474         mem_cgroup_split_huge_fixup(head);
2475
2476         if (PageAnon(head) && PageSwapCache(head)) {
2477                 swp_entry_t entry = { .val = page_private(head) };
2478
2479                 offset = swp_offset(entry);
2480                 swap_cache = swap_address_space(entry);
2481                 xa_lock(&swap_cache->i_pages);
2482         }
2483
2484         /* lock lru list/PageCompound, ref freezed by page_ref_freeze */
2485         lruvec = lock_page_lruvec(head);
2486
2487         for (i = nr - 1; i >= 1; i--) {
2488                 __split_huge_page_tail(head, i, lruvec, list);
2489                 /* Some pages can be beyond i_size: drop them from page cache */
2490                 if (head[i].index >= end) {
2491                         ClearPageDirty(head + i);
2492                         __delete_from_page_cache(head + i, NULL);
2493                         if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2494                                 shmem_uncharge(head->mapping->host, 1);
2495                         put_page(head + i);
2496                 } else if (!PageAnon(page)) {
2497                         __xa_store(&head->mapping->i_pages, head[i].index,
2498                                         head + i, 0);
2499                 } else if (swap_cache) {
2500                         __xa_store(&swap_cache->i_pages, offset + i,
2501                                         head + i, 0);
2502                 }
2503         }
2504
2505         ClearPageCompound(head);
2506         unlock_page_lruvec(lruvec);
2507         /* Caller disabled irqs, so they are still disabled here */
2508
2509         split_page_owner(head, nr);
2510
2511         /* See comment in __split_huge_page_tail() */
2512         if (PageAnon(head)) {
2513                 /* Additional pin to swap cache */
2514                 if (PageSwapCache(head)) {
2515                         page_ref_add(head, 2);
2516                         xa_unlock(&swap_cache->i_pages);
2517                 } else {
2518                         page_ref_inc(head);
2519                 }
2520         } else {
2521                 /* Additional pin to page cache */
2522                 page_ref_add(head, 2);
2523                 xa_unlock(&head->mapping->i_pages);
2524         }
2525         local_irq_enable();
2526
2527         remap_page(head, nr);
2528
2529         if (PageSwapCache(head)) {
2530                 swp_entry_t entry = { .val = page_private(head) };
2531
2532                 split_swap_cluster(entry);
2533         }
2534
2535         for (i = 0; i < nr; i++) {
2536                 struct page *subpage = head + i;
2537                 if (subpage == page)
2538                         continue;
2539                 unlock_page(subpage);
2540
2541                 /*
2542                  * Subpages may be freed if there wasn't any mapping
2543                  * like if add_to_swap() is running on a lru page that
2544                  * had its mapping zapped. And freeing these pages
2545                  * requires taking the lru_lock so we do the put_page
2546                  * of the tail pages after the split is complete.
2547                  */
2548                 put_page(subpage);
2549         }
2550 }
2551
2552 int total_mapcount(struct page *page)
2553 {
2554         int i, compound, nr, ret;
2555
2556         VM_BUG_ON_PAGE(PageTail(page), page);
2557
2558         if (likely(!PageCompound(page)))
2559                 return atomic_read(&page->_mapcount) + 1;
2560
2561         compound = compound_mapcount(page);
2562         nr = compound_nr(page);
2563         if (PageHuge(page))
2564                 return compound;
2565         ret = compound;
2566         for (i = 0; i < nr; i++)
2567                 ret += atomic_read(&page[i]._mapcount) + 1;
2568         /* File pages has compound_mapcount included in _mapcount */
2569         if (!PageAnon(page))
2570                 return ret - compound * nr;
2571         if (PageDoubleMap(page))
2572                 ret -= nr;
2573         return ret;
2574 }
2575
2576 /*
2577  * This calculates accurately how many mappings a transparent hugepage
2578  * has (unlike page_mapcount() which isn't fully accurate). This full
2579  * accuracy is primarily needed to know if copy-on-write faults can
2580  * reuse the page and change the mapping to read-write instead of
2581  * copying them. At the same time this returns the total_mapcount too.
2582  *
2583  * The function returns the highest mapcount any one of the subpages
2584  * has. If the return value is one, even if different processes are
2585  * mapping different subpages of the transparent hugepage, they can
2586  * all reuse it, because each process is reusing a different subpage.
2587  *
2588  * The total_mapcount is instead counting all virtual mappings of the
2589  * subpages. If the total_mapcount is equal to "one", it tells the
2590  * caller all mappings belong to the same "mm" and in turn the
2591  * anon_vma of the transparent hugepage can become the vma->anon_vma
2592  * local one as no other process may be mapping any of the subpages.
2593  *
2594  * It would be more accurate to replace page_mapcount() with
2595  * page_trans_huge_mapcount(), however we only use
2596  * page_trans_huge_mapcount() in the copy-on-write faults where we
2597  * need full accuracy to avoid breaking page pinning, because
2598  * page_trans_huge_mapcount() is slower than page_mapcount().
2599  */
2600 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2601 {
2602         int i, ret, _total_mapcount, mapcount;
2603
2604         /* hugetlbfs shouldn't call it */
2605         VM_BUG_ON_PAGE(PageHuge(page), page);
2606
2607         if (likely(!PageTransCompound(page))) {
2608                 mapcount = atomic_read(&page->_mapcount) + 1;
2609                 if (total_mapcount)
2610                         *total_mapcount = mapcount;
2611                 return mapcount;
2612         }
2613
2614         page = compound_head(page);
2615
2616         _total_mapcount = ret = 0;
2617         for (i = 0; i < thp_nr_pages(page); i++) {
2618                 mapcount = atomic_read(&page[i]._mapcount) + 1;
2619                 ret = max(ret, mapcount);
2620                 _total_mapcount += mapcount;
2621         }
2622         if (PageDoubleMap(page)) {
2623                 ret -= 1;
2624                 _total_mapcount -= thp_nr_pages(page);
2625         }
2626         mapcount = compound_mapcount(page);
2627         ret += mapcount;
2628         _total_mapcount += mapcount;
2629         if (total_mapcount)
2630                 *total_mapcount = _total_mapcount;
2631         return ret;
2632 }
2633
2634 /* Racy check whether the huge page can be split */
2635 bool can_split_huge_page(struct page *page, int *pextra_pins)
2636 {
2637         int extra_pins;
2638
2639         /* Additional pins from page cache */
2640         if (PageAnon(page))
2641                 extra_pins = PageSwapCache(page) ? thp_nr_pages(page) : 0;
2642         else
2643                 extra_pins = thp_nr_pages(page);
2644         if (pextra_pins)
2645                 *pextra_pins = extra_pins;
2646         return total_mapcount(page) == page_count(page) - extra_pins - 1;
2647 }
2648
2649 /*
2650  * This function splits huge page into normal pages. @page can point to any
2651  * subpage of huge page to split. Split doesn't change the position of @page.
2652  *
2653  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2654  * The huge page must be locked.
2655  *
2656  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2657  *
2658  * Both head page and tail pages will inherit mapping, flags, and so on from
2659  * the hugepage.
2660  *
2661  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2662  * they are not mapped.
2663  *
2664  * Returns 0 if the hugepage is split successfully.
2665  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2666  * us.
2667  */
2668 int split_huge_page_to_list(struct page *page, struct list_head *list)
2669 {
2670         struct page *head = compound_head(page);
2671         struct deferred_split *ds_queue = get_deferred_split_queue(head);
2672         struct anon_vma *anon_vma = NULL;
2673         struct address_space *mapping = NULL;
2674         int count, mapcount, extra_pins, ret;
2675         pgoff_t end;
2676
2677         VM_BUG_ON_PAGE(is_huge_zero_page(head), head);
2678         VM_BUG_ON_PAGE(!PageLocked(head), head);
2679         VM_BUG_ON_PAGE(!PageCompound(head), head);
2680
2681         if (PageWriteback(head))
2682                 return -EBUSY;
2683
2684         if (PageAnon(head)) {
2685                 /*
2686                  * The caller does not necessarily hold an mmap_lock that would
2687                  * prevent the anon_vma disappearing so we first we take a
2688                  * reference to it and then lock the anon_vma for write. This
2689                  * is similar to page_lock_anon_vma_read except the write lock
2690                  * is taken to serialise against parallel split or collapse
2691                  * operations.
2692                  */
2693                 anon_vma = page_get_anon_vma(head);
2694                 if (!anon_vma) {
2695                         ret = -EBUSY;
2696                         goto out;
2697                 }
2698                 end = -1;
2699                 mapping = NULL;
2700                 anon_vma_lock_write(anon_vma);
2701         } else {
2702                 mapping = head->mapping;
2703
2704                 /* Truncated ? */
2705                 if (!mapping) {
2706                         ret = -EBUSY;
2707                         goto out;
2708                 }
2709
2710                 anon_vma = NULL;
2711                 i_mmap_lock_read(mapping);
2712
2713                 /*
2714                  *__split_huge_page() may need to trim off pages beyond EOF:
2715                  * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2716                  * which cannot be nested inside the page tree lock. So note
2717                  * end now: i_size itself may be changed at any moment, but
2718                  * head page lock is good enough to serialize the trimming.
2719                  */
2720                 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2721         }
2722
2723         /*
2724          * Racy check if we can split the page, before unmap_page() will
2725          * split PMDs
2726          */
2727         if (!can_split_huge_page(head, &extra_pins)) {
2728                 ret = -EBUSY;
2729                 goto out_unlock;
2730         }
2731
2732         unmap_page(head);
2733         VM_BUG_ON_PAGE(compound_mapcount(head), head);
2734
2735         /* block interrupt reentry in xa_lock and spinlock */
2736         local_irq_disable();
2737         if (mapping) {
2738                 XA_STATE(xas, &mapping->i_pages, page_index(head));
2739
2740                 /*
2741                  * Check if the head page is present in page cache.
2742                  * We assume all tail are present too, if head is there.
2743                  */
2744                 xa_lock(&mapping->i_pages);
2745                 if (xas_load(&xas) != head)
2746                         goto fail;
2747         }
2748
2749         /* Prevent deferred_split_scan() touching ->_refcount */
2750         spin_lock(&ds_queue->split_queue_lock);
2751         count = page_count(head);
2752         mapcount = total_mapcount(head);
2753         if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2754                 if (!list_empty(page_deferred_list(head))) {
2755                         ds_queue->split_queue_len--;
2756                         list_del(page_deferred_list(head));
2757                 }
2758                 spin_unlock(&ds_queue->split_queue_lock);
2759                 if (mapping) {
2760                         int nr = thp_nr_pages(head);
2761
2762                         if (PageSwapBacked(head))
2763                                 __mod_lruvec_page_state(head, NR_SHMEM_THPS,
2764                                                         -nr);
2765                         else
2766                                 __mod_lruvec_page_state(head, NR_FILE_THPS,
2767                                                         -nr);
2768                 }
2769
2770                 __split_huge_page(page, list, end);
2771                 ret = 0;
2772         } else {
2773                 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2774                         pr_alert("total_mapcount: %u, page_count(): %u\n",
2775                                         mapcount, count);
2776                         if (PageTail(page))
2777                                 dump_page(head, NULL);
2778                         dump_page(page, "total_mapcount(head) > 0");
2779                         BUG();
2780                 }
2781                 spin_unlock(&ds_queue->split_queue_lock);
2782 fail:           if (mapping)
2783                         xa_unlock(&mapping->i_pages);
2784                 local_irq_enable();
2785                 remap_page(head, thp_nr_pages(head));
2786                 ret = -EBUSY;
2787         }
2788
2789 out_unlock:
2790         if (anon_vma) {
2791                 anon_vma_unlock_write(anon_vma);
2792                 put_anon_vma(anon_vma);
2793         }
2794         if (mapping)
2795                 i_mmap_unlock_read(mapping);
2796 out:
2797         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2798         return ret;
2799 }
2800
2801 void free_transhuge_page(struct page *page)
2802 {
2803         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2804         unsigned long flags;
2805
2806         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2807         if (!list_empty(page_deferred_list(page))) {
2808                 ds_queue->split_queue_len--;
2809                 list_del(page_deferred_list(page));
2810         }
2811         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2812         free_compound_page(page);
2813 }
2814
2815 void deferred_split_huge_page(struct page *page)
2816 {
2817         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2818 #ifdef CONFIG_MEMCG
2819         struct mem_cgroup *memcg = page_memcg(compound_head(page));
2820 #endif
2821         unsigned long flags;
2822
2823         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2824
2825         /*
2826          * The try_to_unmap() in page reclaim path might reach here too,
2827          * this may cause a race condition to corrupt deferred split queue.
2828          * And, if page reclaim is already handling the same page, it is
2829          * unnecessary to handle it again in shrinker.
2830          *
2831          * Check PageSwapCache to determine if the page is being
2832          * handled by page reclaim since THP swap would add the page into
2833          * swap cache before calling try_to_unmap().
2834          */
2835         if (PageSwapCache(page))
2836                 return;
2837
2838         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2839         if (list_empty(page_deferred_list(page))) {
2840                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2841                 list_add_tail(page_deferred_list(page), &ds_queue->split_queue);
2842                 ds_queue->split_queue_len++;
2843 #ifdef CONFIG_MEMCG
2844                 if (memcg)
2845                         memcg_set_shrinker_bit(memcg, page_to_nid(page),
2846                                                deferred_split_shrinker.id);
2847 #endif
2848         }
2849         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2850 }
2851
2852 static unsigned long deferred_split_count(struct shrinker *shrink,
2853                 struct shrink_control *sc)
2854 {
2855         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2856         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2857
2858 #ifdef CONFIG_MEMCG
2859         if (sc->memcg)
2860                 ds_queue = &sc->memcg->deferred_split_queue;
2861 #endif
2862         return READ_ONCE(ds_queue->split_queue_len);
2863 }
2864
2865 static unsigned long deferred_split_scan(struct shrinker *shrink,
2866                 struct shrink_control *sc)
2867 {
2868         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2869         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2870         unsigned long flags;
2871         LIST_HEAD(list), *pos, *next;
2872         struct page *page;
2873         int split = 0;
2874
2875 #ifdef CONFIG_MEMCG
2876         if (sc->memcg)
2877                 ds_queue = &sc->memcg->deferred_split_queue;
2878 #endif
2879
2880         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2881         /* Take pin on all head pages to avoid freeing them under us */
2882         list_for_each_safe(pos, next, &ds_queue->split_queue) {
2883                 page = list_entry((void *)pos, struct page, mapping);
2884                 page = compound_head(page);
2885                 if (get_page_unless_zero(page)) {
2886                         list_move(page_deferred_list(page), &list);
2887                 } else {
2888                         /* We lost race with put_compound_page() */
2889                         list_del_init(page_deferred_list(page));
2890                         ds_queue->split_queue_len--;
2891                 }
2892                 if (!--sc->nr_to_scan)
2893                         break;
2894         }
2895         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2896
2897         list_for_each_safe(pos, next, &list) {
2898                 page = list_entry((void *)pos, struct page, mapping);
2899                 if (!trylock_page(page))
2900                         goto next;
2901                 /* split_huge_page() removes page from list on success */
2902                 if (!split_huge_page(page))
2903                         split++;
2904                 unlock_page(page);
2905 next:
2906                 put_page(page);
2907         }
2908
2909         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2910         list_splice_tail(&list, &ds_queue->split_queue);
2911         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2912
2913         /*
2914          * Stop shrinker if we didn't split any page, but the queue is empty.
2915          * This can happen if pages were freed under us.
2916          */
2917         if (!split && list_empty(&ds_queue->split_queue))
2918                 return SHRINK_STOP;
2919         return split;
2920 }
2921
2922 static struct shrinker deferred_split_shrinker = {
2923         .count_objects = deferred_split_count,
2924         .scan_objects = deferred_split_scan,
2925         .seeks = DEFAULT_SEEKS,
2926         .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
2927                  SHRINKER_NONSLAB,
2928 };
2929
2930 #ifdef CONFIG_DEBUG_FS
2931 static int split_huge_pages_set(void *data, u64 val)
2932 {
2933         struct zone *zone;
2934         struct page *page;
2935         unsigned long pfn, max_zone_pfn;
2936         unsigned long total = 0, split = 0;
2937
2938         if (val != 1)
2939                 return -EINVAL;
2940
2941         for_each_populated_zone(zone) {
2942                 max_zone_pfn = zone_end_pfn(zone);
2943                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2944                         if (!pfn_valid(pfn))
2945                                 continue;
2946
2947                         page = pfn_to_page(pfn);
2948                         if (!get_page_unless_zero(page))
2949                                 continue;
2950
2951                         if (zone != page_zone(page))
2952                                 goto next;
2953
2954                         if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2955                                 goto next;
2956
2957                         total++;
2958                         lock_page(page);
2959                         if (!split_huge_page(page))
2960                                 split++;
2961                         unlock_page(page);
2962 next:
2963                         put_page(page);
2964                 }
2965         }
2966
2967         pr_info("%lu of %lu THP split\n", split, total);
2968
2969         return 0;
2970 }
2971 DEFINE_DEBUGFS_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2972                 "%llu\n");
2973
2974 static int __init split_huge_pages_debugfs(void)
2975 {
2976         debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2977                             &split_huge_pages_fops);
2978         return 0;
2979 }
2980 late_initcall(split_huge_pages_debugfs);
2981 #endif
2982
2983 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2984 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
2985                 struct page *page)
2986 {
2987         struct vm_area_struct *vma = pvmw->vma;
2988         struct mm_struct *mm = vma->vm_mm;
2989         unsigned long address = pvmw->address;
2990         pmd_t pmdval;
2991         swp_entry_t entry;
2992         pmd_t pmdswp;
2993
2994         if (!(pvmw->pmd && !pvmw->pte))
2995                 return;
2996
2997         flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
2998         pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
2999         if (pmd_dirty(pmdval))
3000                 set_page_dirty(page);
3001         entry = make_migration_entry(page, pmd_write(pmdval));
3002         pmdswp = swp_entry_to_pmd(entry);
3003         if (pmd_soft_dirty(pmdval))
3004                 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
3005         set_pmd_at(mm, address, pvmw->pmd, pmdswp);
3006         page_remove_rmap(page, true);
3007         put_page(page);
3008 }
3009
3010 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
3011 {
3012         struct vm_area_struct *vma = pvmw->vma;
3013         struct mm_struct *mm = vma->vm_mm;
3014         unsigned long address = pvmw->address;
3015         unsigned long mmun_start = address & HPAGE_PMD_MASK;
3016         pmd_t pmde;
3017         swp_entry_t entry;
3018
3019         if (!(pvmw->pmd && !pvmw->pte))
3020                 return;
3021
3022         entry = pmd_to_swp_entry(*pvmw->pmd);
3023         get_page(new);
3024         pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
3025         if (pmd_swp_soft_dirty(*pvmw->pmd))
3026                 pmde = pmd_mksoft_dirty(pmde);
3027         if (is_write_migration_entry(entry))
3028                 pmde = maybe_pmd_mkwrite(pmde, vma);
3029
3030         flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
3031         if (PageAnon(new))
3032                 page_add_anon_rmap(new, vma, mmun_start, true);
3033         else
3034                 page_add_file_rmap(new, true);
3035         set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
3036         if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new))
3037                 mlock_vma_page(new);
3038         update_mmu_cache_pmd(vma, address, pvmw->pmd);
3039 }
3040 #endif