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         if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
 167                 return sprintf(buf, "[always] madvise never\n");
 168         else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
 169                 return sprintf(buf, "always [madvise] never\n");
 170         else
 171                 return sprintf(buf, "always madvise [never]\n");
 172 }
 173
 174 static ssize_t enabled_store(struct kobject *kobj,
 175                              struct kobj_attribute *attr,
 176                              const char *buf, size_t count)
 177 {
 178         ssize_t ret = count;
 179
 180         if (sysfs_streq(buf, "always")) {
 181                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 182                 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 183         } else if (sysfs_streq(buf, "madvise")) {
 184                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 185                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 186         } else if (sysfs_streq(buf, "never")) {
 187                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 188                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 189         } else
 190                 ret = -EINVAL;
 191
 192         if (ret > 0) {
 193                 int err = start_stop_khugepaged();
 194                 if (err)
 195                         ret = err;
 196         }
 197         return ret;
 198 }
 199 static struct kobj_attribute enabled_attr =
 200         __ATTR(enabled, 0644, enabled_show, enabled_store);
 201
 202 ssize_t single_hugepage_flag_show(struct kobject *kobj,
 203                                 struct kobj_attribute *attr, char *buf,
 204                                 enum transparent_hugepage_flag flag)
 205 {
 206         return sprintf(buf, "%d\n",
 207                        !!test_bit(flag, &transparent_hugepage_flags));
 208 }
 209
 210 ssize_t single_hugepage_flag_store(struct kobject *kobj,
 211                                  struct kobj_attribute *attr,
 212                                  const char *buf, size_t count,
 213                                  enum transparent_hugepage_flag flag)
 214 {
 215         unsigned long value;
 216         int ret;
 217
 218         ret = kstrtoul(buf, 10, &value);
 219         if (ret < 0)
 220                 return ret;
 221         if (value > 1)
 222                 return -EINVAL;
 223
 224         if (value)
 225                 set_bit(flag, &transparent_hugepage_flags);
 226         else
 227                 clear_bit(flag, &transparent_hugepage_flags);
 228
 229         return count;
 230 }
 231
 232 static ssize_t defrag_show(struct kobject *kobj,
 233                            struct kobj_attribute *attr, char *buf)
 234 {
 235         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 236                 return sprintf(buf, "[always] defer defer+madvise madvise never\n");
 237         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 238                 return sprintf(buf, "always [defer] defer+madvise madvise never\n");
 239         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 240                 return sprintf(buf, "always defer [defer+madvise] madvise never\n");
 241         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 242                 return sprintf(buf, "always defer defer+madvise [madvise] never\n");
 243         return sprintf(buf, "always defer defer+madvise madvise [never]\n");
 244 }
 245
 246 static ssize_t defrag_store(struct kobject *kobj,
 247                             struct kobj_attribute *attr,
 248                             const char *buf, size_t count)
 249 {
 250         if (sysfs_streq(buf, "always")) {
 251                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 252                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 253                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 254                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 255         } else if (sysfs_streq(buf, "defer+madvise")) {
 256                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 257                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 258                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 259                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 260         } else if (sysfs_streq(buf, "defer")) {
 261                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 262                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 263                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 264                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 265         } else if (sysfs_streq(buf, "madvise")) {
 266                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 267                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 268                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 269                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 270         } else if (sysfs_streq(buf, "never")) {
 271                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 272                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 273                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 274                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 275         } else
 276                 return -EINVAL;
 277
 278         return count;
 279 }
 280 static struct kobj_attribute defrag_attr =
 281         __ATTR(defrag, 0644, defrag_show, defrag_store);
 282
 283 static ssize_t use_zero_page_show(struct kobject *kobj,
 284                 struct kobj_attribute *attr, char *buf)
 285 {
 286         return single_hugepage_flag_show(kobj, attr, buf,
 287                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 288 }
 289 static ssize_t use_zero_page_store(struct kobject *kobj,
 290                 struct kobj_attribute *attr, const char *buf, size_t count)
 291 {
 292         return single_hugepage_flag_store(kobj, attr, buf, count,
 293                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 294 }
 295 static struct kobj_attribute use_zero_page_attr =
 296         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 297
 298 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
 299                 struct kobj_attribute *attr, char *buf)
 300 {
 301         return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE);
 302 }
 303 static struct kobj_attribute hpage_pmd_size_attr =
 304         __ATTR_RO(hpage_pmd_size);
 305
 306 static struct attribute *hugepage_attr[] = {
 307         &enabled_attr.attr,
 308         &defrag_attr.attr,
 309         &use_zero_page_attr.attr,
 310         &hpage_pmd_size_attr.attr,
 311 #ifdef CONFIG_SHMEM
 312         &shmem_enabled_attr.attr,
 313 #endif
 314         NULL,
 315 };
 316
 317 static const struct attribute_group hugepage_attr_group = {
 318         .attrs = hugepage_attr,
 319 };
 320
 321 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 322 {
 323         int err;
 324
 325         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 326         if (unlikely(!*hugepage_kobj)) {
 327                 pr_err("failed to create transparent hugepage kobject\n");
 328                 return -ENOMEM;
 329         }
 330
 331         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 332         if (err) {
 333                 pr_err("failed to register transparent hugepage group\n");
 334                 goto delete_obj;
 335         }
 336
 337         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 338         if (err) {
 339                 pr_err("failed to register transparent hugepage group\n");
 340                 goto remove_hp_group;
 341         }
 342
 343         return 0;
 344
 345 remove_hp_group:
 346         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 347 delete_obj:
 348         kobject_put(*hugepage_kobj);
 349         return err;
 350 }
 351
 352 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 353 {
 354         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 355         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 356         kobject_put(hugepage_kobj);
 357 }
 358 #else
 359 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 360 {
 361         return 0;
 362 }
 363
 364 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 365 {
 366 }
 367 #endif /* CONFIG_SYSFS */
 368
 369 static int __init hugepage_init(void)
 370 {
 371         int err;
 372         struct kobject *hugepage_kobj;
 373
 374         if (!has_transparent_hugepage()) {
 375                 transparent_hugepage_flags = 0;
 376                 return -EINVAL;
 377         }
 378
 379         /*
 380          * hugepages can't be allocated by the buddy allocator
 381          */
 382         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
 383         /*
 384          * we use page->mapping and page->index in second tail page
 385          * as list_head: assuming THP order >= 2
 386          */
 387         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 388
 389         err = hugepage_init_sysfs(&hugepage_kobj);
 390         if (err)
 391                 goto err_sysfs;
 392
 393         err = khugepaged_init();
 394         if (err)
 395                 goto err_slab;
 396
 397         err = register_shrinker(&huge_zero_page_shrinker);
 398         if (err)
 399                 goto err_hzp_shrinker;
 400         err = register_shrinker(&deferred_split_shrinker);
 401         if (err)
 402                 goto err_split_shrinker;
 403
 404         /*
 405          * By default disable transparent hugepages on smaller systems,
 406          * where the extra memory used could hurt more than TLB overhead
 407          * is likely to save.  The admin can still enable it through /sys.
 408          */
 409         if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
 410                 transparent_hugepage_flags = 0;
 411                 return 0;
 412         }
 413
 414         err = start_stop_khugepaged();
 415         if (err)
 416                 goto err_khugepaged;
 417
 418         return 0;
 419 err_khugepaged:
 420         unregister_shrinker(&deferred_split_shrinker);
 421 err_split_shrinker:
 422         unregister_shrinker(&huge_zero_page_shrinker);
 423 err_hzp_shrinker:
 424         khugepaged_destroy();
 425 err_slab:
 426         hugepage_exit_sysfs(hugepage_kobj);
 427 err_sysfs:
 428         return err;
 429 }
 430 subsys_initcall(hugepage_init);
 431
 432 static int __init setup_transparent_hugepage(char *str)
 433 {
 434         int ret = 0;
 435         if (!str)
 436                 goto out;
 437         if (!strcmp(str, "always")) {
 438                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 439                         &transparent_hugepage_flags);
 440                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 441                           &transparent_hugepage_flags);
 442                 ret = 1;
 443         } else if (!strcmp(str, "madvise")) {
 444                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 445                           &transparent_hugepage_flags);
 446                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 447                         &transparent_hugepage_flags);
 448                 ret = 1;
 449         } else if (!strcmp(str, "never")) {
 450                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 451                           &transparent_hugepage_flags);
 452                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 453                           &transparent_hugepage_flags);
 454                 ret = 1;
 455         }
 456 out:
 457         if (!ret)
 458                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 459         return ret;
 460 }
 461 __setup("transparent_hugepage=", setup_transparent_hugepage);
 462
 463 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 464 {
 465         if (likely(vma->vm_flags & VM_WRITE))
 466                 pmd = pmd_mkwrite(pmd);
 467         return pmd;
 468 }
 469
 470 #ifdef CONFIG_MEMCG
 471 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
 472 {
 473         struct mem_cgroup *memcg = compound_head(page)->mem_cgroup;
 474         struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
 475
 476         if (memcg)
 477                 return &memcg->deferred_split_queue;
 478         else
 479                 return &pgdat->deferred_split_queue;
 480 }
 481 #else
 482 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
 483 {
 484         struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
 485
 486         return &pgdat->deferred_split_queue;
 487 }
 488 #endif
 489
 490 void prep_transhuge_page(struct page *page)
 491 {
 492         /*
 493          * we use page->mapping and page->indexlru in second tail page
 494          * as list_head: assuming THP order >= 2
 495          */
 496
 497         INIT_LIST_HEAD(page_deferred_list(page));
 498         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 499 }
 500
 501 bool is_transparent_hugepage(struct page *page)
 502 {
 503         if (!PageCompound(page))
 504                 return false;
 505
 506         page = compound_head(page);
 507         return is_huge_zero_page(page) ||
 508                page[1].compound_dtor == TRANSHUGE_PAGE_DTOR;
 509 }
 510 EXPORT_SYMBOL_GPL(is_transparent_hugepage);
 511
 512 static unsigned long __thp_get_unmapped_area(struct file *filp,
 513                 unsigned long addr, unsigned long len,
 514                 loff_t off, unsigned long flags, unsigned long size)
 515 {
 516         loff_t off_end = off + len;
 517         loff_t off_align = round_up(off, size);
 518         unsigned long len_pad, ret;
 519
 520         if (off_end <= off_align || (off_end - off_align) < size)
 521                 return 0;
 522
 523         len_pad = len + size;
 524         if (len_pad < len || (off + len_pad) < off)
 525                 return 0;
 526
 527         ret = current->mm->get_unmapped_area(filp, addr, len_pad,
 528                                               off >> PAGE_SHIFT, flags);
 529
 530         /*
 531          * The failure might be due to length padding. The caller will retry
 532          * without the padding.
 533          */
 534         if (IS_ERR_VALUE(ret))
 535                 return 0;
 536
 537         /*
 538          * Do not try to align to THP boundary if allocation at the address
 539          * hint succeeds.
 540          */
 541         if (ret == addr)
 542                 return addr;
 543
 544         ret += (off - ret) & (size - 1);
 545         return ret;
 546 }
 547
 548 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 549                 unsigned long len, unsigned long pgoff, unsigned long flags)
 550 {
 551         unsigned long ret;
 552         loff_t off = (loff_t)pgoff << PAGE_SHIFT;
 553
 554         if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
 555                 goto out;
 556
 557         ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE);
 558         if (ret)
 559                 return ret;
 560 out:
 561         return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
 562 }
 563 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
 564
 565 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
 566                         struct page *page, gfp_t gfp)
 567 {
 568         struct vm_area_struct *vma = vmf->vma;
 569         pgtable_t pgtable;
 570         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 571         vm_fault_t ret = 0;
 572
 573         VM_BUG_ON_PAGE(!PageCompound(page), page);
 574
 575         if (mem_cgroup_charge(page, vma->vm_mm, gfp)) {
 576                 put_page(page);
 577                 count_vm_event(THP_FAULT_FALLBACK);
 578                 count_vm_event(THP_FAULT_FALLBACK_CHARGE);
 579                 return VM_FAULT_FALLBACK;
 580         }
 581         cgroup_throttle_swaprate(page, gfp);
 582
 583         pgtable = pte_alloc_one(vma->vm_mm);
 584         if (unlikely(!pgtable)) {
 585                 ret = VM_FAULT_OOM;
 586                 goto release;
 587         }
 588
 589         clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
 590         /*
 591          * The memory barrier inside __SetPageUptodate makes sure that
 592          * clear_huge_page writes become visible before the set_pmd_at()
 593          * write.
 594          */
 595         __SetPageUptodate(page);
 596
 597         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 598         if (unlikely(!pmd_none(*vmf->pmd))) {
 599                 goto unlock_release;
 600         } else {
 601                 pmd_t entry;
 602
 603                 ret = check_stable_address_space(vma->vm_mm);
 604                 if (ret)
 605                         goto unlock_release;
 606
 607                 /* Deliver the page fault to userland */
 608                 if (userfaultfd_missing(vma)) {
 609                         vm_fault_t ret2;
 610
 611                         spin_unlock(vmf->ptl);
 612                         put_page(page);
 613                         pte_free(vma->vm_mm, pgtable);
 614                         ret2 = handle_userfault(vmf, VM_UFFD_MISSING);
 615                         VM_BUG_ON(ret2 & VM_FAULT_FALLBACK);
 616                         return ret2;
 617                 }
 618
 619                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 620                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 621                 page_add_new_anon_rmap(page, vma, haddr, true);
 622                 lru_cache_add_inactive_or_unevictable(page, vma);
 623                 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
 624                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
 625                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 626                 mm_inc_nr_ptes(vma->vm_mm);
 627                 spin_unlock(vmf->ptl);
 628                 count_vm_event(THP_FAULT_ALLOC);
 629                 count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC);
 630         }
 631
 632         return 0;
 633 unlock_release:
 634         spin_unlock(vmf->ptl);
 635 release:
 636         if (pgtable)
 637                 pte_free(vma->vm_mm, pgtable);
 638         put_page(page);
 639         return ret;
 640
 641 }
 642
 643 /*
 644  * always: directly stall for all thp allocations
 645  * defer: wake kswapd and fail if not immediately available
 646  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
 647  *                fail if not immediately available
 648  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
 649  *          available
 650  * never: never stall for any thp allocation
 651  */
 652 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
 653 {
 654         const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
 655
 656         /* Always do synchronous compaction */
 657         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 658                 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
 659
 660         /* Kick kcompactd and fail quickly */
 661         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 662                 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
 663
 664         /* Synchronous compaction if madvised, otherwise kick kcompactd */
 665         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 666                 return GFP_TRANSHUGE_LIGHT |
 667                         (vma_madvised ? __GFP_DIRECT_RECLAIM :
 668                                         __GFP_KSWAPD_RECLAIM);
 669
 670         /* Only do synchronous compaction if madvised */
 671         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 672                 return GFP_TRANSHUGE_LIGHT |
 673                        (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
 674
 675         return GFP_TRANSHUGE_LIGHT;
 676 }
 677
 678 /* Caller must hold page table lock. */
 679 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 680                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 681                 struct page *zero_page)
 682 {
 683         pmd_t entry;
 684         if (!pmd_none(*pmd))
 685                 return false;
 686         entry = mk_pmd(zero_page, vma->vm_page_prot);
 687         entry = pmd_mkhuge(entry);
 688         if (pgtable)
 689                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 690         set_pmd_at(mm, haddr, pmd, entry);
 691         mm_inc_nr_ptes(mm);
 692         return true;
 693 }
 694
 695 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
 696 {
 697         struct vm_area_struct *vma = vmf->vma;
 698         gfp_t gfp;
 699         struct page *page;
 700         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 701
 702         if (!transhuge_vma_suitable(vma, haddr))
 703                 return VM_FAULT_FALLBACK;
 704         if (unlikely(anon_vma_prepare(vma)))
 705                 return VM_FAULT_OOM;
 706         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 707                 return VM_FAULT_OOM;
 708         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
 709                         !mm_forbids_zeropage(vma->vm_mm) &&
 710                         transparent_hugepage_use_zero_page()) {
 711                 pgtable_t pgtable;
 712                 struct page *zero_page;
 713                 vm_fault_t ret;
 714                 pgtable = pte_alloc_one(vma->vm_mm);
 715                 if (unlikely(!pgtable))
 716                         return VM_FAULT_OOM;
 717                 zero_page = mm_get_huge_zero_page(vma->vm_mm);
 718                 if (unlikely(!zero_page)) {
 719                         pte_free(vma->vm_mm, pgtable);
 720                         count_vm_event(THP_FAULT_FALLBACK);
 721                         return VM_FAULT_FALLBACK;
 722                 }
 723                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 724                 ret = 0;
 725                 if (pmd_none(*vmf->pmd)) {
 726                         ret = check_stable_address_space(vma->vm_mm);
 727                         if (ret) {
 728                                 spin_unlock(vmf->ptl);
 729                                 pte_free(vma->vm_mm, pgtable);
 730                         } else if (userfaultfd_missing(vma)) {
 731                                 spin_unlock(vmf->ptl);
 732                                 pte_free(vma->vm_mm, pgtable);
 733                                 ret = handle_userfault(vmf, VM_UFFD_MISSING);
 734                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 735                         } else {
 736                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
 737                                                    haddr, vmf->pmd, zero_page);
 738                                 spin_unlock(vmf->ptl);
 739                         }
 740                 } else {
 741                         spin_unlock(vmf->ptl);
 742                         pte_free(vma->vm_mm, pgtable);
 743                 }
 744                 return ret;
 745         }
 746         gfp = alloc_hugepage_direct_gfpmask(vma);
 747         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 748         if (unlikely(!page)) {
 749                 count_vm_event(THP_FAULT_FALLBACK);
 750                 return VM_FAULT_FALLBACK;
 751         }
 752         prep_transhuge_page(page);
 753         return __do_huge_pmd_anonymous_page(vmf, page, gfp);
 754 }
 755
 756 #ifndef CONFIG_FINEGRAINED_THP
 757 vm_fault_t do_huge_pte_anonymous_page(struct vm_fault *vmf)
 758 {
 759         return VM_FAULT_FALLBACK;
 760 }
 761 #endif /* CONFIG_FINEGRAINED_THP */
 762
 763 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 764                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
 765                 pgtable_t pgtable)
 766 {
 767         struct mm_struct *mm = vma->vm_mm;
 768         pmd_t entry;
 769         spinlock_t *ptl;
 770
 771         ptl = pmd_lock(mm, pmd);
 772         if (!pmd_none(*pmd)) {
 773                 if (write) {
 774                         if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
 775                                 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
 776                                 goto out_unlock;
 777                         }
 778                         entry = pmd_mkyoung(*pmd);
 779                         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 780                         if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
 781                                 update_mmu_cache_pmd(vma, addr, pmd);
 782                 }
 783
 784                 goto out_unlock;
 785         }
 786
 787         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 788         if (pfn_t_devmap(pfn))
 789                 entry = pmd_mkdevmap(entry);
 790         if (write) {
 791                 entry = pmd_mkyoung(pmd_mkdirty(entry));
 792                 entry = maybe_pmd_mkwrite(entry, vma);
 793         }
 794
 795         if (pgtable) {
 796                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 797                 mm_inc_nr_ptes(mm);
 798                 pgtable = NULL;
 799         }
 800
 801         set_pmd_at(mm, addr, pmd, entry);
 802         update_mmu_cache_pmd(vma, addr, pmd);
 803
 804 out_unlock:
 805         spin_unlock(ptl);
 806         if (pgtable)
 807                 pte_free(mm, pgtable);
 808 }
 809
 810 /**
 811  * vmf_insert_pfn_pmd_prot - insert a pmd size pfn
 812  * @vmf: Structure describing the fault
 813  * @pfn: pfn to insert
 814  * @pgprot: page protection to use
 815  * @write: whether it's a write fault
 816  *
 817  * Insert a pmd size pfn. See vmf_insert_pfn() for additional info and
 818  * also consult the vmf_insert_mixed_prot() documentation when
 819  * @pgprot != @vmf->vma->vm_page_prot.
 820  *
 821  * Return: vm_fault_t value.
 822  */
 823 vm_fault_t vmf_insert_pfn_pmd_prot(struct vm_fault *vmf, pfn_t pfn,
 824                                    pgprot_t pgprot, bool write)
 825 {
 826         unsigned long addr = vmf->address & PMD_MASK;
 827         struct vm_area_struct *vma = vmf->vma;
 828         pgtable_t pgtable = NULL;
 829
 830         /*
 831          * If we had pmd_special, we could avoid all these restrictions,
 832          * but we need to be consistent with PTEs and architectures that
 833          * can't support a 'special' bit.
 834          */
 835         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 836                         !pfn_t_devmap(pfn));
 837         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 838                                                 (VM_PFNMAP|VM_MIXEDMAP));
 839         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 840
 841         if (addr < vma->vm_start || addr >= vma->vm_end)
 842                 return VM_FAULT_SIGBUS;
 843
 844         if (arch_needs_pgtable_deposit()) {
 845                 pgtable = pte_alloc_one(vma->vm_mm);
 846                 if (!pgtable)
 847                         return VM_FAULT_OOM;
 848         }
 849
 850         track_pfn_insert(vma, &pgprot, pfn);
 851
 852         insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
 853         return VM_FAULT_NOPAGE;
 854 }
 855 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd_prot);
 856
 857 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 858 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
 859 {
 860         if (likely(vma->vm_flags & VM_WRITE))
 861                 pud = pud_mkwrite(pud);
 862         return pud;
 863 }
 864
 865 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 866                 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
 867 {
 868         struct mm_struct *mm = vma->vm_mm;
 869         pud_t entry;
 870         spinlock_t *ptl;
 871
 872         ptl = pud_lock(mm, pud);
 873         if (!pud_none(*pud)) {
 874                 if (write) {
 875                         if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
 876                                 WARN_ON_ONCE(!is_huge_zero_pud(*pud));
 877                                 goto out_unlock;
 878                         }
 879                         entry = pud_mkyoung(*pud);
 880                         entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
 881                         if (pudp_set_access_flags(vma, addr, pud, entry, 1))
 882                                 update_mmu_cache_pud(vma, addr, pud);
 883                 }
 884                 goto out_unlock;
 885         }
 886
 887         entry = pud_mkhuge(pfn_t_pud(pfn, prot));
 888         if (pfn_t_devmap(pfn))
 889                 entry = pud_mkdevmap(entry);
 890         if (write) {
 891                 entry = pud_mkyoung(pud_mkdirty(entry));
 892                 entry = maybe_pud_mkwrite(entry, vma);
 893         }
 894         set_pud_at(mm, addr, pud, entry);
 895         update_mmu_cache_pud(vma, addr, pud);
 896
 897 out_unlock:
 898         spin_unlock(ptl);
 899 }
 900
 901 /**
 902  * vmf_insert_pfn_pud_prot - insert a pud size pfn
 903  * @vmf: Structure describing the fault
 904  * @pfn: pfn to insert
 905  * @pgprot: page protection to use
 906  * @write: whether it's a write fault
 907  *
 908  * Insert a pud size pfn. See vmf_insert_pfn() for additional info and
 909  * also consult the vmf_insert_mixed_prot() documentation when
 910  * @pgprot != @vmf->vma->vm_page_prot.
 911  *
 912  * Return: vm_fault_t value.
 913  */
 914 vm_fault_t vmf_insert_pfn_pud_prot(struct vm_fault *vmf, pfn_t pfn,
 915                                    pgprot_t pgprot, bool write)
 916 {
 917         unsigned long addr = vmf->address & PUD_MASK;
 918         struct vm_area_struct *vma = vmf->vma;
 919
 920         /*
 921          * If we had pud_special, we could avoid all these restrictions,
 922          * but we need to be consistent with PTEs and architectures that
 923          * can't support a 'special' bit.
 924          */
 925         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 926                         !pfn_t_devmap(pfn));
 927         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 928                                                 (VM_PFNMAP|VM_MIXEDMAP));
 929         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 930
 931         if (addr < vma->vm_start || addr >= vma->vm_end)
 932                 return VM_FAULT_SIGBUS;
 933
 934         track_pfn_insert(vma, &pgprot, pfn);
 935
 936         insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
 937         return VM_FAULT_NOPAGE;
 938 }
 939 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud_prot);
 940 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
 941
 942 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
 943                 pmd_t *pmd, int flags)
 944 {
 945         pmd_t _pmd;
 946
 947         _pmd = pmd_mkyoung(*pmd);
 948         if (flags & FOLL_WRITE)
 949                 _pmd = pmd_mkdirty(_pmd);
 950         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
 951                                 pmd, _pmd, flags & FOLL_WRITE))
 952                 update_mmu_cache_pmd(vma, addr, pmd);
 953 }
 954
 955 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 956                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
 957 {
 958         unsigned long pfn = pmd_pfn(*pmd);
 959         struct mm_struct *mm = vma->vm_mm;
 960         struct page *page;
 961
 962         assert_spin_locked(pmd_lockptr(mm, pmd));
 963
 964         /*
 965          * When we COW a devmap PMD entry, we split it into PTEs, so we should
 966          * not be in this function with `flags & FOLL_COW` set.
 967          */
 968         WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
 969
 970         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
 971         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
 972                          (FOLL_PIN | FOLL_GET)))
 973                 return NULL;
 974
 975         if (flags & FOLL_WRITE && !pmd_write(*pmd))
 976                 return NULL;
 977
 978         if (pmd_present(*pmd) && pmd_devmap(*pmd))
 979                 /* pass */;
 980         else
 981                 return NULL;
 982
 983         if (flags & FOLL_TOUCH)
 984                 touch_pmd(vma, addr, pmd, flags);
 985
 986         /*
 987          * device mapped pages can only be returned if the
 988          * caller will manage the page reference count.
 989          */
 990         if (!(flags & (FOLL_GET | FOLL_PIN)))
 991                 return ERR_PTR(-EEXIST);
 992
 993         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
 994         *pgmap = get_dev_pagemap(pfn, *pgmap);
 995         if (!*pgmap)
 996                 return ERR_PTR(-EFAULT);
 997         page = pfn_to_page(pfn);
 998         if (!try_grab_page(page, flags))
 999                 page = ERR_PTR(-ENOMEM);
1000
1001         return page;
1002 }
1003
1004 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1005                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1006                   struct vm_area_struct *vma)
1007 {
1008         spinlock_t *dst_ptl, *src_ptl;
1009         struct page *src_page;
1010         pmd_t pmd;
1011         pgtable_t pgtable = NULL;
1012         int ret = -ENOMEM;
1013
1014         /* Skip if can be re-fill on fault */
1015         if (!vma_is_anonymous(vma))
1016                 return 0;
1017
1018         pgtable = pte_alloc_one(dst_mm);
1019         if (unlikely(!pgtable))
1020                 goto out;
1021
1022         dst_ptl = pmd_lock(dst_mm, dst_pmd);
1023         src_ptl = pmd_lockptr(src_mm, src_pmd);
1024         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1025
1026         ret = -EAGAIN;
1027         pmd = *src_pmd;
1028
1029         /*
1030          * Make sure the _PAGE_UFFD_WP bit is cleared if the new VMA
1031          * does not have the VM_UFFD_WP, which means that the uffd
1032          * fork event is not enabled.
1033          */
1034         if (!(vma->vm_flags & VM_UFFD_WP))
1035                 pmd = pmd_clear_uffd_wp(pmd);
1036
1037 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1038         if (unlikely(is_swap_pmd(pmd))) {
1039                 swp_entry_t entry = pmd_to_swp_entry(pmd);
1040
1041                 VM_BUG_ON(!is_pmd_migration_entry(pmd));
1042                 if (is_write_migration_entry(entry)) {
1043                         make_migration_entry_read(&entry);
1044                         pmd = swp_entry_to_pmd(entry);
1045                         if (pmd_swp_soft_dirty(*src_pmd))
1046                                 pmd = pmd_swp_mksoft_dirty(pmd);
1047                         set_pmd_at(src_mm, addr, src_pmd, pmd);
1048                 }
1049                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1050                 mm_inc_nr_ptes(dst_mm);
1051                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1052                 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1053                 ret = 0;
1054                 goto out_unlock;
1055         }
1056 #endif
1057
1058         if (unlikely(!pmd_trans_huge(pmd))) {
1059                 pte_free(dst_mm, pgtable);
1060                 goto out_unlock;
1061         }
1062         /*
1063          * When page table lock is held, the huge zero pmd should not be
1064          * under splitting since we don't split the page itself, only pmd to
1065          * a page table.
1066          */
1067         if (is_huge_zero_pmd(pmd)) {
1068                 struct page *zero_page;
1069                 /*
1070                  * get_huge_zero_page() will never allocate a new page here,
1071                  * since we already have a zero page to copy. It just takes a
1072                  * reference.
1073                  */
1074                 zero_page = mm_get_huge_zero_page(dst_mm);
1075                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1076                                 zero_page);
1077                 ret = 0;
1078                 goto out_unlock;
1079         }
1080
1081         src_page = pmd_page(pmd);
1082         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1083
1084         /*
1085          * If this page is a potentially pinned page, split and retry the fault
1086          * with smaller page size.  Normally this should not happen because the
1087          * userspace should use MADV_DONTFORK upon pinned regions.  This is a
1088          * best effort that the pinned pages won't be replaced by another
1089          * random page during the coming copy-on-write.
1090          */
1091         if (unlikely(is_cow_mapping(vma->vm_flags) &&
1092                      atomic_read(&src_mm->has_pinned) &&
1093                      page_maybe_dma_pinned(src_page))) {
1094                 pte_free(dst_mm, pgtable);
1095                 spin_unlock(src_ptl);
1096                 spin_unlock(dst_ptl);
1097                 __split_huge_pmd(vma, src_pmd, addr, false, NULL);
1098                 return -EAGAIN;
1099         }
1100
1101         get_page(src_page);
1102         page_dup_rmap(src_page, true);
1103         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1104         mm_inc_nr_ptes(dst_mm);
1105         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1106
1107         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1108         pmd = pmd_mkold(pmd_wrprotect(pmd));
1109         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1110
1111         ret = 0;
1112 out_unlock:
1113         spin_unlock(src_ptl);
1114         spin_unlock(dst_ptl);
1115 out:
1116         return ret;
1117 }
1118
1119 #ifdef CONFIG_FINEGRAINED_THP
1120 #endif /* CONFIG_FINEGRAINED_THP */
1121
1122 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1123 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1124                 pud_t *pud, int flags)
1125 {
1126         pud_t _pud;
1127
1128         _pud = pud_mkyoung(*pud);
1129         if (flags & FOLL_WRITE)
1130                 _pud = pud_mkdirty(_pud);
1131         if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1132                                 pud, _pud, flags & FOLL_WRITE))
1133                 update_mmu_cache_pud(vma, addr, pud);
1134 }
1135
1136 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1137                 pud_t *pud, int flags, struct dev_pagemap **pgmap)
1138 {
1139         unsigned long pfn = pud_pfn(*pud);
1140         struct mm_struct *mm = vma->vm_mm;
1141         struct page *page;
1142
1143         assert_spin_locked(pud_lockptr(mm, pud));
1144
1145         if (flags & FOLL_WRITE && !pud_write(*pud))
1146                 return NULL;
1147
1148         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
1149         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
1150                          (FOLL_PIN | FOLL_GET)))
1151                 return NULL;
1152
1153         if (pud_present(*pud) && pud_devmap(*pud))
1154                 /* pass */;
1155         else
1156                 return NULL;
1157
1158         if (flags & FOLL_TOUCH)
1159                 touch_pud(vma, addr, pud, flags);
1160
1161         /*
1162          * device mapped pages can only be returned if the
1163          * caller will manage the page reference count.
1164          *
1165          * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here:
1166          */
1167         if (!(flags & (FOLL_GET | FOLL_PIN)))
1168                 return ERR_PTR(-EEXIST);
1169
1170         pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1171         *pgmap = get_dev_pagemap(pfn, *pgmap);
1172         if (!*pgmap)
1173                 return ERR_PTR(-EFAULT);
1174         page = pfn_to_page(pfn);
1175         if (!try_grab_page(page, flags))
1176                 page = ERR_PTR(-ENOMEM);
1177
1178         return page;
1179 }
1180
1181 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1182                   pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1183                   struct vm_area_struct *vma)
1184 {
1185         spinlock_t *dst_ptl, *src_ptl;
1186         pud_t pud;
1187         int ret;
1188
1189         dst_ptl = pud_lock(dst_mm, dst_pud);
1190         src_ptl = pud_lockptr(src_mm, src_pud);
1191         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1192
1193         ret = -EAGAIN;
1194         pud = *src_pud;
1195         if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1196                 goto out_unlock;
1197
1198         /*
1199          * When page table lock is held, the huge zero pud should not be
1200          * under splitting since we don't split the page itself, only pud to
1201          * a page table.
1202          */
1203         if (is_huge_zero_pud(pud)) {
1204                 /* No huge zero pud yet */
1205         }
1206
1207         /* Please refer to comments in copy_huge_pmd() */
1208         if (unlikely(is_cow_mapping(vma->vm_flags) &&
1209                      atomic_read(&src_mm->has_pinned) &&
1210                      page_maybe_dma_pinned(pud_page(pud)))) {
1211                 spin_unlock(src_ptl);
1212                 spin_unlock(dst_ptl);
1213                 __split_huge_pud(vma, src_pud, addr);
1214                 return -EAGAIN;
1215         }
1216
1217         pudp_set_wrprotect(src_mm, addr, src_pud);
1218         pud = pud_mkold(pud_wrprotect(pud));
1219         set_pud_at(dst_mm, addr, dst_pud, pud);
1220
1221         ret = 0;
1222 out_unlock:
1223         spin_unlock(src_ptl);
1224         spin_unlock(dst_ptl);
1225         return ret;
1226 }
1227
1228 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1229 {
1230         pud_t entry;
1231         unsigned long haddr;
1232         bool write = vmf->flags & FAULT_FLAG_WRITE;
1233
1234         vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1235         if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1236                 goto unlock;
1237
1238         entry = pud_mkyoung(orig_pud);
1239         if (write)
1240                 entry = pud_mkdirty(entry);
1241         haddr = vmf->address & HPAGE_PUD_MASK;
1242         if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1243                 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1244
1245 unlock:
1246         spin_unlock(vmf->ptl);
1247 }
1248 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1249
1250 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1251 {
1252         pmd_t entry;
1253         unsigned long haddr;
1254         bool write = vmf->flags & FAULT_FLAG_WRITE;
1255
1256         vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1257         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1258                 goto unlock;
1259
1260         entry = pmd_mkyoung(orig_pmd);
1261         if (write)
1262                 entry = pmd_mkdirty(entry);
1263         haddr = vmf->address & HPAGE_PMD_MASK;
1264         if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1265                 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1266
1267 unlock:
1268         spin_unlock(vmf->ptl);
1269 }
1270
1271 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1272 {
1273         struct vm_area_struct *vma = vmf->vma;
1274         struct page *page;
1275         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1276
1277         vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1278         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1279
1280         if (is_huge_zero_pmd(orig_pmd))
1281                 goto fallback;
1282
1283         spin_lock(vmf->ptl);
1284
1285         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1286                 spin_unlock(vmf->ptl);
1287                 return 0;
1288         }
1289
1290         page = pmd_page(orig_pmd);
1291         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1292
1293         /* Lock page for reuse_swap_page() */
1294         if (!trylock_page(page)) {
1295                 get_page(page);
1296                 spin_unlock(vmf->ptl);
1297                 lock_page(page);
1298                 spin_lock(vmf->ptl);
1299                 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1300                         spin_unlock(vmf->ptl);
1301                         unlock_page(page);
1302                         put_page(page);
1303                         return 0;
1304                 }
1305                 put_page(page);
1306         }
1307
1308         /*
1309          * We can only reuse the page if nobody else maps the huge page or it's
1310          * part.
1311          */
1312         if (reuse_swap_page(page, NULL)) {
1313                 pmd_t entry;
1314                 entry = pmd_mkyoung(orig_pmd);
1315                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1316                 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1317                         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1318                 unlock_page(page);
1319                 spin_unlock(vmf->ptl);
1320                 return VM_FAULT_WRITE;
1321         }
1322
1323         unlock_page(page);
1324         spin_unlock(vmf->ptl);
1325 fallback:
1326         __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL);
1327         return VM_FAULT_FALLBACK;
1328 }
1329
1330 /*
1331  * FOLL_FORCE can write to even unwritable pmd's, but only
1332  * after we've gone through a COW cycle and they are dirty.
1333  */
1334 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1335 {
1336         return pmd_write(pmd) ||
1337                ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1338 }
1339
1340 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1341                                    unsigned long addr,
1342                                    pmd_t *pmd,
1343                                    unsigned int flags)
1344 {
1345         struct mm_struct *mm = vma->vm_mm;
1346         struct page *page = NULL;
1347
1348         assert_spin_locked(pmd_lockptr(mm, pmd));
1349
1350         if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1351                 goto out;
1352
1353         /* Avoid dumping huge zero page */
1354         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1355                 return ERR_PTR(-EFAULT);
1356
1357         /* Full NUMA hinting faults to serialise migration in fault paths */
1358         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1359                 goto out;
1360
1361         page = pmd_page(*pmd);
1362         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1363
1364         if (!try_grab_page(page, flags))
1365                 return ERR_PTR(-ENOMEM);
1366
1367         if (flags & FOLL_TOUCH)
1368                 touch_pmd(vma, addr, pmd, flags);
1369
1370         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1371                 /*
1372                  * We don't mlock() pte-mapped THPs. This way we can avoid
1373                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1374                  *
1375                  * For anon THP:
1376                  *
1377                  * In most cases the pmd is the only mapping of the page as we
1378                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1379                  * writable private mappings in populate_vma_page_range().
1380                  *
1381                  * The only scenario when we have the page shared here is if we
1382                  * mlocking read-only mapping shared over fork(). We skip
1383                  * mlocking such pages.
1384                  *
1385                  * For file THP:
1386                  *
1387                  * We can expect PageDoubleMap() to be stable under page lock:
1388                  * for file pages we set it in page_add_file_rmap(), which
1389                  * requires page to be locked.
1390                  */
1391
1392                 if (PageAnon(page) && compound_mapcount(page) != 1)
1393                         goto skip_mlock;
1394                 if (PageDoubleMap(page) || !page->mapping)
1395                         goto skip_mlock;
1396                 if (!trylock_page(page))
1397                         goto skip_mlock;
1398                 if (page->mapping && !PageDoubleMap(page))
1399                         mlock_vma_page(page);
1400                 unlock_page(page);
1401         }
1402 skip_mlock:
1403         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1404         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1405
1406 out:
1407         return page;
1408 }
1409
1410 /* NUMA hinting page fault entry point for trans huge pmds */
1411 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1412 {
1413         struct vm_area_struct *vma = vmf->vma;
1414         struct anon_vma *anon_vma = NULL;
1415         struct page *page;
1416         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1417         int page_nid = NUMA_NO_NODE, this_nid = numa_node_id();
1418         int target_nid, last_cpupid = -1;
1419         bool page_locked;
1420         bool migrated = false;
1421         bool was_writable;
1422         int flags = 0;
1423
1424         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1425         if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1426                 goto out_unlock;
1427
1428         /*
1429          * If there are potential migrations, wait for completion and retry
1430          * without disrupting NUMA hinting information. Do not relock and
1431          * check_same as the page may no longer be mapped.
1432          */
1433         if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1434                 page = pmd_page(*vmf->pmd);
1435                 if (!get_page_unless_zero(page))
1436                         goto out_unlock;
1437                 spin_unlock(vmf->ptl);
1438                 put_and_wait_on_page_locked(page);
1439                 goto out;
1440         }
1441
1442         page = pmd_page(pmd);
1443         BUG_ON(is_huge_zero_page(page));
1444         page_nid = page_to_nid(page);
1445         last_cpupid = page_cpupid_last(page);
1446         count_vm_numa_event(NUMA_HINT_FAULTS);
1447         if (page_nid == this_nid) {
1448                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1449                 flags |= TNF_FAULT_LOCAL;
1450         }
1451
1452         /* See similar comment in do_numa_page for explanation */
1453         if (!pmd_savedwrite(pmd))
1454                 flags |= TNF_NO_GROUP;
1455
1456         /*
1457          * Acquire the page lock to serialise THP migrations but avoid dropping
1458          * page_table_lock if at all possible
1459          */
1460         page_locked = trylock_page(page);
1461         target_nid = mpol_misplaced(page, vma, haddr);
1462         if (target_nid == NUMA_NO_NODE) {
1463                 /* If the page was locked, there are no parallel migrations */
1464                 if (page_locked)
1465                         goto clear_pmdnuma;
1466         }
1467
1468         /* Migration could have started since the pmd_trans_migrating check */
1469         if (!page_locked) {
1470                 page_nid = NUMA_NO_NODE;
1471                 if (!get_page_unless_zero(page))
1472                         goto out_unlock;
1473                 spin_unlock(vmf->ptl);
1474                 put_and_wait_on_page_locked(page);
1475                 goto out;
1476         }
1477
1478         /*
1479          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1480          * to serialises splits
1481          */
1482         get_page(page);
1483         spin_unlock(vmf->ptl);
1484         anon_vma = page_lock_anon_vma_read(page);
1485
1486         /* Confirm the PMD did not change while page_table_lock was released */
1487         spin_lock(vmf->ptl);
1488         if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1489                 unlock_page(page);
1490                 put_page(page);
1491                 page_nid = NUMA_NO_NODE;
1492                 goto out_unlock;
1493         }
1494
1495         /* Bail if we fail to protect against THP splits for any reason */
1496         if (unlikely(!anon_vma)) {
1497                 put_page(page);
1498                 page_nid = NUMA_NO_NODE;
1499                 goto clear_pmdnuma;
1500         }
1501
1502         /*
1503          * Since we took the NUMA fault, we must have observed the !accessible
1504          * bit. Make sure all other CPUs agree with that, to avoid them
1505          * modifying the page we're about to migrate.
1506          *
1507          * Must be done under PTL such that we'll observe the relevant
1508          * inc_tlb_flush_pending().
1509          *
1510          * We are not sure a pending tlb flush here is for a huge page
1511          * mapping or not. Hence use the tlb range variant
1512          */
1513         if (mm_tlb_flush_pending(vma->vm_mm)) {
1514                 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1515                 /*
1516                  * change_huge_pmd() released the pmd lock before
1517                  * invalidating the secondary MMUs sharing the primary
1518                  * MMU pagetables (with ->invalidate_range()). The
1519                  * mmu_notifier_invalidate_range_end() (which
1520                  * internally calls ->invalidate_range()) in
1521                  * change_pmd_range() will run after us, so we can't
1522                  * rely on it here and we need an explicit invalidate.
1523                  */
1524                 mmu_notifier_invalidate_range(vma->vm_mm, haddr,
1525                                               haddr + HPAGE_PMD_SIZE);
1526         }
1527
1528         /*
1529          * Migrate the THP to the requested node, returns with page unlocked
1530          * and access rights restored.
1531          */
1532         spin_unlock(vmf->ptl);
1533
1534         migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1535                                 vmf->pmd, pmd, vmf->address, page, target_nid);
1536         if (migrated) {
1537                 flags |= TNF_MIGRATED;
1538                 page_nid = target_nid;
1539         } else
1540                 flags |= TNF_MIGRATE_FAIL;
1541
1542         goto out;
1543 clear_pmdnuma:
1544         BUG_ON(!PageLocked(page));
1545         was_writable = pmd_savedwrite(pmd);
1546         pmd = pmd_modify(pmd, vma->vm_page_prot);
1547         pmd = pmd_mkyoung(pmd);
1548         if (was_writable)
1549                 pmd = pmd_mkwrite(pmd);
1550         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1551         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1552         unlock_page(page);
1553 out_unlock:
1554         spin_unlock(vmf->ptl);
1555
1556 out:
1557         if (anon_vma)
1558                 page_unlock_anon_vma_read(anon_vma);
1559
1560         if (page_nid != NUMA_NO_NODE)
1561                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1562                                 flags);
1563
1564         return 0;
1565 }
1566
1567 /*
1568  * Return true if we do MADV_FREE successfully on entire pmd page.
1569  * Otherwise, return false.
1570  */
1571 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1572                 pmd_t *pmd, unsigned long addr, unsigned long next)
1573 {
1574         spinlock_t *ptl;
1575         pmd_t orig_pmd;
1576         struct page *page;
1577         struct mm_struct *mm = tlb->mm;
1578         bool ret = false;
1579
1580         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1581
1582         ptl = pmd_trans_huge_lock(pmd, vma);
1583         if (!ptl)
1584                 goto out_unlocked;
1585
1586         orig_pmd = *pmd;
1587         if (is_huge_zero_pmd(orig_pmd))
1588                 goto out;
1589
1590         if (unlikely(!pmd_present(orig_pmd))) {
1591                 VM_BUG_ON(thp_migration_supported() &&
1592                                   !is_pmd_migration_entry(orig_pmd));
1593                 goto out;
1594         }
1595
1596         page = pmd_page(orig_pmd);
1597         /*
1598          * If other processes are mapping this page, we couldn't discard
1599          * the page unless they all do MADV_FREE so let's skip the page.
1600          */
1601         if (page_mapcount(page) != 1)
1602                 goto out;
1603
1604         if (!trylock_page(page))
1605                 goto out;
1606
1607         /*
1608          * If user want to discard part-pages of THP, split it so MADV_FREE
1609          * will deactivate only them.
1610          */
1611         if (next - addr != HPAGE_PMD_SIZE) {
1612                 get_page(page);
1613                 spin_unlock(ptl);
1614                 split_huge_page(page);
1615                 unlock_page(page);
1616                 put_page(page);
1617                 goto out_unlocked;
1618         }
1619
1620         if (PageDirty(page))
1621                 ClearPageDirty(page);
1622         unlock_page(page);
1623
1624         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1625                 pmdp_invalidate(vma, addr, pmd);
1626                 orig_pmd = pmd_mkold(orig_pmd);
1627                 orig_pmd = pmd_mkclean(orig_pmd);
1628
1629                 set_pmd_at(mm, addr, pmd, orig_pmd);
1630                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1631         }
1632
1633         mark_page_lazyfree(page);
1634         ret = true;
1635 out:
1636         spin_unlock(ptl);
1637 out_unlocked:
1638         return ret;
1639 }
1640
1641 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1642 {
1643         pgtable_t pgtable;
1644
1645         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1646         pte_free(mm, pgtable);
1647         mm_dec_nr_ptes(mm);
1648 }
1649
1650 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1651                  pmd_t *pmd, unsigned long addr)
1652 {
1653         pmd_t orig_pmd;
1654         spinlock_t *ptl;
1655
1656         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1657
1658         ptl = __pmd_trans_huge_lock(pmd, vma);
1659         if (!ptl)
1660                 return 0;
1661         /*
1662          * For architectures like ppc64 we look at deposited pgtable
1663          * when calling pmdp_huge_get_and_clear. So do the
1664          * pgtable_trans_huge_withdraw after finishing pmdp related
1665          * operations.
1666          */
1667         orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd,
1668                                                 tlb->fullmm);
1669         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1670         if (vma_is_special_huge(vma)) {
1671                 if (arch_needs_pgtable_deposit())
1672                         zap_deposited_table(tlb->mm, pmd);
1673                 atomic_long_add(-HPAGE_PMD_NR, &nr_phys_huge_pmd_pages);
1674                 spin_unlock(ptl);
1675                 if (is_huge_zero_pmd(orig_pmd))
1676                         tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1677         } else if (is_huge_zero_pmd(orig_pmd)) {
1678                 zap_deposited_table(tlb->mm, pmd);
1679                 spin_unlock(ptl);
1680                 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1681         } else {
1682                 struct page *page = NULL;
1683                 int flush_needed = 1;
1684
1685                 if (pmd_present(orig_pmd)) {
1686                         page = pmd_page(orig_pmd);
1687                         page_remove_rmap(page, true);
1688                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1689                         VM_BUG_ON_PAGE(!PageHead(page), page);
1690                 } else if (thp_migration_supported()) {
1691                         swp_entry_t entry;
1692
1693                         VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1694                         entry = pmd_to_swp_entry(orig_pmd);
1695                         page = pfn_to_page(swp_offset(entry));
1696                         flush_needed = 0;
1697                 } else
1698                         WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1699
1700                 if (PageAnon(page)) {
1701                         zap_deposited_table(tlb->mm, pmd);
1702                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1703                 } else {
1704                         if (arch_needs_pgtable_deposit())
1705                                 zap_deposited_table(tlb->mm, pmd);
1706                         add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1707                 }
1708
1709                 spin_unlock(ptl);
1710                 if (flush_needed)
1711                         tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1712         }
1713         return 1;
1714 }
1715
1716 #ifndef pmd_move_must_withdraw
1717 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1718                                          spinlock_t *old_pmd_ptl,
1719                                          struct vm_area_struct *vma)
1720 {
1721         /*
1722          * With split pmd lock we also need to move preallocated
1723          * PTE page table if new_pmd is on different PMD page table.
1724          *
1725          * We also don't deposit and withdraw tables for file pages.
1726          */
1727         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1728 }
1729 #endif
1730
1731 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1732 {
1733 #ifdef CONFIG_MEM_SOFT_DIRTY
1734         if (unlikely(is_pmd_migration_entry(pmd)))
1735                 pmd = pmd_swp_mksoft_dirty(pmd);
1736         else if (pmd_present(pmd))
1737                 pmd = pmd_mksoft_dirty(pmd);
1738 #endif
1739         return pmd;
1740 }
1741
1742 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1743                   unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
1744 {
1745         spinlock_t *old_ptl, *new_ptl;
1746         pmd_t pmd;
1747         struct mm_struct *mm = vma->vm_mm;
1748         bool force_flush = false;
1749
1750         /*
1751          * The destination pmd shouldn't be established, free_pgtables()
1752          * should have release it.
1753          */
1754         if (WARN_ON(!pmd_none(*new_pmd))) {
1755                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1756                 return false;
1757         }
1758
1759         /*
1760          * We don't have to worry about the ordering of src and dst
1761          * ptlocks because exclusive mmap_lock prevents deadlock.
1762          */
1763         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1764         if (old_ptl) {
1765                 new_ptl = pmd_lockptr(mm, new_pmd);
1766                 if (new_ptl != old_ptl)
1767                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1768                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1769                 if (pmd_present(pmd))
1770                         force_flush = true;
1771                 VM_BUG_ON(!pmd_none(*new_pmd));
1772
1773                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1774                         pgtable_t pgtable;
1775                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1776                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1777                 }
1778                 pmd = move_soft_dirty_pmd(pmd);
1779                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1780                 if (force_flush)
1781                         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1782                 if (new_ptl != old_ptl)
1783                         spin_unlock(new_ptl);
1784                 spin_unlock(old_ptl);
1785                 return true;
1786         }
1787         return false;
1788 }
1789
1790 /*
1791  * Returns
1792  *  - 0 if PMD could not be locked
1793  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1794  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1795  */
1796 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1797                 unsigned long addr, pgprot_t newprot, unsigned long cp_flags)
1798 {
1799         struct mm_struct *mm = vma->vm_mm;
1800         spinlock_t *ptl;
1801         pmd_t entry;
1802         bool preserve_write;
1803         int ret;
1804         bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
1805         bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
1806         bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
1807
1808         ptl = __pmd_trans_huge_lock(pmd, vma);
1809         if (!ptl)
1810                 return 0;
1811
1812         preserve_write = prot_numa && pmd_write(*pmd);
1813         ret = 1;
1814
1815 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1816         if (is_swap_pmd(*pmd)) {
1817                 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1818
1819                 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1820                 if (is_write_migration_entry(entry)) {
1821                         pmd_t newpmd;
1822                         /*
1823                          * A protection check is difficult so
1824                          * just be safe and disable write
1825                          */
1826                         make_migration_entry_read(&entry);
1827                         newpmd = swp_entry_to_pmd(entry);
1828                         if (pmd_swp_soft_dirty(*pmd))
1829                                 newpmd = pmd_swp_mksoft_dirty(newpmd);
1830                         set_pmd_at(mm, addr, pmd, newpmd);
1831                 }
1832                 goto unlock;
1833         }
1834 #endif
1835
1836         /*
1837          * Avoid trapping faults against the zero page. The read-only
1838          * data is likely to be read-cached on the local CPU and
1839          * local/remote hits to the zero page are not interesting.
1840          */
1841         if (prot_numa && is_huge_zero_pmd(*pmd))
1842                 goto unlock;
1843
1844         if (prot_numa && pmd_protnone(*pmd))
1845                 goto unlock;
1846
1847         /*
1848          * In case prot_numa, we are under mmap_read_lock(mm). It's critical
1849          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1850          * which is also under mmap_read_lock(mm):
1851          *
1852          *      CPU0:                           CPU1:
1853          *                              change_huge_pmd(prot_numa=1)
1854          *                               pmdp_huge_get_and_clear_notify()
1855          * madvise_dontneed()
1856          *  zap_pmd_range()
1857          *   pmd_trans_huge(*pmd) == 0 (without ptl)
1858          *   // skip the pmd
1859          *                               set_pmd_at();
1860          *                               // pmd is re-established
1861          *
1862          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1863          * which may break userspace.
1864          *
1865          * pmdp_invalidate() is required to make sure we don't miss
1866          * dirty/young flags set by hardware.
1867          */
1868         entry = pmdp_invalidate(vma, addr, pmd);
1869
1870         entry = pmd_modify(entry, newprot);
1871         if (preserve_write)
1872                 entry = pmd_mk_savedwrite(entry);
1873         if (uffd_wp) {
1874                 entry = pmd_wrprotect(entry);
1875                 entry = pmd_mkuffd_wp(entry);
1876         } else if (uffd_wp_resolve) {
1877                 /*
1878                  * Leave the write bit to be handled by PF interrupt
1879                  * handler, then things like COW could be properly
1880                  * handled.
1881                  */
1882                 entry = pmd_clear_uffd_wp(entry);
1883         }
1884         ret = HPAGE_PMD_NR;
1885         set_pmd_at(mm, addr, pmd, entry);
1886         BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1887 unlock:
1888         spin_unlock(ptl);
1889         return ret;
1890 }
1891
1892 /*
1893  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1894  *
1895  * Note that if it returns page table lock pointer, this routine returns without
1896  * unlocking page table lock. So callers must unlock it.
1897  */
1898 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1899 {
1900         spinlock_t *ptl;
1901         ptl = pmd_lock(vma->vm_mm, pmd);
1902         if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1903                         pmd_devmap(*pmd)))
1904                 return ptl;
1905         spin_unlock(ptl);
1906         return NULL;
1907 }
1908
1909 /*
1910  * Returns true if a given pud maps a thp, false otherwise.
1911  *
1912  * Note that if it returns true, this routine returns without unlocking page
1913  * table lock. So callers must unlock it.
1914  */
1915 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1916 {
1917         spinlock_t *ptl;
1918
1919         ptl = pud_lock(vma->vm_mm, pud);
1920         if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1921                 return ptl;
1922         spin_unlock(ptl);
1923         return NULL;
1924 }
1925
1926 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1927 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1928                  pud_t *pud, unsigned long addr)
1929 {
1930         spinlock_t *ptl;
1931
1932         ptl = __pud_trans_huge_lock(pud, vma);
1933         if (!ptl)
1934                 return 0;
1935         /*
1936          * For architectures like ppc64 we look at deposited pgtable
1937          * when calling pudp_huge_get_and_clear. So do the
1938          * pgtable_trans_huge_withdraw after finishing pudp related
1939          * operations.
1940          */
1941         pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
1942         tlb_remove_pud_tlb_entry(tlb, pud, addr);
1943         if (vma_is_special_huge(vma)) {
1944                 spin_unlock(ptl);
1945                 /* No zero page support yet */
1946         } else {
1947                 /* No support for anonymous PUD pages yet */
1948                 BUG();
1949         }
1950         return 1;
1951 }
1952
1953 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1954                 unsigned long haddr)
1955 {
1956         VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1957         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1958         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
1959         VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
1960
1961         count_vm_event(THP_SPLIT_PUD);
1962
1963         pudp_huge_clear_flush_notify(vma, haddr, pud);
1964 }
1965
1966 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
1967                 unsigned long address)
1968 {
1969         spinlock_t *ptl;
1970         struct mmu_notifier_range range;
1971
1972         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1973                                 address & HPAGE_PUD_MASK,
1974                                 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
1975         mmu_notifier_invalidate_range_start(&range);
1976         ptl = pud_lock(vma->vm_mm, pud);
1977         if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
1978                 goto out;
1979         __split_huge_pud_locked(vma, pud, range.start);
1980
1981 out:
1982         spin_unlock(ptl);
1983         /*
1984          * No need to double call mmu_notifier->invalidate_range() callback as
1985          * the above pudp_huge_clear_flush_notify() did already call it.
1986          */
1987         mmu_notifier_invalidate_range_only_end(&range);
1988 }
1989 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1990
1991 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1992                 unsigned long haddr, pmd_t *pmd)
1993 {
1994         struct mm_struct *mm = vma->vm_mm;
1995         pgtable_t pgtable;
1996         pmd_t _pmd;
1997         int i;
1998
1999         /*
2000          * Leave pmd empty until pte is filled note that it is fine to delay
2001          * notification until mmu_notifier_invalidate_range_end() as we are
2002          * replacing a zero pmd write protected page with a zero pte write
2003          * protected page.
2004          *
2005          * See Documentation/vm/mmu_notifier.rst
2006          */
2007         pmdp_huge_clear_flush(vma, haddr, pmd);
2008
2009         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2010         pmd_populate(mm, &_pmd, pgtable);
2011
2012         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2013                 pte_t *pte, entry;
2014                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2015                 entry = pte_mkspecial(entry);
2016                 pte = pte_offset_map(&_pmd, haddr);
2017                 VM_BUG_ON(!pte_none(*pte));
2018                 set_pte_at(mm, haddr, pte, entry);
2019                 pte_unmap(pte);
2020         }
2021         smp_wmb(); /* make pte visible before pmd */
2022         pmd_populate(mm, pmd, pgtable);
2023 }
2024
2025 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2026                 unsigned long haddr, bool freeze)
2027 {
2028         struct mm_struct *mm = vma->vm_mm;
2029         struct page *page;
2030         pgtable_t pgtable;
2031         pmd_t old_pmd, _pmd;
2032         bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false;
2033         unsigned long addr;
2034         int i;
2035
2036         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2037         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2038         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2039         VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2040                                 && !pmd_devmap(*pmd));
2041
2042         count_vm_event(THP_SPLIT_PMD);
2043
2044         if (!vma_is_anonymous(vma)) {
2045                 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2046                 /*
2047                  * We are going to unmap this huge page. So
2048                  * just go ahead and zap it
2049                  */
2050                 if (arch_needs_pgtable_deposit())
2051                         zap_deposited_table(mm, pmd);
2052                 if (vma_is_special_huge(vma))
2053                         return;
2054                 page = pmd_page(_pmd);
2055                 if (!PageDirty(page) && pmd_dirty(_pmd))
2056                         set_page_dirty(page);
2057                 if (!PageReferenced(page) && pmd_young(_pmd))
2058                         SetPageReferenced(page);
2059                 page_remove_rmap(page, true);
2060                 put_page(page);
2061                 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2062                 return;
2063         } else if (pmd_trans_huge(*pmd) && is_huge_zero_pmd(*pmd)) {
2064                 /*
2065                  * FIXME: Do we want to invalidate secondary mmu by calling
2066                  * mmu_notifier_invalidate_range() see comments below inside
2067                  * __split_huge_pmd() ?
2068                  *
2069                  * We are going from a zero huge page write protected to zero
2070                  * small page also write protected so it does not seems useful
2071                  * to invalidate secondary mmu at this time.
2072                  */
2073                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2074         }
2075
2076         /*
2077          * Up to this point the pmd is present and huge and userland has the
2078          * whole access to the hugepage during the split (which happens in
2079          * place). If we overwrite the pmd with the not-huge version pointing
2080          * to the pte here (which of course we could if all CPUs were bug
2081          * free), userland could trigger a small page size TLB miss on the
2082          * small sized TLB while the hugepage TLB entry is still established in
2083          * the huge TLB. Some CPU doesn't like that.
2084          * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum
2085          * 383 on page 105. Intel should be safe but is also warns that it's
2086          * only safe if the permission and cache attributes of the two entries
2087          * loaded in the two TLB is identical (which should be the case here).
2088          * But it is generally safer to never allow small and huge TLB entries
2089          * for the same virtual address to be loaded simultaneously. So instead
2090          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2091          * current pmd notpresent (atomically because here the pmd_trans_huge
2092          * must remain set at all times on the pmd until the split is complete
2093          * for this pmd), then we flush the SMP TLB and finally we write the
2094          * non-huge version of the pmd entry with pmd_populate.
2095          */
2096         old_pmd = pmdp_invalidate(vma, haddr, pmd);
2097
2098         pmd_migration = is_pmd_migration_entry(old_pmd);
2099         if (unlikely(pmd_migration)) {
2100                 swp_entry_t entry;
2101
2102                 entry = pmd_to_swp_entry(old_pmd);
2103                 page = pfn_to_page(swp_offset(entry));
2104                 write = is_write_migration_entry(entry);
2105                 young = false;
2106                 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2107                 uffd_wp = pmd_swp_uffd_wp(old_pmd);
2108         } else {
2109                 page = pmd_page(old_pmd);
2110                 if (pmd_dirty(old_pmd))
2111                         SetPageDirty(page);
2112                 write = pmd_write(old_pmd);
2113                 young = pmd_young(old_pmd);
2114                 soft_dirty = pmd_soft_dirty(old_pmd);
2115                 uffd_wp = pmd_uffd_wp(old_pmd);
2116         }
2117         VM_BUG_ON_PAGE(!page_count(page), page);
2118         page_ref_add(page, HPAGE_PMD_NR - 1);
2119
2120         /*
2121          * Withdraw the table only after we mark the pmd entry invalid.
2122          * This's critical for some architectures (Power).
2123          */
2124         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2125         pmd_populate(mm, &_pmd, pgtable);
2126
2127         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2128                 pte_t entry, *pte;
2129                 /*
2130                  * Note that NUMA hinting access restrictions are not
2131                  * transferred to avoid any possibility of altering
2132                  * permissions across VMAs.
2133                  */
2134                 if (freeze || pmd_migration) {
2135                         swp_entry_t swp_entry;
2136                         swp_entry = make_migration_entry(page + i, write);
2137                         entry = swp_entry_to_pte(swp_entry);
2138                         if (soft_dirty)
2139                                 entry = pte_swp_mksoft_dirty(entry);
2140                         if (uffd_wp)
2141                                 entry = pte_swp_mkuffd_wp(entry);
2142                 } else {
2143                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2144                         entry = maybe_mkwrite(entry, vma);
2145                         if (!write)
2146                                 entry = pte_wrprotect(entry);
2147                         if (!young)
2148                                 entry = pte_mkold(entry);
2149                         if (soft_dirty)
2150                                 entry = pte_mksoft_dirty(entry);
2151                         if (uffd_wp)
2152                                 entry = pte_mkuffd_wp(entry);
2153                 }
2154                 pte = pte_offset_map(&_pmd, addr);
2155                 BUG_ON(!pte_none(*pte));
2156                 set_pte_at(mm, addr, pte, entry);
2157                 if (!pmd_migration)
2158                         atomic_inc(&page[i]._mapcount);
2159                 pte_unmap(pte);
2160         }
2161
2162         if (!pmd_migration) {
2163                 /*
2164                  * Set PG_double_map before dropping compound_mapcount to avoid
2165                  * false-negative page_mapped().
2166                  */
2167                 if (compound_mapcount(page) > 1 &&
2168                     !TestSetPageDoubleMap(page)) {
2169                         for (i = 0; i < HPAGE_PMD_NR; i++)
2170                                 atomic_inc(&page[i]._mapcount);
2171                 }
2172
2173                 lock_page_memcg(page);
2174                 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2175                         /* Last compound_mapcount is gone. */
2176                         __dec_lruvec_page_state(page, NR_ANON_THPS);
2177                         if (TestClearPageDoubleMap(page)) {
2178                                 /* No need in mapcount reference anymore */
2179                                 for (i = 0; i < HPAGE_PMD_NR; i++)
2180                                         atomic_dec(&page[i]._mapcount);
2181                         }
2182                 }
2183                 unlock_page_memcg(page);
2184         }
2185
2186         smp_wmb(); /* make pte visible before pmd */
2187         pmd_populate(mm, pmd, pgtable);
2188
2189         if (freeze) {
2190                 for (i = 0; i < HPAGE_PMD_NR; i++) {
2191                         page_remove_rmap(page + i, false);
2192                         put_page(page + i);
2193                 }
2194         }
2195 }
2196
2197 #ifdef CONFIG_FINEGRAINED_THP
2198 static int thp_pte_alloc_locked(struct mm_struct *mm, pmd_t *pmd)
2199 {
2200         pgtable_t new = pte_alloc_one(mm);
2201         if (!new)
2202                 return -ENOMEM;
2203
2204         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
2205                 mm_inc_nr_ptes(mm);
2206                 pmd_populate(mm, pmd, new);
2207                 new = NULL;
2208         }
2209         if (new)
2210                 pte_free(mm, new);
2211         return 0;
2212 }
2213
2214 static int thp_remap_pte_range_locked(struct mm_struct *mm, pmd_t *pmd,
2215                         unsigned long addr, unsigned long end,
2216                         unsigned long pfn, pgprot_t prot)
2217 {
2218         pte_t *pte;
2219         int err = 0;
2220
2221         err = thp_pte_alloc_locked(mm, pmd);
2222         if (err)
2223                 return err;
2224
2225         pte = pte_offset_map(pmd, addr);
2226         if (!pte)
2227                 return -ENOMEM;
2228
2229         arch_enter_lazy_mmu_mode();
2230         do {
2231                 BUG_ON(!pte_none(*pte));
2232                 if (!pfn_modify_allowed(pfn, prot)) {
2233                         err = -EACCES;
2234                         break;
2235                 }
2236
2237                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2238                 pfn++;
2239                 pte++;
2240                 addr += PAGE_SIZE;
2241         } while (addr != end);
2242         arch_leave_lazy_mmu_mode();
2243         return err;
2244 }
2245
2246 static inline pgprot_t thp_pmd_pgprot(pmd_t pmd)
2247 {
2248         unsigned long pfn = pmd_pfn(pmd);
2249
2250         return __pgprot(pmd_val(pfn_pmd(pfn, __pgprot(0))) ^ pmd_val(pmd));
2251 }
2252 #endif
2253
2254 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2255                 unsigned long address, bool freeze, struct page *page)
2256 {
2257         spinlock_t *ptl;
2258         struct mmu_notifier_range range;
2259         bool do_unlock_page = false;
2260         pmd_t _pmd;
2261
2262         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2263                                 address & HPAGE_PMD_MASK,
2264                                 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2265         mmu_notifier_invalidate_range_start(&range);
2266         ptl = pmd_lock(vma->vm_mm, pmd);
2267
2268         /*
2269          * If caller asks to setup a migration entries, we need a page to check
2270          * pmd against. Otherwise we can end up replacing wrong page.
2271          */
2272         VM_BUG_ON(freeze && !page);
2273         if (page) {
2274                 VM_WARN_ON_ONCE(!PageLocked(page));
2275                 if (page != pmd_page(*pmd))
2276                         goto out;
2277         }
2278
2279 repeat:
2280 #ifdef CONFIG_FINEGRAINED_THP
2281         if (pmd_trans_huge(*pmd) && !vm_normal_page_pmd(vma, address, *pmd) && !is_huge_zero_pmd(*pmd)) {
2282                 struct mm_struct *mm = vma->vm_mm;
2283                 unsigned long haddr = address & HPAGE_PMD_MASK;
2284                 pmd_t orig_pmd;
2285
2286                 orig_pmd = pmdp_huge_get_and_clear_full(vma, haddr, pmd, 0);
2287                 atomic_long_add(-HPAGE_PMD_NR, &nr_phys_huge_pmd_pages);
2288                 thp_remap_pte_range_locked(mm, pmd, haddr,
2289                                            haddr + HPAGE_PMD_SIZE,
2290                                            pmd_pfn(orig_pmd),
2291                                            thp_pmd_pgprot(orig_pmd));
2292                 goto out;
2293         } else
2294 #endif /* CONFIG_FINEGRAINED_THP */
2295         if (pmd_trans_huge(*pmd) && vm_normal_page_pmd(vma, address, *pmd)) {
2296                 if (!page) {
2297                         page = pmd_page(*pmd);
2298                         /*
2299                          * An anonymous page must be locked, to ensure that a
2300                          * concurrent reuse_swap_page() sees stable mapcount;
2301                          * but reuse_swap_page() is not used on shmem or file,
2302                          * and page lock must not be taken when zap_pmd_range()
2303                          * calls __split_huge_pmd() while i_mmap_lock is held.
2304                          */
2305                         if (PageAnon(page)) {
2306                                 if (unlikely(!trylock_page(page))) {
2307                                         get_page(page);
2308                                         _pmd = *pmd;
2309                                         spin_unlock(ptl);
2310                                         lock_page(page);
2311                                         spin_lock(ptl);
2312                                         if (unlikely(!pmd_same(*pmd, _pmd))) {
2313                                                 unlock_page(page);
2314                                                 put_page(page);
2315                                                 page = NULL;
2316                                                 goto repeat;
2317                                         }
2318                                         put_page(page);
2319                                 }
2320                                 do_unlock_page = true;
2321                         }
2322                 }
2323                 if (PageMlocked(page))
2324                         clear_page_mlock(page);
2325         } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2326                 goto out;
2327         __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2328 out:
2329         spin_unlock(ptl);
2330         if (do_unlock_page)
2331                 unlock_page(page);
2332         /*
2333          * No need to double call mmu_notifier->invalidate_range() callback.
2334          * They are 3 cases to consider inside __split_huge_pmd_locked():
2335          *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2336          *  2) __split_huge_zero_page_pmd() read only zero page and any write
2337          *    fault will trigger a flush_notify before pointing to a new page
2338          *    (it is fine if the secondary mmu keeps pointing to the old zero
2339          *    page in the meantime)
2340          *  3) Split a huge pmd into pte pointing to the same page. No need
2341          *     to invalidate secondary tlb entry they are all still valid.
2342          *     any further changes to individual pte will notify. So no need
2343          *     to call mmu_notifier->invalidate_range()
2344          */
2345         mmu_notifier_invalidate_range_only_end(&range);
2346 }
2347
2348 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2349                 bool freeze, struct page *page)
2350 {
2351         pgd_t *pgd;
2352         p4d_t *p4d;
2353         pud_t *pud;
2354         pmd_t *pmd;
2355
2356         pgd = pgd_offset(vma->vm_mm, address);
2357         if (!pgd_present(*pgd))
2358                 return;
2359
2360         p4d = p4d_offset(pgd, address);
2361         if (!p4d_present(*p4d))
2362                 return;
2363
2364         pud = pud_offset(p4d, address);
2365         if (!pud_present(*pud))
2366                 return;
2367
2368         pmd = pmd_offset(pud, address);
2369
2370         __split_huge_pmd(vma, pmd, address, freeze, page);
2371 }
2372
2373 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2374                              unsigned long start,
2375                              unsigned long end,
2376                              long adjust_next)
2377 {
2378         /*
2379          * If the new start address isn't hpage aligned and it could
2380          * previously contain an hugepage: check if we need to split
2381          * an huge pmd.
2382          */
2383         if (start & ~HPAGE_PMD_MASK &&
2384             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2385             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2386                 split_huge_pmd_address(vma, start, false, NULL);
2387 #ifdef CONFIG_FINEGRAINED_THP
2388         if (start & ~HPAGE_CONT_PTE_MASK &&
2389                 (start & HPAGE_CONT_PTE_MASK) >= vma->vm_start &&
2390                 (start & HPAGE_CONT_PTE_MASK) + HPAGE_CONT_PTE_SIZE <= vma->vm_end)
2391                 split_huge_pte_address(vma, start, false, NULL);
2392 #endif
2393         /*
2394          * If the new end address isn't hpage aligned and it could
2395          * previously contain an hugepage: check if we need to split
2396          * an huge pmd.
2397          */
2398         if (end & ~HPAGE_PMD_MASK &&
2399             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2400             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2401                 split_huge_pmd_address(vma, end, false, NULL);
2402 #ifdef CONFIG_FINEGRAINED_THP
2403         if (end & ~HPAGE_CONT_PTE_MASK &&
2404                 (end & HPAGE_CONT_PTE_MASK) >= vma->vm_start &&
2405                 (end & HPAGE_CONT_PTE_MASK) + HPAGE_CONT_PTE_SIZE <= vma->vm_end)
2406                 split_huge_pte_address(vma, end, false, NULL);
2407 #endif
2408
2409         /*
2410          * If we're also updating the vma->vm_next->vm_start, if the new
2411          * vm_next->vm_start isn't hpage aligned and it could previously
2412          * contain an hugepage: check if we need to split an huge pmd.
2413          */
2414         if (adjust_next > 0) {
2415                 struct vm_area_struct *next = vma->vm_next;
2416                 unsigned long nstart = next->vm_start;
2417                 nstart += adjust_next;
2418                 if (nstart & ~HPAGE_PMD_MASK &&
2419                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2420                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2421                         split_huge_pmd_address(next, nstart, false, NULL);
2422 #ifdef CONFIG_FINEGRAINED_THP
2423                 if (nstart & ~HPAGE_CONT_PTE_MASK &&
2424                         (nstart & HPAGE_CONT_PTE_MASK) >= next->vm_start &&
2425                         (nstart & HPAGE_CONT_PTE_MASK) + HPAGE_CONT_PTE_SIZE <= next->vm_end)
2426                         split_huge_pte_address(next, nstart, false, NULL);
2427 #endif
2428         }
2429 }
2430
2431 static void unmap_page(struct page *page)
2432 {
2433 #ifdef CONFIG_FINEGRAINED_THP
2434         enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK |
2435                 TTU_RMAP_LOCKED;
2436 #else
2437         enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK |
2438                 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2439 #endif
2440         bool unmap_success;
2441
2442         VM_BUG_ON_PAGE(!PageHead(page), page);
2443
2444 #ifdef CONFIG_FINEGRAINED_THP
2445         if (compound_order(page) == HPAGE_PMD_ORDER)
2446                 ttu_flags |= TTU_SPLIT_HUGE_PMD;
2447         else
2448                 ttu_flags |= TTU_SPLIT_HUGE_PTE;
2449 #endif /* CONFIG_FINEGRAINED_THP */
2450         if (PageAnon(page))
2451                 ttu_flags |= TTU_SPLIT_FREEZE;
2452
2453         unmap_success = try_to_unmap(page, ttu_flags);
2454         VM_BUG_ON_PAGE(!unmap_success, page);
2455 }
2456
2457 static void remap_page(struct page *page, unsigned int nr)
2458 {
2459         int i;
2460         if (PageTransHuge(page)) {
2461                 remove_migration_ptes(page, page, true);
2462         } else {
2463                 for (i = 0; i < nr; i++)
2464                         remove_migration_ptes(page + i, page + i, true);
2465         }
2466 }
2467
2468 static void __split_huge_page_tail(struct page *head, int tail,
2469                 struct lruvec *lruvec, struct list_head *list)
2470 {
2471         struct page *page_tail = head + tail;
2472
2473         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2474
2475         /*
2476          * Clone page flags before unfreezing refcount.
2477          *
2478          * After successful get_page_unless_zero() might follow flags change,
2479          * for exmaple lock_page() which set PG_waiters.
2480          */
2481         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2482         page_tail->flags |= (head->flags &
2483                         ((1L << PG_referenced) |
2484                          (1L << PG_swapbacked) |
2485                          (1L << PG_swapcache) |
2486                          (1L << PG_mlocked) |
2487                          (1L << PG_uptodate) |
2488                          (1L << PG_active) |
2489                          (1L << PG_workingset) |
2490                          (1L << PG_locked) |
2491                          (1L << PG_unevictable) |
2492 #ifdef CONFIG_64BIT
2493                          (1L << PG_arch_2) |
2494 #endif
2495                          (1L << PG_dirty)));
2496
2497         /* ->mapping in first tail page is compound_mapcount */
2498         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2499                         page_tail);
2500         page_tail->mapping = head->mapping;
2501         page_tail->index = head->index + tail;
2502
2503         /* Page flags must be visible before we make the page non-compound. */
2504         smp_wmb();
2505
2506         /*
2507          * Clear PageTail before unfreezing page refcount.
2508          *
2509          * After successful get_page_unless_zero() might follow put_page()
2510          * which needs correct compound_head().
2511          */
2512         clear_compound_head(page_tail);
2513
2514         /* Finally unfreeze refcount. Additional reference from page cache. */
2515         page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2516                                           PageSwapCache(head)));
2517
2518         if (page_is_young(head))
2519                 set_page_young(page_tail);
2520         if (page_is_idle(head))
2521                 set_page_idle(page_tail);
2522
2523         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2524
2525         /*
2526          * always add to the tail because some iterators expect new
2527          * pages to show after the currently processed elements - e.g.
2528          * migrate_pages
2529          */
2530         lru_add_page_tail(head, page_tail, lruvec, list);
2531 }
2532
2533 static void __split_huge_page(struct page *page, struct list_head *list,
2534                 pgoff_t end, unsigned long flags)
2535 {
2536         struct page *head = compound_head(page);
2537         pg_data_t *pgdat = page_pgdat(head);
2538         struct lruvec *lruvec;
2539         struct address_space *swap_cache = NULL;
2540         unsigned long offset = 0;
2541         unsigned int nr = thp_nr_pages(head);
2542         int i;
2543
2544         lruvec = mem_cgroup_page_lruvec(head, pgdat);
2545
2546         /* complete memcg works before add pages to LRU */
2547         mem_cgroup_split_huge_fixup(head);
2548
2549         if (PageAnon(head) && PageSwapCache(head)) {
2550                 swp_entry_t entry = { .val = page_private(head) };
2551
2552                 offset = swp_offset(entry);
2553                 swap_cache = swap_address_space(entry);
2554                 xa_lock(&swap_cache->i_pages);
2555         }
2556
2557         for (i = nr - 1; i >= 1; i--) {
2558                 __split_huge_page_tail(head, i, lruvec, list);
2559                 /* Some pages can be beyond i_size: drop them from page cache */
2560                 if (head[i].index >= end) {
2561                         ClearPageDirty(head + i);
2562                         __delete_from_page_cache(head + i, NULL);
2563                         if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2564                                 shmem_uncharge(head->mapping->host, 1);
2565                         put_page(head + i);
2566                 } else if (!PageAnon(page)) {
2567                         __xa_store(&head->mapping->i_pages, head[i].index,
2568                                         head + i, 0);
2569                 } else if (swap_cache) {
2570                         __xa_store(&swap_cache->i_pages, offset + i,
2571                                         head + i, 0);
2572                 }
2573         }
2574
2575         ClearPageCompound(head);
2576
2577         split_page_owner(head, nr);
2578
2579         /* See comment in __split_huge_page_tail() */
2580         if (PageAnon(head)) {
2581                 /* Additional pin to swap cache */
2582                 if (PageSwapCache(head)) {
2583                         page_ref_add(head, 2);
2584                         xa_unlock(&swap_cache->i_pages);
2585                 } else {
2586                         page_ref_inc(head);
2587                 }
2588         } else {
2589                 /* Additional pin to page cache */
2590                 page_ref_add(head, 2);
2591                 xa_unlock(&head->mapping->i_pages);
2592         }
2593
2594         spin_unlock_irqrestore(&pgdat->lru_lock, flags);
2595
2596         remap_page(head, nr);
2597
2598         if (PageSwapCache(head)) {
2599                 swp_entry_t entry = { .val = page_private(head) };
2600
2601                 split_swap_cluster(entry);
2602         }
2603
2604         for (i = 0; i < nr; i++) {
2605                 struct page *subpage = head + i;
2606                 if (subpage == page)
2607                         continue;
2608                 unlock_page(subpage);
2609
2610                 /*
2611                  * Subpages may be freed if there wasn't any mapping
2612                  * like if add_to_swap() is running on a lru page that
2613                  * had its mapping zapped. And freeing these pages
2614                  * requires taking the lru_lock so we do the put_page
2615                  * of the tail pages after the split is complete.
2616                  */
2617                 put_page(subpage);
2618         }
2619 }
2620
2621 int total_mapcount(struct page *page)
2622 {
2623         int i, compound, nr, ret;
2624
2625         VM_BUG_ON_PAGE(PageTail(page), page);
2626
2627         if (likely(!PageCompound(page)))
2628                 return atomic_read(&page->_mapcount) + 1;
2629
2630         compound = compound_mapcount(page);
2631         nr = compound_nr(page);
2632         if (PageHuge(page))
2633                 return compound;
2634         ret = compound;
2635         for (i = 0; i < nr; i++)
2636                 ret += atomic_read(&page[i]._mapcount) + 1;
2637         /* File pages has compound_mapcount included in _mapcount */
2638         if (!PageAnon(page))
2639                 return ret - compound * nr;
2640         if (PageDoubleMap(page))
2641                 ret -= nr;
2642         return ret;
2643 }
2644
2645 /*
2646  * This calculates accurately how many mappings a transparent hugepage
2647  * has (unlike page_mapcount() which isn't fully accurate). This full
2648  * accuracy is primarily needed to know if copy-on-write faults can
2649  * reuse the page and change the mapping to read-write instead of
2650  * copying them. At the same time this returns the total_mapcount too.
2651  *
2652  * The function returns the highest mapcount any one of the subpages
2653  * has. If the return value is one, even if different processes are
2654  * mapping different subpages of the transparent hugepage, they can
2655  * all reuse it, because each process is reusing a different subpage.
2656  *
2657  * The total_mapcount is instead counting all virtual mappings of the
2658  * subpages. If the total_mapcount is equal to "one", it tells the
2659  * caller all mappings belong to the same "mm" and in turn the
2660  * anon_vma of the transparent hugepage can become the vma->anon_vma
2661  * local one as no other process may be mapping any of the subpages.
2662  *
2663  * It would be more accurate to replace page_mapcount() with
2664  * page_trans_huge_mapcount(), however we only use
2665  * page_trans_huge_mapcount() in the copy-on-write faults where we
2666  * need full accuracy to avoid breaking page pinning, because
2667  * page_trans_huge_mapcount() is slower than page_mapcount().
2668  */
2669 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2670 {
2671         int i, ret, _total_mapcount, mapcount;
2672
2673         /* hugetlbfs shouldn't call it */
2674         VM_BUG_ON_PAGE(PageHuge(page), page);
2675
2676         if (likely(!PageTransCompound(page))) {
2677                 mapcount = atomic_read(&page->_mapcount) + 1;
2678                 if (total_mapcount)
2679                         *total_mapcount = mapcount;
2680                 return mapcount;
2681         }
2682
2683         page = compound_head(page);
2684
2685         _total_mapcount = ret = 0;
2686         for (i = 0; i < thp_nr_pages(page); i++) {
2687                 mapcount = atomic_read(&page[i]._mapcount) + 1;
2688                 ret = max(ret, mapcount);
2689                 _total_mapcount += mapcount;
2690         }
2691         if (PageDoubleMap(page)) {
2692                 ret -= 1;
2693                 _total_mapcount -= thp_nr_pages(page);
2694         }
2695         mapcount = compound_mapcount(page);
2696         ret += mapcount;
2697         _total_mapcount += mapcount;
2698         if (total_mapcount)
2699                 *total_mapcount = _total_mapcount;
2700         return ret;
2701 }
2702
2703 /* Racy check whether the huge page can be split */
2704 bool can_split_huge_page(struct page *page, int *pextra_pins)
2705 {
2706         int extra_pins;
2707
2708         /* Additional pins from page cache */
2709         if (PageAnon(page))
2710                 extra_pins = PageSwapCache(page) ? thp_nr_pages(page) : 0;
2711         else
2712                 extra_pins = thp_nr_pages(page);
2713         if (pextra_pins)
2714                 *pextra_pins = extra_pins;
2715         return total_mapcount(page) == page_count(page) - extra_pins - 1;
2716 }
2717
2718 /*
2719  * This function splits huge page into normal pages. @page can point to any
2720  * subpage of huge page to split. Split doesn't change the position of @page.
2721  *
2722  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2723  * The huge page must be locked.
2724  *
2725  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2726  *
2727  * Both head page and tail pages will inherit mapping, flags, and so on from
2728  * the hugepage.
2729  *
2730  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2731  * they are not mapped.
2732  *
2733  * Returns 0 if the hugepage is split successfully.
2734  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2735  * us.
2736  */
2737 int split_huge_page_to_list(struct page *page, struct list_head *list)
2738 {
2739         struct page *head = compound_head(page);
2740         struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2741         struct deferred_split *ds_queue = get_deferred_split_queue(head);
2742         struct anon_vma *anon_vma = NULL;
2743         struct address_space *mapping = NULL;
2744         int count, mapcount, extra_pins, ret;
2745         unsigned long flags;
2746         pgoff_t end;
2747
2748         VM_BUG_ON_PAGE(is_huge_zero_page(head), head);
2749         VM_BUG_ON_PAGE(!PageLocked(head), head);
2750         VM_BUG_ON_PAGE(!PageCompound(head), head);
2751
2752         if (PageWriteback(head))
2753                 return -EBUSY;
2754
2755         if (PageAnon(head)) {
2756                 /*
2757                  * The caller does not necessarily hold an mmap_lock that would
2758                  * prevent the anon_vma disappearing so we first we take a
2759                  * reference to it and then lock the anon_vma for write. This
2760                  * is similar to page_lock_anon_vma_read except the write lock
2761                  * is taken to serialise against parallel split or collapse
2762                  * operations.
2763                  */
2764                 anon_vma = page_get_anon_vma(head);
2765                 if (!anon_vma) {
2766                         ret = -EBUSY;
2767                         goto out;
2768                 }
2769                 end = -1;
2770                 mapping = NULL;
2771                 anon_vma_lock_write(anon_vma);
2772         } else {
2773                 mapping = head->mapping;
2774
2775                 /* Truncated ? */
2776                 if (!mapping) {
2777                         ret = -EBUSY;
2778                         goto out;
2779                 }
2780
2781                 anon_vma = NULL;
2782                 i_mmap_lock_read(mapping);
2783
2784                 /*
2785                  *__split_huge_page() may need to trim off pages beyond EOF:
2786                  * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2787                  * which cannot be nested inside the page tree lock. So note
2788                  * end now: i_size itself may be changed at any moment, but
2789                  * head page lock is good enough to serialize the trimming.
2790                  */
2791                 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2792         }
2793
2794         /*
2795          * Racy check if we can split the page, before unmap_page() will
2796          * split PMDs
2797          */
2798         if (!can_split_huge_page(head, &extra_pins)) {
2799                 ret = -EBUSY;
2800                 goto out_unlock;
2801         }
2802
2803         unmap_page(head);
2804         VM_BUG_ON_PAGE(compound_mapcount(head), head);
2805
2806         /* prevent PageLRU to go away from under us, and freeze lru stats */
2807         spin_lock_irqsave(&pgdata->lru_lock, flags);
2808
2809         if (mapping) {
2810                 XA_STATE(xas, &mapping->i_pages, page_index(head));
2811
2812                 /*
2813                  * Check if the head page is present in page cache.
2814                  * We assume all tail are present too, if head is there.
2815                  */
2816                 xa_lock(&mapping->i_pages);
2817                 if (xas_load(&xas) != head)
2818                         goto fail;
2819         }
2820
2821         /* Prevent deferred_split_scan() touching ->_refcount */
2822         spin_lock(&ds_queue->split_queue_lock);
2823         count = page_count(head);
2824         mapcount = total_mapcount(head);
2825         if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2826                 if (!list_empty(page_deferred_list(head))) {
2827                         ds_queue->split_queue_len--;
2828                         list_del(page_deferred_list(head));
2829                 }
2830                 spin_unlock(&ds_queue->split_queue_lock);
2831                 if (mapping) {
2832                         if (PageSwapBacked(head)) {
2833 #ifdef CONFIG_FINEGRAINED_THP
2834                                 if (thp_nr_pages(head) == HPAGE_CONT_PTE_NR)
2835                                         __dec_node_page_state(head, NR_SHMEM_64KB_THPS);
2836                                 else
2837 #endif /* CONFIG_FINEGRAINED_THP */
2838                                         __dec_node_page_state(head, NR_SHMEM_THPS);
2839                         } else {
2840 #ifdef CONFIG_FINEGRAINED_THP
2841                                 if (thp_nr_pages(head) == HPAGE_CONT_PTE_NR)
2842                                         __dec_node_page_state(head, NR_FILE_64KB_THPS);
2843                                 else
2844 #endif /* CONFIG_FINEGRAINED_THP */
2845                                 __dec_node_page_state(head, NR_FILE_THPS);
2846                         }
2847                 }
2848
2849                 __split_huge_page(page, list, end, flags);
2850                 ret = 0;
2851         } else {
2852                 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2853                         pr_alert("total_mapcount: %u, page_count(): %u\n",
2854                                         mapcount, count);
2855                         if (PageTail(page))
2856                                 dump_page(head, NULL);
2857                         dump_page(page, "total_mapcount(head) > 0");
2858                         BUG();
2859                 }
2860                 spin_unlock(&ds_queue->split_queue_lock);
2861 fail:           if (mapping)
2862                         xa_unlock(&mapping->i_pages);
2863                 spin_unlock_irqrestore(&pgdata->lru_lock, flags);
2864                 remap_page(head, thp_nr_pages(head));
2865                 ret = -EBUSY;
2866         }
2867
2868 out_unlock:
2869         if (anon_vma) {
2870                 anon_vma_unlock_write(anon_vma);
2871                 put_anon_vma(anon_vma);
2872         }
2873         if (mapping)
2874                 i_mmap_unlock_read(mapping);
2875 out:
2876         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2877         return ret;
2878 }
2879
2880 void free_transhuge_page(struct page *page)
2881 {
2882         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2883         unsigned long flags;
2884
2885         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2886         if (!list_empty(page_deferred_list(page))) {
2887                 ds_queue->split_queue_len--;
2888                 list_del(page_deferred_list(page));
2889         }
2890         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2891         free_compound_page(page);
2892 }
2893
2894 void deferred_split_huge_page(struct page *page)
2895 {
2896         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2897 #ifdef CONFIG_MEMCG
2898         struct mem_cgroup *memcg = compound_head(page)->mem_cgroup;
2899 #endif
2900         unsigned long flags;
2901
2902         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2903
2904         /*
2905          * The try_to_unmap() in page reclaim path might reach here too,
2906          * this may cause a race condition to corrupt deferred split queue.
2907          * And, if page reclaim is already handling the same page, it is
2908          * unnecessary to handle it again in shrinker.
2909          *
2910          * Check PageSwapCache to determine if the page is being
2911          * handled by page reclaim since THP swap would add the page into
2912          * swap cache before calling try_to_unmap().
2913          */
2914         if (PageSwapCache(page))
2915                 return;
2916
2917         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2918         if (list_empty(page_deferred_list(page))) {
2919                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2920                 list_add_tail(page_deferred_list(page), &ds_queue->split_queue);
2921                 ds_queue->split_queue_len++;
2922 #ifdef CONFIG_MEMCG
2923                 if (memcg)
2924                         memcg_set_shrinker_bit(memcg, page_to_nid(page),
2925                                                deferred_split_shrinker.id);
2926 #endif
2927         }
2928         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2929 }
2930
2931 static unsigned long deferred_split_count(struct shrinker *shrink,
2932                 struct shrink_control *sc)
2933 {
2934         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2935         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2936
2937 #ifdef CONFIG_MEMCG
2938         if (sc->memcg)
2939                 ds_queue = &sc->memcg->deferred_split_queue;
2940 #endif
2941         return READ_ONCE(ds_queue->split_queue_len);
2942 }
2943
2944 static unsigned long deferred_split_scan(struct shrinker *shrink,
2945                 struct shrink_control *sc)
2946 {
2947         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2948         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2949         unsigned long flags;
2950         LIST_HEAD(list), *pos, *next;
2951         struct page *page;
2952         int split = 0;
2953
2954 #ifdef CONFIG_MEMCG
2955         if (sc->memcg)
2956                 ds_queue = &sc->memcg->deferred_split_queue;
2957 #endif
2958
2959         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2960         /* Take pin on all head pages to avoid freeing them under us */
2961         list_for_each_safe(pos, next, &ds_queue->split_queue) {
2962                 page = list_entry((void *)pos, struct page, mapping);
2963                 page = compound_head(page);
2964                 if (get_page_unless_zero(page)) {
2965                         list_move(page_deferred_list(page), &list);
2966                 } else {
2967                         /* We lost race with put_compound_page() */
2968                         list_del_init(page_deferred_list(page));
2969                         ds_queue->split_queue_len--;
2970                 }
2971                 if (!--sc->nr_to_scan)
2972                         break;
2973         }
2974         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2975
2976         list_for_each_safe(pos, next, &list) {
2977                 page = list_entry((void *)pos, struct page, mapping);
2978                 if (!trylock_page(page))
2979                         goto next;
2980                 /* split_huge_page() removes page from list on success */
2981                 if (!split_huge_page(page))
2982                         split++;
2983                 unlock_page(page);
2984 next:
2985                 put_page(page);
2986         }
2987
2988         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2989         list_splice_tail(&list, &ds_queue->split_queue);
2990         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2991
2992         /*
2993          * Stop shrinker if we didn't split any page, but the queue is empty.
2994          * This can happen if pages were freed under us.
2995          */
2996         if (!split && list_empty(&ds_queue->split_queue))
2997                 return SHRINK_STOP;
2998         return split;
2999 }
3000
3001 static struct shrinker deferred_split_shrinker = {
3002         .count_objects = deferred_split_count,
3003         .scan_objects = deferred_split_scan,
3004         .seeks = DEFAULT_SEEKS,
3005         .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
3006                  SHRINKER_NONSLAB,
3007 };
3008
3009 #ifdef CONFIG_DEBUG_FS
3010 static int split_huge_pages_set(void *data, u64 val)
3011 {
3012         struct zone *zone;
3013         struct page *page;
3014         unsigned long pfn, max_zone_pfn;
3015         unsigned long total = 0, split = 0;
3016
3017         if (val != 1)
3018                 return -EINVAL;
3019
3020         for_each_populated_zone(zone) {
3021                 max_zone_pfn = zone_end_pfn(zone);
3022                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
3023                         if (!pfn_valid(pfn))
3024                                 continue;
3025
3026                         page = pfn_to_page(pfn);
3027                         if (!get_page_unless_zero(page))
3028                                 continue;
3029
3030                         if (zone != page_zone(page))
3031                                 goto next;
3032
3033                         if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
3034                                 goto next;
3035
3036                         total++;
3037                         lock_page(page);
3038                         if (!split_huge_page(page))
3039                                 split++;
3040                         unlock_page(page);
3041 next:
3042                         put_page(page);
3043                 }
3044         }
3045
3046         pr_info("%lu of %lu THP split\n", split, total);
3047
3048         return 0;
3049 }
3050 DEFINE_DEBUGFS_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
3051                 "%llu\n");
3052
3053 static int __init split_huge_pages_debugfs(void)
3054 {
3055         debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3056                             &split_huge_pages_fops);
3057         return 0;
3058 }
3059 late_initcall(split_huge_pages_debugfs);
3060 #endif
3061
3062 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
3063 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
3064                 struct page *page)
3065 {
3066         struct vm_area_struct *vma = pvmw->vma;
3067         struct mm_struct *mm = vma->vm_mm;
3068         unsigned long address = pvmw->address;
3069         pmd_t pmdval;
3070         swp_entry_t entry;
3071         pmd_t pmdswp;
3072
3073         if (!(pvmw->pmd && !pvmw->pte))
3074                 return;
3075
3076         flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
3077         pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
3078         if (pmd_dirty(pmdval))
3079                 set_page_dirty(page);
3080         entry = make_migration_entry(page, pmd_write(pmdval));
3081         pmdswp = swp_entry_to_pmd(entry);
3082         if (pmd_soft_dirty(pmdval))
3083                 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
3084         set_pmd_at(mm, address, pvmw->pmd, pmdswp);
3085         page_remove_rmap(page, true);
3086         put_page(page);
3087 }
3088
3089 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
3090 {
3091         struct vm_area_struct *vma = pvmw->vma;
3092         struct mm_struct *mm = vma->vm_mm;
3093         unsigned long address = pvmw->address;
3094         unsigned long mmun_start = address & HPAGE_PMD_MASK;
3095         pmd_t pmde;
3096         swp_entry_t entry;
3097
3098         if (!(pvmw->pmd && !pvmw->pte))
3099                 return;
3100
3101         entry = pmd_to_swp_entry(*pvmw->pmd);
3102         get_page(new);
3103         pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
3104         if (pmd_swp_soft_dirty(*pvmw->pmd))
3105                 pmde = pmd_mksoft_dirty(pmde);
3106         if (is_write_migration_entry(entry))
3107                 pmde = maybe_pmd_mkwrite(pmde, vma);
3108
3109         flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
3110         if (PageAnon(new))
3111                 page_add_anon_rmap(new, vma, mmun_start, true);
3112         else
3113                 page_add_file_rmap(new, true);
3114         set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
3115         if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new))
3116                 mlock_vma_page(new);
3117         update_mmu_cache_pmd(vma, address, pvmw->pmd);
3118 }
3119 #endif