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 unsigned long huge_zero_pfn __read_mostly = ~0UL;
  65
  66 static inline bool file_thp_enabled(struct vm_area_struct *vma)
  67 {
  68         return transhuge_vma_enabled(vma, vma->vm_flags) && vma->vm_file &&
  69                !inode_is_open_for_write(vma->vm_file->f_inode) &&
  70                (vma->vm_flags & VM_EXEC);
  71 }
  72
  73 bool transparent_hugepage_active(struct vm_area_struct *vma)
  74 {
  75         /* The addr is used to check if the vma size fits */
  76         unsigned long addr = (vma->vm_end & HPAGE_PMD_MASK) - HPAGE_PMD_SIZE;
  77
  78         if (!transhuge_vma_suitable(vma, addr))
  79                 return false;
  80         if (vma_is_anonymous(vma))
  81                 return __transparent_hugepage_enabled(vma);
  82         if (vma_is_shmem(vma))
  83                 return shmem_huge_enabled(vma);
  84         if (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS))
  85                 return file_thp_enabled(vma);
  86
  87         return false;
  88 }
  89
  90 static struct page *get_huge_zero_page(void)
  91 {
  92         struct page *zero_page;
  93 retry:
  94         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
  95                 return READ_ONCE(huge_zero_page);
  96
  97         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
  98                         HPAGE_PMD_ORDER);
  99         if (!zero_page) {
 100                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
 101                 return NULL;
 102         }
 103         count_vm_event(THP_ZERO_PAGE_ALLOC);
 104         preempt_disable();
 105         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
 106                 preempt_enable();
 107                 __free_pages(zero_page, compound_order(zero_page));
 108                 goto retry;
 109         }
 110         WRITE_ONCE(huge_zero_pfn, page_to_pfn(zero_page));
 111
 112         /* We take additional reference here. It will be put back by shrinker */
 113         atomic_set(&huge_zero_refcount, 2);
 114         preempt_enable();
 115         return READ_ONCE(huge_zero_page);
 116 }
 117
 118 static void put_huge_zero_page(void)
 119 {
 120         /*
 121          * Counter should never go to zero here. Only shrinker can put
 122          * last reference.
 123          */
 124         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 125 }
 126
 127 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
 128 {
 129         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 130                 return READ_ONCE(huge_zero_page);
 131
 132         if (!get_huge_zero_page())
 133                 return NULL;
 134
 135         if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 136                 put_huge_zero_page();
 137
 138         return READ_ONCE(huge_zero_page);
 139 }
 140
 141 void mm_put_huge_zero_page(struct mm_struct *mm)
 142 {
 143         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 144                 put_huge_zero_page();
 145 }
 146
 147 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 148                                         struct shrink_control *sc)
 149 {
 150         /* we can free zero page only if last reference remains */
 151         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 152 }
 153
 154 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 155                                        struct shrink_control *sc)
 156 {
 157         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 158                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 159                 BUG_ON(zero_page == NULL);
 160                 WRITE_ONCE(huge_zero_pfn, ~0UL);
 161                 __free_pages(zero_page, compound_order(zero_page));
 162                 return HPAGE_PMD_NR;
 163         }
 164
 165         return 0;
 166 }
 167
 168 static struct shrinker huge_zero_page_shrinker = {
 169         .count_objects = shrink_huge_zero_page_count,
 170         .scan_objects = shrink_huge_zero_page_scan,
 171         .seeks = DEFAULT_SEEKS,
 172 };
 173
 174 #ifdef CONFIG_SYSFS
 175 static ssize_t enabled_show(struct kobject *kobj,
 176                             struct kobj_attribute *attr, char *buf)
 177 {
 178         if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
 179                 return sprintf(buf, "[always] madvise never\n");
 180         else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
 181                 return sprintf(buf, "always [madvise] never\n");
 182         else
 183                 return sprintf(buf, "always madvise [never]\n");
 184 }
 185
 186 static ssize_t enabled_store(struct kobject *kobj,
 187                              struct kobj_attribute *attr,
 188                              const char *buf, size_t count)
 189 {
 190         ssize_t ret = count;
 191
 192         if (sysfs_streq(buf, "always")) {
 193                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 194                 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 195         } else if (sysfs_streq(buf, "madvise")) {
 196                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 197                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 198         } else if (sysfs_streq(buf, "never")) {
 199                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 200                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 201         } else
 202                 ret = -EINVAL;
 203
 204         if (ret > 0) {
 205                 int err = start_stop_khugepaged();
 206                 if (err)
 207                         ret = err;
 208         }
 209         return ret;
 210 }
 211 static struct kobj_attribute enabled_attr =
 212         __ATTR(enabled, 0644, enabled_show, enabled_store);
 213
 214 ssize_t single_hugepage_flag_show(struct kobject *kobj,
 215                                 struct kobj_attribute *attr, char *buf,
 216                                 enum transparent_hugepage_flag flag)
 217 {
 218         return sprintf(buf, "%d\n",
 219                        !!test_bit(flag, &transparent_hugepage_flags));
 220 }
 221
 222 ssize_t single_hugepage_flag_store(struct kobject *kobj,
 223                                  struct kobj_attribute *attr,
 224                                  const char *buf, size_t count,
 225                                  enum transparent_hugepage_flag flag)
 226 {
 227         unsigned long value;
 228         int ret;
 229
 230         ret = kstrtoul(buf, 10, &value);
 231         if (ret < 0)
 232                 return ret;
 233         if (value > 1)
 234                 return -EINVAL;
 235
 236         if (value)
 237                 set_bit(flag, &transparent_hugepage_flags);
 238         else
 239                 clear_bit(flag, &transparent_hugepage_flags);
 240
 241         return count;
 242 }
 243
 244 static ssize_t defrag_show(struct kobject *kobj,
 245                            struct kobj_attribute *attr, char *buf)
 246 {
 247         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 248                 return sprintf(buf, "[always] defer defer+madvise madvise never\n");
 249         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 250                 return sprintf(buf, "always [defer] defer+madvise madvise never\n");
 251         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 252                 return sprintf(buf, "always defer [defer+madvise] madvise never\n");
 253         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 254                 return sprintf(buf, "always defer defer+madvise [madvise] never\n");
 255         return sprintf(buf, "always defer defer+madvise madvise [never]\n");
 256 }
 257
 258 static ssize_t defrag_store(struct kobject *kobj,
 259                             struct kobj_attribute *attr,
 260                             const char *buf, size_t count)
 261 {
 262         if (sysfs_streq(buf, "always")) {
 263                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 264                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 265                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 266                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 267         } else if (sysfs_streq(buf, "defer+madvise")) {
 268                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 269                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 270                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 271                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 272         } else if (sysfs_streq(buf, "defer")) {
 273                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 274                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 275                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 276                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 277         } else if (sysfs_streq(buf, "madvise")) {
 278                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 279                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 280                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 281                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 282         } else if (sysfs_streq(buf, "never")) {
 283                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 284                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 285                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 286                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 287         } else
 288                 return -EINVAL;
 289
 290         return count;
 291 }
 292 static struct kobj_attribute defrag_attr =
 293         __ATTR(defrag, 0644, defrag_show, defrag_store);
 294
 295 static ssize_t use_zero_page_show(struct kobject *kobj,
 296                 struct kobj_attribute *attr, char *buf)
 297 {
 298         return single_hugepage_flag_show(kobj, attr, buf,
 299                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 300 }
 301 static ssize_t use_zero_page_store(struct kobject *kobj,
 302                 struct kobj_attribute *attr, const char *buf, size_t count)
 303 {
 304         return single_hugepage_flag_store(kobj, attr, buf, count,
 305                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 306 }
 307 static struct kobj_attribute use_zero_page_attr =
 308         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 309
 310 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
 311                 struct kobj_attribute *attr, char *buf)
 312 {
 313         return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE);
 314 }
 315 static struct kobj_attribute hpage_pmd_size_attr =
 316         __ATTR_RO(hpage_pmd_size);
 317
 318 static struct attribute *hugepage_attr[] = {
 319         &enabled_attr.attr,
 320         &defrag_attr.attr,
 321         &use_zero_page_attr.attr,
 322         &hpage_pmd_size_attr.attr,
 323 #ifdef CONFIG_SHMEM
 324         &shmem_enabled_attr.attr,
 325 #endif
 326         NULL,
 327 };
 328
 329 static const struct attribute_group hugepage_attr_group = {
 330         .attrs = hugepage_attr,
 331 };
 332
 333 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 334 {
 335         int err;
 336
 337         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 338         if (unlikely(!*hugepage_kobj)) {
 339                 pr_err("failed to create transparent hugepage kobject\n");
 340                 return -ENOMEM;
 341         }
 342
 343         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 344         if (err) {
 345                 pr_err("failed to register transparent hugepage group\n");
 346                 goto delete_obj;
 347         }
 348
 349         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 350         if (err) {
 351                 pr_err("failed to register transparent hugepage group\n");
 352                 goto remove_hp_group;
 353         }
 354
 355         return 0;
 356
 357 remove_hp_group:
 358         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 359 delete_obj:
 360         kobject_put(*hugepage_kobj);
 361         return err;
 362 }
 363
 364 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 365 {
 366         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 367         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 368         kobject_put(hugepage_kobj);
 369 }
 370 #else
 371 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 372 {
 373         return 0;
 374 }
 375
 376 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 377 {
 378 }
 379 #endif /* CONFIG_SYSFS */
 380
 381 static int __init hugepage_init(void)
 382 {
 383         int err;
 384         struct kobject *hugepage_kobj;
 385
 386         if (!has_transparent_hugepage()) {
 387                 /*
 388                  * Hardware doesn't support hugepages, hence disable
 389                  * DAX PMD support.
 390                  */
 391                 transparent_hugepage_flags = 1 << TRANSPARENT_HUGEPAGE_NEVER_DAX;
 392                 return -EINVAL;
 393         }
 394
 395         /*
 396          * hugepages can't be allocated by the buddy allocator
 397          */
 398         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
 399         /*
 400          * we use page->mapping and page->index in second tail page
 401          * as list_head: assuming THP order >= 2
 402          */
 403         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 404
 405         err = hugepage_init_sysfs(&hugepage_kobj);
 406         if (err)
 407                 goto err_sysfs;
 408
 409         err = khugepaged_init();
 410         if (err)
 411                 goto err_slab;
 412
 413         err = register_shrinker(&huge_zero_page_shrinker);
 414         if (err)
 415                 goto err_hzp_shrinker;
 416         err = register_shrinker(&deferred_split_shrinker);
 417         if (err)
 418                 goto err_split_shrinker;
 419
 420         /*
 421          * By default disable transparent hugepages on smaller systems,
 422          * where the extra memory used could hurt more than TLB overhead
 423          * is likely to save.  The admin can still enable it through /sys.
 424          */
 425         if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
 426                 transparent_hugepage_flags = 0;
 427                 return 0;
 428         }
 429
 430         err = start_stop_khugepaged();
 431         if (err)
 432                 goto err_khugepaged;
 433
 434         return 0;
 435 err_khugepaged:
 436         unregister_shrinker(&deferred_split_shrinker);
 437 err_split_shrinker:
 438         unregister_shrinker(&huge_zero_page_shrinker);
 439 err_hzp_shrinker:
 440         khugepaged_destroy();
 441 err_slab:
 442         hugepage_exit_sysfs(hugepage_kobj);
 443 err_sysfs:
 444         return err;
 445 }
 446 subsys_initcall(hugepage_init);
 447
 448 static int __init setup_transparent_hugepage(char *str)
 449 {
 450         int ret = 0;
 451         if (!str)
 452                 goto out;
 453         if (!strcmp(str, "always")) {
 454                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 455                         &transparent_hugepage_flags);
 456                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 457                           &transparent_hugepage_flags);
 458                 ret = 1;
 459         } else if (!strcmp(str, "madvise")) {
 460                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 461                           &transparent_hugepage_flags);
 462                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 463                         &transparent_hugepage_flags);
 464                 ret = 1;
 465         } else if (!strcmp(str, "never")) {
 466                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 467                           &transparent_hugepage_flags);
 468                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 469                           &transparent_hugepage_flags);
 470                 ret = 1;
 471         }
 472 out:
 473         if (!ret)
 474                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 475         return ret;
 476 }
 477 __setup("transparent_hugepage=", setup_transparent_hugepage);
 478
 479 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 480 {
 481         if (likely(vma->vm_flags & VM_WRITE))
 482                 pmd = pmd_mkwrite(pmd);
 483         return pmd;
 484 }
 485
 486 #ifdef CONFIG_MEMCG
 487 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
 488 {
 489         struct mem_cgroup *memcg = compound_head(page)->mem_cgroup;
 490         struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
 491
 492         if (memcg)
 493                 return &memcg->deferred_split_queue;
 494         else
 495                 return &pgdat->deferred_split_queue;
 496 }
 497 #else
 498 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
 499 {
 500         struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
 501
 502         return &pgdat->deferred_split_queue;
 503 }
 504 #endif
 505
 506 void prep_transhuge_page(struct page *page)
 507 {
 508         /*
 509          * we use page->mapping and page->indexlru in second tail page
 510          * as list_head: assuming THP order >= 2
 511          */
 512
 513         INIT_LIST_HEAD(page_deferred_list(page));
 514         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 515 }
 516
 517 bool is_transparent_hugepage(struct page *page)
 518 {
 519         if (!PageCompound(page))
 520                 return false;
 521
 522         page = compound_head(page);
 523         return is_huge_zero_page(page) ||
 524                page[1].compound_dtor == TRANSHUGE_PAGE_DTOR;
 525 }
 526 EXPORT_SYMBOL_GPL(is_transparent_hugepage);
 527
 528 static unsigned long __thp_get_unmapped_area(struct file *filp,
 529                 unsigned long addr, unsigned long len,
 530                 loff_t off, unsigned long flags, unsigned long size)
 531 {
 532         loff_t off_end = off + len;
 533         loff_t off_align = round_up(off, size);
 534         unsigned long len_pad, ret;
 535
 536         if (off_end <= off_align || (off_end - off_align) < size)
 537                 return 0;
 538
 539         len_pad = len + size;
 540         if (len_pad < len || (off + len_pad) < off)
 541                 return 0;
 542
 543         ret = current->mm->get_unmapped_area(filp, addr, len_pad,
 544                                               off >> PAGE_SHIFT, flags);
 545
 546         /*
 547          * The failure might be due to length padding. The caller will retry
 548          * without the padding.
 549          */
 550         if (IS_ERR_VALUE(ret))
 551                 return 0;
 552
 553         /*
 554          * Do not try to align to THP boundary if allocation at the address
 555          * hint succeeds.
 556          */
 557         if (ret == addr)
 558                 return addr;
 559
 560         ret += (off - ret) & (size - 1);
 561         return ret;
 562 }
 563
 564 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 565                 unsigned long len, unsigned long pgoff, unsigned long flags)
 566 {
 567         unsigned long ret;
 568         loff_t off = (loff_t)pgoff << PAGE_SHIFT;
 569
 570         if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
 571                 goto out;
 572
 573         ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE);
 574         if (ret)
 575                 return ret;
 576 out:
 577         return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
 578 }
 579 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
 580
 581 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
 582                         struct page *page, gfp_t gfp)
 583 {
 584         struct vm_area_struct *vma = vmf->vma;
 585         pgtable_t pgtable;
 586         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 587         vm_fault_t ret = 0;
 588
 589         VM_BUG_ON_PAGE(!PageCompound(page), page);
 590
 591         if (mem_cgroup_charge(page, vma->vm_mm, gfp)) {
 592                 put_page(page);
 593                 count_vm_event(THP_FAULT_FALLBACK);
 594                 count_vm_event(THP_FAULT_FALLBACK_CHARGE);
 595                 return VM_FAULT_FALLBACK;
 596         }
 597         cgroup_throttle_swaprate(page, gfp);
 598
 599         pgtable = pte_alloc_one(vma->vm_mm);
 600         if (unlikely(!pgtable)) {
 601                 ret = VM_FAULT_OOM;
 602                 goto release;
 603         }
 604
 605         clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
 606         /*
 607          * The memory barrier inside __SetPageUptodate makes sure that
 608          * clear_huge_page writes become visible before the set_pmd_at()
 609          * write.
 610          */
 611         __SetPageUptodate(page);
 612
 613         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 614         if (unlikely(!pmd_none(*vmf->pmd))) {
 615                 goto unlock_release;
 616         } else {
 617                 pmd_t entry;
 618
 619                 ret = check_stable_address_space(vma->vm_mm);
 620                 if (ret)
 621                         goto unlock_release;
 622
 623                 /* Deliver the page fault to userland */
 624                 if (userfaultfd_missing(vma)) {
 625                         vm_fault_t ret2;
 626
 627                         spin_unlock(vmf->ptl);
 628                         put_page(page);
 629                         pte_free(vma->vm_mm, pgtable);
 630                         ret2 = handle_userfault(vmf, VM_UFFD_MISSING);
 631                         VM_BUG_ON(ret2 & VM_FAULT_FALLBACK);
 632                         return ret2;
 633                 }
 634
 635                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 636                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 637                 page_add_new_anon_rmap(page, vma, haddr, true);
 638                 lru_cache_add_inactive_or_unevictable(page, vma);
 639                 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
 640                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
 641                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 642                 mm_inc_nr_ptes(vma->vm_mm);
 643                 spin_unlock(vmf->ptl);
 644                 count_vm_event(THP_FAULT_ALLOC);
 645                 count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC);
 646         }
 647
 648         return 0;
 649 unlock_release:
 650         spin_unlock(vmf->ptl);
 651 release:
 652         if (pgtable)
 653                 pte_free(vma->vm_mm, pgtable);
 654         put_page(page);
 655         return ret;
 656
 657 }
 658
 659 /*
 660  * always: directly stall for all thp allocations
 661  * defer: wake kswapd and fail if not immediately available
 662  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
 663  *                fail if not immediately available
 664  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
 665  *          available
 666  * never: never stall for any thp allocation
 667  */
 668 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
 669 {
 670         const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
 671
 672         /* Always do synchronous compaction */
 673         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 674                 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
 675
 676         /* Kick kcompactd and fail quickly */
 677         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 678                 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
 679
 680         /* Synchronous compaction if madvised, otherwise kick kcompactd */
 681         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 682                 return GFP_TRANSHUGE_LIGHT |
 683                         (vma_madvised ? __GFP_DIRECT_RECLAIM :
 684                                         __GFP_KSWAPD_RECLAIM);
 685
 686         /* Only do synchronous compaction if madvised */
 687         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 688                 return GFP_TRANSHUGE_LIGHT |
 689                        (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
 690
 691         return GFP_TRANSHUGE_LIGHT;
 692 }
 693
 694 /* Caller must hold page table lock. */
 695 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 696                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 697                 struct page *zero_page)
 698 {
 699         pmd_t entry;
 700         if (!pmd_none(*pmd))
 701                 return false;
 702         entry = mk_pmd(zero_page, vma->vm_page_prot);
 703         entry = pmd_mkhuge(entry);
 704         if (pgtable)
 705                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 706         set_pmd_at(mm, haddr, pmd, entry);
 707         mm_inc_nr_ptes(mm);
 708         return true;
 709 }
 710
 711 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
 712 {
 713         struct vm_area_struct *vma = vmf->vma;
 714         gfp_t gfp;
 715         struct page *page;
 716         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 717
 718         if (!transhuge_vma_suitable(vma, haddr))
 719                 return VM_FAULT_FALLBACK;
 720         if (unlikely(anon_vma_prepare(vma)))
 721                 return VM_FAULT_OOM;
 722         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 723                 return VM_FAULT_OOM;
 724         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
 725                         !mm_forbids_zeropage(vma->vm_mm) &&
 726                         transparent_hugepage_use_zero_page()) {
 727                 pgtable_t pgtable;
 728                 struct page *zero_page;
 729                 vm_fault_t ret;
 730                 pgtable = pte_alloc_one(vma->vm_mm);
 731                 if (unlikely(!pgtable))
 732                         return VM_FAULT_OOM;
 733                 zero_page = mm_get_huge_zero_page(vma->vm_mm);
 734                 if (unlikely(!zero_page)) {
 735                         pte_free(vma->vm_mm, pgtable);
 736                         count_vm_event(THP_FAULT_FALLBACK);
 737                         return VM_FAULT_FALLBACK;
 738                 }
 739                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 740                 ret = 0;
 741                 if (pmd_none(*vmf->pmd)) {
 742                         ret = check_stable_address_space(vma->vm_mm);
 743                         if (ret) {
 744                                 spin_unlock(vmf->ptl);
 745                                 pte_free(vma->vm_mm, pgtable);
 746                         } else if (userfaultfd_missing(vma)) {
 747                                 spin_unlock(vmf->ptl);
 748                                 pte_free(vma->vm_mm, pgtable);
 749                                 ret = handle_userfault(vmf, VM_UFFD_MISSING);
 750                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 751                         } else {
 752                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
 753                                                    haddr, vmf->pmd, zero_page);
 754                                 spin_unlock(vmf->ptl);
 755                         }
 756                 } else {
 757                         spin_unlock(vmf->ptl);
 758                         pte_free(vma->vm_mm, pgtable);
 759                 }
 760                 return ret;
 761         }
 762         gfp = alloc_hugepage_direct_gfpmask(vma);
 763         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 764         if (unlikely(!page)) {
 765                 count_vm_event(THP_FAULT_FALLBACK);
 766                 return VM_FAULT_FALLBACK;
 767         }
 768         prep_transhuge_page(page);
 769         return __do_huge_pmd_anonymous_page(vmf, page, gfp);
 770 }
 771
 772 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 773                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
 774                 pgtable_t pgtable)
 775 {
 776         struct mm_struct *mm = vma->vm_mm;
 777         pmd_t entry;
 778         spinlock_t *ptl;
 779
 780         ptl = pmd_lock(mm, pmd);
 781         if (!pmd_none(*pmd)) {
 782                 if (write) {
 783                         if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
 784                                 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
 785                                 goto out_unlock;
 786                         }
 787                         entry = pmd_mkyoung(*pmd);
 788                         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 789                         if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
 790                                 update_mmu_cache_pmd(vma, addr, pmd);
 791                 }
 792
 793                 goto out_unlock;
 794         }
 795
 796         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 797         if (pfn_t_devmap(pfn))
 798                 entry = pmd_mkdevmap(entry);
 799         if (write) {
 800                 entry = pmd_mkyoung(pmd_mkdirty(entry));
 801                 entry = maybe_pmd_mkwrite(entry, vma);
 802         }
 803
 804         if (pgtable) {
 805                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 806                 mm_inc_nr_ptes(mm);
 807                 pgtable = NULL;
 808         }
 809
 810         set_pmd_at(mm, addr, pmd, entry);
 811         update_mmu_cache_pmd(vma, addr, pmd);
 812
 813 out_unlock:
 814         spin_unlock(ptl);
 815         if (pgtable)
 816                 pte_free(mm, pgtable);
 817 }
 818
 819 /**
 820  * vmf_insert_pfn_pmd_prot - insert a pmd size pfn
 821  * @vmf: Structure describing the fault
 822  * @pfn: pfn to insert
 823  * @pgprot: page protection to use
 824  * @write: whether it's a write fault
 825  *
 826  * Insert a pmd size pfn. See vmf_insert_pfn() for additional info and
 827  * also consult the vmf_insert_mixed_prot() documentation when
 828  * @pgprot != @vmf->vma->vm_page_prot.
 829  *
 830  * Return: vm_fault_t value.
 831  */
 832 vm_fault_t vmf_insert_pfn_pmd_prot(struct vm_fault *vmf, pfn_t pfn,
 833                                    pgprot_t pgprot, bool write)
 834 {
 835         unsigned long addr = vmf->address & PMD_MASK;
 836         struct vm_area_struct *vma = vmf->vma;
 837         pgtable_t pgtable = NULL;
 838
 839         /*
 840          * If we had pmd_special, we could avoid all these restrictions,
 841          * but we need to be consistent with PTEs and architectures that
 842          * can't support a 'special' bit.
 843          */
 844         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 845                         !pfn_t_devmap(pfn));
 846         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 847                                                 (VM_PFNMAP|VM_MIXEDMAP));
 848         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 849
 850         if (addr < vma->vm_start || addr >= vma->vm_end)
 851                 return VM_FAULT_SIGBUS;
 852
 853         if (arch_needs_pgtable_deposit()) {
 854                 pgtable = pte_alloc_one(vma->vm_mm);
 855                 if (!pgtable)
 856                         return VM_FAULT_OOM;
 857         }
 858
 859         track_pfn_insert(vma, &pgprot, pfn);
 860
 861         insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
 862         return VM_FAULT_NOPAGE;
 863 }
 864 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd_prot);
 865
 866 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 867 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
 868 {
 869         if (likely(vma->vm_flags & VM_WRITE))
 870                 pud = pud_mkwrite(pud);
 871         return pud;
 872 }
 873
 874 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 875                 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
 876 {
 877         struct mm_struct *mm = vma->vm_mm;
 878         pud_t entry;
 879         spinlock_t *ptl;
 880
 881         ptl = pud_lock(mm, pud);
 882         if (!pud_none(*pud)) {
 883                 if (write) {
 884                         if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
 885                                 WARN_ON_ONCE(!is_huge_zero_pud(*pud));
 886                                 goto out_unlock;
 887                         }
 888                         entry = pud_mkyoung(*pud);
 889                         entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
 890                         if (pudp_set_access_flags(vma, addr, pud, entry, 1))
 891                                 update_mmu_cache_pud(vma, addr, pud);
 892                 }
 893                 goto out_unlock;
 894         }
 895
 896         entry = pud_mkhuge(pfn_t_pud(pfn, prot));
 897         if (pfn_t_devmap(pfn))
 898                 entry = pud_mkdevmap(entry);
 899         if (write) {
 900                 entry = pud_mkyoung(pud_mkdirty(entry));
 901                 entry = maybe_pud_mkwrite(entry, vma);
 902         }
 903         set_pud_at(mm, addr, pud, entry);
 904         update_mmu_cache_pud(vma, addr, pud);
 905
 906 out_unlock:
 907         spin_unlock(ptl);
 908 }
 909
 910 /**
 911  * vmf_insert_pfn_pud_prot - insert a pud size pfn
 912  * @vmf: Structure describing the fault
 913  * @pfn: pfn to insert
 914  * @pgprot: page protection to use
 915  * @write: whether it's a write fault
 916  *
 917  * Insert a pud size pfn. See vmf_insert_pfn() for additional info and
 918  * also consult the vmf_insert_mixed_prot() documentation when
 919  * @pgprot != @vmf->vma->vm_page_prot.
 920  *
 921  * Return: vm_fault_t value.
 922  */
 923 vm_fault_t vmf_insert_pfn_pud_prot(struct vm_fault *vmf, pfn_t pfn,
 924                                    pgprot_t pgprot, bool write)
 925 {
 926         unsigned long addr = vmf->address & PUD_MASK;
 927         struct vm_area_struct *vma = vmf->vma;
 928
 929         /*
 930          * If we had pud_special, we could avoid all these restrictions,
 931          * but we need to be consistent with PTEs and architectures that
 932          * can't support a 'special' bit.
 933          */
 934         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 935                         !pfn_t_devmap(pfn));
 936         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 937                                                 (VM_PFNMAP|VM_MIXEDMAP));
 938         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 939
 940         if (addr < vma->vm_start || addr >= vma->vm_end)
 941                 return VM_FAULT_SIGBUS;
 942
 943         track_pfn_insert(vma, &pgprot, pfn);
 944
 945         insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
 946         return VM_FAULT_NOPAGE;
 947 }
 948 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud_prot);
 949 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
 950
 951 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
 952                 pmd_t *pmd, int flags)
 953 {
 954         pmd_t _pmd;
 955
 956         _pmd = pmd_mkyoung(*pmd);
 957         if (flags & FOLL_WRITE)
 958                 _pmd = pmd_mkdirty(_pmd);
 959         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
 960                                 pmd, _pmd, flags & FOLL_WRITE))
 961                 update_mmu_cache_pmd(vma, addr, pmd);
 962 }
 963
 964 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 965                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
 966 {
 967         unsigned long pfn = pmd_pfn(*pmd);
 968         struct mm_struct *mm = vma->vm_mm;
 969         struct page *page;
 970
 971         assert_spin_locked(pmd_lockptr(mm, pmd));
 972
 973         /*
 974          * When we COW a devmap PMD entry, we split it into PTEs, so we should
 975          * not be in this function with `flags & FOLL_COW` set.
 976          */
 977         WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
 978
 979         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
 980         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
 981                          (FOLL_PIN | FOLL_GET)))
 982                 return NULL;
 983
 984         if (flags & FOLL_WRITE && !pmd_write(*pmd))
 985                 return NULL;
 986
 987         if (pmd_present(*pmd) && pmd_devmap(*pmd))
 988                 /* pass */;
 989         else
 990                 return NULL;
 991
 992         if (flags & FOLL_TOUCH)
 993                 touch_pmd(vma, addr, pmd, flags);
 994
 995         /*
 996          * device mapped pages can only be returned if the
 997          * caller will manage the page reference count.
 998          */
 999         if (!(flags & (FOLL_GET | FOLL_PIN)))
1000                 return ERR_PTR(-EEXIST);
1001
1002         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1003         *pgmap = get_dev_pagemap(pfn, *pgmap);
1004         if (!*pgmap)
1005                 return ERR_PTR(-EFAULT);
1006         page = pfn_to_page(pfn);
1007         if (!try_grab_page(page, flags))
1008                 page = ERR_PTR(-ENOMEM);
1009
1010         return page;
1011 }
1012
1013 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1014                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1015                   struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1016 {
1017         spinlock_t *dst_ptl, *src_ptl;
1018         struct page *src_page;
1019         pmd_t pmd;
1020         pgtable_t pgtable = NULL;
1021         int ret = -ENOMEM;
1022
1023         /* Skip if can be re-fill on fault */
1024         if (!vma_is_anonymous(dst_vma))
1025                 return 0;
1026
1027         pgtable = pte_alloc_one(dst_mm);
1028         if (unlikely(!pgtable))
1029                 goto out;
1030
1031         dst_ptl = pmd_lock(dst_mm, dst_pmd);
1032         src_ptl = pmd_lockptr(src_mm, src_pmd);
1033         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1034
1035         ret = -EAGAIN;
1036         pmd = *src_pmd;
1037
1038 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1039         if (unlikely(is_swap_pmd(pmd))) {
1040                 swp_entry_t entry = pmd_to_swp_entry(pmd);
1041
1042                 VM_BUG_ON(!is_pmd_migration_entry(pmd));
1043                 if (is_write_migration_entry(entry)) {
1044                         make_migration_entry_read(&entry);
1045                         pmd = swp_entry_to_pmd(entry);
1046                         if (pmd_swp_soft_dirty(*src_pmd))
1047                                 pmd = pmd_swp_mksoft_dirty(pmd);
1048                         if (pmd_swp_uffd_wp(*src_pmd))
1049                                 pmd = pmd_swp_mkuffd_wp(pmd);
1050                         set_pmd_at(src_mm, addr, src_pmd, pmd);
1051                 }
1052                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1053                 mm_inc_nr_ptes(dst_mm);
1054                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1055                 if (!userfaultfd_wp(dst_vma))
1056                         pmd = pmd_swp_clear_uffd_wp(pmd);
1057                 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1058                 ret = 0;
1059                 goto out_unlock;
1060         }
1061 #endif
1062
1063         if (unlikely(!pmd_trans_huge(pmd))) {
1064                 pte_free(dst_mm, pgtable);
1065                 goto out_unlock;
1066         }
1067         /*
1068          * When page table lock is held, the huge zero pmd should not be
1069          * under splitting since we don't split the page itself, only pmd to
1070          * a page table.
1071          */
1072         if (is_huge_zero_pmd(pmd)) {
1073                 /*
1074                  * get_huge_zero_page() will never allocate a new page here,
1075                  * since we already have a zero page to copy. It just takes a
1076                  * reference.
1077                  */
1078                 mm_get_huge_zero_page(dst_mm);
1079                 goto out_zero_page;
1080         }
1081
1082         src_page = pmd_page(pmd);
1083         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1084
1085         /*
1086          * If this page is a potentially pinned page, split and retry the fault
1087          * with smaller page size.  Normally this should not happen because the
1088          * userspace should use MADV_DONTFORK upon pinned regions.  This is a
1089          * best effort that the pinned pages won't be replaced by another
1090          * random page during the coming copy-on-write.
1091          */
1092         if (unlikely(is_cow_mapping(src_vma->vm_flags) &&
1093                      atomic_read(&src_mm->has_pinned) &&
1094                      page_maybe_dma_pinned(src_page))) {
1095                 pte_free(dst_mm, pgtable);
1096                 spin_unlock(src_ptl);
1097                 spin_unlock(dst_ptl);
1098                 __split_huge_pmd(src_vma, src_pmd, addr, false, NULL);
1099                 return -EAGAIN;
1100         }
1101
1102         get_page(src_page);
1103         page_dup_rmap(src_page, true);
1104         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1105 out_zero_page:
1106         mm_inc_nr_ptes(dst_mm);
1107         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1108         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1109         if (!userfaultfd_wp(dst_vma))
1110                 pmd = pmd_clear_uffd_wp(pmd);
1111         pmd = pmd_mkold(pmd_wrprotect(pmd));
1112         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1113
1114         ret = 0;
1115 out_unlock:
1116         spin_unlock(src_ptl);
1117         spin_unlock(dst_ptl);
1118 out:
1119         return ret;
1120 }
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 (total_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                 spin_unlock(ptl);
1674                 if (is_huge_zero_pmd(orig_pmd))
1675                         tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1676         } else if (is_huge_zero_pmd(orig_pmd)) {
1677                 zap_deposited_table(tlb->mm, pmd);
1678                 spin_unlock(ptl);
1679                 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1680         } else {
1681                 struct page *page = NULL;
1682                 int flush_needed = 1;
1683
1684                 if (pmd_present(orig_pmd)) {
1685                         page = pmd_page(orig_pmd);
1686                         page_remove_rmap(page, true);
1687                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1688                         VM_BUG_ON_PAGE(!PageHead(page), page);
1689                 } else if (thp_migration_supported()) {
1690                         swp_entry_t entry;
1691
1692                         VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1693                         entry = pmd_to_swp_entry(orig_pmd);
1694                         page = pfn_to_page(swp_offset(entry));
1695                         flush_needed = 0;
1696                 } else
1697                         WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1698
1699                 if (PageAnon(page)) {
1700                         zap_deposited_table(tlb->mm, pmd);
1701                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1702                 } else {
1703                         if (arch_needs_pgtable_deposit())
1704                                 zap_deposited_table(tlb->mm, pmd);
1705                         add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1706                 }
1707
1708                 spin_unlock(ptl);
1709                 if (flush_needed)
1710                         tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1711         }
1712         return 1;
1713 }
1714
1715 #ifndef pmd_move_must_withdraw
1716 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1717                                          spinlock_t *old_pmd_ptl,
1718                                          struct vm_area_struct *vma)
1719 {
1720         /*
1721          * With split pmd lock we also need to move preallocated
1722          * PTE page table if new_pmd is on different PMD page table.
1723          *
1724          * We also don't deposit and withdraw tables for file pages.
1725          */
1726         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1727 }
1728 #endif
1729
1730 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1731 {
1732 #ifdef CONFIG_MEM_SOFT_DIRTY
1733         if (unlikely(is_pmd_migration_entry(pmd)))
1734                 pmd = pmd_swp_mksoft_dirty(pmd);
1735         else if (pmd_present(pmd))
1736                 pmd = pmd_mksoft_dirty(pmd);
1737 #endif
1738         return pmd;
1739 }
1740
1741 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1742                   unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
1743 {
1744         spinlock_t *old_ptl, *new_ptl;
1745         pmd_t pmd;
1746         struct mm_struct *mm = vma->vm_mm;
1747         bool force_flush = false;
1748
1749         /*
1750          * The destination pmd shouldn't be established, free_pgtables()
1751          * should have release it.
1752          */
1753         if (WARN_ON(!pmd_none(*new_pmd))) {
1754                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1755                 return false;
1756         }
1757
1758         /*
1759          * We don't have to worry about the ordering of src and dst
1760          * ptlocks because exclusive mmap_lock prevents deadlock.
1761          */
1762         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1763         if (old_ptl) {
1764                 new_ptl = pmd_lockptr(mm, new_pmd);
1765                 if (new_ptl != old_ptl)
1766                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1767                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1768                 if (pmd_present(pmd))
1769                         force_flush = true;
1770                 VM_BUG_ON(!pmd_none(*new_pmd));
1771
1772                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1773                         pgtable_t pgtable;
1774                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1775                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1776                 }
1777                 pmd = move_soft_dirty_pmd(pmd);
1778                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1779                 if (force_flush)
1780                         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1781                 if (new_ptl != old_ptl)
1782                         spin_unlock(new_ptl);
1783                 spin_unlock(old_ptl);
1784                 return true;
1785         }
1786         return false;
1787 }
1788
1789 /*
1790  * Returns
1791  *  - 0 if PMD could not be locked
1792  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1793  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1794  */
1795 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1796                 unsigned long addr, pgprot_t newprot, unsigned long cp_flags)
1797 {
1798         struct mm_struct *mm = vma->vm_mm;
1799         spinlock_t *ptl;
1800         pmd_t entry;
1801         bool preserve_write;
1802         int ret;
1803         bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
1804         bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
1805         bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
1806
1807         ptl = __pmd_trans_huge_lock(pmd, vma);
1808         if (!ptl)
1809                 return 0;
1810
1811         preserve_write = prot_numa && pmd_write(*pmd);
1812         ret = 1;
1813
1814 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1815         if (is_swap_pmd(*pmd)) {
1816                 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1817
1818                 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1819                 if (is_write_migration_entry(entry)) {
1820                         pmd_t newpmd;
1821                         /*
1822                          * A protection check is difficult so
1823                          * just be safe and disable write
1824                          */
1825                         make_migration_entry_read(&entry);
1826                         newpmd = swp_entry_to_pmd(entry);
1827                         if (pmd_swp_soft_dirty(*pmd))
1828                                 newpmd = pmd_swp_mksoft_dirty(newpmd);
1829                         if (pmd_swp_uffd_wp(*pmd))
1830                                 newpmd = pmd_swp_mkuffd_wp(newpmd);
1831                         set_pmd_at(mm, addr, pmd, newpmd);
1832                 }
1833                 goto unlock;
1834         }
1835 #endif
1836
1837         /*
1838          * Avoid trapping faults against the zero page. The read-only
1839          * data is likely to be read-cached on the local CPU and
1840          * local/remote hits to the zero page are not interesting.
1841          */
1842         if (prot_numa && is_huge_zero_pmd(*pmd))
1843                 goto unlock;
1844
1845         if (prot_numa && pmd_protnone(*pmd))
1846                 goto unlock;
1847
1848         /*
1849          * In case prot_numa, we are under mmap_read_lock(mm). It's critical
1850          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1851          * which is also under mmap_read_lock(mm):
1852          *
1853          *      CPU0:                           CPU1:
1854          *                              change_huge_pmd(prot_numa=1)
1855          *                               pmdp_huge_get_and_clear_notify()
1856          * madvise_dontneed()
1857          *  zap_pmd_range()
1858          *   pmd_trans_huge(*pmd) == 0 (without ptl)
1859          *   // skip the pmd
1860          *                               set_pmd_at();
1861          *                               // pmd is re-established
1862          *
1863          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1864          * which may break userspace.
1865          *
1866          * pmdp_invalidate() is required to make sure we don't miss
1867          * dirty/young flags set by hardware.
1868          */
1869         entry = pmdp_invalidate(vma, addr, pmd);
1870
1871         entry = pmd_modify(entry, newprot);
1872         if (preserve_write)
1873                 entry = pmd_mk_savedwrite(entry);
1874         if (uffd_wp) {
1875                 entry = pmd_wrprotect(entry);
1876                 entry = pmd_mkuffd_wp(entry);
1877         } else if (uffd_wp_resolve) {
1878                 /*
1879                  * Leave the write bit to be handled by PF interrupt
1880                  * handler, then things like COW could be properly
1881                  * handled.
1882                  */
1883                 entry = pmd_clear_uffd_wp(entry);
1884         }
1885         ret = HPAGE_PMD_NR;
1886         set_pmd_at(mm, addr, pmd, entry);
1887         BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1888 unlock:
1889         spin_unlock(ptl);
1890         return ret;
1891 }
1892
1893 /*
1894  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1895  *
1896  * Note that if it returns page table lock pointer, this routine returns without
1897  * unlocking page table lock. So callers must unlock it.
1898  */
1899 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1900 {
1901         spinlock_t *ptl;
1902         ptl = pmd_lock(vma->vm_mm, pmd);
1903         if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1904                         pmd_devmap(*pmd)))
1905                 return ptl;
1906         spin_unlock(ptl);
1907         return NULL;
1908 }
1909
1910 /*
1911  * Returns true if a given pud maps a thp, false otherwise.
1912  *
1913  * Note that if it returns true, this routine returns without unlocking page
1914  * table lock. So callers must unlock it.
1915  */
1916 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1917 {
1918         spinlock_t *ptl;
1919
1920         ptl = pud_lock(vma->vm_mm, pud);
1921         if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1922                 return ptl;
1923         spin_unlock(ptl);
1924         return NULL;
1925 }
1926
1927 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1928 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1929                  pud_t *pud, unsigned long addr)
1930 {
1931         spinlock_t *ptl;
1932
1933         ptl = __pud_trans_huge_lock(pud, vma);
1934         if (!ptl)
1935                 return 0;
1936         /*
1937          * For architectures like ppc64 we look at deposited pgtable
1938          * when calling pudp_huge_get_and_clear. So do the
1939          * pgtable_trans_huge_withdraw after finishing pudp related
1940          * operations.
1941          */
1942         pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
1943         tlb_remove_pud_tlb_entry(tlb, pud, addr);
1944         if (vma_is_special_huge(vma)) {
1945                 spin_unlock(ptl);
1946                 /* No zero page support yet */
1947         } else {
1948                 /* No support for anonymous PUD pages yet */
1949                 BUG();
1950         }
1951         return 1;
1952 }
1953
1954 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1955                 unsigned long haddr)
1956 {
1957         VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1958         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1959         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
1960         VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
1961
1962         count_vm_event(THP_SPLIT_PUD);
1963
1964         pudp_huge_clear_flush_notify(vma, haddr, pud);
1965 }
1966
1967 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
1968                 unsigned long address)
1969 {
1970         spinlock_t *ptl;
1971         struct mmu_notifier_range range;
1972
1973         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1974                                 address & HPAGE_PUD_MASK,
1975                                 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
1976         mmu_notifier_invalidate_range_start(&range);
1977         ptl = pud_lock(vma->vm_mm, pud);
1978         if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
1979                 goto out;
1980         __split_huge_pud_locked(vma, pud, range.start);
1981
1982 out:
1983         spin_unlock(ptl);
1984         /*
1985          * No need to double call mmu_notifier->invalidate_range() callback as
1986          * the above pudp_huge_clear_flush_notify() did already call it.
1987          */
1988         mmu_notifier_invalidate_range_only_end(&range);
1989 }
1990 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1991
1992 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1993                 unsigned long haddr, pmd_t *pmd)
1994 {
1995         struct mm_struct *mm = vma->vm_mm;
1996         pgtable_t pgtable;
1997         pmd_t _pmd;
1998         int i;
1999
2000         /*
2001          * Leave pmd empty until pte is filled note that it is fine to delay
2002          * notification until mmu_notifier_invalidate_range_end() as we are
2003          * replacing a zero pmd write protected page with a zero pte write
2004          * protected page.
2005          *
2006          * See Documentation/vm/mmu_notifier.rst
2007          */
2008         pmdp_huge_clear_flush(vma, haddr, pmd);
2009
2010         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2011         pmd_populate(mm, &_pmd, pgtable);
2012
2013         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2014                 pte_t *pte, entry;
2015                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2016                 entry = pte_mkspecial(entry);
2017                 pte = pte_offset_map(&_pmd, haddr);
2018                 VM_BUG_ON(!pte_none(*pte));
2019                 set_pte_at(mm, haddr, pte, entry);
2020                 pte_unmap(pte);
2021         }
2022         smp_wmb(); /* make pte visible before pmd */
2023         pmd_populate(mm, pmd, pgtable);
2024 }
2025
2026 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2027                 unsigned long haddr, bool freeze)
2028 {
2029         struct mm_struct *mm = vma->vm_mm;
2030         struct page *page;
2031         pgtable_t pgtable;
2032         pmd_t old_pmd, _pmd;
2033         bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false;
2034         unsigned long addr;
2035         int i;
2036
2037         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2038         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2039         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2040         VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2041                                 && !pmd_devmap(*pmd));
2042
2043         count_vm_event(THP_SPLIT_PMD);
2044
2045         if (!vma_is_anonymous(vma)) {
2046                 old_pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2047                 /*
2048                  * We are going to unmap this huge page. So
2049                  * just go ahead and zap it
2050                  */
2051                 if (arch_needs_pgtable_deposit())
2052                         zap_deposited_table(mm, pmd);
2053                 if (vma_is_special_huge(vma))
2054                         return;
2055                 if (unlikely(is_pmd_migration_entry(old_pmd))) {
2056                         swp_entry_t entry;
2057
2058                         entry = pmd_to_swp_entry(old_pmd);
2059                         page = migration_entry_to_page(entry);
2060                 } else {
2061                         page = pmd_page(old_pmd);
2062                         if (!PageDirty(page) && pmd_dirty(old_pmd))
2063                                 set_page_dirty(page);
2064                         if (!PageReferenced(page) && pmd_young(old_pmd))
2065                                 SetPageReferenced(page);
2066                         page_remove_rmap(page, true);
2067                         put_page(page);
2068                 }
2069                 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2070                 return;
2071         }
2072
2073         if (is_huge_zero_pmd(*pmd)) {
2074                 /*
2075                  * FIXME: Do we want to invalidate secondary mmu by calling
2076                  * mmu_notifier_invalidate_range() see comments below inside
2077                  * __split_huge_pmd() ?
2078                  *
2079                  * We are going from a zero huge page write protected to zero
2080                  * small page also write protected so it does not seems useful
2081                  * to invalidate secondary mmu at this time.
2082                  */
2083                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2084         }
2085
2086         /*
2087          * Up to this point the pmd is present and huge and userland has the
2088          * whole access to the hugepage during the split (which happens in
2089          * place). If we overwrite the pmd with the not-huge version pointing
2090          * to the pte here (which of course we could if all CPUs were bug
2091          * free), userland could trigger a small page size TLB miss on the
2092          * small sized TLB while the hugepage TLB entry is still established in
2093          * the huge TLB. Some CPU doesn't like that.
2094          * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum
2095          * 383 on page 105. Intel should be safe but is also warns that it's
2096          * only safe if the permission and cache attributes of the two entries
2097          * loaded in the two TLB is identical (which should be the case here).
2098          * But it is generally safer to never allow small and huge TLB entries
2099          * for the same virtual address to be loaded simultaneously. So instead
2100          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2101          * current pmd notpresent (atomically because here the pmd_trans_huge
2102          * must remain set at all times on the pmd until the split is complete
2103          * for this pmd), then we flush the SMP TLB and finally we write the
2104          * non-huge version of the pmd entry with pmd_populate.
2105          */
2106         old_pmd = pmdp_invalidate(vma, haddr, pmd);
2107
2108         pmd_migration = is_pmd_migration_entry(old_pmd);
2109         if (unlikely(pmd_migration)) {
2110                 swp_entry_t entry;
2111
2112                 entry = pmd_to_swp_entry(old_pmd);
2113                 page = pfn_to_page(swp_offset(entry));
2114                 write = is_write_migration_entry(entry);
2115                 young = false;
2116                 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2117                 uffd_wp = pmd_swp_uffd_wp(old_pmd);
2118         } else {
2119                 page = pmd_page(old_pmd);
2120                 if (pmd_dirty(old_pmd))
2121                         SetPageDirty(page);
2122                 write = pmd_write(old_pmd);
2123                 young = pmd_young(old_pmd);
2124                 soft_dirty = pmd_soft_dirty(old_pmd);
2125                 uffd_wp = pmd_uffd_wp(old_pmd);
2126         }
2127         VM_BUG_ON_PAGE(!page_count(page), page);
2128         page_ref_add(page, HPAGE_PMD_NR - 1);
2129
2130         /*
2131          * Withdraw the table only after we mark the pmd entry invalid.
2132          * This's critical for some architectures (Power).
2133          */
2134         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2135         pmd_populate(mm, &_pmd, pgtable);
2136
2137         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2138                 pte_t entry, *pte;
2139                 /*
2140                  * Note that NUMA hinting access restrictions are not
2141                  * transferred to avoid any possibility of altering
2142                  * permissions across VMAs.
2143                  */
2144                 if (freeze || pmd_migration) {
2145                         swp_entry_t swp_entry;
2146                         swp_entry = make_migration_entry(page + i, write);
2147                         entry = swp_entry_to_pte(swp_entry);
2148                         if (soft_dirty)
2149                                 entry = pte_swp_mksoft_dirty(entry);
2150                         if (uffd_wp)
2151                                 entry = pte_swp_mkuffd_wp(entry);
2152                 } else {
2153                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2154                         entry = maybe_mkwrite(entry, vma);
2155                         if (!write)
2156                                 entry = pte_wrprotect(entry);
2157                         if (!young)
2158                                 entry = pte_mkold(entry);
2159                         if (soft_dirty)
2160                                 entry = pte_mksoft_dirty(entry);
2161                         if (uffd_wp)
2162                                 entry = pte_mkuffd_wp(entry);
2163                 }
2164                 pte = pte_offset_map(&_pmd, addr);
2165                 BUG_ON(!pte_none(*pte));
2166                 set_pte_at(mm, addr, pte, entry);
2167                 if (!pmd_migration)
2168                         atomic_inc(&page[i]._mapcount);
2169                 pte_unmap(pte);
2170         }
2171
2172         if (!pmd_migration) {
2173                 /*
2174                  * Set PG_double_map before dropping compound_mapcount to avoid
2175                  * false-negative page_mapped().
2176                  */
2177                 if (compound_mapcount(page) > 1 &&
2178                     !TestSetPageDoubleMap(page)) {
2179                         for (i = 0; i < HPAGE_PMD_NR; i++)
2180                                 atomic_inc(&page[i]._mapcount);
2181                 }
2182
2183                 lock_page_memcg(page);
2184                 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2185                         /* Last compound_mapcount is gone. */
2186                         __dec_lruvec_page_state(page, NR_ANON_THPS);
2187                         if (TestClearPageDoubleMap(page)) {
2188                                 /* No need in mapcount reference anymore */
2189                                 for (i = 0; i < HPAGE_PMD_NR; i++)
2190                                         atomic_dec(&page[i]._mapcount);
2191                         }
2192                 }
2193                 unlock_page_memcg(page);
2194         }
2195
2196         smp_wmb(); /* make pte visible before pmd */
2197         pmd_populate(mm, pmd, pgtable);
2198
2199         if (freeze) {
2200                 for (i = 0; i < HPAGE_PMD_NR; i++) {
2201                         page_remove_rmap(page + i, false);
2202                         put_page(page + i);
2203                 }
2204         }
2205 }
2206
2207 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2208                 unsigned long address, bool freeze, struct page *page)
2209 {
2210         spinlock_t *ptl;
2211         struct mmu_notifier_range range;
2212         bool do_unlock_page = false;
2213         pmd_t _pmd;
2214
2215         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2216                                 address & HPAGE_PMD_MASK,
2217                                 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2218         mmu_notifier_invalidate_range_start(&range);
2219         ptl = pmd_lock(vma->vm_mm, pmd);
2220
2221         /*
2222          * If caller asks to setup a migration entries, we need a page to check
2223          * pmd against. Otherwise we can end up replacing wrong page.
2224          */
2225         VM_BUG_ON(freeze && !page);
2226         if (page) {
2227                 VM_WARN_ON_ONCE(!PageLocked(page));
2228                 if (page != pmd_page(*pmd))
2229                         goto out;
2230         }
2231
2232 repeat:
2233         if (pmd_trans_huge(*pmd)) {
2234                 if (!page) {
2235                         page = pmd_page(*pmd);
2236                         /*
2237                          * An anonymous page must be locked, to ensure that a
2238                          * concurrent reuse_swap_page() sees stable mapcount;
2239                          * but reuse_swap_page() is not used on shmem or file,
2240                          * and page lock must not be taken when zap_pmd_range()
2241                          * calls __split_huge_pmd() while i_mmap_lock is held.
2242                          */
2243                         if (PageAnon(page)) {
2244                                 if (unlikely(!trylock_page(page))) {
2245                                         get_page(page);
2246                                         _pmd = *pmd;
2247                                         spin_unlock(ptl);
2248                                         lock_page(page);
2249                                         spin_lock(ptl);
2250                                         if (unlikely(!pmd_same(*pmd, _pmd))) {
2251                                                 unlock_page(page);
2252                                                 put_page(page);
2253                                                 page = NULL;
2254                                                 goto repeat;
2255                                         }
2256                                         put_page(page);
2257                                 }
2258                                 do_unlock_page = true;
2259                         }
2260                 }
2261                 if (PageMlocked(page))
2262                         clear_page_mlock(page);
2263         } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2264                 goto out;
2265         __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2266 out:
2267         spin_unlock(ptl);
2268         if (do_unlock_page)
2269                 unlock_page(page);
2270         /*
2271          * No need to double call mmu_notifier->invalidate_range() callback.
2272          * They are 3 cases to consider inside __split_huge_pmd_locked():
2273          *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2274          *  2) __split_huge_zero_page_pmd() read only zero page and any write
2275          *    fault will trigger a flush_notify before pointing to a new page
2276          *    (it is fine if the secondary mmu keeps pointing to the old zero
2277          *    page in the meantime)
2278          *  3) Split a huge pmd into pte pointing to the same page. No need
2279          *     to invalidate secondary tlb entry they are all still valid.
2280          *     any further changes to individual pte will notify. So no need
2281          *     to call mmu_notifier->invalidate_range()
2282          */
2283         mmu_notifier_invalidate_range_only_end(&range);
2284 }
2285
2286 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2287                 bool freeze, struct page *page)
2288 {
2289         pgd_t *pgd;
2290         p4d_t *p4d;
2291         pud_t *pud;
2292         pmd_t *pmd;
2293
2294         pgd = pgd_offset(vma->vm_mm, address);
2295         if (!pgd_present(*pgd))
2296                 return;
2297
2298         p4d = p4d_offset(pgd, address);
2299         if (!p4d_present(*p4d))
2300                 return;
2301
2302         pud = pud_offset(p4d, address);
2303         if (!pud_present(*pud))
2304                 return;
2305
2306         pmd = pmd_offset(pud, address);
2307
2308         __split_huge_pmd(vma, pmd, address, freeze, page);
2309 }
2310
2311 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2312                              unsigned long start,
2313                              unsigned long end,
2314                              long adjust_next)
2315 {
2316         /*
2317          * If the new start address isn't hpage aligned and it could
2318          * previously contain an hugepage: check if we need to split
2319          * an huge pmd.
2320          */
2321         if (start & ~HPAGE_PMD_MASK &&
2322             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2323             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2324                 split_huge_pmd_address(vma, start, false, NULL);
2325
2326         /*
2327          * If the new end address isn't hpage aligned and it could
2328          * previously contain an hugepage: check if we need to split
2329          * an huge pmd.
2330          */
2331         if (end & ~HPAGE_PMD_MASK &&
2332             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2333             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2334                 split_huge_pmd_address(vma, end, false, NULL);
2335
2336         /*
2337          * If we're also updating the vma->vm_next->vm_start, if the new
2338          * vm_next->vm_start isn't hpage aligned and it could previously
2339          * contain an hugepage: check if we need to split an huge pmd.
2340          */
2341         if (adjust_next > 0) {
2342                 struct vm_area_struct *next = vma->vm_next;
2343                 unsigned long nstart = next->vm_start;
2344                 nstart += adjust_next;
2345                 if (nstart & ~HPAGE_PMD_MASK &&
2346                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2347                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2348                         split_huge_pmd_address(next, nstart, false, NULL);
2349         }
2350 }
2351
2352 static void unmap_page(struct page *page)
2353 {
2354         enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_SYNC |
2355                 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2356
2357         VM_BUG_ON_PAGE(!PageHead(page), page);
2358
2359         if (PageAnon(page))
2360                 ttu_flags |= TTU_SPLIT_FREEZE;
2361
2362         try_to_unmap(page, ttu_flags);
2363
2364         VM_WARN_ON_ONCE_PAGE(page_mapped(page), page);
2365 }
2366
2367 static void remap_page(struct page *page, unsigned int nr)
2368 {
2369         int i;
2370         if (PageTransHuge(page)) {
2371                 remove_migration_ptes(page, page, true);
2372         } else {
2373                 for (i = 0; i < nr; i++)
2374                         remove_migration_ptes(page + i, page + i, true);
2375         }
2376 }
2377
2378 static void __split_huge_page_tail(struct page *head, int tail,
2379                 struct lruvec *lruvec, struct list_head *list)
2380 {
2381         struct page *page_tail = head + tail;
2382
2383         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2384
2385         /*
2386          * Clone page flags before unfreezing refcount.
2387          *
2388          * After successful get_page_unless_zero() might follow flags change,
2389          * for exmaple lock_page() which set PG_waiters.
2390          */
2391         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2392         page_tail->flags |= (head->flags &
2393                         ((1L << PG_referenced) |
2394                          (1L << PG_swapbacked) |
2395                          (1L << PG_swapcache) |
2396                          (1L << PG_mlocked) |
2397                          (1L << PG_uptodate) |
2398                          (1L << PG_active) |
2399                          (1L << PG_workingset) |
2400                          (1L << PG_locked) |
2401                          (1L << PG_unevictable) |
2402 #ifdef CONFIG_64BIT
2403                          (1L << PG_arch_2) |
2404 #endif
2405                          (1L << PG_dirty)));
2406
2407         /* ->mapping in first tail page is compound_mapcount */
2408         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2409                         page_tail);
2410         page_tail->mapping = head->mapping;
2411         page_tail->index = head->index + tail;
2412
2413         /* Page flags must be visible before we make the page non-compound. */
2414         smp_wmb();
2415
2416         /*
2417          * Clear PageTail before unfreezing page refcount.
2418          *
2419          * After successful get_page_unless_zero() might follow put_page()
2420          * which needs correct compound_head().
2421          */
2422         clear_compound_head(page_tail);
2423
2424         /* Finally unfreeze refcount. Additional reference from page cache. */
2425         page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2426                                           PageSwapCache(head)));
2427
2428         if (page_is_young(head))
2429                 set_page_young(page_tail);
2430         if (page_is_idle(head))
2431                 set_page_idle(page_tail);
2432
2433         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2434
2435         /*
2436          * always add to the tail because some iterators expect new
2437          * pages to show after the currently processed elements - e.g.
2438          * migrate_pages
2439          */
2440         lru_add_page_tail(head, page_tail, lruvec, list);
2441 }
2442
2443 static void __split_huge_page(struct page *page, struct list_head *list,
2444                 pgoff_t end, unsigned long flags)
2445 {
2446         struct page *head = compound_head(page);
2447         pg_data_t *pgdat = page_pgdat(head);
2448         struct lruvec *lruvec;
2449         struct address_space *swap_cache = NULL;
2450         unsigned long offset = 0;
2451         unsigned int nr = thp_nr_pages(head);
2452         int i;
2453
2454         lruvec = mem_cgroup_page_lruvec(head, pgdat);
2455
2456         /* complete memcg works before add pages to LRU */
2457         split_page_memcg(head, nr);
2458
2459         if (PageAnon(head) && PageSwapCache(head)) {
2460                 swp_entry_t entry = { .val = page_private(head) };
2461
2462                 offset = swp_offset(entry);
2463                 swap_cache = swap_address_space(entry);
2464                 xa_lock(&swap_cache->i_pages);
2465         }
2466
2467         for (i = nr - 1; i >= 1; i--) {
2468                 __split_huge_page_tail(head, i, lruvec, list);
2469                 /* Some pages can be beyond i_size: drop them from page cache */
2470                 if (head[i].index >= end) {
2471                         ClearPageDirty(head + i);
2472                         __delete_from_page_cache(head + i, NULL);
2473                         if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2474                                 shmem_uncharge(head->mapping->host, 1);
2475                         put_page(head + i);
2476                 } else if (!PageAnon(page)) {
2477                         __xa_store(&head->mapping->i_pages, head[i].index,
2478                                         head + i, 0);
2479                 } else if (swap_cache) {
2480                         __xa_store(&swap_cache->i_pages, offset + i,
2481                                         head + i, 0);
2482                 }
2483         }
2484
2485         ClearPageCompound(head);
2486
2487         split_page_owner(head, nr);
2488
2489         /* See comment in __split_huge_page_tail() */
2490         if (PageAnon(head)) {
2491                 /* Additional pin to swap cache */
2492                 if (PageSwapCache(head)) {
2493                         page_ref_add(head, 2);
2494                         xa_unlock(&swap_cache->i_pages);
2495                 } else {
2496                         page_ref_inc(head);
2497                 }
2498         } else {
2499                 /* Additional pin to page cache */
2500                 page_ref_add(head, 2);
2501                 xa_unlock(&head->mapping->i_pages);
2502         }
2503
2504         spin_unlock_irqrestore(&pgdat->lru_lock, flags);
2505
2506         remap_page(head, nr);
2507
2508         if (PageSwapCache(head)) {
2509                 swp_entry_t entry = { .val = page_private(head) };
2510
2511                 split_swap_cluster(entry);
2512         }
2513
2514         for (i = 0; i < nr; i++) {
2515                 struct page *subpage = head + i;
2516                 if (subpage == page)
2517                         continue;
2518                 unlock_page(subpage);
2519
2520                 /*
2521                  * Subpages may be freed if there wasn't any mapping
2522                  * like if add_to_swap() is running on a lru page that
2523                  * had its mapping zapped. And freeing these pages
2524                  * requires taking the lru_lock so we do the put_page
2525                  * of the tail pages after the split is complete.
2526                  */
2527                 put_page(subpage);
2528         }
2529 }
2530
2531 int total_mapcount(struct page *page)
2532 {
2533         int i, compound, nr, ret;
2534
2535         VM_BUG_ON_PAGE(PageTail(page), page);
2536
2537         if (likely(!PageCompound(page)))
2538                 return atomic_read(&page->_mapcount) + 1;
2539
2540         compound = compound_mapcount(page);
2541         nr = compound_nr(page);
2542         if (PageHuge(page))
2543                 return compound;
2544         ret = compound;
2545         for (i = 0; i < nr; i++)
2546                 ret += atomic_read(&page[i]._mapcount) + 1;
2547         /* File pages has compound_mapcount included in _mapcount */
2548         if (!PageAnon(page))
2549                 return ret - compound * nr;
2550         if (PageDoubleMap(page))
2551                 ret -= nr;
2552         return ret;
2553 }
2554
2555 /*
2556  * This calculates accurately how many mappings a transparent hugepage
2557  * has (unlike page_mapcount() which isn't fully accurate). This full
2558  * accuracy is primarily needed to know if copy-on-write faults can
2559  * reuse the page and change the mapping to read-write instead of
2560  * copying them. At the same time this returns the total_mapcount too.
2561  *
2562  * The function returns the highest mapcount any one of the subpages
2563  * has. If the return value is one, even if different processes are
2564  * mapping different subpages of the transparent hugepage, they can
2565  * all reuse it, because each process is reusing a different subpage.
2566  *
2567  * The total_mapcount is instead counting all virtual mappings of the
2568  * subpages. If the total_mapcount is equal to "one", it tells the
2569  * caller all mappings belong to the same "mm" and in turn the
2570  * anon_vma of the transparent hugepage can become the vma->anon_vma
2571  * local one as no other process may be mapping any of the subpages.
2572  *
2573  * It would be more accurate to replace page_mapcount() with
2574  * page_trans_huge_mapcount(), however we only use
2575  * page_trans_huge_mapcount() in the copy-on-write faults where we
2576  * need full accuracy to avoid breaking page pinning, because
2577  * page_trans_huge_mapcount() is slower than page_mapcount().
2578  */
2579 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2580 {
2581         int i, ret, _total_mapcount, mapcount;
2582
2583         /* hugetlbfs shouldn't call it */
2584         VM_BUG_ON_PAGE(PageHuge(page), page);
2585
2586         if (likely(!PageTransCompound(page))) {
2587                 mapcount = atomic_read(&page->_mapcount) + 1;
2588                 if (total_mapcount)
2589                         *total_mapcount = mapcount;
2590                 return mapcount;
2591         }
2592
2593         page = compound_head(page);
2594
2595         _total_mapcount = ret = 0;
2596         for (i = 0; i < thp_nr_pages(page); i++) {
2597                 mapcount = atomic_read(&page[i]._mapcount) + 1;
2598                 ret = max(ret, mapcount);
2599                 _total_mapcount += mapcount;
2600         }
2601         if (PageDoubleMap(page)) {
2602                 ret -= 1;
2603                 _total_mapcount -= thp_nr_pages(page);
2604         }
2605         mapcount = compound_mapcount(page);
2606         ret += mapcount;
2607         _total_mapcount += mapcount;
2608         if (total_mapcount)
2609                 *total_mapcount = _total_mapcount;
2610         return ret;
2611 }
2612
2613 /* Racy check whether the huge page can be split */
2614 bool can_split_huge_page(struct page *page, int *pextra_pins)
2615 {
2616         int extra_pins;
2617
2618         /* Additional pins from page cache */
2619         if (PageAnon(page))
2620                 extra_pins = PageSwapCache(page) ? thp_nr_pages(page) : 0;
2621         else
2622                 extra_pins = thp_nr_pages(page);
2623         if (pextra_pins)
2624                 *pextra_pins = extra_pins;
2625         return total_mapcount(page) == page_count(page) - extra_pins - 1;
2626 }
2627
2628 /*
2629  * This function splits huge page into normal pages. @page can point to any
2630  * subpage of huge page to split. Split doesn't change the position of @page.
2631  *
2632  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2633  * The huge page must be locked.
2634  *
2635  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2636  *
2637  * Both head page and tail pages will inherit mapping, flags, and so on from
2638  * the hugepage.
2639  *
2640  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2641  * they are not mapped.
2642  *
2643  * Returns 0 if the hugepage is split successfully.
2644  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2645  * us.
2646  */
2647 int split_huge_page_to_list(struct page *page, struct list_head *list)
2648 {
2649         struct page *head = compound_head(page);
2650         struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2651         struct deferred_split *ds_queue = get_deferred_split_queue(head);
2652         struct anon_vma *anon_vma = NULL;
2653         struct address_space *mapping = NULL;
2654         int extra_pins, ret;
2655         unsigned long flags;
2656         pgoff_t end;
2657
2658         VM_BUG_ON_PAGE(is_huge_zero_page(head), head);
2659         VM_BUG_ON_PAGE(!PageLocked(head), head);
2660         VM_BUG_ON_PAGE(!PageCompound(head), head);
2661
2662         if (PageWriteback(head))
2663                 return -EBUSY;
2664
2665         if (PageAnon(head)) {
2666                 /*
2667                  * The caller does not necessarily hold an mmap_lock that would
2668                  * prevent the anon_vma disappearing so we first we take a
2669                  * reference to it and then lock the anon_vma for write. This
2670                  * is similar to page_lock_anon_vma_read except the write lock
2671                  * is taken to serialise against parallel split or collapse
2672                  * operations.
2673                  */
2674                 anon_vma = page_get_anon_vma(head);
2675                 if (!anon_vma) {
2676                         ret = -EBUSY;
2677                         goto out;
2678                 }
2679                 end = -1;
2680                 mapping = NULL;
2681                 anon_vma_lock_write(anon_vma);
2682         } else {
2683                 mapping = head->mapping;
2684
2685                 /* Truncated ? */
2686                 if (!mapping) {
2687                         ret = -EBUSY;
2688                         goto out;
2689                 }
2690
2691                 anon_vma = NULL;
2692                 i_mmap_lock_read(mapping);
2693
2694                 /*
2695                  *__split_huge_page() may need to trim off pages beyond EOF:
2696                  * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2697                  * which cannot be nested inside the page tree lock. So note
2698                  * end now: i_size itself may be changed at any moment, but
2699                  * head page lock is good enough to serialize the trimming.
2700                  */
2701                 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2702         }
2703
2704         /*
2705          * Racy check if we can split the page, before unmap_page() will
2706          * split PMDs
2707          */
2708         if (!can_split_huge_page(head, &extra_pins)) {
2709                 ret = -EBUSY;
2710                 goto out_unlock;
2711         }
2712
2713         unmap_page(head);
2714
2715         /* prevent PageLRU to go away from under us, and freeze lru stats */
2716         spin_lock_irqsave(&pgdata->lru_lock, flags);
2717
2718         if (mapping) {
2719                 XA_STATE(xas, &mapping->i_pages, page_index(head));
2720
2721                 /*
2722                  * Check if the head page is present in page cache.
2723                  * We assume all tail are present too, if head is there.
2724                  */
2725                 xa_lock(&mapping->i_pages);
2726                 if (xas_load(&xas) != head)
2727                         goto fail;
2728         }
2729
2730         /* Prevent deferred_split_scan() touching ->_refcount */
2731         spin_lock(&ds_queue->split_queue_lock);
2732         if (page_ref_freeze(head, 1 + extra_pins)) {
2733                 if (!list_empty(page_deferred_list(head))) {
2734                         ds_queue->split_queue_len--;
2735                         list_del(page_deferred_list(head));
2736                 }
2737                 spin_unlock(&ds_queue->split_queue_lock);
2738                 if (mapping) {
2739                         if (PageSwapBacked(head))
2740                                 __dec_node_page_state(head, NR_SHMEM_THPS);
2741                         else
2742                                 __dec_node_page_state(head, NR_FILE_THPS);
2743                 }
2744
2745                 __split_huge_page(page, list, end, flags);
2746                 ret = 0;
2747         } else {
2748                 spin_unlock(&ds_queue->split_queue_lock);
2749 fail:
2750                 if (mapping)
2751                         xa_unlock(&mapping->i_pages);
2752                 spin_unlock_irqrestore(&pgdata->lru_lock, flags);
2753                 remap_page(head, thp_nr_pages(head));
2754                 ret = -EBUSY;
2755         }
2756
2757 out_unlock:
2758         if (anon_vma) {
2759                 anon_vma_unlock_write(anon_vma);
2760                 put_anon_vma(anon_vma);
2761         }
2762         if (mapping)
2763                 i_mmap_unlock_read(mapping);
2764 out:
2765         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2766         return ret;
2767 }
2768
2769 void free_transhuge_page(struct page *page)
2770 {
2771         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2772         unsigned long flags;
2773
2774         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2775         if (!list_empty(page_deferred_list(page))) {
2776                 ds_queue->split_queue_len--;
2777                 list_del(page_deferred_list(page));
2778         }
2779         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2780         free_compound_page(page);
2781 }
2782
2783 void deferred_split_huge_page(struct page *page)
2784 {
2785         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2786 #ifdef CONFIG_MEMCG
2787         struct mem_cgroup *memcg = compound_head(page)->mem_cgroup;
2788 #endif
2789         unsigned long flags;
2790
2791         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2792
2793         /*
2794          * The try_to_unmap() in page reclaim path might reach here too,
2795          * this may cause a race condition to corrupt deferred split queue.
2796          * And, if page reclaim is already handling the same page, it is
2797          * unnecessary to handle it again in shrinker.
2798          *
2799          * Check PageSwapCache to determine if the page is being
2800          * handled by page reclaim since THP swap would add the page into
2801          * swap cache before calling try_to_unmap().
2802          */
2803         if (PageSwapCache(page))
2804                 return;
2805
2806         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2807         if (list_empty(page_deferred_list(page))) {
2808                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2809                 list_add_tail(page_deferred_list(page), &ds_queue->split_queue);
2810                 ds_queue->split_queue_len++;
2811 #ifdef CONFIG_MEMCG
2812                 if (memcg)
2813                         memcg_set_shrinker_bit(memcg, page_to_nid(page),
2814                                                deferred_split_shrinker.id);
2815 #endif
2816         }
2817         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2818 }
2819
2820 static unsigned long deferred_split_count(struct shrinker *shrink,
2821                 struct shrink_control *sc)
2822 {
2823         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2824         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2825
2826 #ifdef CONFIG_MEMCG
2827         if (sc->memcg)
2828                 ds_queue = &sc->memcg->deferred_split_queue;
2829 #endif
2830         return READ_ONCE(ds_queue->split_queue_len);
2831 }
2832
2833 static unsigned long deferred_split_scan(struct shrinker *shrink,
2834                 struct shrink_control *sc)
2835 {
2836         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2837         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2838         unsigned long flags;
2839         LIST_HEAD(list), *pos, *next;
2840         struct page *page;
2841         int split = 0;
2842
2843 #ifdef CONFIG_MEMCG
2844         if (sc->memcg)
2845                 ds_queue = &sc->memcg->deferred_split_queue;
2846 #endif
2847
2848         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2849         /* Take pin on all head pages to avoid freeing them under us */
2850         list_for_each_safe(pos, next, &ds_queue->split_queue) {
2851                 page = list_entry((void *)pos, struct page, mapping);
2852                 page = compound_head(page);
2853                 if (get_page_unless_zero(page)) {
2854                         list_move(page_deferred_list(page), &list);
2855                 } else {
2856                         /* We lost race with put_compound_page() */
2857                         list_del_init(page_deferred_list(page));
2858                         ds_queue->split_queue_len--;
2859                 }
2860                 if (!--sc->nr_to_scan)
2861                         break;
2862         }
2863         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2864
2865         list_for_each_safe(pos, next, &list) {
2866                 page = list_entry((void *)pos, struct page, mapping);
2867                 if (!trylock_page(page))
2868                         goto next;
2869                 /* split_huge_page() removes page from list on success */
2870                 if (!split_huge_page(page))
2871                         split++;
2872                 unlock_page(page);
2873 next:
2874                 put_page(page);
2875         }
2876
2877         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2878         list_splice_tail(&list, &ds_queue->split_queue);
2879         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2880
2881         /*
2882          * Stop shrinker if we didn't split any page, but the queue is empty.
2883          * This can happen if pages were freed under us.
2884          */
2885         if (!split && list_empty(&ds_queue->split_queue))
2886                 return SHRINK_STOP;
2887         return split;
2888 }
2889
2890 static struct shrinker deferred_split_shrinker = {
2891         .count_objects = deferred_split_count,
2892         .scan_objects = deferred_split_scan,
2893         .seeks = DEFAULT_SEEKS,
2894         .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
2895                  SHRINKER_NONSLAB,
2896 };
2897
2898 #ifdef CONFIG_DEBUG_FS
2899 static int split_huge_pages_set(void *data, u64 val)
2900 {
2901         struct zone *zone;
2902         struct page *page;
2903         unsigned long pfn, max_zone_pfn;
2904         unsigned long total = 0, split = 0;
2905
2906         if (val != 1)
2907                 return -EINVAL;
2908
2909         for_each_populated_zone(zone) {
2910                 max_zone_pfn = zone_end_pfn(zone);
2911                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2912                         if (!pfn_valid(pfn))
2913                                 continue;
2914
2915                         page = pfn_to_page(pfn);
2916                         if (!get_page_unless_zero(page))
2917                                 continue;
2918
2919                         if (zone != page_zone(page))
2920                                 goto next;
2921
2922                         if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2923                                 goto next;
2924
2925                         total++;
2926                         lock_page(page);
2927                         if (!split_huge_page(page))
2928                                 split++;
2929                         unlock_page(page);
2930 next:
2931                         put_page(page);
2932                 }
2933         }
2934
2935         pr_info("%lu of %lu THP split\n", split, total);
2936
2937         return 0;
2938 }
2939 DEFINE_DEBUGFS_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2940                 "%llu\n");
2941
2942 static int __init split_huge_pages_debugfs(void)
2943 {
2944         debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2945                             &split_huge_pages_fops);
2946         return 0;
2947 }
2948 late_initcall(split_huge_pages_debugfs);
2949 #endif
2950
2951 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2952 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
2953                 struct page *page)
2954 {
2955         struct vm_area_struct *vma = pvmw->vma;
2956         struct mm_struct *mm = vma->vm_mm;
2957         unsigned long address = pvmw->address;
2958         pmd_t pmdval;
2959         swp_entry_t entry;
2960         pmd_t pmdswp;
2961
2962         if (!(pvmw->pmd && !pvmw->pte))
2963                 return;
2964
2965         flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
2966         pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
2967         if (pmd_dirty(pmdval))
2968                 set_page_dirty(page);
2969         entry = make_migration_entry(page, pmd_write(pmdval));
2970         pmdswp = swp_entry_to_pmd(entry);
2971         if (pmd_soft_dirty(pmdval))
2972                 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
2973         set_pmd_at(mm, address, pvmw->pmd, pmdswp);
2974         page_remove_rmap(page, true);
2975         put_page(page);
2976 }
2977
2978 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
2979 {
2980         struct vm_area_struct *vma = pvmw->vma;
2981         struct mm_struct *mm = vma->vm_mm;
2982         unsigned long address = pvmw->address;
2983         unsigned long mmun_start = address & HPAGE_PMD_MASK;
2984         pmd_t pmde;
2985         swp_entry_t entry;
2986
2987         if (!(pvmw->pmd && !pvmw->pte))
2988                 return;
2989
2990         entry = pmd_to_swp_entry(*pvmw->pmd);
2991         get_page(new);
2992         pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
2993         if (pmd_swp_soft_dirty(*pvmw->pmd))
2994                 pmde = pmd_mksoft_dirty(pmde);
2995         if (is_write_migration_entry(entry))
2996                 pmde = maybe_pmd_mkwrite(pmde, vma);
2997         if (pmd_swp_uffd_wp(*pvmw->pmd))
2998                 pmde = pmd_wrprotect(pmd_mkuffd_wp(pmde));
2999
3000         flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
3001         if (PageAnon(new))
3002                 page_add_anon_rmap(new, vma, mmun_start, true);
3003         else
3004                 page_add_file_rmap(new, true);
3005         set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
3006         if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new))
3007                 mlock_vma_page(new);
3008         update_mmu_cache_pmd(vma, address, pvmw->pmd);
3009 }
3010 #endif