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 bool transparent_hugepage_enabled(struct vm_area_struct *vma)
  67 {
  68         /* The addr is used to check if the vma size fits */
  69         unsigned long addr = (vma->vm_end & HPAGE_PMD_MASK) - HPAGE_PMD_SIZE;
  70
  71         if (!transhuge_vma_suitable(vma, addr))
  72                 return false;
  73         if (vma_is_anonymous(vma))
  74                 return __transparent_hugepage_enabled(vma);
  75         if (vma_is_shmem(vma))
  76                 return shmem_huge_enabled(vma);
  77
  78         return false;
  79 }
  80
  81 static struct page *get_huge_zero_page(void)
  82 {
  83         struct page *zero_page;
  84 retry:
  85         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
  86                 return READ_ONCE(huge_zero_page);
  87
  88         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
  89                         HPAGE_PMD_ORDER);
  90         if (!zero_page) {
  91                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
  92                 return NULL;
  93         }
  94         count_vm_event(THP_ZERO_PAGE_ALLOC);
  95         preempt_disable();
  96         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
  97                 preempt_enable();
  98                 __free_pages(zero_page, compound_order(zero_page));
  99                 goto retry;
 100         }
 101         WRITE_ONCE(huge_zero_pfn, page_to_pfn(zero_page));
 102
 103         /* We take additional reference here. It will be put back by shrinker */
 104         atomic_set(&huge_zero_refcount, 2);
 105         preempt_enable();
 106         return READ_ONCE(huge_zero_page);
 107 }
 108
 109 static void put_huge_zero_page(void)
 110 {
 111         /*
 112          * Counter should never go to zero here. Only shrinker can put
 113          * last reference.
 114          */
 115         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 116 }
 117
 118 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
 119 {
 120         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 121                 return READ_ONCE(huge_zero_page);
 122
 123         if (!get_huge_zero_page())
 124                 return NULL;
 125
 126         if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 127                 put_huge_zero_page();
 128
 129         return READ_ONCE(huge_zero_page);
 130 }
 131
 132 void mm_put_huge_zero_page(struct mm_struct *mm)
 133 {
 134         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 135                 put_huge_zero_page();
 136 }
 137
 138 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 139                                         struct shrink_control *sc)
 140 {
 141         /* we can free zero page only if last reference remains */
 142         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 143 }
 144
 145 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 146                                        struct shrink_control *sc)
 147 {
 148         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 149                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 150                 BUG_ON(zero_page == NULL);
 151                 WRITE_ONCE(huge_zero_pfn, ~0UL);
 152                 __free_pages(zero_page, compound_order(zero_page));
 153                 return HPAGE_PMD_NR;
 154         }
 155
 156         return 0;
 157 }
 158
 159 static struct shrinker huge_zero_page_shrinker = {
 160         .count_objects = shrink_huge_zero_page_count,
 161         .scan_objects = shrink_huge_zero_page_scan,
 162         .seeks = DEFAULT_SEEKS,
 163 };
 164
 165 #ifdef CONFIG_SYSFS
 166 static ssize_t enabled_show(struct kobject *kobj,
 167                             struct kobj_attribute *attr, char *buf)
 168 {
 169         if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
 170                 return sprintf(buf, "[always] madvise never\n");
 171         else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
 172                 return sprintf(buf, "always [madvise] never\n");
 173         else
 174                 return sprintf(buf, "always madvise [never]\n");
 175 }
 176
 177 static ssize_t enabled_store(struct kobject *kobj,
 178                              struct kobj_attribute *attr,
 179                              const char *buf, size_t count)
 180 {
 181         ssize_t ret = count;
 182
 183         if (sysfs_streq(buf, "always")) {
 184                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 185                 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 186         } else if (sysfs_streq(buf, "madvise")) {
 187                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 188                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 189         } else if (sysfs_streq(buf, "never")) {
 190                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 191                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 192         } else
 193                 ret = -EINVAL;
 194
 195         if (ret > 0) {
 196                 int err = start_stop_khugepaged();
 197                 if (err)
 198                         ret = err;
 199         }
 200         return ret;
 201 }
 202 static struct kobj_attribute enabled_attr =
 203         __ATTR(enabled, 0644, enabled_show, enabled_store);
 204
 205 ssize_t single_hugepage_flag_show(struct kobject *kobj,
 206                                 struct kobj_attribute *attr, char *buf,
 207                                 enum transparent_hugepage_flag flag)
 208 {
 209         return sprintf(buf, "%d\n",
 210                        !!test_bit(flag, &transparent_hugepage_flags));
 211 }
 212
 213 ssize_t single_hugepage_flag_store(struct kobject *kobj,
 214                                  struct kobj_attribute *attr,
 215                                  const char *buf, size_t count,
 216                                  enum transparent_hugepage_flag flag)
 217 {
 218         unsigned long value;
 219         int ret;
 220
 221         ret = kstrtoul(buf, 10, &value);
 222         if (ret < 0)
 223                 return ret;
 224         if (value > 1)
 225                 return -EINVAL;
 226
 227         if (value)
 228                 set_bit(flag, &transparent_hugepage_flags);
 229         else
 230                 clear_bit(flag, &transparent_hugepage_flags);
 231
 232         return count;
 233 }
 234
 235 static ssize_t defrag_show(struct kobject *kobj,
 236                            struct kobj_attribute *attr, char *buf)
 237 {
 238         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 239                 return sprintf(buf, "[always] defer defer+madvise madvise never\n");
 240         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 241                 return sprintf(buf, "always [defer] defer+madvise madvise never\n");
 242         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 243                 return sprintf(buf, "always defer [defer+madvise] madvise never\n");
 244         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 245                 return sprintf(buf, "always defer defer+madvise [madvise] never\n");
 246         return sprintf(buf, "always defer defer+madvise madvise [never]\n");
 247 }
 248
 249 static ssize_t defrag_store(struct kobject *kobj,
 250                             struct kobj_attribute *attr,
 251                             const char *buf, size_t count)
 252 {
 253         if (sysfs_streq(buf, "always")) {
 254                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 255                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 256                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 257                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 258         } else if (sysfs_streq(buf, "defer+madvise")) {
 259                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 260                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 261                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 262                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 263         } else if (sysfs_streq(buf, "defer")) {
 264                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 265                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 266                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 267                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 268         } else if (sysfs_streq(buf, "madvise")) {
 269                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 270                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 271                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 272                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 273         } else if (sysfs_streq(buf, "never")) {
 274                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 275                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 276                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 277                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 278         } else
 279                 return -EINVAL;
 280
 281         return count;
 282 }
 283 static struct kobj_attribute defrag_attr =
 284         __ATTR(defrag, 0644, defrag_show, defrag_store);
 285
 286 static ssize_t use_zero_page_show(struct kobject *kobj,
 287                 struct kobj_attribute *attr, char *buf)
 288 {
 289         return single_hugepage_flag_show(kobj, attr, buf,
 290                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 291 }
 292 static ssize_t use_zero_page_store(struct kobject *kobj,
 293                 struct kobj_attribute *attr, const char *buf, size_t count)
 294 {
 295         return single_hugepage_flag_store(kobj, attr, buf, count,
 296                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 297 }
 298 static struct kobj_attribute use_zero_page_attr =
 299         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 300
 301 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
 302                 struct kobj_attribute *attr, char *buf)
 303 {
 304         return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE);
 305 }
 306 static struct kobj_attribute hpage_pmd_size_attr =
 307         __ATTR_RO(hpage_pmd_size);
 308
 309 #ifdef CONFIG_DEBUG_VM
 310 static ssize_t debug_cow_show(struct kobject *kobj,
 311                                 struct kobj_attribute *attr, char *buf)
 312 {
 313         return single_hugepage_flag_show(kobj, attr, buf,
 314                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 315 }
 316 static ssize_t debug_cow_store(struct kobject *kobj,
 317                                struct kobj_attribute *attr,
 318                                const char *buf, size_t count)
 319 {
 320         return single_hugepage_flag_store(kobj, attr, buf, count,
 321                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 322 }
 323 static struct kobj_attribute debug_cow_attr =
 324         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 325 #endif /* CONFIG_DEBUG_VM */
 326
 327 static struct attribute *hugepage_attr[] = {
 328         &enabled_attr.attr,
 329         &defrag_attr.attr,
 330         &use_zero_page_attr.attr,
 331         &hpage_pmd_size_attr.attr,
 332 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
 333         &shmem_enabled_attr.attr,
 334 #endif
 335 #ifdef CONFIG_DEBUG_VM
 336         &debug_cow_attr.attr,
 337 #endif
 338         NULL,
 339 };
 340
 341 static const struct attribute_group hugepage_attr_group = {
 342         .attrs = hugepage_attr,
 343 };
 344
 345 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 346 {
 347         int err;
 348
 349         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 350         if (unlikely(!*hugepage_kobj)) {
 351                 pr_err("failed to create transparent hugepage kobject\n");
 352                 return -ENOMEM;
 353         }
 354
 355         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 356         if (err) {
 357                 pr_err("failed to register transparent hugepage group\n");
 358                 goto delete_obj;
 359         }
 360
 361         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 362         if (err) {
 363                 pr_err("failed to register transparent hugepage group\n");
 364                 goto remove_hp_group;
 365         }
 366
 367         return 0;
 368
 369 remove_hp_group:
 370         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 371 delete_obj:
 372         kobject_put(*hugepage_kobj);
 373         return err;
 374 }
 375
 376 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 377 {
 378         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 379         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 380         kobject_put(hugepage_kobj);
 381 }
 382 #else
 383 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 384 {
 385         return 0;
 386 }
 387
 388 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 389 {
 390 }
 391 #endif /* CONFIG_SYSFS */
 392
 393 static int __init hugepage_init(void)
 394 {
 395         int err;
 396         struct kobject *hugepage_kobj;
 397
 398         if (!has_transparent_hugepage()) {
 399                 transparent_hugepage_flags = 0;
 400                 return -EINVAL;
 401         }
 402
 403         /*
 404          * hugepages can't be allocated by the buddy allocator
 405          */
 406         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
 407         /*
 408          * we use page->mapping and page->index in second tail page
 409          * as list_head: assuming THP order >= 2
 410          */
 411         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 412
 413         err = hugepage_init_sysfs(&hugepage_kobj);
 414         if (err)
 415                 goto err_sysfs;
 416
 417         err = khugepaged_init();
 418         if (err)
 419                 goto err_slab;
 420
 421         err = register_shrinker(&huge_zero_page_shrinker);
 422         if (err)
 423                 goto err_hzp_shrinker;
 424         err = register_shrinker(&deferred_split_shrinker);
 425         if (err)
 426                 goto err_split_shrinker;
 427
 428         /*
 429          * By default disable transparent hugepages on smaller systems,
 430          * where the extra memory used could hurt more than TLB overhead
 431          * is likely to save.  The admin can still enable it through /sys.
 432          */
 433         if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
 434                 transparent_hugepage_flags = 0;
 435                 return 0;
 436         }
 437
 438         err = start_stop_khugepaged();
 439         if (err)
 440                 goto err_khugepaged;
 441
 442         return 0;
 443 err_khugepaged:
 444         unregister_shrinker(&deferred_split_shrinker);
 445 err_split_shrinker:
 446         unregister_shrinker(&huge_zero_page_shrinker);
 447 err_hzp_shrinker:
 448         khugepaged_destroy();
 449 err_slab:
 450         hugepage_exit_sysfs(hugepage_kobj);
 451 err_sysfs:
 452         return err;
 453 }
 454 subsys_initcall(hugepage_init);
 455
 456 static int __init setup_transparent_hugepage(char *str)
 457 {
 458         int ret = 0;
 459         if (!str)
 460                 goto out;
 461         if (!strcmp(str, "always")) {
 462                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 463                         &transparent_hugepage_flags);
 464                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 465                           &transparent_hugepage_flags);
 466                 ret = 1;
 467         } else if (!strcmp(str, "madvise")) {
 468                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 469                           &transparent_hugepage_flags);
 470                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 471                         &transparent_hugepage_flags);
 472                 ret = 1;
 473         } else if (!strcmp(str, "never")) {
 474                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 475                           &transparent_hugepage_flags);
 476                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 477                           &transparent_hugepage_flags);
 478                 ret = 1;
 479         }
 480 out:
 481         if (!ret)
 482                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 483         return ret;
 484 }
 485 __setup("transparent_hugepage=", setup_transparent_hugepage);
 486
 487 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 488 {
 489         if (likely(vma->vm_flags & VM_WRITE))
 490                 pmd = pmd_mkwrite(pmd);
 491         return pmd;
 492 }
 493
 494 #ifdef CONFIG_MEMCG
 495 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
 496 {
 497         struct mem_cgroup *memcg = compound_head(page)->mem_cgroup;
 498         struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
 499
 500         if (memcg)
 501                 return &memcg->deferred_split_queue;
 502         else
 503                 return &pgdat->deferred_split_queue;
 504 }
 505 #else
 506 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
 507 {
 508         struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
 509
 510         return &pgdat->deferred_split_queue;
 511 }
 512 #endif
 513
 514 void prep_transhuge_page(struct page *page)
 515 {
 516         /*
 517          * we use page->mapping and page->indexlru in second tail page
 518          * as list_head: assuming THP order >= 2
 519          */
 520
 521         INIT_LIST_HEAD(page_deferred_list(page));
 522         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 523 }
 524
 525 static unsigned long __thp_get_unmapped_area(struct file *filp,
 526                 unsigned long addr, unsigned long len,
 527                 loff_t off, unsigned long flags, unsigned long size)
 528 {
 529         loff_t off_end = off + len;
 530         loff_t off_align = round_up(off, size);
 531         unsigned long len_pad, ret;
 532
 533         if (off_end <= off_align || (off_end - off_align) < size)
 534                 return 0;
 535
 536         len_pad = len + size;
 537         if (len_pad < len || (off + len_pad) < off)
 538                 return 0;
 539
 540         ret = current->mm->get_unmapped_area(filp, addr, len_pad,
 541                                               off >> PAGE_SHIFT, flags);
 542
 543         /*
 544          * The failure might be due to length padding. The caller will retry
 545          * without the padding.
 546          */
 547         if (IS_ERR_VALUE(ret))
 548                 return 0;
 549
 550         /*
 551          * Do not try to align to THP boundary if allocation at the address
 552          * hint succeeds.
 553          */
 554         if (ret == addr)
 555                 return addr;
 556
 557         ret += (off - ret) & (size - 1);
 558         return ret;
 559 }
 560
 561 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 562                 unsigned long len, unsigned long pgoff, unsigned long flags)
 563 {
 564         unsigned long ret;
 565         loff_t off = (loff_t)pgoff << PAGE_SHIFT;
 566
 567         if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
 568                 goto out;
 569
 570         ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE);
 571         if (ret)
 572                 return ret;
 573 out:
 574         return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
 575 }
 576 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
 577
 578 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
 579                         struct page *page, gfp_t gfp)
 580 {
 581         struct vm_area_struct *vma = vmf->vma;
 582         struct mem_cgroup *memcg;
 583         pgtable_t pgtable;
 584         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 585         vm_fault_t ret = 0;
 586
 587         VM_BUG_ON_PAGE(!PageCompound(page), page);
 588
 589         if (mem_cgroup_try_charge_delay(page, vma->vm_mm, gfp, &memcg, true)) {
 590                 put_page(page);
 591                 count_vm_event(THP_FAULT_FALLBACK);
 592                 return VM_FAULT_FALLBACK;
 593         }
 594
 595         pgtable = pte_alloc_one(vma->vm_mm);
 596         if (unlikely(!pgtable)) {
 597                 ret = VM_FAULT_OOM;
 598                 goto release;
 599         }
 600
 601         clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
 602         /*
 603          * The memory barrier inside __SetPageUptodate makes sure that
 604          * clear_huge_page writes become visible before the set_pmd_at()
 605          * write.
 606          */
 607         __SetPageUptodate(page);
 608
 609         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 610         if (unlikely(!pmd_none(*vmf->pmd))) {
 611                 goto unlock_release;
 612         } else {
 613                 pmd_t entry;
 614
 615                 ret = check_stable_address_space(vma->vm_mm);
 616                 if (ret)
 617                         goto unlock_release;
 618
 619                 /* Deliver the page fault to userland */
 620                 if (userfaultfd_missing(vma)) {
 621                         vm_fault_t ret2;
 622
 623                         spin_unlock(vmf->ptl);
 624                         mem_cgroup_cancel_charge(page, memcg, true);
 625                         put_page(page);
 626                         pte_free(vma->vm_mm, pgtable);
 627                         ret2 = handle_userfault(vmf, VM_UFFD_MISSING);
 628                         VM_BUG_ON(ret2 & VM_FAULT_FALLBACK);
 629                         return ret2;
 630                 }
 631
 632                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 633                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 634                 page_add_new_anon_rmap(page, vma, haddr, true);
 635                 mem_cgroup_commit_charge(page, memcg, false, true);
 636                 lru_cache_add_active_or_unevictable(page, vma);
 637                 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
 638                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
 639                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 640                 mm_inc_nr_ptes(vma->vm_mm);
 641                 spin_unlock(vmf->ptl);
 642                 count_vm_event(THP_FAULT_ALLOC);
 643                 count_memcg_events(memcg, THP_FAULT_ALLOC, 1);
 644         }
 645
 646         return 0;
 647 unlock_release:
 648         spin_unlock(vmf->ptl);
 649 release:
 650         if (pgtable)
 651                 pte_free(vma->vm_mm, pgtable);
 652         mem_cgroup_cancel_charge(page, memcg, true);
 653         put_page(page);
 654         return ret;
 655
 656 }
 657
 658 /*
 659  * always: directly stall for all thp allocations
 660  * defer: wake kswapd and fail if not immediately available
 661  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
 662  *                fail if not immediately available
 663  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
 664  *          available
 665  * never: never stall for any thp allocation
 666  */
 667 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
 668 {
 669         const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
 670
 671         /* Always do synchronous compaction */
 672         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 673                 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
 674
 675         /* Kick kcompactd and fail quickly */
 676         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 677                 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
 678
 679         /* Synchronous compaction if madvised, otherwise kick kcompactd */
 680         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 681                 return GFP_TRANSHUGE_LIGHT |
 682                         (vma_madvised ? __GFP_DIRECT_RECLAIM :
 683                                         __GFP_KSWAPD_RECLAIM);
 684
 685         /* Only do synchronous compaction if madvised */
 686         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 687                 return GFP_TRANSHUGE_LIGHT |
 688                        (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
 689
 690         return GFP_TRANSHUGE_LIGHT;
 691 }
 692
 693 /* Caller must hold page table lock. */
 694 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 695                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 696                 struct page *zero_page)
 697 {
 698         pmd_t entry;
 699         if (!pmd_none(*pmd))
 700                 return false;
 701         entry = mk_pmd(zero_page, vma->vm_page_prot);
 702         entry = pmd_mkhuge(entry);
 703         if (pgtable)
 704                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 705         set_pmd_at(mm, haddr, pmd, entry);
 706         mm_inc_nr_ptes(mm);
 707         return true;
 708 }
 709
 710 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
 711 {
 712         struct vm_area_struct *vma = vmf->vma;
 713         gfp_t gfp;
 714         struct page *page;
 715         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 716
 717         if (!transhuge_vma_suitable(vma, haddr))
 718                 return VM_FAULT_FALLBACK;
 719         if (unlikely(anon_vma_prepare(vma)))
 720                 return VM_FAULT_OOM;
 721         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 722                 return VM_FAULT_OOM;
 723         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
 724                         !mm_forbids_zeropage(vma->vm_mm) &&
 725                         transparent_hugepage_use_zero_page()) {
 726                 pgtable_t pgtable;
 727                 struct page *zero_page;
 728                 vm_fault_t ret;
 729                 pgtable = pte_alloc_one(vma->vm_mm);
 730                 if (unlikely(!pgtable))
 731                         return VM_FAULT_OOM;
 732                 zero_page = mm_get_huge_zero_page(vma->vm_mm);
 733                 if (unlikely(!zero_page)) {
 734                         pte_free(vma->vm_mm, pgtable);
 735                         count_vm_event(THP_FAULT_FALLBACK);
 736                         return VM_FAULT_FALLBACK;
 737                 }
 738                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 739                 ret = 0;
 740                 if (pmd_none(*vmf->pmd)) {
 741                         ret = check_stable_address_space(vma->vm_mm);
 742                         if (ret) {
 743                                 spin_unlock(vmf->ptl);
 744                                 pte_free(vma->vm_mm, pgtable);
 745                         } else if (userfaultfd_missing(vma)) {
 746                                 spin_unlock(vmf->ptl);
 747                                 pte_free(vma->vm_mm, pgtable);
 748                                 ret = handle_userfault(vmf, VM_UFFD_MISSING);
 749                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 750                         } else {
 751                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
 752                                                    haddr, vmf->pmd, zero_page);
 753                                 spin_unlock(vmf->ptl);
 754                         }
 755                 } else {
 756                         spin_unlock(vmf->ptl);
 757                         pte_free(vma->vm_mm, pgtable);
 758                 }
 759                 return ret;
 760         }
 761         gfp = alloc_hugepage_direct_gfpmask(vma);
 762         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 763         if (unlikely(!page)) {
 764                 count_vm_event(THP_FAULT_FALLBACK);
 765                 return VM_FAULT_FALLBACK;
 766         }
 767         prep_transhuge_page(page);
 768         return __do_huge_pmd_anonymous_page(vmf, page, gfp);
 769 }
 770
 771 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 772                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
 773                 pgtable_t pgtable)
 774 {
 775         struct mm_struct *mm = vma->vm_mm;
 776         pmd_t entry;
 777         spinlock_t *ptl;
 778
 779         ptl = pmd_lock(mm, pmd);
 780         if (!pmd_none(*pmd)) {
 781                 if (write) {
 782                         if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
 783                                 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
 784                                 goto out_unlock;
 785                         }
 786                         entry = pmd_mkyoung(*pmd);
 787                         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 788                         if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
 789                                 update_mmu_cache_pmd(vma, addr, pmd);
 790                 }
 791
 792                 goto out_unlock;
 793         }
 794
 795         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 796         if (pfn_t_devmap(pfn))
 797                 entry = pmd_mkdevmap(entry);
 798         if (write) {
 799                 entry = pmd_mkyoung(pmd_mkdirty(entry));
 800                 entry = maybe_pmd_mkwrite(entry, vma);
 801         }
 802
 803         if (pgtable) {
 804                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 805                 mm_inc_nr_ptes(mm);
 806                 pgtable = NULL;
 807         }
 808
 809         set_pmd_at(mm, addr, pmd, entry);
 810         update_mmu_cache_pmd(vma, addr, pmd);
 811
 812 out_unlock:
 813         spin_unlock(ptl);
 814         if (pgtable)
 815                 pte_free(mm, pgtable);
 816 }
 817
 818 vm_fault_t vmf_insert_pfn_pmd(struct vm_fault *vmf, pfn_t pfn, bool write)
 819 {
 820         unsigned long addr = vmf->address & PMD_MASK;
 821         struct vm_area_struct *vma = vmf->vma;
 822         pgprot_t pgprot = vma->vm_page_prot;
 823         pgtable_t pgtable = NULL;
 824
 825         /*
 826          * If we had pmd_special, we could avoid all these restrictions,
 827          * but we need to be consistent with PTEs and architectures that
 828          * can't support a 'special' bit.
 829          */
 830         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 831                         !pfn_t_devmap(pfn));
 832         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 833                                                 (VM_PFNMAP|VM_MIXEDMAP));
 834         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 835
 836         if (addr < vma->vm_start || addr >= vma->vm_end)
 837                 return VM_FAULT_SIGBUS;
 838
 839         if (arch_needs_pgtable_deposit()) {
 840                 pgtable = pte_alloc_one(vma->vm_mm);
 841                 if (!pgtable)
 842                         return VM_FAULT_OOM;
 843         }
 844
 845         track_pfn_insert(vma, &pgprot, pfn);
 846
 847         insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
 848         return VM_FAULT_NOPAGE;
 849 }
 850 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
 851
 852 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 853 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
 854 {
 855         if (likely(vma->vm_flags & VM_WRITE))
 856                 pud = pud_mkwrite(pud);
 857         return pud;
 858 }
 859
 860 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 861                 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
 862 {
 863         struct mm_struct *mm = vma->vm_mm;
 864         pud_t entry;
 865         spinlock_t *ptl;
 866
 867         ptl = pud_lock(mm, pud);
 868         if (!pud_none(*pud)) {
 869                 if (write) {
 870                         if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
 871                                 WARN_ON_ONCE(!is_huge_zero_pud(*pud));
 872                                 goto out_unlock;
 873                         }
 874                         entry = pud_mkyoung(*pud);
 875                         entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
 876                         if (pudp_set_access_flags(vma, addr, pud, entry, 1))
 877                                 update_mmu_cache_pud(vma, addr, pud);
 878                 }
 879                 goto out_unlock;
 880         }
 881
 882         entry = pud_mkhuge(pfn_t_pud(pfn, prot));
 883         if (pfn_t_devmap(pfn))
 884                 entry = pud_mkdevmap(entry);
 885         if (write) {
 886                 entry = pud_mkyoung(pud_mkdirty(entry));
 887                 entry = maybe_pud_mkwrite(entry, vma);
 888         }
 889         set_pud_at(mm, addr, pud, entry);
 890         update_mmu_cache_pud(vma, addr, pud);
 891
 892 out_unlock:
 893         spin_unlock(ptl);
 894 }
 895
 896 vm_fault_t vmf_insert_pfn_pud(struct vm_fault *vmf, pfn_t pfn, bool write)
 897 {
 898         unsigned long addr = vmf->address & PUD_MASK;
 899         struct vm_area_struct *vma = vmf->vma;
 900         pgprot_t pgprot = vma->vm_page_prot;
 901
 902         /*
 903          * If we had pud_special, we could avoid all these restrictions,
 904          * but we need to be consistent with PTEs and architectures that
 905          * can't support a 'special' bit.
 906          */
 907         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 908                         !pfn_t_devmap(pfn));
 909         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 910                                                 (VM_PFNMAP|VM_MIXEDMAP));
 911         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 912
 913         if (addr < vma->vm_start || addr >= vma->vm_end)
 914                 return VM_FAULT_SIGBUS;
 915
 916         track_pfn_insert(vma, &pgprot, pfn);
 917
 918         insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
 919         return VM_FAULT_NOPAGE;
 920 }
 921 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
 922 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
 923
 924 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
 925                 pmd_t *pmd, int flags)
 926 {
 927         pmd_t _pmd;
 928
 929         _pmd = pmd_mkyoung(*pmd);
 930         if (flags & FOLL_WRITE)
 931                 _pmd = pmd_mkdirty(_pmd);
 932         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
 933                                 pmd, _pmd, flags & FOLL_WRITE))
 934                 update_mmu_cache_pmd(vma, addr, pmd);
 935 }
 936
 937 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 938                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
 939 {
 940         unsigned long pfn = pmd_pfn(*pmd);
 941         struct mm_struct *mm = vma->vm_mm;
 942         struct page *page;
 943
 944         assert_spin_locked(pmd_lockptr(mm, pmd));
 945
 946         /*
 947          * When we COW a devmap PMD entry, we split it into PTEs, so we should
 948          * not be in this function with `flags & FOLL_COW` set.
 949          */
 950         WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
 951
 952         if (flags & FOLL_WRITE && !pmd_write(*pmd))
 953                 return NULL;
 954
 955         if (pmd_present(*pmd) && pmd_devmap(*pmd))
 956                 /* pass */;
 957         else
 958                 return NULL;
 959
 960         if (flags & FOLL_TOUCH)
 961                 touch_pmd(vma, addr, pmd, flags);
 962
 963         /*
 964          * device mapped pages can only be returned if the
 965          * caller will manage the page reference count.
 966          */
 967         if (!(flags & FOLL_GET))
 968                 return ERR_PTR(-EEXIST);
 969
 970         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
 971         *pgmap = get_dev_pagemap(pfn, *pgmap);
 972         if (!*pgmap)
 973                 return ERR_PTR(-EFAULT);
 974         page = pfn_to_page(pfn);
 975         get_page(page);
 976
 977         return page;
 978 }
 979
 980 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 981                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 982                   struct vm_area_struct *vma)
 983 {
 984         spinlock_t *dst_ptl, *src_ptl;
 985         struct page *src_page;
 986         pmd_t pmd;
 987         pgtable_t pgtable = NULL;
 988         int ret = -ENOMEM;
 989
 990         /* Skip if can be re-fill on fault */
 991         if (!vma_is_anonymous(vma))
 992                 return 0;
 993
 994         pgtable = pte_alloc_one(dst_mm);
 995         if (unlikely(!pgtable))
 996                 goto out;
 997
 998         dst_ptl = pmd_lock(dst_mm, dst_pmd);
 999         src_ptl = pmd_lockptr(src_mm, src_pmd);
1000         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1001
1002         ret = -EAGAIN;
1003         pmd = *src_pmd;
1004
1005 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1006         if (unlikely(is_swap_pmd(pmd))) {
1007                 swp_entry_t entry = pmd_to_swp_entry(pmd);
1008
1009                 VM_BUG_ON(!is_pmd_migration_entry(pmd));
1010                 if (is_write_migration_entry(entry)) {
1011                         make_migration_entry_read(&entry);
1012                         pmd = swp_entry_to_pmd(entry);
1013                         if (pmd_swp_soft_dirty(*src_pmd))
1014                                 pmd = pmd_swp_mksoft_dirty(pmd);
1015                         set_pmd_at(src_mm, addr, src_pmd, pmd);
1016                 }
1017                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1018                 mm_inc_nr_ptes(dst_mm);
1019                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1020                 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1021                 ret = 0;
1022                 goto out_unlock;
1023         }
1024 #endif
1025
1026         if (unlikely(!pmd_trans_huge(pmd))) {
1027                 pte_free(dst_mm, pgtable);
1028                 goto out_unlock;
1029         }
1030         /*
1031          * When page table lock is held, the huge zero pmd should not be
1032          * under splitting since we don't split the page itself, only pmd to
1033          * a page table.
1034          */
1035         if (is_huge_zero_pmd(pmd)) {
1036                 struct page *zero_page;
1037                 /*
1038                  * get_huge_zero_page() will never allocate a new page here,
1039                  * since we already have a zero page to copy. It just takes a
1040                  * reference.
1041                  */
1042                 zero_page = mm_get_huge_zero_page(dst_mm);
1043                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1044                                 zero_page);
1045                 ret = 0;
1046                 goto out_unlock;
1047         }
1048
1049         src_page = pmd_page(pmd);
1050         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1051         get_page(src_page);
1052         page_dup_rmap(src_page, true);
1053         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1054         mm_inc_nr_ptes(dst_mm);
1055         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1056
1057         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1058         pmd = pmd_mkold(pmd_wrprotect(pmd));
1059         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1060
1061         ret = 0;
1062 out_unlock:
1063         spin_unlock(src_ptl);
1064         spin_unlock(dst_ptl);
1065 out:
1066         return ret;
1067 }
1068
1069 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1070 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1071                 pud_t *pud, int flags)
1072 {
1073         pud_t _pud;
1074
1075         _pud = pud_mkyoung(*pud);
1076         if (flags & FOLL_WRITE)
1077                 _pud = pud_mkdirty(_pud);
1078         if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1079                                 pud, _pud, flags & FOLL_WRITE))
1080                 update_mmu_cache_pud(vma, addr, pud);
1081 }
1082
1083 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1084                 pud_t *pud, int flags, struct dev_pagemap **pgmap)
1085 {
1086         unsigned long pfn = pud_pfn(*pud);
1087         struct mm_struct *mm = vma->vm_mm;
1088         struct page *page;
1089
1090         assert_spin_locked(pud_lockptr(mm, pud));
1091
1092         if (flags & FOLL_WRITE && !pud_write(*pud))
1093                 return NULL;
1094
1095         if (pud_present(*pud) && pud_devmap(*pud))
1096                 /* pass */;
1097         else
1098                 return NULL;
1099
1100         if (flags & FOLL_TOUCH)
1101                 touch_pud(vma, addr, pud, flags);
1102
1103         /*
1104          * device mapped pages can only be returned if the
1105          * caller will manage the page reference count.
1106          */
1107         if (!(flags & FOLL_GET))
1108                 return ERR_PTR(-EEXIST);
1109
1110         pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1111         *pgmap = get_dev_pagemap(pfn, *pgmap);
1112         if (!*pgmap)
1113                 return ERR_PTR(-EFAULT);
1114         page = pfn_to_page(pfn);
1115         get_page(page);
1116
1117         return page;
1118 }
1119
1120 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1121                   pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1122                   struct vm_area_struct *vma)
1123 {
1124         spinlock_t *dst_ptl, *src_ptl;
1125         pud_t pud;
1126         int ret;
1127
1128         dst_ptl = pud_lock(dst_mm, dst_pud);
1129         src_ptl = pud_lockptr(src_mm, src_pud);
1130         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1131
1132         ret = -EAGAIN;
1133         pud = *src_pud;
1134         if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1135                 goto out_unlock;
1136
1137         /*
1138          * When page table lock is held, the huge zero pud should not be
1139          * under splitting since we don't split the page itself, only pud to
1140          * a page table.
1141          */
1142         if (is_huge_zero_pud(pud)) {
1143                 /* No huge zero pud yet */
1144         }
1145
1146         pudp_set_wrprotect(src_mm, addr, src_pud);
1147         pud = pud_mkold(pud_wrprotect(pud));
1148         set_pud_at(dst_mm, addr, dst_pud, pud);
1149
1150         ret = 0;
1151 out_unlock:
1152         spin_unlock(src_ptl);
1153         spin_unlock(dst_ptl);
1154         return ret;
1155 }
1156
1157 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1158 {
1159         pud_t entry;
1160         unsigned long haddr;
1161         bool write = vmf->flags & FAULT_FLAG_WRITE;
1162
1163         vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1164         if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1165                 goto unlock;
1166
1167         entry = pud_mkyoung(orig_pud);
1168         if (write)
1169                 entry = pud_mkdirty(entry);
1170         haddr = vmf->address & HPAGE_PUD_MASK;
1171         if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1172                 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1173
1174 unlock:
1175         spin_unlock(vmf->ptl);
1176 }
1177 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1178
1179 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1180 {
1181         pmd_t entry;
1182         unsigned long haddr;
1183         bool write = vmf->flags & FAULT_FLAG_WRITE;
1184
1185         vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1186         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1187                 goto unlock;
1188
1189         entry = pmd_mkyoung(orig_pmd);
1190         if (write)
1191                 entry = pmd_mkdirty(entry);
1192         haddr = vmf->address & HPAGE_PMD_MASK;
1193         if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1194                 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1195
1196 unlock:
1197         spin_unlock(vmf->ptl);
1198 }
1199
1200 static vm_fault_t do_huge_pmd_wp_page_fallback(struct vm_fault *vmf,
1201                         pmd_t orig_pmd, struct page *page)
1202 {
1203         struct vm_area_struct *vma = vmf->vma;
1204         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1205         struct mem_cgroup *memcg;
1206         pgtable_t pgtable;
1207         pmd_t _pmd;
1208         int i;
1209         vm_fault_t ret = 0;
1210         struct page **pages;
1211         struct mmu_notifier_range range;
1212
1213         pages = kmalloc_array(HPAGE_PMD_NR, sizeof(struct page *),
1214                               GFP_KERNEL);
1215         if (unlikely(!pages)) {
1216                 ret |= VM_FAULT_OOM;
1217                 goto out;
1218         }
1219
1220         for (i = 0; i < HPAGE_PMD_NR; i++) {
1221                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
1222                                                vmf->address, page_to_nid(page));
1223                 if (unlikely(!pages[i] ||
1224                              mem_cgroup_try_charge_delay(pages[i], vma->vm_mm,
1225                                      GFP_KERNEL, &memcg, false))) {
1226                         if (pages[i])
1227                                 put_page(pages[i]);
1228                         while (--i >= 0) {
1229                                 memcg = (void *)page_private(pages[i]);
1230                                 set_page_private(pages[i], 0);
1231                                 mem_cgroup_cancel_charge(pages[i], memcg,
1232                                                 false);
1233                                 put_page(pages[i]);
1234                         }
1235                         kfree(pages);
1236                         ret |= VM_FAULT_OOM;
1237                         goto out;
1238                 }
1239                 set_page_private(pages[i], (unsigned long)memcg);
1240         }
1241
1242         for (i = 0; i < HPAGE_PMD_NR; i++) {
1243                 copy_user_highpage(pages[i], page + i,
1244                                    haddr + PAGE_SIZE * i, vma);
1245                 __SetPageUptodate(pages[i]);
1246                 cond_resched();
1247         }
1248
1249         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1250                                 haddr, haddr + HPAGE_PMD_SIZE);
1251         mmu_notifier_invalidate_range_start(&range);
1252
1253         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1254         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1255                 goto out_free_pages;
1256         VM_BUG_ON_PAGE(!PageHead(page), page);
1257
1258         /*
1259          * Leave pmd empty until pte is filled note we must notify here as
1260          * concurrent CPU thread might write to new page before the call to
1261          * mmu_notifier_invalidate_range_end() happens which can lead to a
1262          * device seeing memory write in different order than CPU.
1263          *
1264          * See Documentation/vm/mmu_notifier.rst
1265          */
1266         pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1267
1268         pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
1269         pmd_populate(vma->vm_mm, &_pmd, pgtable);
1270
1271         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1272                 pte_t entry;
1273                 entry = mk_pte(pages[i], vma->vm_page_prot);
1274                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1275                 memcg = (void *)page_private(pages[i]);
1276                 set_page_private(pages[i], 0);
1277                 page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
1278                 mem_cgroup_commit_charge(pages[i], memcg, false, false);
1279                 lru_cache_add_active_or_unevictable(pages[i], vma);
1280                 vmf->pte = pte_offset_map(&_pmd, haddr);
1281                 VM_BUG_ON(!pte_none(*vmf->pte));
1282                 set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
1283                 pte_unmap(vmf->pte);
1284         }
1285         kfree(pages);
1286
1287         smp_wmb(); /* make pte visible before pmd */
1288         pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
1289         page_remove_rmap(page, true);
1290         spin_unlock(vmf->ptl);
1291
1292         /*
1293          * No need to double call mmu_notifier->invalidate_range() callback as
1294          * the above pmdp_huge_clear_flush_notify() did already call it.
1295          */
1296         mmu_notifier_invalidate_range_only_end(&range);
1297
1298         ret |= VM_FAULT_WRITE;
1299         put_page(page);
1300
1301 out:
1302         return ret;
1303
1304 out_free_pages:
1305         spin_unlock(vmf->ptl);
1306         mmu_notifier_invalidate_range_end(&range);
1307         for (i = 0; i < HPAGE_PMD_NR; i++) {
1308                 memcg = (void *)page_private(pages[i]);
1309                 set_page_private(pages[i], 0);
1310                 mem_cgroup_cancel_charge(pages[i], memcg, false);
1311                 put_page(pages[i]);
1312         }
1313         kfree(pages);
1314         goto out;
1315 }
1316
1317 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1318 {
1319         struct vm_area_struct *vma = vmf->vma;
1320         struct page *page = NULL, *new_page;
1321         struct mem_cgroup *memcg;
1322         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1323         struct mmu_notifier_range range;
1324         gfp_t huge_gfp;                 /* for allocation and charge */
1325         vm_fault_t ret = 0;
1326
1327         vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1328         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1329         if (is_huge_zero_pmd(orig_pmd))
1330                 goto alloc;
1331         spin_lock(vmf->ptl);
1332         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1333                 goto out_unlock;
1334
1335         page = pmd_page(orig_pmd);
1336         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1337         /*
1338          * We can only reuse the page if nobody else maps the huge page or it's
1339          * part.
1340          */
1341         if (!trylock_page(page)) {
1342                 get_page(page);
1343                 spin_unlock(vmf->ptl);
1344                 lock_page(page);
1345                 spin_lock(vmf->ptl);
1346                 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1347                         unlock_page(page);
1348                         put_page(page);
1349                         goto out_unlock;
1350                 }
1351                 put_page(page);
1352         }
1353         if (reuse_swap_page(page, NULL)) {
1354                 pmd_t entry;
1355                 entry = pmd_mkyoung(orig_pmd);
1356                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1357                 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry,  1))
1358                         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1359                 ret |= VM_FAULT_WRITE;
1360                 unlock_page(page);
1361                 goto out_unlock;
1362         }
1363         unlock_page(page);
1364         get_page(page);
1365         spin_unlock(vmf->ptl);
1366 alloc:
1367         if (__transparent_hugepage_enabled(vma) &&
1368             !transparent_hugepage_debug_cow()) {
1369                 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1370                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1371         } else
1372                 new_page = NULL;
1373
1374         if (likely(new_page)) {
1375                 prep_transhuge_page(new_page);
1376         } else {
1377                 if (!page) {
1378                         split_huge_pmd(vma, vmf->pmd, vmf->address);
1379                         ret |= VM_FAULT_FALLBACK;
1380                 } else {
1381                         ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
1382                         if (ret & VM_FAULT_OOM) {
1383                                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1384                                 ret |= VM_FAULT_FALLBACK;
1385                         }
1386                         put_page(page);
1387                 }
1388                 count_vm_event(THP_FAULT_FALLBACK);
1389                 goto out;
1390         }
1391
1392         if (unlikely(mem_cgroup_try_charge_delay(new_page, vma->vm_mm,
1393                                         huge_gfp, &memcg, true))) {
1394                 put_page(new_page);
1395                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1396                 if (page)
1397                         put_page(page);
1398                 ret |= VM_FAULT_FALLBACK;
1399                 count_vm_event(THP_FAULT_FALLBACK);
1400                 goto out;
1401         }
1402
1403         count_vm_event(THP_FAULT_ALLOC);
1404         count_memcg_events(memcg, THP_FAULT_ALLOC, 1);
1405
1406         if (!page)
1407                 clear_huge_page(new_page, vmf->address, HPAGE_PMD_NR);
1408         else
1409                 copy_user_huge_page(new_page, page, vmf->address,
1410                                     vma, HPAGE_PMD_NR);
1411         __SetPageUptodate(new_page);
1412
1413         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1414                                 haddr, haddr + HPAGE_PMD_SIZE);
1415         mmu_notifier_invalidate_range_start(&range);
1416
1417         spin_lock(vmf->ptl);
1418         if (page)
1419                 put_page(page);
1420         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1421                 spin_unlock(vmf->ptl);
1422                 mem_cgroup_cancel_charge(new_page, memcg, true);
1423                 put_page(new_page);
1424                 goto out_mn;
1425         } else {
1426                 pmd_t entry;
1427                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1428                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1429                 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1430                 page_add_new_anon_rmap(new_page, vma, haddr, true);
1431                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1432                 lru_cache_add_active_or_unevictable(new_page, vma);
1433                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1434                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1435                 if (!page) {
1436                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1437                 } else {
1438                         VM_BUG_ON_PAGE(!PageHead(page), page);
1439                         page_remove_rmap(page, true);
1440                         put_page(page);
1441                 }
1442                 ret |= VM_FAULT_WRITE;
1443         }
1444         spin_unlock(vmf->ptl);
1445 out_mn:
1446         /*
1447          * No need to double call mmu_notifier->invalidate_range() callback as
1448          * the above pmdp_huge_clear_flush_notify() did already call it.
1449          */
1450         mmu_notifier_invalidate_range_only_end(&range);
1451 out:
1452         return ret;
1453 out_unlock:
1454         spin_unlock(vmf->ptl);
1455         return ret;
1456 }
1457
1458 /*
1459  * FOLL_FORCE or a forced COW break can write even to unwritable pmd's,
1460  * but only after we've gone through a COW cycle and they are dirty.
1461  */
1462 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1463 {
1464         return pmd_write(pmd) || ((flags & FOLL_COW) && pmd_dirty(pmd));
1465 }
1466
1467 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1468                                    unsigned long addr,
1469                                    pmd_t *pmd,
1470                                    unsigned int flags)
1471 {
1472         struct mm_struct *mm = vma->vm_mm;
1473         struct page *page = NULL;
1474
1475         assert_spin_locked(pmd_lockptr(mm, pmd));
1476
1477         if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1478                 goto out;
1479
1480         /* Avoid dumping huge zero page */
1481         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1482                 return ERR_PTR(-EFAULT);
1483
1484         /* Full NUMA hinting faults to serialise migration in fault paths */
1485         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1486                 goto out;
1487
1488         page = pmd_page(*pmd);
1489         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1490         if (flags & FOLL_TOUCH)
1491                 touch_pmd(vma, addr, pmd, flags);
1492         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1493                 /*
1494                  * We don't mlock() pte-mapped THPs. This way we can avoid
1495                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1496                  *
1497                  * For anon THP:
1498                  *
1499                  * In most cases the pmd is the only mapping of the page as we
1500                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1501                  * writable private mappings in populate_vma_page_range().
1502                  *
1503                  * The only scenario when we have the page shared here is if we
1504                  * mlocking read-only mapping shared over fork(). We skip
1505                  * mlocking such pages.
1506                  *
1507                  * For file THP:
1508                  *
1509                  * We can expect PageDoubleMap() to be stable under page lock:
1510                  * for file pages we set it in page_add_file_rmap(), which
1511                  * requires page to be locked.
1512                  */
1513
1514                 if (PageAnon(page) && compound_mapcount(page) != 1)
1515                         goto skip_mlock;
1516                 if (PageDoubleMap(page) || !page->mapping)
1517                         goto skip_mlock;
1518                 if (!trylock_page(page))
1519                         goto skip_mlock;
1520                 lru_add_drain();
1521                 if (page->mapping && !PageDoubleMap(page))
1522                         mlock_vma_page(page);
1523                 unlock_page(page);
1524         }
1525 skip_mlock:
1526         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1527         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1528         if (flags & FOLL_GET)
1529                 get_page(page);
1530
1531 out:
1532         return page;
1533 }
1534
1535 /* NUMA hinting page fault entry point for trans huge pmds */
1536 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1537 {
1538         struct vm_area_struct *vma = vmf->vma;
1539         struct anon_vma *anon_vma = NULL;
1540         struct page *page;
1541         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1542         int page_nid = NUMA_NO_NODE, this_nid = numa_node_id();
1543         int target_nid, last_cpupid = -1;
1544         bool page_locked;
1545         bool migrated = false;
1546         bool was_writable;
1547         int flags = 0;
1548
1549         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1550         if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1551                 goto out_unlock;
1552
1553         /*
1554          * If there are potential migrations, wait for completion and retry
1555          * without disrupting NUMA hinting information. Do not relock and
1556          * check_same as the page may no longer be mapped.
1557          */
1558         if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1559                 page = pmd_page(*vmf->pmd);
1560                 if (!get_page_unless_zero(page))
1561                         goto out_unlock;
1562                 spin_unlock(vmf->ptl);
1563                 put_and_wait_on_page_locked(page);
1564                 goto out;
1565         }
1566
1567         page = pmd_page(pmd);
1568         BUG_ON(is_huge_zero_page(page));
1569         page_nid = page_to_nid(page);
1570         last_cpupid = page_cpupid_last(page);
1571         count_vm_numa_event(NUMA_HINT_FAULTS);
1572         if (page_nid == this_nid) {
1573                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1574                 flags |= TNF_FAULT_LOCAL;
1575         }
1576
1577         /* See similar comment in do_numa_page for explanation */
1578         if (!pmd_savedwrite(pmd))
1579                 flags |= TNF_NO_GROUP;
1580
1581         /*
1582          * Acquire the page lock to serialise THP migrations but avoid dropping
1583          * page_table_lock if at all possible
1584          */
1585         page_locked = trylock_page(page);
1586         target_nid = mpol_misplaced(page, vma, haddr);
1587         if (target_nid == NUMA_NO_NODE) {
1588                 /* If the page was locked, there are no parallel migrations */
1589                 if (page_locked)
1590                         goto clear_pmdnuma;
1591         }
1592
1593         /* Migration could have started since the pmd_trans_migrating check */
1594         if (!page_locked) {
1595                 page_nid = NUMA_NO_NODE;
1596                 if (!get_page_unless_zero(page))
1597                         goto out_unlock;
1598                 spin_unlock(vmf->ptl);
1599                 put_and_wait_on_page_locked(page);
1600                 goto out;
1601         }
1602
1603         /*
1604          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1605          * to serialises splits
1606          */
1607         get_page(page);
1608         spin_unlock(vmf->ptl);
1609         anon_vma = page_lock_anon_vma_read(page);
1610
1611         /* Confirm the PMD did not change while page_table_lock was released */
1612         spin_lock(vmf->ptl);
1613         if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1614                 unlock_page(page);
1615                 put_page(page);
1616                 page_nid = NUMA_NO_NODE;
1617                 goto out_unlock;
1618         }
1619
1620         /* Bail if we fail to protect against THP splits for any reason */
1621         if (unlikely(!anon_vma)) {
1622                 put_page(page);
1623                 page_nid = NUMA_NO_NODE;
1624                 goto clear_pmdnuma;
1625         }
1626
1627         /*
1628          * Since we took the NUMA fault, we must have observed the !accessible
1629          * bit. Make sure all other CPUs agree with that, to avoid them
1630          * modifying the page we're about to migrate.
1631          *
1632          * Must be done under PTL such that we'll observe the relevant
1633          * inc_tlb_flush_pending().
1634          *
1635          * We are not sure a pending tlb flush here is for a huge page
1636          * mapping or not. Hence use the tlb range variant
1637          */
1638         if (mm_tlb_flush_pending(vma->vm_mm)) {
1639                 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1640                 /*
1641                  * change_huge_pmd() released the pmd lock before
1642                  * invalidating the secondary MMUs sharing the primary
1643                  * MMU pagetables (with ->invalidate_range()). The
1644                  * mmu_notifier_invalidate_range_end() (which
1645                  * internally calls ->invalidate_range()) in
1646                  * change_pmd_range() will run after us, so we can't
1647                  * rely on it here and we need an explicit invalidate.
1648                  */
1649                 mmu_notifier_invalidate_range(vma->vm_mm, haddr,
1650                                               haddr + HPAGE_PMD_SIZE);
1651         }
1652
1653         /*
1654          * Migrate the THP to the requested node, returns with page unlocked
1655          * and access rights restored.
1656          */
1657         spin_unlock(vmf->ptl);
1658
1659         migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1660                                 vmf->pmd, pmd, vmf->address, page, target_nid);
1661         if (migrated) {
1662                 flags |= TNF_MIGRATED;
1663                 page_nid = target_nid;
1664         } else
1665                 flags |= TNF_MIGRATE_FAIL;
1666
1667         goto out;
1668 clear_pmdnuma:
1669         BUG_ON(!PageLocked(page));
1670         was_writable = pmd_savedwrite(pmd);
1671         pmd = pmd_modify(pmd, vma->vm_page_prot);
1672         pmd = pmd_mkyoung(pmd);
1673         if (was_writable)
1674                 pmd = pmd_mkwrite(pmd);
1675         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1676         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1677         unlock_page(page);
1678 out_unlock:
1679         spin_unlock(vmf->ptl);
1680
1681 out:
1682         if (anon_vma)
1683                 page_unlock_anon_vma_read(anon_vma);
1684
1685         if (page_nid != NUMA_NO_NODE)
1686                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1687                                 flags);
1688
1689         return 0;
1690 }
1691
1692 /*
1693  * Return true if we do MADV_FREE successfully on entire pmd page.
1694  * Otherwise, return false.
1695  */
1696 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1697                 pmd_t *pmd, unsigned long addr, unsigned long next)
1698 {
1699         spinlock_t *ptl;
1700         pmd_t orig_pmd;
1701         struct page *page;
1702         struct mm_struct *mm = tlb->mm;
1703         bool ret = false;
1704
1705         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1706
1707         ptl = pmd_trans_huge_lock(pmd, vma);
1708         if (!ptl)
1709                 goto out_unlocked;
1710
1711         orig_pmd = *pmd;
1712         if (is_huge_zero_pmd(orig_pmd))
1713                 goto out;
1714
1715         if (unlikely(!pmd_present(orig_pmd))) {
1716                 VM_BUG_ON(thp_migration_supported() &&
1717                                   !is_pmd_migration_entry(orig_pmd));
1718                 goto out;
1719         }
1720
1721         page = pmd_page(orig_pmd);
1722         /*
1723          * If other processes are mapping this page, we couldn't discard
1724          * the page unless they all do MADV_FREE so let's skip the page.
1725          */
1726         if (total_mapcount(page) != 1)
1727                 goto out;
1728
1729         if (!trylock_page(page))
1730                 goto out;
1731
1732         /*
1733          * If user want to discard part-pages of THP, split it so MADV_FREE
1734          * will deactivate only them.
1735          */
1736         if (next - addr != HPAGE_PMD_SIZE) {
1737                 get_page(page);
1738                 spin_unlock(ptl);
1739                 split_huge_page(page);
1740                 unlock_page(page);
1741                 put_page(page);
1742                 goto out_unlocked;
1743         }
1744
1745         if (PageDirty(page))
1746                 ClearPageDirty(page);
1747         unlock_page(page);
1748
1749         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1750                 pmdp_invalidate(vma, addr, pmd);
1751                 orig_pmd = pmd_mkold(orig_pmd);
1752                 orig_pmd = pmd_mkclean(orig_pmd);
1753
1754                 set_pmd_at(mm, addr, pmd, orig_pmd);
1755                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1756         }
1757
1758         mark_page_lazyfree(page);
1759         ret = true;
1760 out:
1761         spin_unlock(ptl);
1762 out_unlocked:
1763         return ret;
1764 }
1765
1766 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1767 {
1768         pgtable_t pgtable;
1769
1770         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1771         pte_free(mm, pgtable);
1772         mm_dec_nr_ptes(mm);
1773 }
1774
1775 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1776                  pmd_t *pmd, unsigned long addr)
1777 {
1778         pmd_t orig_pmd;
1779         spinlock_t *ptl;
1780
1781         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1782
1783         ptl = __pmd_trans_huge_lock(pmd, vma);
1784         if (!ptl)
1785                 return 0;
1786         /*
1787          * For architectures like ppc64 we look at deposited pgtable
1788          * when calling pmdp_huge_get_and_clear. So do the
1789          * pgtable_trans_huge_withdraw after finishing pmdp related
1790          * operations.
1791          */
1792         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1793                         tlb->fullmm);
1794         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1795         if (vma_is_dax(vma)) {
1796                 if (arch_needs_pgtable_deposit())
1797                         zap_deposited_table(tlb->mm, pmd);
1798                 spin_unlock(ptl);
1799                 if (is_huge_zero_pmd(orig_pmd))
1800                         tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1801         } else if (is_huge_zero_pmd(orig_pmd)) {
1802                 zap_deposited_table(tlb->mm, pmd);
1803                 spin_unlock(ptl);
1804                 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1805         } else {
1806                 struct page *page = NULL;
1807                 int flush_needed = 1;
1808
1809                 if (pmd_present(orig_pmd)) {
1810                         page = pmd_page(orig_pmd);
1811                         page_remove_rmap(page, true);
1812                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1813                         VM_BUG_ON_PAGE(!PageHead(page), page);
1814                 } else if (thp_migration_supported()) {
1815                         swp_entry_t entry;
1816
1817                         VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1818                         entry = pmd_to_swp_entry(orig_pmd);
1819                         page = pfn_to_page(swp_offset(entry));
1820                         flush_needed = 0;
1821                 } else
1822                         WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1823
1824                 if (PageAnon(page)) {
1825                         zap_deposited_table(tlb->mm, pmd);
1826                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1827                 } else {
1828                         if (arch_needs_pgtable_deposit())
1829                                 zap_deposited_table(tlb->mm, pmd);
1830                         add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1831                 }
1832
1833                 spin_unlock(ptl);
1834                 if (flush_needed)
1835                         tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1836         }
1837         return 1;
1838 }
1839
1840 #ifndef pmd_move_must_withdraw
1841 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1842                                          spinlock_t *old_pmd_ptl,
1843                                          struct vm_area_struct *vma)
1844 {
1845         /*
1846          * With split pmd lock we also need to move preallocated
1847          * PTE page table if new_pmd is on different PMD page table.
1848          *
1849          * We also don't deposit and withdraw tables for file pages.
1850          */
1851         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1852 }
1853 #endif
1854
1855 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1856 {
1857 #ifdef CONFIG_MEM_SOFT_DIRTY
1858         if (unlikely(is_pmd_migration_entry(pmd)))
1859                 pmd = pmd_swp_mksoft_dirty(pmd);
1860         else if (pmd_present(pmd))
1861                 pmd = pmd_mksoft_dirty(pmd);
1862 #endif
1863         return pmd;
1864 }
1865
1866 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1867                   unsigned long new_addr, unsigned long old_end,
1868                   pmd_t *old_pmd, pmd_t *new_pmd)
1869 {
1870         spinlock_t *old_ptl, *new_ptl;
1871         pmd_t pmd;
1872         struct mm_struct *mm = vma->vm_mm;
1873         bool force_flush = false;
1874
1875         if ((old_addr & ~HPAGE_PMD_MASK) ||
1876             (new_addr & ~HPAGE_PMD_MASK) ||
1877             old_end - old_addr < HPAGE_PMD_SIZE)
1878                 return false;
1879
1880         /*
1881          * The destination pmd shouldn't be established, free_pgtables()
1882          * should have release it.
1883          */
1884         if (WARN_ON(!pmd_none(*new_pmd))) {
1885                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1886                 return false;
1887         }
1888
1889         /*
1890          * We don't have to worry about the ordering of src and dst
1891          * ptlocks because exclusive mmap_sem prevents deadlock.
1892          */
1893         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1894         if (old_ptl) {
1895                 new_ptl = pmd_lockptr(mm, new_pmd);
1896                 if (new_ptl != old_ptl)
1897                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1898                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1899                 if (pmd_present(pmd))
1900                         force_flush = true;
1901                 VM_BUG_ON(!pmd_none(*new_pmd));
1902
1903                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1904                         pgtable_t pgtable;
1905                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1906                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1907                 }
1908                 pmd = move_soft_dirty_pmd(pmd);
1909                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1910                 if (force_flush)
1911                         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1912                 if (new_ptl != old_ptl)
1913                         spin_unlock(new_ptl);
1914                 spin_unlock(old_ptl);
1915                 return true;
1916         }
1917         return false;
1918 }
1919
1920 /*
1921  * Returns
1922  *  - 0 if PMD could not be locked
1923  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1924  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1925  */
1926 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1927                 unsigned long addr, pgprot_t newprot, int prot_numa)
1928 {
1929         struct mm_struct *mm = vma->vm_mm;
1930         spinlock_t *ptl;
1931         pmd_t entry;
1932         bool preserve_write;
1933         int ret;
1934
1935         ptl = __pmd_trans_huge_lock(pmd, vma);
1936         if (!ptl)
1937                 return 0;
1938
1939         preserve_write = prot_numa && pmd_write(*pmd);
1940         ret = 1;
1941
1942 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1943         if (is_swap_pmd(*pmd)) {
1944                 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1945
1946                 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1947                 if (is_write_migration_entry(entry)) {
1948                         pmd_t newpmd;
1949                         /*
1950                          * A protection check is difficult so
1951                          * just be safe and disable write
1952                          */
1953                         make_migration_entry_read(&entry);
1954                         newpmd = swp_entry_to_pmd(entry);
1955                         if (pmd_swp_soft_dirty(*pmd))
1956                                 newpmd = pmd_swp_mksoft_dirty(newpmd);
1957                         set_pmd_at(mm, addr, pmd, newpmd);
1958                 }
1959                 goto unlock;
1960         }
1961 #endif
1962
1963         /*
1964          * Avoid trapping faults against the zero page. The read-only
1965          * data is likely to be read-cached on the local CPU and
1966          * local/remote hits to the zero page are not interesting.
1967          */
1968         if (prot_numa && is_huge_zero_pmd(*pmd))
1969                 goto unlock;
1970
1971         if (prot_numa && pmd_protnone(*pmd))
1972                 goto unlock;
1973
1974         /*
1975          * In case prot_numa, we are under down_read(mmap_sem). It's critical
1976          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1977          * which is also under down_read(mmap_sem):
1978          *
1979          *      CPU0:                           CPU1:
1980          *                              change_huge_pmd(prot_numa=1)
1981          *                               pmdp_huge_get_and_clear_notify()
1982          * madvise_dontneed()
1983          *  zap_pmd_range()
1984          *   pmd_trans_huge(*pmd) == 0 (without ptl)
1985          *   // skip the pmd
1986          *                               set_pmd_at();
1987          *                               // pmd is re-established
1988          *
1989          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1990          * which may break userspace.
1991          *
1992          * pmdp_invalidate() is required to make sure we don't miss
1993          * dirty/young flags set by hardware.
1994          */
1995         entry = pmdp_invalidate(vma, addr, pmd);
1996
1997         entry = pmd_modify(entry, newprot);
1998         if (preserve_write)
1999                 entry = pmd_mk_savedwrite(entry);
2000         ret = HPAGE_PMD_NR;
2001         set_pmd_at(mm, addr, pmd, entry);
2002         BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
2003 unlock:
2004         spin_unlock(ptl);
2005         return ret;
2006 }
2007
2008 /*
2009  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
2010  *
2011  * Note that if it returns page table lock pointer, this routine returns without
2012  * unlocking page table lock. So callers must unlock it.
2013  */
2014 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
2015 {
2016         spinlock_t *ptl;
2017         ptl = pmd_lock(vma->vm_mm, pmd);
2018         if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
2019                         pmd_devmap(*pmd)))
2020                 return ptl;
2021         spin_unlock(ptl);
2022         return NULL;
2023 }
2024
2025 /*
2026  * Returns true if a given pud maps a thp, false otherwise.
2027  *
2028  * Note that if it returns true, this routine returns without unlocking page
2029  * table lock. So callers must unlock it.
2030  */
2031 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
2032 {
2033         spinlock_t *ptl;
2034
2035         ptl = pud_lock(vma->vm_mm, pud);
2036         if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
2037                 return ptl;
2038         spin_unlock(ptl);
2039         return NULL;
2040 }
2041
2042 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
2043 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
2044                  pud_t *pud, unsigned long addr)
2045 {
2046         spinlock_t *ptl;
2047
2048         ptl = __pud_trans_huge_lock(pud, vma);
2049         if (!ptl)
2050                 return 0;
2051         /*
2052          * For architectures like ppc64 we look at deposited pgtable
2053          * when calling pudp_huge_get_and_clear. So do the
2054          * pgtable_trans_huge_withdraw after finishing pudp related
2055          * operations.
2056          */
2057         pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
2058         tlb_remove_pud_tlb_entry(tlb, pud, addr);
2059         if (vma_is_dax(vma)) {
2060                 spin_unlock(ptl);
2061                 /* No zero page support yet */
2062         } else {
2063                 /* No support for anonymous PUD pages yet */
2064                 BUG();
2065         }
2066         return 1;
2067 }
2068
2069 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
2070                 unsigned long haddr)
2071 {
2072         VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
2073         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2074         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
2075         VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
2076
2077         count_vm_event(THP_SPLIT_PUD);
2078
2079         pudp_huge_clear_flush_notify(vma, haddr, pud);
2080 }
2081
2082 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2083                 unsigned long address)
2084 {
2085         spinlock_t *ptl;
2086         struct mmu_notifier_range range;
2087
2088         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2089                                 address & HPAGE_PUD_MASK,
2090                                 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
2091         mmu_notifier_invalidate_range_start(&range);
2092         ptl = pud_lock(vma->vm_mm, pud);
2093         if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2094                 goto out;
2095         __split_huge_pud_locked(vma, pud, range.start);
2096
2097 out:
2098         spin_unlock(ptl);
2099         /*
2100          * No need to double call mmu_notifier->invalidate_range() callback as
2101          * the above pudp_huge_clear_flush_notify() did already call it.
2102          */
2103         mmu_notifier_invalidate_range_only_end(&range);
2104 }
2105 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2106
2107 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2108                 unsigned long haddr, pmd_t *pmd)
2109 {
2110         struct mm_struct *mm = vma->vm_mm;
2111         pgtable_t pgtable;
2112         pmd_t _pmd;
2113         int i;
2114
2115         /*
2116          * Leave pmd empty until pte is filled note that it is fine to delay
2117          * notification until mmu_notifier_invalidate_range_end() as we are
2118          * replacing a zero pmd write protected page with a zero pte write
2119          * protected page.
2120          *
2121          * See Documentation/vm/mmu_notifier.rst
2122          */
2123         pmdp_huge_clear_flush(vma, haddr, pmd);
2124
2125         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2126         pmd_populate(mm, &_pmd, pgtable);
2127
2128         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2129                 pte_t *pte, entry;
2130                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2131                 entry = pte_mkspecial(entry);
2132                 pte = pte_offset_map(&_pmd, haddr);
2133                 VM_BUG_ON(!pte_none(*pte));
2134                 set_pte_at(mm, haddr, pte, entry);
2135                 pte_unmap(pte);
2136         }
2137         smp_wmb(); /* make pte visible before pmd */
2138         pmd_populate(mm, pmd, pgtable);
2139 }
2140
2141 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2142                 unsigned long haddr, bool freeze)
2143 {
2144         struct mm_struct *mm = vma->vm_mm;
2145         struct page *page;
2146         pgtable_t pgtable;
2147         pmd_t old_pmd, _pmd;
2148         bool young, write, soft_dirty, pmd_migration = false;
2149         unsigned long addr;
2150         int i;
2151
2152         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2153         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2154         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2155         VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2156                                 && !pmd_devmap(*pmd));
2157
2158         count_vm_event(THP_SPLIT_PMD);
2159
2160         if (!vma_is_anonymous(vma)) {
2161                 old_pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2162                 /*
2163                  * We are going to unmap this huge page. So
2164                  * just go ahead and zap it
2165                  */
2166                 if (arch_needs_pgtable_deposit())
2167                         zap_deposited_table(mm, pmd);
2168                 if (vma_is_dax(vma))
2169                         return;
2170                 if (unlikely(is_pmd_migration_entry(old_pmd))) {
2171                         swp_entry_t entry;
2172
2173                         entry = pmd_to_swp_entry(old_pmd);
2174                         page = migration_entry_to_page(entry);
2175                 } else {
2176                         page = pmd_page(old_pmd);
2177                         if (!PageDirty(page) && pmd_dirty(old_pmd))
2178                                 set_page_dirty(page);
2179                         if (!PageReferenced(page) && pmd_young(old_pmd))
2180                                 SetPageReferenced(page);
2181                         page_remove_rmap(page, true);
2182                         put_page(page);
2183                 }
2184                 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2185                 return;
2186         }
2187
2188         if (is_huge_zero_pmd(*pmd)) {
2189                 /*
2190                  * FIXME: Do we want to invalidate secondary mmu by calling
2191                  * mmu_notifier_invalidate_range() see comments below inside
2192                  * __split_huge_pmd() ?
2193                  *
2194                  * We are going from a zero huge page write protected to zero
2195                  * small page also write protected so it does not seems useful
2196                  * to invalidate secondary mmu at this time.
2197                  */
2198                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2199         }
2200
2201         /*
2202          * Up to this point the pmd is present and huge and userland has the
2203          * whole access to the hugepage during the split (which happens in
2204          * place). If we overwrite the pmd with the not-huge version pointing
2205          * to the pte here (which of course we could if all CPUs were bug
2206          * free), userland could trigger a small page size TLB miss on the
2207          * small sized TLB while the hugepage TLB entry is still established in
2208          * the huge TLB. Some CPU doesn't like that.
2209          * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2210          * 383 on page 93. Intel should be safe but is also warns that it's
2211          * only safe if the permission and cache attributes of the two entries
2212          * loaded in the two TLB is identical (which should be the case here).
2213          * But it is generally safer to never allow small and huge TLB entries
2214          * for the same virtual address to be loaded simultaneously. So instead
2215          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2216          * current pmd notpresent (atomically because here the pmd_trans_huge
2217          * must remain set at all times on the pmd until the split is complete
2218          * for this pmd), then we flush the SMP TLB and finally we write the
2219          * non-huge version of the pmd entry with pmd_populate.
2220          */
2221         old_pmd = pmdp_invalidate(vma, haddr, pmd);
2222
2223         pmd_migration = is_pmd_migration_entry(old_pmd);
2224         if (unlikely(pmd_migration)) {
2225                 swp_entry_t entry;
2226
2227                 entry = pmd_to_swp_entry(old_pmd);
2228                 page = pfn_to_page(swp_offset(entry));
2229                 write = is_write_migration_entry(entry);
2230                 young = false;
2231                 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2232         } else {
2233                 page = pmd_page(old_pmd);
2234                 if (pmd_dirty(old_pmd))
2235                         SetPageDirty(page);
2236                 write = pmd_write(old_pmd);
2237                 young = pmd_young(old_pmd);
2238                 soft_dirty = pmd_soft_dirty(old_pmd);
2239         }
2240         VM_BUG_ON_PAGE(!page_count(page), page);
2241         page_ref_add(page, HPAGE_PMD_NR - 1);
2242
2243         /*
2244          * Withdraw the table only after we mark the pmd entry invalid.
2245          * This's critical for some architectures (Power).
2246          */
2247         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2248         pmd_populate(mm, &_pmd, pgtable);
2249
2250         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2251                 pte_t entry, *pte;
2252                 /*
2253                  * Note that NUMA hinting access restrictions are not
2254                  * transferred to avoid any possibility of altering
2255                  * permissions across VMAs.
2256                  */
2257                 if (freeze || pmd_migration) {
2258                         swp_entry_t swp_entry;
2259                         swp_entry = make_migration_entry(page + i, write);
2260                         entry = swp_entry_to_pte(swp_entry);
2261                         if (soft_dirty)
2262                                 entry = pte_swp_mksoft_dirty(entry);
2263                 } else {
2264                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2265                         entry = maybe_mkwrite(entry, vma);
2266                         if (!write)
2267                                 entry = pte_wrprotect(entry);
2268                         if (!young)
2269                                 entry = pte_mkold(entry);
2270                         if (soft_dirty)
2271                                 entry = pte_mksoft_dirty(entry);
2272                 }
2273                 pte = pte_offset_map(&_pmd, addr);
2274                 BUG_ON(!pte_none(*pte));
2275                 set_pte_at(mm, addr, pte, entry);
2276                 if (!pmd_migration)
2277                         atomic_inc(&page[i]._mapcount);
2278                 pte_unmap(pte);
2279         }
2280
2281         if (!pmd_migration) {
2282                 /*
2283                  * Set PG_double_map before dropping compound_mapcount to avoid
2284                  * false-negative page_mapped().
2285                  */
2286                 if (compound_mapcount(page) > 1 &&
2287                     !TestSetPageDoubleMap(page)) {
2288                         for (i = 0; i < HPAGE_PMD_NR; i++)
2289                                 atomic_inc(&page[i]._mapcount);
2290                 }
2291
2292                 lock_page_memcg(page);
2293                 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2294                         /* Last compound_mapcount is gone. */
2295                         __dec_lruvec_page_state(page, NR_ANON_THPS);
2296                         if (TestClearPageDoubleMap(page)) {
2297                                 /* No need in mapcount reference anymore */
2298                                 for (i = 0; i < HPAGE_PMD_NR; i++)
2299                                         atomic_dec(&page[i]._mapcount);
2300                         }
2301                 }
2302                 unlock_page_memcg(page);
2303         }
2304
2305         smp_wmb(); /* make pte visible before pmd */
2306         pmd_populate(mm, pmd, pgtable);
2307
2308         if (freeze) {
2309                 for (i = 0; i < HPAGE_PMD_NR; i++) {
2310                         page_remove_rmap(page + i, false);
2311                         put_page(page + i);
2312                 }
2313         }
2314 }
2315
2316 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2317                 unsigned long address, bool freeze, struct page *page)
2318 {
2319         spinlock_t *ptl;
2320         struct mmu_notifier_range range;
2321         bool do_unlock_page = false;
2322         pmd_t _pmd;
2323
2324         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2325                                 address & HPAGE_PMD_MASK,
2326                                 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2327         mmu_notifier_invalidate_range_start(&range);
2328         ptl = pmd_lock(vma->vm_mm, pmd);
2329
2330         /*
2331          * If caller asks to setup a migration entries, we need a page to check
2332          * pmd against. Otherwise we can end up replacing wrong page.
2333          */
2334         VM_BUG_ON(freeze && !page);
2335         if (page) {
2336                 VM_WARN_ON_ONCE(!PageLocked(page));
2337                 if (page != pmd_page(*pmd))
2338                         goto out;
2339         }
2340
2341 repeat:
2342         if (pmd_trans_huge(*pmd)) {
2343                 if (!page) {
2344                         page = pmd_page(*pmd);
2345                         /*
2346                          * An anonymous page must be locked, to ensure that a
2347                          * concurrent reuse_swap_page() sees stable mapcount;
2348                          * but reuse_swap_page() is not used on shmem or file,
2349                          * and page lock must not be taken when zap_pmd_range()
2350                          * calls __split_huge_pmd() while i_mmap_lock is held.
2351                          */
2352                         if (PageAnon(page)) {
2353                                 if (unlikely(!trylock_page(page))) {
2354                                         get_page(page);
2355                                         _pmd = *pmd;
2356                                         spin_unlock(ptl);
2357                                         lock_page(page);
2358                                         spin_lock(ptl);
2359                                         if (unlikely(!pmd_same(*pmd, _pmd))) {
2360                                                 unlock_page(page);
2361                                                 put_page(page);
2362                                                 page = NULL;
2363                                                 goto repeat;
2364                                         }
2365                                         put_page(page);
2366                                 }
2367                                 do_unlock_page = true;
2368                         }
2369                 }
2370                 if (PageMlocked(page))
2371                         clear_page_mlock(page);
2372         } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2373                 goto out;
2374         __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2375 out:
2376         spin_unlock(ptl);
2377         if (do_unlock_page)
2378                 unlock_page(page);
2379         /*
2380          * No need to double call mmu_notifier->invalidate_range() callback.
2381          * They are 3 cases to consider inside __split_huge_pmd_locked():
2382          *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2383          *  2) __split_huge_zero_page_pmd() read only zero page and any write
2384          *    fault will trigger a flush_notify before pointing to a new page
2385          *    (it is fine if the secondary mmu keeps pointing to the old zero
2386          *    page in the meantime)
2387          *  3) Split a huge pmd into pte pointing to the same page. No need
2388          *     to invalidate secondary tlb entry they are all still valid.
2389          *     any further changes to individual pte will notify. So no need
2390          *     to call mmu_notifier->invalidate_range()
2391          */
2392         mmu_notifier_invalidate_range_only_end(&range);
2393 }
2394
2395 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2396                 bool freeze, struct page *page)
2397 {
2398         pgd_t *pgd;
2399         p4d_t *p4d;
2400         pud_t *pud;
2401         pmd_t *pmd;
2402
2403         pgd = pgd_offset(vma->vm_mm, address);
2404         if (!pgd_present(*pgd))
2405                 return;
2406
2407         p4d = p4d_offset(pgd, address);
2408         if (!p4d_present(*p4d))
2409                 return;
2410
2411         pud = pud_offset(p4d, address);
2412         if (!pud_present(*pud))
2413                 return;
2414
2415         pmd = pmd_offset(pud, address);
2416
2417         __split_huge_pmd(vma, pmd, address, freeze, page);
2418 }
2419
2420 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2421                              unsigned long start,
2422                              unsigned long end,
2423                              long adjust_next)
2424 {
2425         /*
2426          * If the new start address isn't hpage aligned and it could
2427          * previously contain an hugepage: check if we need to split
2428          * an huge pmd.
2429          */
2430         if (start & ~HPAGE_PMD_MASK &&
2431             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2432             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2433                 split_huge_pmd_address(vma, start, false, NULL);
2434
2435         /*
2436          * If the new end address isn't hpage aligned and it could
2437          * previously contain an hugepage: check if we need to split
2438          * an huge pmd.
2439          */
2440         if (end & ~HPAGE_PMD_MASK &&
2441             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2442             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2443                 split_huge_pmd_address(vma, end, false, NULL);
2444
2445         /*
2446          * If we're also updating the vma->vm_next->vm_start, if the new
2447          * vm_next->vm_start isn't page aligned and it could previously
2448          * contain an hugepage: check if we need to split an huge pmd.
2449          */
2450         if (adjust_next > 0) {
2451                 struct vm_area_struct *next = vma->vm_next;
2452                 unsigned long nstart = next->vm_start;
2453                 nstart += adjust_next << PAGE_SHIFT;
2454                 if (nstart & ~HPAGE_PMD_MASK &&
2455                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2456                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2457                         split_huge_pmd_address(next, nstart, false, NULL);
2458         }
2459 }
2460
2461 static void unmap_page(struct page *page)
2462 {
2463         enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
2464                 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD | TTU_SYNC;
2465
2466         VM_BUG_ON_PAGE(!PageHead(page), page);
2467
2468         if (PageAnon(page))
2469                 ttu_flags |= TTU_SPLIT_FREEZE;
2470
2471         try_to_unmap(page, ttu_flags);
2472
2473         VM_WARN_ON_ONCE_PAGE(page_mapped(page), page);
2474 }
2475
2476 static void remap_page(struct page *page)
2477 {
2478         int i;
2479         if (PageTransHuge(page)) {
2480                 remove_migration_ptes(page, page, true);
2481         } else {
2482                 for (i = 0; i < HPAGE_PMD_NR; i++)
2483                         remove_migration_ptes(page + i, page + i, true);
2484         }
2485 }
2486
2487 static void __split_huge_page_tail(struct page *head, int tail,
2488                 struct lruvec *lruvec, struct list_head *list)
2489 {
2490         struct page *page_tail = head + tail;
2491
2492         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2493
2494         /*
2495          * Clone page flags before unfreezing refcount.
2496          *
2497          * After successful get_page_unless_zero() might follow flags change,
2498          * for exmaple lock_page() which set PG_waiters.
2499          */
2500         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2501         page_tail->flags |= (head->flags &
2502                         ((1L << PG_referenced) |
2503                          (1L << PG_swapbacked) |
2504                          (1L << PG_swapcache) |
2505                          (1L << PG_mlocked) |
2506                          (1L << PG_uptodate) |
2507                          (1L << PG_active) |
2508                          (1L << PG_workingset) |
2509                          (1L << PG_locked) |
2510                          (1L << PG_unevictable) |
2511                          (1L << PG_dirty)));
2512
2513         /* ->mapping in first tail page is compound_mapcount */
2514         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2515                         page_tail);
2516         page_tail->mapping = head->mapping;
2517         page_tail->index = head->index + tail;
2518
2519         /* Page flags must be visible before we make the page non-compound. */
2520         smp_wmb();
2521
2522         /*
2523          * Clear PageTail before unfreezing page refcount.
2524          *
2525          * After successful get_page_unless_zero() might follow put_page()
2526          * which needs correct compound_head().
2527          */
2528         clear_compound_head(page_tail);
2529
2530         /* Finally unfreeze refcount. Additional reference from page cache. */
2531         page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2532                                           PageSwapCache(head)));
2533
2534         if (page_is_young(head))
2535                 set_page_young(page_tail);
2536         if (page_is_idle(head))
2537                 set_page_idle(page_tail);
2538
2539         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2540
2541         /*
2542          * always add to the tail because some iterators expect new
2543          * pages to show after the currently processed elements - e.g.
2544          * migrate_pages
2545          */
2546         lru_add_page_tail(head, page_tail, lruvec, list);
2547 }
2548
2549 static void __split_huge_page(struct page *page, struct list_head *list,
2550                 pgoff_t end, unsigned long flags)
2551 {
2552         struct page *head = compound_head(page);
2553         pg_data_t *pgdat = page_pgdat(head);
2554         struct lruvec *lruvec;
2555         struct address_space *swap_cache = NULL;
2556         unsigned long offset = 0;
2557         int i;
2558
2559         lruvec = mem_cgroup_page_lruvec(head, pgdat);
2560
2561         /* complete memcg works before add pages to LRU */
2562         mem_cgroup_split_huge_fixup(head);
2563
2564         if (PageAnon(head) && PageSwapCache(head)) {
2565                 swp_entry_t entry = { .val = page_private(head) };
2566
2567                 offset = swp_offset(entry);
2568                 swap_cache = swap_address_space(entry);
2569                 xa_lock(&swap_cache->i_pages);
2570         }
2571
2572         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
2573                 __split_huge_page_tail(head, i, lruvec, list);
2574                 /* Some pages can be beyond i_size: drop them from page cache */
2575                 if (head[i].index >= end) {
2576                         ClearPageDirty(head + i);
2577                         __delete_from_page_cache(head + i, NULL);
2578                         if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2579                                 shmem_uncharge(head->mapping->host, 1);
2580                         put_page(head + i);
2581                 } else if (!PageAnon(page)) {
2582                         __xa_store(&head->mapping->i_pages, head[i].index,
2583                                         head + i, 0);
2584                 } else if (swap_cache) {
2585                         __xa_store(&swap_cache->i_pages, offset + i,
2586                                         head + i, 0);
2587                 }
2588         }
2589
2590         ClearPageCompound(head);
2591
2592         split_page_owner(head, HPAGE_PMD_NR);
2593
2594         /* See comment in __split_huge_page_tail() */
2595         if (PageAnon(head)) {
2596                 /* Additional pin to swap cache */
2597                 if (PageSwapCache(head)) {
2598                         page_ref_add(head, 2);
2599                         xa_unlock(&swap_cache->i_pages);
2600                 } else {
2601                         page_ref_inc(head);
2602                 }
2603         } else {
2604                 /* Additional pin to page cache */
2605                 page_ref_add(head, 2);
2606                 xa_unlock(&head->mapping->i_pages);
2607         }
2608
2609         spin_unlock_irqrestore(&pgdat->lru_lock, flags);
2610
2611         remap_page(head);
2612
2613         for (i = 0; i < HPAGE_PMD_NR; i++) {
2614                 struct page *subpage = head + i;
2615                 if (subpage == page)
2616                         continue;
2617                 unlock_page(subpage);
2618
2619                 /*
2620                  * Subpages may be freed if there wasn't any mapping
2621                  * like if add_to_swap() is running on a lru page that
2622                  * had its mapping zapped. And freeing these pages
2623                  * requires taking the lru_lock so we do the put_page
2624                  * of the tail pages after the split is complete.
2625                  */
2626                 put_page(subpage);
2627         }
2628 }
2629
2630 int total_mapcount(struct page *page)
2631 {
2632         int i, compound, ret;
2633
2634         VM_BUG_ON_PAGE(PageTail(page), page);
2635
2636         if (likely(!PageCompound(page)))
2637                 return atomic_read(&page->_mapcount) + 1;
2638
2639         compound = compound_mapcount(page);
2640         if (PageHuge(page))
2641                 return compound;
2642         ret = compound;
2643         for (i = 0; i < HPAGE_PMD_NR; i++)
2644                 ret += atomic_read(&page[i]._mapcount) + 1;
2645         /* File pages has compound_mapcount included in _mapcount */
2646         if (!PageAnon(page))
2647                 return ret - compound * HPAGE_PMD_NR;
2648         if (PageDoubleMap(page))
2649                 ret -= HPAGE_PMD_NR;
2650         return ret;
2651 }
2652
2653 /*
2654  * This calculates accurately how many mappings a transparent hugepage
2655  * has (unlike page_mapcount() which isn't fully accurate). This full
2656  * accuracy is primarily needed to know if copy-on-write faults can
2657  * reuse the page and change the mapping to read-write instead of
2658  * copying them. At the same time this returns the total_mapcount too.
2659  *
2660  * The function returns the highest mapcount any one of the subpages
2661  * has. If the return value is one, even if different processes are
2662  * mapping different subpages of the transparent hugepage, they can
2663  * all reuse it, because each process is reusing a different subpage.
2664  *
2665  * The total_mapcount is instead counting all virtual mappings of the
2666  * subpages. If the total_mapcount is equal to "one", it tells the
2667  * caller all mappings belong to the same "mm" and in turn the
2668  * anon_vma of the transparent hugepage can become the vma->anon_vma
2669  * local one as no other process may be mapping any of the subpages.
2670  *
2671  * It would be more accurate to replace page_mapcount() with
2672  * page_trans_huge_mapcount(), however we only use
2673  * page_trans_huge_mapcount() in the copy-on-write faults where we
2674  * need full accuracy to avoid breaking page pinning, because
2675  * page_trans_huge_mapcount() is slower than page_mapcount().
2676  */
2677 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2678 {
2679         int i, ret, _total_mapcount, mapcount;
2680
2681         /* hugetlbfs shouldn't call it */
2682         VM_BUG_ON_PAGE(PageHuge(page), page);
2683
2684         if (likely(!PageTransCompound(page))) {
2685                 mapcount = atomic_read(&page->_mapcount) + 1;
2686                 if (total_mapcount)
2687                         *total_mapcount = mapcount;
2688                 return mapcount;
2689         }
2690
2691         page = compound_head(page);
2692
2693         _total_mapcount = ret = 0;
2694         for (i = 0; i < HPAGE_PMD_NR; i++) {
2695                 mapcount = atomic_read(&page[i]._mapcount) + 1;
2696                 ret = max(ret, mapcount);
2697                 _total_mapcount += mapcount;
2698         }
2699         if (PageDoubleMap(page)) {
2700                 ret -= 1;
2701                 _total_mapcount -= HPAGE_PMD_NR;
2702         }
2703         mapcount = compound_mapcount(page);
2704         ret += mapcount;
2705         _total_mapcount += mapcount;
2706         if (total_mapcount)
2707                 *total_mapcount = _total_mapcount;
2708         return ret;
2709 }
2710
2711 /* Racy check whether the huge page can be split */
2712 bool can_split_huge_page(struct page *page, int *pextra_pins)
2713 {
2714         int extra_pins;
2715
2716         /* Additional pins from page cache */
2717         if (PageAnon(page))
2718                 extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0;
2719         else
2720                 extra_pins = HPAGE_PMD_NR;
2721         if (pextra_pins)
2722                 *pextra_pins = extra_pins;
2723         return total_mapcount(page) == page_count(page) - extra_pins - 1;
2724 }
2725
2726 /*
2727  * This function splits huge page into normal pages. @page can point to any
2728  * subpage of huge page to split. Split doesn't change the position of @page.
2729  *
2730  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2731  * The huge page must be locked.
2732  *
2733  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2734  *
2735  * Both head page and tail pages will inherit mapping, flags, and so on from
2736  * the hugepage.
2737  *
2738  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2739  * they are not mapped.
2740  *
2741  * Returns 0 if the hugepage is split successfully.
2742  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2743  * us.
2744  */
2745 int split_huge_page_to_list(struct page *page, struct list_head *list)
2746 {
2747         struct page *head = compound_head(page);
2748         struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2749         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2750         struct anon_vma *anon_vma = NULL;
2751         struct address_space *mapping = NULL;
2752         int extra_pins, ret;
2753         bool mlocked;
2754         unsigned long flags;
2755         pgoff_t end;
2756
2757         VM_BUG_ON_PAGE(is_huge_zero_page(head), head);
2758         VM_BUG_ON_PAGE(!PageLocked(page), page);
2759         VM_BUG_ON_PAGE(!PageCompound(page), page);
2760
2761         if (PageWriteback(page))
2762                 return -EBUSY;
2763
2764         if (PageAnon(head)) {
2765                 /*
2766                  * The caller does not necessarily hold an mmap_sem that would
2767                  * prevent the anon_vma disappearing so we first we take a
2768                  * reference to it and then lock the anon_vma for write. This
2769                  * is similar to page_lock_anon_vma_read except the write lock
2770                  * is taken to serialise against parallel split or collapse
2771                  * operations.
2772                  */
2773                 anon_vma = page_get_anon_vma(head);
2774                 if (!anon_vma) {
2775                         ret = -EBUSY;
2776                         goto out;
2777                 }
2778                 end = -1;
2779                 mapping = NULL;
2780                 anon_vma_lock_write(anon_vma);
2781         } else {
2782                 mapping = head->mapping;
2783
2784                 /* Truncated ? */
2785                 if (!mapping) {
2786                         ret = -EBUSY;
2787                         goto out;
2788                 }
2789
2790                 anon_vma = NULL;
2791                 i_mmap_lock_read(mapping);
2792
2793                 /*
2794                  *__split_huge_page() may need to trim off pages beyond EOF:
2795                  * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2796                  * which cannot be nested inside the page tree lock. So note
2797                  * end now: i_size itself may be changed at any moment, but
2798                  * head page lock is good enough to serialize the trimming.
2799                  */
2800                 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2801         }
2802
2803         /*
2804          * Racy check if we can split the page, before unmap_page() will
2805          * split PMDs
2806          */
2807         if (!can_split_huge_page(head, &extra_pins)) {
2808                 ret = -EBUSY;
2809                 goto out_unlock;
2810         }
2811
2812         mlocked = PageMlocked(page);
2813         unmap_page(head);
2814
2815         /* Make sure the page is not on per-CPU pagevec as it takes pin */
2816         if (mlocked)
2817                 lru_add_drain();
2818
2819         /* prevent PageLRU to go away from under us, and freeze lru stats */
2820         spin_lock_irqsave(&pgdata->lru_lock, flags);
2821
2822         if (mapping) {
2823                 XA_STATE(xas, &mapping->i_pages, page_index(head));
2824
2825                 /*
2826                  * Check if the head page is present in page cache.
2827                  * We assume all tail are present too, if head is there.
2828                  */
2829                 xa_lock(&mapping->i_pages);
2830                 if (xas_load(&xas) != head)
2831                         goto fail;
2832         }
2833
2834         /* Prevent deferred_split_scan() touching ->_refcount */
2835         spin_lock(&ds_queue->split_queue_lock);
2836         if (page_ref_freeze(head, 1 + extra_pins)) {
2837                 if (!list_empty(page_deferred_list(head))) {
2838                         ds_queue->split_queue_len--;
2839                         list_del(page_deferred_list(head));
2840                 }
2841                 if (mapping) {
2842                         if (PageSwapBacked(page))
2843                                 __dec_node_page_state(page, NR_SHMEM_THPS);
2844                         else
2845                                 __dec_node_page_state(page, NR_FILE_THPS);
2846                 }
2847
2848                 spin_unlock(&ds_queue->split_queue_lock);
2849                 __split_huge_page(page, list, end, flags);
2850                 if (PageSwapCache(head)) {
2851                         swp_entry_t entry = { .val = page_private(head) };
2852
2853                         ret = split_swap_cluster(entry);
2854                 } else
2855                         ret = 0;
2856         } else {
2857                 spin_unlock(&ds_queue->split_queue_lock);
2858 fail:
2859                 if (mapping)
2860                         xa_unlock(&mapping->i_pages);
2861                 spin_unlock_irqrestore(&pgdata->lru_lock, flags);
2862                 remap_page(head);
2863                 ret = -EBUSY;
2864         }
2865
2866 out_unlock:
2867         if (anon_vma) {
2868                 anon_vma_unlock_write(anon_vma);
2869                 put_anon_vma(anon_vma);
2870         }
2871         if (mapping)
2872                 i_mmap_unlock_read(mapping);
2873 out:
2874         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2875         return ret;
2876 }
2877
2878 void free_transhuge_page(struct page *page)
2879 {
2880         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2881         unsigned long flags;
2882
2883         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2884         if (!list_empty(page_deferred_list(page))) {
2885                 ds_queue->split_queue_len--;
2886                 list_del(page_deferred_list(page));
2887         }
2888         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2889         free_compound_page(page);
2890 }
2891
2892 void deferred_split_huge_page(struct page *page)
2893 {
2894         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2895 #ifdef CONFIG_MEMCG
2896         struct mem_cgroup *memcg = compound_head(page)->mem_cgroup;
2897 #endif
2898         unsigned long flags;
2899
2900         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2901
2902         /*
2903          * The try_to_unmap() in page reclaim path might reach here too,
2904          * this may cause a race condition to corrupt deferred split queue.
2905          * And, if page reclaim is already handling the same page, it is
2906          * unnecessary to handle it again in shrinker.
2907          *
2908          * Check PageSwapCache to determine if the page is being
2909          * handled by page reclaim since THP swap would add the page into
2910          * swap cache before calling try_to_unmap().
2911          */
2912         if (PageSwapCache(page))
2913                 return;
2914
2915         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2916         if (list_empty(page_deferred_list(page))) {
2917                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2918                 list_add_tail(page_deferred_list(page), &ds_queue->split_queue);
2919                 ds_queue->split_queue_len++;
2920 #ifdef CONFIG_MEMCG
2921                 if (memcg)
2922                         memcg_set_shrinker_bit(memcg, page_to_nid(page),
2923                                                deferred_split_shrinker.id);
2924 #endif
2925         }
2926         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2927 }
2928
2929 static unsigned long deferred_split_count(struct shrinker *shrink,
2930                 struct shrink_control *sc)
2931 {
2932         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2933         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2934
2935 #ifdef CONFIG_MEMCG
2936         if (sc->memcg)
2937                 ds_queue = &sc->memcg->deferred_split_queue;
2938 #endif
2939         return READ_ONCE(ds_queue->split_queue_len);
2940 }
2941
2942 static unsigned long deferred_split_scan(struct shrinker *shrink,
2943                 struct shrink_control *sc)
2944 {
2945         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2946         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2947         unsigned long flags;
2948         LIST_HEAD(list), *pos, *next;
2949         struct page *page;
2950         int split = 0;
2951
2952 #ifdef CONFIG_MEMCG
2953         if (sc->memcg)
2954                 ds_queue = &sc->memcg->deferred_split_queue;
2955 #endif
2956
2957         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2958         /* Take pin on all head pages to avoid freeing them under us */
2959         list_for_each_safe(pos, next, &ds_queue->split_queue) {
2960                 page = list_entry((void *)pos, struct page, mapping);
2961                 page = compound_head(page);
2962                 if (get_page_unless_zero(page)) {
2963                         list_move(page_deferred_list(page), &list);
2964                 } else {
2965                         /* We lost race with put_compound_page() */
2966                         list_del_init(page_deferred_list(page));
2967                         ds_queue->split_queue_len--;
2968                 }
2969                 if (!--sc->nr_to_scan)
2970                         break;
2971         }
2972         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2973
2974         list_for_each_safe(pos, next, &list) {
2975                 page = list_entry((void *)pos, struct page, mapping);
2976                 if (!trylock_page(page))
2977                         goto next;
2978                 /* split_huge_page() removes page from list on success */
2979                 if (!split_huge_page(page))
2980                         split++;
2981                 unlock_page(page);
2982 next:
2983                 put_page(page);
2984         }
2985
2986         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2987         list_splice_tail(&list, &ds_queue->split_queue);
2988         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2989
2990         /*
2991          * Stop shrinker if we didn't split any page, but the queue is empty.
2992          * This can happen if pages were freed under us.
2993          */
2994         if (!split && list_empty(&ds_queue->split_queue))
2995                 return SHRINK_STOP;
2996         return split;
2997 }
2998
2999 static struct shrinker deferred_split_shrinker = {
3000         .count_objects = deferred_split_count,
3001         .scan_objects = deferred_split_scan,
3002         .seeks = DEFAULT_SEEKS,
3003         .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
3004                  SHRINKER_NONSLAB,
3005 };
3006
3007 #ifdef CONFIG_DEBUG_FS
3008 static int split_huge_pages_set(void *data, u64 val)
3009 {
3010         struct zone *zone;
3011         struct page *page;
3012         unsigned long pfn, max_zone_pfn;
3013         unsigned long total = 0, split = 0;
3014
3015         if (val != 1)
3016                 return -EINVAL;
3017
3018         for_each_populated_zone(zone) {
3019                 max_zone_pfn = zone_end_pfn(zone);
3020                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
3021                         if (!pfn_valid(pfn))
3022                                 continue;
3023
3024                         page = pfn_to_page(pfn);
3025                         if (!get_page_unless_zero(page))
3026                                 continue;
3027
3028                         if (zone != page_zone(page))
3029                                 goto next;
3030
3031                         if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
3032                                 goto next;
3033
3034                         total++;
3035                         lock_page(page);
3036                         if (!split_huge_page(page))
3037                                 split++;
3038                         unlock_page(page);
3039 next:
3040                         put_page(page);
3041                 }
3042         }
3043
3044         pr_info("%lu of %lu THP split\n", split, total);
3045
3046         return 0;
3047 }
3048 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
3049                 "%llu\n");
3050
3051 static int __init split_huge_pages_debugfs(void)
3052 {
3053         debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3054                             &split_huge_pages_fops);
3055         return 0;
3056 }
3057 late_initcall(split_huge_pages_debugfs);
3058 #endif
3059
3060 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
3061 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
3062                 struct page *page)
3063 {
3064         struct vm_area_struct *vma = pvmw->vma;
3065         struct mm_struct *mm = vma->vm_mm;
3066         unsigned long address = pvmw->address;
3067         pmd_t pmdval;
3068         swp_entry_t entry;
3069         pmd_t pmdswp;
3070
3071         if (!(pvmw->pmd && !pvmw->pte))
3072                 return;
3073
3074         flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
3075         pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
3076         if (pmd_dirty(pmdval))
3077                 set_page_dirty(page);
3078         entry = make_migration_entry(page, pmd_write(pmdval));
3079         pmdswp = swp_entry_to_pmd(entry);
3080         if (pmd_soft_dirty(pmdval))
3081                 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
3082         set_pmd_at(mm, address, pvmw->pmd, pmdswp);
3083         page_remove_rmap(page, true);
3084         put_page(page);
3085 }
3086
3087 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
3088 {
3089         struct vm_area_struct *vma = pvmw->vma;
3090         struct mm_struct *mm = vma->vm_mm;
3091         unsigned long address = pvmw->address;
3092         unsigned long mmun_start = address & HPAGE_PMD_MASK;
3093         pmd_t pmde;
3094         swp_entry_t entry;
3095
3096         if (!(pvmw->pmd && !pvmw->pte))
3097                 return;
3098
3099         entry = pmd_to_swp_entry(*pvmw->pmd);
3100         get_page(new);
3101         pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
3102         if (pmd_swp_soft_dirty(*pvmw->pmd))
3103                 pmde = pmd_mksoft_dirty(pmde);
3104         if (is_write_migration_entry(entry))
3105                 pmde = maybe_pmd_mkwrite(pmde, vma);
3106
3107         flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
3108         if (PageAnon(new))
3109                 page_add_anon_rmap(new, vma, mmun_start, true);
3110         else
3111                 page_add_file_rmap(new, true);
3112         set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
3113         if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new))
3114                 mlock_vma_page(new);
3115         update_mmu_cache_pmd(vma, address, pvmw->pmd);
3116 }
3117 #endif