mm/huge_memory.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *  Copyright (C) 2009  Red Hat, Inc.
   4  */
   5
   6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   7
   8 #include <linux/mm.h>
   9 #include <linux/sched.h>
  10 #include <linux/sched/coredump.h>
  11 #include <linux/sched/numa_balancing.h>
  12 #include <linux/highmem.h>
  13 #include <linux/hugetlb.h>
  14 #include <linux/mmu_notifier.h>
  15 #include <linux/rmap.h>
  16 #include <linux/swap.h>
  17 #include <linux/shrinker.h>
  18 #include <linux/mm_inline.h>
  19 #include <linux/swapops.h>
  20 #include <linux/dax.h>
  21 #include <linux/khugepaged.h>
  22 #include <linux/freezer.h>
  23 #include <linux/pfn_t.h>
  24 #include <linux/mman.h>
  25 #include <linux/memremap.h>
  26 #include <linux/pagemap.h>
  27 #include <linux/debugfs.h>
  28 #include <linux/migrate.h>
  29 #include <linux/hashtable.h>
  30 #include <linux/userfaultfd_k.h>
  31 #include <linux/page_idle.h>
  32 #include <linux/shmem_fs.h>
  33 #include <linux/oom.h>
  34 #include <linux/numa.h>
  35 #include <linux/page_owner.h>
  36
  37 #include <asm/tlb.h>
  38 #include <asm/pgalloc.h>
  39 #include "internal.h"
  40
  41 /*
  42  * By default, transparent hugepage support is disabled in order to avoid
  43  * risking an increased memory footprint for applications that are not
  44  * guaranteed to benefit from it. When transparent hugepage support is
  45  * enabled, it is for all mappings, and khugepaged scans all mappings.
  46  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  47  * for all hugepage allocations.
  48  */
  49 unsigned long transparent_hugepage_flags __read_mostly =
  50 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  51         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  52 #endif
  53 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  54         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  55 #endif
  56         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
  57         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  58         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  59
  60 static struct shrinker deferred_split_shrinker;
  61
  62 static atomic_t huge_zero_refcount;
  63 struct page *huge_zero_page __read_mostly;
  64
  65 bool transparent_hugepage_enabled(struct vm_area_struct *vma)
  66 {
  67         /* The addr is used to check if the vma size fits */
  68         unsigned long addr = (vma->vm_end & HPAGE_PMD_MASK) - HPAGE_PMD_SIZE;
  69
  70         if (!transhuge_vma_suitable(vma, addr))
  71                 return false;
  72         if (vma_is_anonymous(vma))
  73                 return __transparent_hugepage_enabled(vma);
  74         if (vma_is_shmem(vma))
  75                 return shmem_huge_enabled(vma);
  76
  77         return false;
  78 }
  79
  80 static struct page *get_huge_zero_page(void)
  81 {
  82         struct page *zero_page;
  83 retry:
  84         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
  85                 return READ_ONCE(huge_zero_page);
  86
  87         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
  88                         HPAGE_PMD_ORDER);
  89         if (!zero_page) {
  90                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
  91                 return NULL;
  92         }
  93         count_vm_event(THP_ZERO_PAGE_ALLOC);
  94         preempt_disable();
  95         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
  96                 preempt_enable();
  97                 __free_pages(zero_page, compound_order(zero_page));
  98                 goto retry;
  99         }
 100
 101         /* We take additional reference here. It will be put back by shrinker */
 102         atomic_set(&huge_zero_refcount, 2);
 103         preempt_enable();
 104         return READ_ONCE(huge_zero_page);
 105 }
 106
 107 static void put_huge_zero_page(void)
 108 {
 109         /*
 110          * Counter should never go to zero here. Only shrinker can put
 111          * last reference.
 112          */
 113         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 114 }
 115
 116 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
 117 {
 118         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 119                 return READ_ONCE(huge_zero_page);
 120
 121         if (!get_huge_zero_page())
 122                 return NULL;
 123
 124         if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 125                 put_huge_zero_page();
 126
 127         return READ_ONCE(huge_zero_page);
 128 }
 129
 130 void mm_put_huge_zero_page(struct mm_struct *mm)
 131 {
 132         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 133                 put_huge_zero_page();
 134 }
 135
 136 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 137                                         struct shrink_control *sc)
 138 {
 139         /* we can free zero page only if last reference remains */
 140         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 141 }
 142
 143 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 144                                        struct shrink_control *sc)
 145 {
 146         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 147                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 148                 BUG_ON(zero_page == NULL);
 149                 __free_pages(zero_page, compound_order(zero_page));
 150                 return HPAGE_PMD_NR;
 151         }
 152
 153         return 0;
 154 }
 155
 156 static struct shrinker huge_zero_page_shrinker = {
 157         .count_objects = shrink_huge_zero_page_count,
 158         .scan_objects = shrink_huge_zero_page_scan,
 159         .seeks = DEFAULT_SEEKS,
 160 };
 161
 162 #ifdef CONFIG_SYSFS
 163 static ssize_t enabled_show(struct kobject *kobj,
 164                             struct kobj_attribute *attr, char *buf)
 165 {
 166         if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
 167                 return sprintf(buf, "[always] madvise never\n");
 168         else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
 169                 return sprintf(buf, "always [madvise] never\n");
 170         else
 171                 return sprintf(buf, "always madvise [never]\n");
 172 }
 173
 174 static ssize_t enabled_store(struct kobject *kobj,
 175                              struct kobj_attribute *attr,
 176                              const char *buf, size_t count)
 177 {
 178         ssize_t ret = count;
 179
 180         if (!memcmp("always", buf,
 181                     min(sizeof("always")-1, count))) {
 182                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 183                 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 184         } else if (!memcmp("madvise", buf,
 185                            min(sizeof("madvise")-1, count))) {
 186                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 187                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 188         } else if (!memcmp("never", buf,
 189                            min(sizeof("never")-1, count))) {
 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 (!memcmp("always", buf,
 254                     min(sizeof("always")-1, count))) {
 255                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 256                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 257                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 258                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 259         } else if (!memcmp("defer+madvise", buf,
 260                     min(sizeof("defer+madvise")-1, count))) {
 261                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 262                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 263                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 264                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 265         } else if (!memcmp("defer", buf,
 266                     min(sizeof("defer")-1, count))) {
 267                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 268                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 269                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 270                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 271         } else if (!memcmp("madvise", buf,
 272                            min(sizeof("madvise")-1, count))) {
 273                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 274                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 275                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 276                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 277         } else if (!memcmp("never", buf,
 278                            min(sizeof("never")-1, count))) {
 279                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 280                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 281                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 282                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 283         } else
 284                 return -EINVAL;
 285
 286         return count;
 287 }
 288 static struct kobj_attribute defrag_attr =
 289         __ATTR(defrag, 0644, defrag_show, defrag_store);
 290
 291 static ssize_t use_zero_page_show(struct kobject *kobj,
 292                 struct kobj_attribute *attr, char *buf)
 293 {
 294         return single_hugepage_flag_show(kobj, attr, buf,
 295                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 296 }
 297 static ssize_t use_zero_page_store(struct kobject *kobj,
 298                 struct kobj_attribute *attr, const char *buf, size_t count)
 299 {
 300         return single_hugepage_flag_store(kobj, attr, buf, count,
 301                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 302 }
 303 static struct kobj_attribute use_zero_page_attr =
 304         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 305
 306 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
 307                 struct kobj_attribute *attr, char *buf)
 308 {
 309         return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE);
 310 }
 311 static struct kobj_attribute hpage_pmd_size_attr =
 312         __ATTR_RO(hpage_pmd_size);
 313
 314 #ifdef CONFIG_DEBUG_VM
 315 static ssize_t debug_cow_show(struct kobject *kobj,
 316                                 struct kobj_attribute *attr, char *buf)
 317 {
 318         return single_hugepage_flag_show(kobj, attr, buf,
 319                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 320 }
 321 static ssize_t debug_cow_store(struct kobject *kobj,
 322                                struct kobj_attribute *attr,
 323                                const char *buf, size_t count)
 324 {
 325         return single_hugepage_flag_store(kobj, attr, buf, count,
 326                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 327 }
 328 static struct kobj_attribute debug_cow_attr =
 329         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 330 #endif /* CONFIG_DEBUG_VM */
 331
 332 static struct attribute *hugepage_attr[] = {
 333         &enabled_attr.attr,
 334         &defrag_attr.attr,
 335         &use_zero_page_attr.attr,
 336         &hpage_pmd_size_attr.attr,
 337 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
 338         &shmem_enabled_attr.attr,
 339 #endif
 340 #ifdef CONFIG_DEBUG_VM
 341         &debug_cow_attr.attr,
 342 #endif
 343         NULL,
 344 };
 345
 346 static const struct attribute_group hugepage_attr_group = {
 347         .attrs = hugepage_attr,
 348 };
 349
 350 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 351 {
 352         int err;
 353
 354         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 355         if (unlikely(!*hugepage_kobj)) {
 356                 pr_err("failed to create transparent hugepage kobject\n");
 357                 return -ENOMEM;
 358         }
 359
 360         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 361         if (err) {
 362                 pr_err("failed to register transparent hugepage group\n");
 363                 goto delete_obj;
 364         }
 365
 366         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 367         if (err) {
 368                 pr_err("failed to register transparent hugepage group\n");
 369                 goto remove_hp_group;
 370         }
 371
 372         return 0;
 373
 374 remove_hp_group:
 375         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 376 delete_obj:
 377         kobject_put(*hugepage_kobj);
 378         return err;
 379 }
 380
 381 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 382 {
 383         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 384         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 385         kobject_put(hugepage_kobj);
 386 }
 387 #else
 388 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 389 {
 390         return 0;
 391 }
 392
 393 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 394 {
 395 }
 396 #endif /* CONFIG_SYSFS */
 397
 398 static int __init hugepage_init(void)
 399 {
 400         int err;
 401         struct kobject *hugepage_kobj;
 402
 403         if (!has_transparent_hugepage()) {
 404                 transparent_hugepage_flags = 0;
 405                 return -EINVAL;
 406         }
 407
 408         /*
 409          * hugepages can't be allocated by the buddy allocator
 410          */
 411         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
 412         /*
 413          * we use page->mapping and page->index in second tail page
 414          * as list_head: assuming THP order >= 2
 415          */
 416         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 417
 418         err = hugepage_init_sysfs(&hugepage_kobj);
 419         if (err)
 420                 goto err_sysfs;
 421
 422         err = khugepaged_init();
 423         if (err)
 424                 goto err_slab;
 425
 426         err = register_shrinker(&huge_zero_page_shrinker);
 427         if (err)
 428                 goto err_hzp_shrinker;
 429         err = register_shrinker(&deferred_split_shrinker);
 430         if (err)
 431                 goto err_split_shrinker;
 432
 433         /*
 434          * By default disable transparent hugepages on smaller systems,
 435          * where the extra memory used could hurt more than TLB overhead
 436          * is likely to save.  The admin can still enable it through /sys.
 437          */
 438         if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
 439                 transparent_hugepage_flags = 0;
 440                 return 0;
 441         }
 442
 443         err = start_stop_khugepaged();
 444         if (err)
 445                 goto err_khugepaged;
 446
 447         return 0;
 448 err_khugepaged:
 449         unregister_shrinker(&deferred_split_shrinker);
 450 err_split_shrinker:
 451         unregister_shrinker(&huge_zero_page_shrinker);
 452 err_hzp_shrinker:
 453         khugepaged_destroy();
 454 err_slab:
 455         hugepage_exit_sysfs(hugepage_kobj);
 456 err_sysfs:
 457         return err;
 458 }
 459 subsys_initcall(hugepage_init);
 460
 461 static int __init setup_transparent_hugepage(char *str)
 462 {
 463         int ret = 0;
 464         if (!str)
 465                 goto out;
 466         if (!strcmp(str, "always")) {
 467                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 468                         &transparent_hugepage_flags);
 469                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 470                           &transparent_hugepage_flags);
 471                 ret = 1;
 472         } else if (!strcmp(str, "madvise")) {
 473                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 474                           &transparent_hugepage_flags);
 475                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 476                         &transparent_hugepage_flags);
 477                 ret = 1;
 478         } else if (!strcmp(str, "never")) {
 479                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 480                           &transparent_hugepage_flags);
 481                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 482                           &transparent_hugepage_flags);
 483                 ret = 1;
 484         }
 485 out:
 486         if (!ret)
 487                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 488         return ret;
 489 }
 490 __setup("transparent_hugepage=", setup_transparent_hugepage);
 491
 492 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 493 {
 494         if (likely(vma->vm_flags & VM_WRITE))
 495                 pmd = pmd_mkwrite(pmd);
 496         return pmd;
 497 }
 498
 499 #ifdef CONFIG_MEMCG
 500 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
 501 {
 502         struct mem_cgroup *memcg = compound_head(page)->mem_cgroup;
 503         struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
 504
 505         if (memcg)
 506                 return &memcg->deferred_split_queue;
 507         else
 508                 return &pgdat->deferred_split_queue;
 509 }
 510 #else
 511 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
 512 {
 513         struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
 514
 515         return &pgdat->deferred_split_queue;
 516 }
 517 #endif
 518
 519 void prep_transhuge_page(struct page *page)
 520 {
 521         /*
 522          * we use page->mapping and page->indexlru in second tail page
 523          * as list_head: assuming THP order >= 2
 524          */
 525
 526         INIT_LIST_HEAD(page_deferred_list(page));
 527         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 528 }
 529
 530 bool is_transparent_hugepage(struct page *page)
 531 {
 532         if (!PageCompound(page))
 533                 return 0;
 534
 535         page = compound_head(page);
 536         return is_huge_zero_page(page) ||
 537                page[1].compound_dtor == TRANSHUGE_PAGE_DTOR;
 538 }
 539 EXPORT_SYMBOL_GPL(is_transparent_hugepage);
 540
 541 static unsigned long __thp_get_unmapped_area(struct file *filp,
 542                 unsigned long addr, unsigned long len,
 543                 loff_t off, unsigned long flags, unsigned long size)
 544 {
 545         loff_t off_end = off + len;
 546         loff_t off_align = round_up(off, size);
 547         unsigned long len_pad, ret;
 548
 549         if (off_end <= off_align || (off_end - off_align) < size)
 550                 return 0;
 551
 552         len_pad = len + size;
 553         if (len_pad < len || (off + len_pad) < off)
 554                 return 0;
 555
 556         ret = current->mm->get_unmapped_area(filp, addr, len_pad,
 557                                               off >> PAGE_SHIFT, flags);
 558
 559         /*
 560          * The failure might be due to length padding. The caller will retry
 561          * without the padding.
 562          */
 563         if (IS_ERR_VALUE(ret))
 564                 return 0;
 565
 566         /*
 567          * Do not try to align to THP boundary if allocation at the address
 568          * hint succeeds.
 569          */
 570         if (ret == addr)
 571                 return addr;
 572
 573         ret += (off - ret) & (size - 1);
 574         return ret;
 575 }
 576
 577 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 578                 unsigned long len, unsigned long pgoff, unsigned long flags)
 579 {
 580         unsigned long ret;
 581         loff_t off = (loff_t)pgoff << PAGE_SHIFT;
 582
 583         if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
 584                 goto out;
 585
 586         ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE);
 587         if (ret)
 588                 return ret;
 589 out:
 590         return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
 591 }
 592 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
 593
 594 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
 595                         struct page *page, gfp_t gfp)
 596 {
 597         struct vm_area_struct *vma = vmf->vma;
 598         struct mem_cgroup *memcg;
 599         pgtable_t pgtable;
 600         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 601         vm_fault_t ret = 0;
 602
 603         VM_BUG_ON_PAGE(!PageCompound(page), page);
 604
 605         if (mem_cgroup_try_charge_delay(page, vma->vm_mm, gfp, &memcg, true)) {
 606                 put_page(page);
 607                 count_vm_event(THP_FAULT_FALLBACK);
 608                 return VM_FAULT_FALLBACK;
 609         }
 610
 611         pgtable = pte_alloc_one(vma->vm_mm);
 612         if (unlikely(!pgtable)) {
 613                 ret = VM_FAULT_OOM;
 614                 goto release;
 615         }
 616
 617         clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
 618         /*
 619          * The memory barrier inside __SetPageUptodate makes sure that
 620          * clear_huge_page writes become visible before the set_pmd_at()
 621          * write.
 622          */
 623         __SetPageUptodate(page);
 624
 625         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 626         if (unlikely(!pmd_none(*vmf->pmd))) {
 627                 goto unlock_release;
 628         } else {
 629                 pmd_t entry;
 630
 631                 ret = check_stable_address_space(vma->vm_mm);
 632                 if (ret)
 633                         goto unlock_release;
 634
 635                 /* Deliver the page fault to userland */
 636                 if (userfaultfd_missing(vma)) {
 637                         vm_fault_t ret2;
 638
 639                         spin_unlock(vmf->ptl);
 640                         mem_cgroup_cancel_charge(page, memcg, true);
 641                         put_page(page);
 642                         pte_free(vma->vm_mm, pgtable);
 643                         ret2 = handle_userfault(vmf, VM_UFFD_MISSING);
 644                         VM_BUG_ON(ret2 & VM_FAULT_FALLBACK);
 645                         return ret2;
 646                 }
 647
 648                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 649                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 650                 page_add_new_anon_rmap(page, vma, haddr, true);
 651                 mem_cgroup_commit_charge(page, memcg, false, true);
 652                 lru_cache_add_active_or_unevictable(page, vma);
 653                 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
 654                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
 655                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 656                 mm_inc_nr_ptes(vma->vm_mm);
 657                 spin_unlock(vmf->ptl);
 658                 count_vm_event(THP_FAULT_ALLOC);
 659                 count_memcg_events(memcg, THP_FAULT_ALLOC, 1);
 660         }
 661
 662         return 0;
 663 unlock_release:
 664         spin_unlock(vmf->ptl);
 665 release:
 666         if (pgtable)
 667                 pte_free(vma->vm_mm, pgtable);
 668         mem_cgroup_cancel_charge(page, memcg, true);
 669         put_page(page);
 670         return ret;
 671
 672 }
 673
 674 /*
 675  * always: directly stall for all thp allocations
 676  * defer: wake kswapd and fail if not immediately available
 677  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
 678  *                fail if not immediately available
 679  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
 680  *          available
 681  * never: never stall for any thp allocation
 682  */
 683 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
 684 {
 685         const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
 686
 687         /* Always do synchronous compaction */
 688         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 689                 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
 690
 691         /* Kick kcompactd and fail quickly */
 692         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 693                 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
 694
 695         /* Synchronous compaction if madvised, otherwise kick kcompactd */
 696         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 697                 return GFP_TRANSHUGE_LIGHT |
 698                         (vma_madvised ? __GFP_DIRECT_RECLAIM :
 699                                         __GFP_KSWAPD_RECLAIM);
 700
 701         /* Only do synchronous compaction if madvised */
 702         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 703                 return GFP_TRANSHUGE_LIGHT |
 704                        (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
 705
 706         return GFP_TRANSHUGE_LIGHT;
 707 }
 708
 709 /* Caller must hold page table lock. */
 710 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 711                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 712                 struct page *zero_page)
 713 {
 714         pmd_t entry;
 715         if (!pmd_none(*pmd))
 716                 return false;
 717         entry = mk_pmd(zero_page, vma->vm_page_prot);
 718         entry = pmd_mkhuge(entry);
 719         if (pgtable)
 720                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 721         set_pmd_at(mm, haddr, pmd, entry);
 722         mm_inc_nr_ptes(mm);
 723         return true;
 724 }
 725
 726 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
 727 {
 728         struct vm_area_struct *vma = vmf->vma;
 729         gfp_t gfp;
 730         struct page *page;
 731         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 732
 733         if (!transhuge_vma_suitable(vma, haddr))
 734                 return VM_FAULT_FALLBACK;
 735         if (unlikely(anon_vma_prepare(vma)))
 736                 return VM_FAULT_OOM;
 737         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 738                 return VM_FAULT_OOM;
 739         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
 740                         !mm_forbids_zeropage(vma->vm_mm) &&
 741                         transparent_hugepage_use_zero_page()) {
 742                 pgtable_t pgtable;
 743                 struct page *zero_page;
 744                 bool set;
 745                 vm_fault_t ret;
 746                 pgtable = pte_alloc_one(vma->vm_mm);
 747                 if (unlikely(!pgtable))
 748                         return VM_FAULT_OOM;
 749                 zero_page = mm_get_huge_zero_page(vma->vm_mm);
 750                 if (unlikely(!zero_page)) {
 751                         pte_free(vma->vm_mm, pgtable);
 752                         count_vm_event(THP_FAULT_FALLBACK);
 753                         return VM_FAULT_FALLBACK;
 754                 }
 755                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 756                 ret = 0;
 757                 set = false;
 758                 if (pmd_none(*vmf->pmd)) {
 759                         ret = check_stable_address_space(vma->vm_mm);
 760                         if (ret) {
 761                                 spin_unlock(vmf->ptl);
 762                         } else if (userfaultfd_missing(vma)) {
 763                                 spin_unlock(vmf->ptl);
 764                                 ret = handle_userfault(vmf, VM_UFFD_MISSING);
 765                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 766                         } else {
 767                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
 768                                                    haddr, vmf->pmd, zero_page);
 769                                 spin_unlock(vmf->ptl);
 770                                 set = true;
 771                         }
 772                 } else
 773                         spin_unlock(vmf->ptl);
 774                 if (!set)
 775                         pte_free(vma->vm_mm, pgtable);
 776                 return ret;
 777         }
 778         gfp = alloc_hugepage_direct_gfpmask(vma);
 779         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 780         if (unlikely(!page)) {
 781                 count_vm_event(THP_FAULT_FALLBACK);
 782                 return VM_FAULT_FALLBACK;
 783         }
 784         prep_transhuge_page(page);
 785         return __do_huge_pmd_anonymous_page(vmf, page, gfp);
 786 }
 787
 788 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 789                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
 790                 pgtable_t pgtable)
 791 {
 792         struct mm_struct *mm = vma->vm_mm;
 793         pmd_t entry;
 794         spinlock_t *ptl;
 795
 796         ptl = pmd_lock(mm, pmd);
 797         if (!pmd_none(*pmd)) {
 798                 if (write) {
 799                         if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
 800                                 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
 801                                 goto out_unlock;
 802                         }
 803                         entry = pmd_mkyoung(*pmd);
 804                         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 805                         if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
 806                                 update_mmu_cache_pmd(vma, addr, pmd);
 807                 }
 808
 809                 goto out_unlock;
 810         }
 811
 812         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 813         if (pfn_t_devmap(pfn))
 814                 entry = pmd_mkdevmap(entry);
 815         if (write) {
 816                 entry = pmd_mkyoung(pmd_mkdirty(entry));
 817                 entry = maybe_pmd_mkwrite(entry, vma);
 818         }
 819
 820         if (pgtable) {
 821                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 822                 mm_inc_nr_ptes(mm);
 823                 pgtable = NULL;
 824         }
 825
 826         set_pmd_at(mm, addr, pmd, entry);
 827         update_mmu_cache_pmd(vma, addr, pmd);
 828
 829 out_unlock:
 830         spin_unlock(ptl);
 831         if (pgtable)
 832                 pte_free(mm, pgtable);
 833 }
 834
 835 vm_fault_t vmf_insert_pfn_pmd(struct vm_fault *vmf, pfn_t pfn, bool write)
 836 {
 837         unsigned long addr = vmf->address & PMD_MASK;
 838         struct vm_area_struct *vma = vmf->vma;
 839         pgprot_t pgprot = vma->vm_page_prot;
 840         pgtable_t pgtable = NULL;
 841
 842         /*
 843          * If we had pmd_special, we could avoid all these restrictions,
 844          * but we need to be consistent with PTEs and architectures that
 845          * can't support a 'special' bit.
 846          */
 847         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 848                         !pfn_t_devmap(pfn));
 849         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 850                                                 (VM_PFNMAP|VM_MIXEDMAP));
 851         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 852
 853         if (addr < vma->vm_start || addr >= vma->vm_end)
 854                 return VM_FAULT_SIGBUS;
 855
 856         if (arch_needs_pgtable_deposit()) {
 857                 pgtable = pte_alloc_one(vma->vm_mm);
 858                 if (!pgtable)
 859                         return VM_FAULT_OOM;
 860         }
 861
 862         track_pfn_insert(vma, &pgprot, pfn);
 863
 864         insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
 865         return VM_FAULT_NOPAGE;
 866 }
 867 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
 868
 869 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 870 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
 871 {
 872         if (likely(vma->vm_flags & VM_WRITE))
 873                 pud = pud_mkwrite(pud);
 874         return pud;
 875 }
 876
 877 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 878                 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
 879 {
 880         struct mm_struct *mm = vma->vm_mm;
 881         pud_t entry;
 882         spinlock_t *ptl;
 883
 884         ptl = pud_lock(mm, pud);
 885         if (!pud_none(*pud)) {
 886                 if (write) {
 887                         if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
 888                                 WARN_ON_ONCE(!is_huge_zero_pud(*pud));
 889                                 goto out_unlock;
 890                         }
 891                         entry = pud_mkyoung(*pud);
 892                         entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
 893                         if (pudp_set_access_flags(vma, addr, pud, entry, 1))
 894                                 update_mmu_cache_pud(vma, addr, pud);
 895                 }
 896                 goto out_unlock;
 897         }
 898
 899         entry = pud_mkhuge(pfn_t_pud(pfn, prot));
 900         if (pfn_t_devmap(pfn))
 901                 entry = pud_mkdevmap(entry);
 902         if (write) {
 903                 entry = pud_mkyoung(pud_mkdirty(entry));
 904                 entry = maybe_pud_mkwrite(entry, vma);
 905         }
 906         set_pud_at(mm, addr, pud, entry);
 907         update_mmu_cache_pud(vma, addr, pud);
 908
 909 out_unlock:
 910         spin_unlock(ptl);
 911 }
 912
 913 vm_fault_t vmf_insert_pfn_pud(struct vm_fault *vmf, pfn_t pfn, bool write)
 914 {
 915         unsigned long addr = vmf->address & PUD_MASK;
 916         struct vm_area_struct *vma = vmf->vma;
 917         pgprot_t pgprot = vma->vm_page_prot;
 918
 919         /*
 920          * If we had pud_special, we could avoid all these restrictions,
 921          * but we need to be consistent with PTEs and architectures that
 922          * can't support a 'special' bit.
 923          */
 924         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 925                         !pfn_t_devmap(pfn));
 926         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 927                                                 (VM_PFNMAP|VM_MIXEDMAP));
 928         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 929
 930         if (addr < vma->vm_start || addr >= vma->vm_end)
 931                 return VM_FAULT_SIGBUS;
 932
 933         track_pfn_insert(vma, &pgprot, pfn);
 934
 935         insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
 936         return VM_FAULT_NOPAGE;
 937 }
 938 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
 939 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
 940
 941 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
 942                 pmd_t *pmd, int flags)
 943 {
 944         pmd_t _pmd;
 945
 946         _pmd = pmd_mkyoung(*pmd);
 947         if (flags & FOLL_WRITE)
 948                 _pmd = pmd_mkdirty(_pmd);
 949         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
 950                                 pmd, _pmd, flags & FOLL_WRITE))
 951                 update_mmu_cache_pmd(vma, addr, pmd);
 952 }
 953
 954 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 955                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
 956 {
 957         unsigned long pfn = pmd_pfn(*pmd);
 958         struct mm_struct *mm = vma->vm_mm;
 959         struct page *page;
 960
 961         assert_spin_locked(pmd_lockptr(mm, pmd));
 962
 963         /*
 964          * When we COW a devmap PMD entry, we split it into PTEs, so we should
 965          * not be in this function with `flags & FOLL_COW` set.
 966          */
 967         WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
 968
 969         if (flags & FOLL_WRITE && !pmd_write(*pmd))
 970                 return NULL;
 971
 972         if (pmd_present(*pmd) && pmd_devmap(*pmd))
 973                 /* pass */;
 974         else
 975                 return NULL;
 976
 977         if (flags & FOLL_TOUCH)
 978                 touch_pmd(vma, addr, pmd, flags);
 979
 980         /*
 981          * device mapped pages can only be returned if the
 982          * caller will manage the page reference count.
 983          */
 984         if (!(flags & FOLL_GET))
 985                 return ERR_PTR(-EEXIST);
 986
 987         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
 988         *pgmap = get_dev_pagemap(pfn, *pgmap);
 989         if (!*pgmap)
 990                 return ERR_PTR(-EFAULT);
 991         page = pfn_to_page(pfn);
 992         get_page(page);
 993
 994         return page;
 995 }
 996
 997 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 998                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 999                   struct vm_area_struct *vma)
1000 {
1001         spinlock_t *dst_ptl, *src_ptl;
1002         struct page *src_page;
1003         pmd_t pmd;
1004         pgtable_t pgtable = NULL;
1005         int ret = -ENOMEM;
1006
1007         /* Skip if can be re-fill on fault */
1008         if (!vma_is_anonymous(vma))
1009                 return 0;
1010
1011         pgtable = pte_alloc_one(dst_mm);
1012         if (unlikely(!pgtable))
1013                 goto out;
1014
1015         dst_ptl = pmd_lock(dst_mm, dst_pmd);
1016         src_ptl = pmd_lockptr(src_mm, src_pmd);
1017         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1018
1019         ret = -EAGAIN;
1020         pmd = *src_pmd;
1021
1022 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1023         if (unlikely(is_swap_pmd(pmd))) {
1024                 swp_entry_t entry = pmd_to_swp_entry(pmd);
1025
1026                 VM_BUG_ON(!is_pmd_migration_entry(pmd));
1027                 if (is_write_migration_entry(entry)) {
1028                         make_migration_entry_read(&entry);
1029                         pmd = swp_entry_to_pmd(entry);
1030                         if (pmd_swp_soft_dirty(*src_pmd))
1031                                 pmd = pmd_swp_mksoft_dirty(pmd);
1032                         set_pmd_at(src_mm, addr, src_pmd, pmd);
1033                 }
1034                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1035                 mm_inc_nr_ptes(dst_mm);
1036                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1037                 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1038                 ret = 0;
1039                 goto out_unlock;
1040         }
1041 #endif
1042
1043         if (unlikely(!pmd_trans_huge(pmd))) {
1044                 pte_free(dst_mm, pgtable);
1045                 goto out_unlock;
1046         }
1047         /*
1048          * When page table lock is held, the huge zero pmd should not be
1049          * under splitting since we don't split the page itself, only pmd to
1050          * a page table.
1051          */
1052         if (is_huge_zero_pmd(pmd)) {
1053                 struct page *zero_page;
1054                 /*
1055                  * get_huge_zero_page() will never allocate a new page here,
1056                  * since we already have a zero page to copy. It just takes a
1057                  * reference.
1058                  */
1059                 zero_page = mm_get_huge_zero_page(dst_mm);
1060                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1061                                 zero_page);
1062                 ret = 0;
1063                 goto out_unlock;
1064         }
1065
1066         src_page = pmd_page(pmd);
1067         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1068         get_page(src_page);
1069         page_dup_rmap(src_page, true);
1070         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1071         mm_inc_nr_ptes(dst_mm);
1072         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1073
1074         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1075         pmd = pmd_mkold(pmd_wrprotect(pmd));
1076         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1077
1078         ret = 0;
1079 out_unlock:
1080         spin_unlock(src_ptl);
1081         spin_unlock(dst_ptl);
1082 out:
1083         return ret;
1084 }
1085
1086 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1087 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1088                 pud_t *pud, int flags)
1089 {
1090         pud_t _pud;
1091
1092         _pud = pud_mkyoung(*pud);
1093         if (flags & FOLL_WRITE)
1094                 _pud = pud_mkdirty(_pud);
1095         if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1096                                 pud, _pud, flags & FOLL_WRITE))
1097                 update_mmu_cache_pud(vma, addr, pud);
1098 }
1099
1100 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1101                 pud_t *pud, int flags, struct dev_pagemap **pgmap)
1102 {
1103         unsigned long pfn = pud_pfn(*pud);
1104         struct mm_struct *mm = vma->vm_mm;
1105         struct page *page;
1106
1107         assert_spin_locked(pud_lockptr(mm, pud));
1108
1109         if (flags & FOLL_WRITE && !pud_write(*pud))
1110                 return NULL;
1111
1112         if (pud_present(*pud) && pud_devmap(*pud))
1113                 /* pass */;
1114         else
1115                 return NULL;
1116
1117         if (flags & FOLL_TOUCH)
1118                 touch_pud(vma, addr, pud, flags);
1119
1120         /*
1121          * device mapped pages can only be returned if the
1122          * caller will manage the page reference count.
1123          */
1124         if (!(flags & FOLL_GET))
1125                 return ERR_PTR(-EEXIST);
1126
1127         pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1128         *pgmap = get_dev_pagemap(pfn, *pgmap);
1129         if (!*pgmap)
1130                 return ERR_PTR(-EFAULT);
1131         page = pfn_to_page(pfn);
1132         get_page(page);
1133
1134         return page;
1135 }
1136
1137 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1138                   pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1139                   struct vm_area_struct *vma)
1140 {
1141         spinlock_t *dst_ptl, *src_ptl;
1142         pud_t pud;
1143         int ret;
1144
1145         dst_ptl = pud_lock(dst_mm, dst_pud);
1146         src_ptl = pud_lockptr(src_mm, src_pud);
1147         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1148
1149         ret = -EAGAIN;
1150         pud = *src_pud;
1151         if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1152                 goto out_unlock;
1153
1154         /*
1155          * When page table lock is held, the huge zero pud should not be
1156          * under splitting since we don't split the page itself, only pud to
1157          * a page table.
1158          */
1159         if (is_huge_zero_pud(pud)) {
1160                 /* No huge zero pud yet */
1161         }
1162
1163         pudp_set_wrprotect(src_mm, addr, src_pud);
1164         pud = pud_mkold(pud_wrprotect(pud));
1165         set_pud_at(dst_mm, addr, dst_pud, pud);
1166
1167         ret = 0;
1168 out_unlock:
1169         spin_unlock(src_ptl);
1170         spin_unlock(dst_ptl);
1171         return ret;
1172 }
1173
1174 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1175 {
1176         pud_t entry;
1177         unsigned long haddr;
1178         bool write = vmf->flags & FAULT_FLAG_WRITE;
1179
1180         vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1181         if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1182                 goto unlock;
1183
1184         entry = pud_mkyoung(orig_pud);
1185         if (write)
1186                 entry = pud_mkdirty(entry);
1187         haddr = vmf->address & HPAGE_PUD_MASK;
1188         if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1189                 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1190
1191 unlock:
1192         spin_unlock(vmf->ptl);
1193 }
1194 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1195
1196 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1197 {
1198         pmd_t entry;
1199         unsigned long haddr;
1200         bool write = vmf->flags & FAULT_FLAG_WRITE;
1201
1202         vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1203         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1204                 goto unlock;
1205
1206         entry = pmd_mkyoung(orig_pmd);
1207         if (write)
1208                 entry = pmd_mkdirty(entry);
1209         haddr = vmf->address & HPAGE_PMD_MASK;
1210         if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1211                 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1212
1213 unlock:
1214         spin_unlock(vmf->ptl);
1215 }
1216
1217 static vm_fault_t do_huge_pmd_wp_page_fallback(struct vm_fault *vmf,
1218                         pmd_t orig_pmd, struct page *page)
1219 {
1220         struct vm_area_struct *vma = vmf->vma;
1221         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1222         struct mem_cgroup *memcg;
1223         pgtable_t pgtable;
1224         pmd_t _pmd;
1225         int i;
1226         vm_fault_t ret = 0;
1227         struct page **pages;
1228         struct mmu_notifier_range range;
1229
1230         pages = kmalloc_array(HPAGE_PMD_NR, sizeof(struct page *),
1231                               GFP_KERNEL);
1232         if (unlikely(!pages)) {
1233                 ret |= VM_FAULT_OOM;
1234                 goto out;
1235         }
1236
1237         for (i = 0; i < HPAGE_PMD_NR; i++) {
1238                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
1239                                                vmf->address, page_to_nid(page));
1240                 if (unlikely(!pages[i] ||
1241                              mem_cgroup_try_charge_delay(pages[i], vma->vm_mm,
1242                                      GFP_KERNEL, &memcg, false))) {
1243                         if (pages[i])
1244                                 put_page(pages[i]);
1245                         while (--i >= 0) {
1246                                 memcg = (void *)page_private(pages[i]);
1247                                 set_page_private(pages[i], 0);
1248                                 mem_cgroup_cancel_charge(pages[i], memcg,
1249                                                 false);
1250                                 put_page(pages[i]);
1251                         }
1252                         kfree(pages);
1253                         ret |= VM_FAULT_OOM;
1254                         goto out;
1255                 }
1256                 set_page_private(pages[i], (unsigned long)memcg);
1257         }
1258
1259         for (i = 0; i < HPAGE_PMD_NR; i++) {
1260                 copy_user_highpage(pages[i], page + i,
1261                                    haddr + PAGE_SIZE * i, vma);
1262                 __SetPageUptodate(pages[i]);
1263                 cond_resched();
1264         }
1265
1266         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1267                                 haddr, haddr + HPAGE_PMD_SIZE);
1268         mmu_notifier_invalidate_range_start(&range);
1269
1270         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1271         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1272                 goto out_free_pages;
1273         VM_BUG_ON_PAGE(!PageHead(page), page);
1274
1275         /*
1276          * Leave pmd empty until pte is filled note we must notify here as
1277          * concurrent CPU thread might write to new page before the call to
1278          * mmu_notifier_invalidate_range_end() happens which can lead to a
1279          * device seeing memory write in different order than CPU.
1280          *
1281          * See Documentation/vm/mmu_notifier.rst
1282          */
1283         pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1284
1285         pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
1286         pmd_populate(vma->vm_mm, &_pmd, pgtable);
1287
1288         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1289                 pte_t entry;
1290                 entry = mk_pte(pages[i], vma->vm_page_prot);
1291                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1292                 memcg = (void *)page_private(pages[i]);
1293                 set_page_private(pages[i], 0);
1294                 page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
1295                 mem_cgroup_commit_charge(pages[i], memcg, false, false);
1296                 lru_cache_add_active_or_unevictable(pages[i], vma);
1297                 vmf->pte = pte_offset_map(&_pmd, haddr);
1298                 VM_BUG_ON(!pte_none(*vmf->pte));
1299                 set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
1300                 pte_unmap(vmf->pte);
1301         }
1302         kfree(pages);
1303
1304         smp_wmb(); /* make pte visible before pmd */
1305         pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
1306         page_remove_rmap(page, true);
1307         spin_unlock(vmf->ptl);
1308
1309         /*
1310          * No need to double call mmu_notifier->invalidate_range() callback as
1311          * the above pmdp_huge_clear_flush_notify() did already call it.
1312          */
1313         mmu_notifier_invalidate_range_only_end(&range);
1314
1315         ret |= VM_FAULT_WRITE;
1316         put_page(page);
1317
1318 out:
1319         return ret;
1320
1321 out_free_pages:
1322         spin_unlock(vmf->ptl);
1323         mmu_notifier_invalidate_range_end(&range);
1324         for (i = 0; i < HPAGE_PMD_NR; i++) {
1325                 memcg = (void *)page_private(pages[i]);
1326                 set_page_private(pages[i], 0);
1327                 mem_cgroup_cancel_charge(pages[i], memcg, false);
1328                 put_page(pages[i]);
1329         }
1330         kfree(pages);
1331         goto out;
1332 }
1333
1334 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1335 {
1336         struct vm_area_struct *vma = vmf->vma;
1337         struct page *page = NULL, *new_page;
1338         struct mem_cgroup *memcg;
1339         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1340         struct mmu_notifier_range range;
1341         gfp_t huge_gfp;                 /* for allocation and charge */
1342         vm_fault_t ret = 0;
1343
1344         vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1345         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1346         if (is_huge_zero_pmd(orig_pmd))
1347                 goto alloc;
1348         spin_lock(vmf->ptl);
1349         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1350                 goto out_unlock;
1351
1352         page = pmd_page(orig_pmd);
1353         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1354         /*
1355          * We can only reuse the page if nobody else maps the huge page or it's
1356          * part.
1357          */
1358         if (!trylock_page(page)) {
1359                 get_page(page);
1360                 spin_unlock(vmf->ptl);
1361                 lock_page(page);
1362                 spin_lock(vmf->ptl);
1363                 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1364                         unlock_page(page);
1365                         put_page(page);
1366                         goto out_unlock;
1367                 }
1368                 put_page(page);
1369         }
1370         if (reuse_swap_page(page, NULL)) {
1371                 pmd_t entry;
1372                 entry = pmd_mkyoung(orig_pmd);
1373                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1374                 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry,  1))
1375                         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1376                 ret |= VM_FAULT_WRITE;
1377                 unlock_page(page);
1378                 goto out_unlock;
1379         }
1380         unlock_page(page);
1381         get_page(page);
1382         spin_unlock(vmf->ptl);
1383 alloc:
1384         if (__transparent_hugepage_enabled(vma) &&
1385             !transparent_hugepage_debug_cow()) {
1386                 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1387                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1388         } else
1389                 new_page = NULL;
1390
1391         if (likely(new_page)) {
1392                 prep_transhuge_page(new_page);
1393         } else {
1394                 if (!page) {
1395                         split_huge_pmd(vma, vmf->pmd, vmf->address);
1396                         ret |= VM_FAULT_FALLBACK;
1397                 } else {
1398                         ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
1399                         if (ret & VM_FAULT_OOM) {
1400                                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1401                                 ret |= VM_FAULT_FALLBACK;
1402                         }
1403                         put_page(page);
1404                 }
1405                 count_vm_event(THP_FAULT_FALLBACK);
1406                 goto out;
1407         }
1408
1409         if (unlikely(mem_cgroup_try_charge_delay(new_page, vma->vm_mm,
1410                                         huge_gfp, &memcg, true))) {
1411                 put_page(new_page);
1412                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1413                 if (page)
1414                         put_page(page);
1415                 ret |= VM_FAULT_FALLBACK;
1416                 count_vm_event(THP_FAULT_FALLBACK);
1417                 goto out;
1418         }
1419
1420         count_vm_event(THP_FAULT_ALLOC);
1421         count_memcg_events(memcg, THP_FAULT_ALLOC, 1);
1422
1423         if (!page)
1424                 clear_huge_page(new_page, vmf->address, HPAGE_PMD_NR);
1425         else
1426                 copy_user_huge_page(new_page, page, vmf->address,
1427                                     vma, HPAGE_PMD_NR);
1428         __SetPageUptodate(new_page);
1429
1430         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1431                                 haddr, haddr + HPAGE_PMD_SIZE);
1432         mmu_notifier_invalidate_range_start(&range);
1433
1434         spin_lock(vmf->ptl);
1435         if (page)
1436                 put_page(page);
1437         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1438                 spin_unlock(vmf->ptl);
1439                 mem_cgroup_cancel_charge(new_page, memcg, true);
1440                 put_page(new_page);
1441                 goto out_mn;
1442         } else {
1443                 pmd_t entry;
1444                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1445                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1446                 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1447                 page_add_new_anon_rmap(new_page, vma, haddr, true);
1448                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1449                 lru_cache_add_active_or_unevictable(new_page, vma);
1450                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1451                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1452                 if (!page) {
1453                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1454                 } else {
1455                         VM_BUG_ON_PAGE(!PageHead(page), page);
1456                         page_remove_rmap(page, true);
1457                         put_page(page);
1458                 }
1459                 ret |= VM_FAULT_WRITE;
1460         }
1461         spin_unlock(vmf->ptl);
1462 out_mn:
1463         /*
1464          * No need to double call mmu_notifier->invalidate_range() callback as
1465          * the above pmdp_huge_clear_flush_notify() did already call it.
1466          */
1467         mmu_notifier_invalidate_range_only_end(&range);
1468 out:
1469         return ret;
1470 out_unlock:
1471         spin_unlock(vmf->ptl);
1472         return ret;
1473 }
1474
1475 /*
1476  * FOLL_FORCE can write to even unwritable pmd's, but only
1477  * after we've gone through a COW cycle and they are dirty.
1478  */
1479 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1480 {
1481         return pmd_write(pmd) ||
1482                ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1483 }
1484
1485 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1486                                    unsigned long addr,
1487                                    pmd_t *pmd,
1488                                    unsigned int flags)
1489 {
1490         struct mm_struct *mm = vma->vm_mm;
1491         struct page *page = NULL;
1492
1493         assert_spin_locked(pmd_lockptr(mm, pmd));
1494
1495         if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1496                 goto out;
1497
1498         /* Avoid dumping huge zero page */
1499         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1500                 return ERR_PTR(-EFAULT);
1501
1502         /* Full NUMA hinting faults to serialise migration in fault paths */
1503         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1504                 goto out;
1505
1506         page = pmd_page(*pmd);
1507         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1508         if (flags & FOLL_TOUCH)
1509                 touch_pmd(vma, addr, pmd, flags);
1510         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1511                 /*
1512                  * We don't mlock() pte-mapped THPs. This way we can avoid
1513                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1514                  *
1515                  * For anon THP:
1516                  *
1517                  * In most cases the pmd is the only mapping of the page as we
1518                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1519                  * writable private mappings in populate_vma_page_range().
1520                  *
1521                  * The only scenario when we have the page shared here is if we
1522                  * mlocking read-only mapping shared over fork(). We skip
1523                  * mlocking such pages.
1524                  *
1525                  * For file THP:
1526                  *
1527                  * We can expect PageDoubleMap() to be stable under page lock:
1528                  * for file pages we set it in page_add_file_rmap(), which
1529                  * requires page to be locked.
1530                  */
1531
1532                 if (PageAnon(page) && compound_mapcount(page) != 1)
1533                         goto skip_mlock;
1534                 if (PageDoubleMap(page) || !page->mapping)
1535                         goto skip_mlock;
1536                 if (!trylock_page(page))
1537                         goto skip_mlock;
1538                 lru_add_drain();
1539                 if (page->mapping && !PageDoubleMap(page))
1540                         mlock_vma_page(page);
1541                 unlock_page(page);
1542         }
1543 skip_mlock:
1544         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1545         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1546         if (flags & FOLL_GET)
1547                 get_page(page);
1548
1549 out:
1550         return page;
1551 }
1552
1553 /* NUMA hinting page fault entry point for trans huge pmds */
1554 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1555 {
1556         struct vm_area_struct *vma = vmf->vma;
1557         struct anon_vma *anon_vma = NULL;
1558         struct page *page;
1559         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1560         int page_nid = NUMA_NO_NODE, this_nid = numa_node_id();
1561         int target_nid, last_cpupid = -1;
1562         bool page_locked;
1563         bool migrated = false;
1564         bool was_writable;
1565         int flags = 0;
1566
1567         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1568         if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1569                 goto out_unlock;
1570
1571         /*
1572          * If there are potential migrations, wait for completion and retry
1573          * without disrupting NUMA hinting information. Do not relock and
1574          * check_same as the page may no longer be mapped.
1575          */
1576         if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1577                 page = pmd_page(*vmf->pmd);
1578                 if (!get_page_unless_zero(page))
1579                         goto out_unlock;
1580                 spin_unlock(vmf->ptl);
1581                 put_and_wait_on_page_locked(page);
1582                 goto out;
1583         }
1584
1585         page = pmd_page(pmd);
1586         BUG_ON(is_huge_zero_page(page));
1587         page_nid = page_to_nid(page);
1588         last_cpupid = page_cpupid_last(page);
1589         count_vm_numa_event(NUMA_HINT_FAULTS);
1590         if (page_nid == this_nid) {
1591                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1592                 flags |= TNF_FAULT_LOCAL;
1593         }
1594
1595         /* See similar comment in do_numa_page for explanation */
1596         if (!pmd_savedwrite(pmd))
1597                 flags |= TNF_NO_GROUP;
1598
1599         /*
1600          * Acquire the page lock to serialise THP migrations but avoid dropping
1601          * page_table_lock if at all possible
1602          */
1603         page_locked = trylock_page(page);
1604         target_nid = mpol_misplaced(page, vma, haddr);
1605         if (target_nid == NUMA_NO_NODE) {
1606                 /* If the page was locked, there are no parallel migrations */
1607                 if (page_locked)
1608                         goto clear_pmdnuma;
1609         }
1610
1611         /* Migration could have started since the pmd_trans_migrating check */
1612         if (!page_locked) {
1613                 page_nid = NUMA_NO_NODE;
1614                 if (!get_page_unless_zero(page))
1615                         goto out_unlock;
1616                 spin_unlock(vmf->ptl);
1617                 put_and_wait_on_page_locked(page);
1618                 goto out;
1619         }
1620
1621         /*
1622          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1623          * to serialises splits
1624          */
1625         get_page(page);
1626         spin_unlock(vmf->ptl);
1627         anon_vma = page_lock_anon_vma_read(page);
1628
1629         /* Confirm the PMD did not change while page_table_lock was released */
1630         spin_lock(vmf->ptl);
1631         if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1632                 unlock_page(page);
1633                 put_page(page);
1634                 page_nid = NUMA_NO_NODE;
1635                 goto out_unlock;
1636         }
1637
1638         /* Bail if we fail to protect against THP splits for any reason */
1639         if (unlikely(!anon_vma)) {
1640                 put_page(page);
1641                 page_nid = NUMA_NO_NODE;
1642                 goto clear_pmdnuma;
1643         }
1644
1645         /*
1646          * Since we took the NUMA fault, we must have observed the !accessible
1647          * bit. Make sure all other CPUs agree with that, to avoid them
1648          * modifying the page we're about to migrate.
1649          *
1650          * Must be done under PTL such that we'll observe the relevant
1651          * inc_tlb_flush_pending().
1652          *
1653          * We are not sure a pending tlb flush here is for a huge page
1654          * mapping or not. Hence use the tlb range variant
1655          */
1656         if (mm_tlb_flush_pending(vma->vm_mm)) {
1657                 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1658                 /*
1659                  * change_huge_pmd() released the pmd lock before
1660                  * invalidating the secondary MMUs sharing the primary
1661                  * MMU pagetables (with ->invalidate_range()). The
1662                  * mmu_notifier_invalidate_range_end() (which
1663                  * internally calls ->invalidate_range()) in
1664                  * change_pmd_range() will run after us, so we can't
1665                  * rely on it here and we need an explicit invalidate.
1666                  */
1667                 mmu_notifier_invalidate_range(vma->vm_mm, haddr,
1668                                               haddr + HPAGE_PMD_SIZE);
1669         }
1670
1671         /*
1672          * Migrate the THP to the requested node, returns with page unlocked
1673          * and access rights restored.
1674          */
1675         spin_unlock(vmf->ptl);
1676
1677         migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1678                                 vmf->pmd, pmd, vmf->address, page, target_nid);
1679         if (migrated) {
1680                 flags |= TNF_MIGRATED;
1681                 page_nid = target_nid;
1682         } else
1683                 flags |= TNF_MIGRATE_FAIL;
1684
1685         goto out;
1686 clear_pmdnuma:
1687         BUG_ON(!PageLocked(page));
1688         was_writable = pmd_savedwrite(pmd);
1689         pmd = pmd_modify(pmd, vma->vm_page_prot);
1690         pmd = pmd_mkyoung(pmd);
1691         if (was_writable)
1692                 pmd = pmd_mkwrite(pmd);
1693         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1694         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1695         unlock_page(page);
1696 out_unlock:
1697         spin_unlock(vmf->ptl);
1698
1699 out:
1700         if (anon_vma)
1701                 page_unlock_anon_vma_read(anon_vma);
1702
1703         if (page_nid != NUMA_NO_NODE)
1704                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1705                                 flags);
1706
1707         return 0;
1708 }
1709
1710 /*
1711  * Return true if we do MADV_FREE successfully on entire pmd page.
1712  * Otherwise, return false.
1713  */
1714 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1715                 pmd_t *pmd, unsigned long addr, unsigned long next)
1716 {
1717         spinlock_t *ptl;
1718         pmd_t orig_pmd;
1719         struct page *page;
1720         struct mm_struct *mm = tlb->mm;
1721         bool ret = false;
1722
1723         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1724
1725         ptl = pmd_trans_huge_lock(pmd, vma);
1726         if (!ptl)
1727                 goto out_unlocked;
1728
1729         orig_pmd = *pmd;
1730         if (is_huge_zero_pmd(orig_pmd))
1731                 goto out;
1732
1733         if (unlikely(!pmd_present(orig_pmd))) {
1734                 VM_BUG_ON(thp_migration_supported() &&
1735                                   !is_pmd_migration_entry(orig_pmd));
1736                 goto out;
1737         }
1738
1739         page = pmd_page(orig_pmd);
1740         /*
1741          * If other processes are mapping this page, we couldn't discard
1742          * the page unless they all do MADV_FREE so let's skip the page.
1743          */
1744         if (page_mapcount(page) != 1)
1745                 goto out;
1746
1747         if (!trylock_page(page))
1748                 goto out;
1749
1750         /*
1751          * If user want to discard part-pages of THP, split it so MADV_FREE
1752          * will deactivate only them.
1753          */
1754         if (next - addr != HPAGE_PMD_SIZE) {
1755                 get_page(page);
1756                 spin_unlock(ptl);
1757                 split_huge_page(page);
1758                 unlock_page(page);
1759                 put_page(page);
1760                 goto out_unlocked;
1761         }
1762
1763         if (PageDirty(page))
1764                 ClearPageDirty(page);
1765         unlock_page(page);
1766
1767         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1768                 pmdp_invalidate(vma, addr, pmd);
1769                 orig_pmd = pmd_mkold(orig_pmd);
1770                 orig_pmd = pmd_mkclean(orig_pmd);
1771
1772                 set_pmd_at(mm, addr, pmd, orig_pmd);
1773                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1774         }
1775
1776         mark_page_lazyfree(page);
1777         ret = true;
1778 out:
1779         spin_unlock(ptl);
1780 out_unlocked:
1781         return ret;
1782 }
1783
1784 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1785 {
1786         pgtable_t pgtable;
1787
1788         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1789         pte_free(mm, pgtable);
1790         mm_dec_nr_ptes(mm);
1791 }
1792
1793 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1794                  pmd_t *pmd, unsigned long addr)
1795 {
1796         pmd_t orig_pmd;
1797         spinlock_t *ptl;
1798
1799         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1800
1801         ptl = __pmd_trans_huge_lock(pmd, vma);
1802         if (!ptl)
1803                 return 0;
1804         /*
1805          * For architectures like ppc64 we look at deposited pgtable
1806          * when calling pmdp_huge_get_and_clear. So do the
1807          * pgtable_trans_huge_withdraw after finishing pmdp related
1808          * operations.
1809          */
1810         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1811                         tlb->fullmm);
1812         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1813         if (vma_is_dax(vma)) {
1814                 if (arch_needs_pgtable_deposit())
1815                         zap_deposited_table(tlb->mm, pmd);
1816                 spin_unlock(ptl);
1817                 if (is_huge_zero_pmd(orig_pmd))
1818                         tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1819         } else if (is_huge_zero_pmd(orig_pmd)) {
1820                 zap_deposited_table(tlb->mm, pmd);
1821                 spin_unlock(ptl);
1822                 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1823         } else {
1824                 struct page *page = NULL;
1825                 int flush_needed = 1;
1826
1827                 if (pmd_present(orig_pmd)) {
1828                         page = pmd_page(orig_pmd);
1829                         page_remove_rmap(page, true);
1830                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1831                         VM_BUG_ON_PAGE(!PageHead(page), page);
1832                 } else if (thp_migration_supported()) {
1833                         swp_entry_t entry;
1834
1835                         VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1836                         entry = pmd_to_swp_entry(orig_pmd);
1837                         page = pfn_to_page(swp_offset(entry));
1838                         flush_needed = 0;
1839                 } else
1840                         WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1841
1842                 if (PageAnon(page)) {
1843                         zap_deposited_table(tlb->mm, pmd);
1844                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1845                 } else {
1846                         if (arch_needs_pgtable_deposit())
1847                                 zap_deposited_table(tlb->mm, pmd);
1848                         add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1849                 }
1850
1851                 spin_unlock(ptl);
1852                 if (flush_needed)
1853                         tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1854         }
1855         return 1;
1856 }
1857
1858 #ifndef pmd_move_must_withdraw
1859 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1860                                          spinlock_t *old_pmd_ptl,
1861                                          struct vm_area_struct *vma)
1862 {
1863         /*
1864          * With split pmd lock we also need to move preallocated
1865          * PTE page table if new_pmd is on different PMD page table.
1866          *
1867          * We also don't deposit and withdraw tables for file pages.
1868          */
1869         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1870 }
1871 #endif
1872
1873 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1874 {
1875 #ifdef CONFIG_MEM_SOFT_DIRTY
1876         if (unlikely(is_pmd_migration_entry(pmd)))
1877                 pmd = pmd_swp_mksoft_dirty(pmd);
1878         else if (pmd_present(pmd))
1879                 pmd = pmd_mksoft_dirty(pmd);
1880 #endif
1881         return pmd;
1882 }
1883
1884 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1885                   unsigned long new_addr, unsigned long old_end,
1886                   pmd_t *old_pmd, pmd_t *new_pmd)
1887 {
1888         spinlock_t *old_ptl, *new_ptl;
1889         pmd_t pmd;
1890         struct mm_struct *mm = vma->vm_mm;
1891         bool force_flush = false;
1892
1893         if ((old_addr & ~HPAGE_PMD_MASK) ||
1894             (new_addr & ~HPAGE_PMD_MASK) ||
1895             old_end - old_addr < HPAGE_PMD_SIZE)
1896                 return false;
1897
1898         /*
1899          * The destination pmd shouldn't be established, free_pgtables()
1900          * should have release it.
1901          */
1902         if (WARN_ON(!pmd_none(*new_pmd))) {
1903                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1904                 return false;
1905         }
1906
1907         /*
1908          * We don't have to worry about the ordering of src and dst
1909          * ptlocks because exclusive mmap_sem prevents deadlock.
1910          */
1911         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1912         if (old_ptl) {
1913                 new_ptl = pmd_lockptr(mm, new_pmd);
1914                 if (new_ptl != old_ptl)
1915                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1916                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1917                 if (pmd_present(pmd))
1918                         force_flush = true;
1919                 VM_BUG_ON(!pmd_none(*new_pmd));
1920
1921                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1922                         pgtable_t pgtable;
1923                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1924                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1925                 }
1926                 pmd = move_soft_dirty_pmd(pmd);
1927                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1928                 if (force_flush)
1929                         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1930                 if (new_ptl != old_ptl)
1931                         spin_unlock(new_ptl);
1932                 spin_unlock(old_ptl);
1933                 return true;
1934         }
1935         return false;
1936 }
1937
1938 /*
1939  * Returns
1940  *  - 0 if PMD could not be locked
1941  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1942  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1943  */
1944 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1945                 unsigned long addr, pgprot_t newprot, int prot_numa)
1946 {
1947         struct mm_struct *mm = vma->vm_mm;
1948         spinlock_t *ptl;
1949         pmd_t entry;
1950         bool preserve_write;
1951         int ret;
1952
1953         ptl = __pmd_trans_huge_lock(pmd, vma);
1954         if (!ptl)
1955                 return 0;
1956
1957         preserve_write = prot_numa && pmd_write(*pmd);
1958         ret = 1;
1959
1960 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1961         if (is_swap_pmd(*pmd)) {
1962                 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1963
1964                 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1965                 if (is_write_migration_entry(entry)) {
1966                         pmd_t newpmd;
1967                         /*
1968                          * A protection check is difficult so
1969                          * just be safe and disable write
1970                          */
1971                         make_migration_entry_read(&entry);
1972                         newpmd = swp_entry_to_pmd(entry);
1973                         if (pmd_swp_soft_dirty(*pmd))
1974                                 newpmd = pmd_swp_mksoft_dirty(newpmd);
1975                         set_pmd_at(mm, addr, pmd, newpmd);
1976                 }
1977                 goto unlock;
1978         }
1979 #endif
1980
1981         /*
1982          * Avoid trapping faults against the zero page. The read-only
1983          * data is likely to be read-cached on the local CPU and
1984          * local/remote hits to the zero page are not interesting.
1985          */
1986         if (prot_numa && is_huge_zero_pmd(*pmd))
1987                 goto unlock;
1988
1989         if (prot_numa && pmd_protnone(*pmd))
1990                 goto unlock;
1991
1992         /*
1993          * In case prot_numa, we are under down_read(mmap_sem). It's critical
1994          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1995          * which is also under down_read(mmap_sem):
1996          *
1997          *      CPU0:                           CPU1:
1998          *                              change_huge_pmd(prot_numa=1)
1999          *                               pmdp_huge_get_and_clear_notify()
2000          * madvise_dontneed()
2001          *  zap_pmd_range()
2002          *   pmd_trans_huge(*pmd) == 0 (without ptl)
2003          *   // skip the pmd
2004          *                               set_pmd_at();
2005          *                               // pmd is re-established
2006          *
2007          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
2008          * which may break userspace.
2009          *
2010          * pmdp_invalidate() is required to make sure we don't miss
2011          * dirty/young flags set by hardware.
2012          */
2013         entry = pmdp_invalidate(vma, addr, pmd);
2014
2015         entry = pmd_modify(entry, newprot);
2016         if (preserve_write)
2017                 entry = pmd_mk_savedwrite(entry);
2018         ret = HPAGE_PMD_NR;
2019         set_pmd_at(mm, addr, pmd, entry);
2020         BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
2021 unlock:
2022         spin_unlock(ptl);
2023         return ret;
2024 }
2025
2026 /*
2027  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
2028  *
2029  * Note that if it returns page table lock pointer, this routine returns without
2030  * unlocking page table lock. So callers must unlock it.
2031  */
2032 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
2033 {
2034         spinlock_t *ptl;
2035         ptl = pmd_lock(vma->vm_mm, pmd);
2036         if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
2037                         pmd_devmap(*pmd)))
2038                 return ptl;
2039         spin_unlock(ptl);
2040         return NULL;
2041 }
2042
2043 /*
2044  * Returns true if a given pud maps a thp, false otherwise.
2045  *
2046  * Note that if it returns true, this routine returns without unlocking page
2047  * table lock. So callers must unlock it.
2048  */
2049 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
2050 {
2051         spinlock_t *ptl;
2052
2053         ptl = pud_lock(vma->vm_mm, pud);
2054         if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
2055                 return ptl;
2056         spin_unlock(ptl);
2057         return NULL;
2058 }
2059
2060 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
2061 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
2062                  pud_t *pud, unsigned long addr)
2063 {
2064         spinlock_t *ptl;
2065
2066         ptl = __pud_trans_huge_lock(pud, vma);
2067         if (!ptl)
2068                 return 0;
2069         /*
2070          * For architectures like ppc64 we look at deposited pgtable
2071          * when calling pudp_huge_get_and_clear. So do the
2072          * pgtable_trans_huge_withdraw after finishing pudp related
2073          * operations.
2074          */
2075         pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
2076         tlb_remove_pud_tlb_entry(tlb, pud, addr);
2077         if (vma_is_dax(vma)) {
2078                 spin_unlock(ptl);
2079                 /* No zero page support yet */
2080         } else {
2081                 /* No support for anonymous PUD pages yet */
2082                 BUG();
2083         }
2084         return 1;
2085 }
2086
2087 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
2088                 unsigned long haddr)
2089 {
2090         VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
2091         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2092         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
2093         VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
2094
2095         count_vm_event(THP_SPLIT_PUD);
2096
2097         pudp_huge_clear_flush_notify(vma, haddr, pud);
2098 }
2099
2100 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2101                 unsigned long address)
2102 {
2103         spinlock_t *ptl;
2104         struct mmu_notifier_range range;
2105
2106         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2107                                 address & HPAGE_PUD_MASK,
2108                                 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
2109         mmu_notifier_invalidate_range_start(&range);
2110         ptl = pud_lock(vma->vm_mm, pud);
2111         if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2112                 goto out;
2113         __split_huge_pud_locked(vma, pud, range.start);
2114
2115 out:
2116         spin_unlock(ptl);
2117         /*
2118          * No need to double call mmu_notifier->invalidate_range() callback as
2119          * the above pudp_huge_clear_flush_notify() did already call it.
2120          */
2121         mmu_notifier_invalidate_range_only_end(&range);
2122 }
2123 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2124
2125 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2126                 unsigned long haddr, pmd_t *pmd)
2127 {
2128         struct mm_struct *mm = vma->vm_mm;
2129         pgtable_t pgtable;
2130         pmd_t _pmd;
2131         int i;
2132
2133         /*
2134          * Leave pmd empty until pte is filled note that it is fine to delay
2135          * notification until mmu_notifier_invalidate_range_end() as we are
2136          * replacing a zero pmd write protected page with a zero pte write
2137          * protected page.
2138          *
2139          * See Documentation/vm/mmu_notifier.rst
2140          */
2141         pmdp_huge_clear_flush(vma, haddr, pmd);
2142
2143         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2144         pmd_populate(mm, &_pmd, pgtable);
2145
2146         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2147                 pte_t *pte, entry;
2148                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2149                 entry = pte_mkspecial(entry);
2150                 pte = pte_offset_map(&_pmd, haddr);
2151                 VM_BUG_ON(!pte_none(*pte));
2152                 set_pte_at(mm, haddr, pte, entry);
2153                 pte_unmap(pte);
2154         }
2155         smp_wmb(); /* make pte visible before pmd */
2156         pmd_populate(mm, pmd, pgtable);
2157 }
2158
2159 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2160                 unsigned long haddr, bool freeze)
2161 {
2162         struct mm_struct *mm = vma->vm_mm;
2163         struct page *page;
2164         pgtable_t pgtable;
2165         pmd_t old_pmd, _pmd;
2166         bool young, write, soft_dirty, pmd_migration = false;
2167         unsigned long addr;
2168         int i;
2169
2170         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2171         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2172         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2173         VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2174                                 && !pmd_devmap(*pmd));
2175
2176         count_vm_event(THP_SPLIT_PMD);
2177
2178         if (!vma_is_anonymous(vma)) {
2179                 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2180                 /*
2181                  * We are going to unmap this huge page. So
2182                  * just go ahead and zap it
2183                  */
2184                 if (arch_needs_pgtable_deposit())
2185                         zap_deposited_table(mm, pmd);
2186                 if (vma_is_dax(vma))
2187                         return;
2188                 page = pmd_page(_pmd);
2189                 if (!PageDirty(page) && pmd_dirty(_pmd))
2190                         set_page_dirty(page);
2191                 if (!PageReferenced(page) && pmd_young(_pmd))
2192                         SetPageReferenced(page);
2193                 page_remove_rmap(page, true);
2194                 put_page(page);
2195                 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2196                 return;
2197         } else if (is_huge_zero_pmd(*pmd)) {
2198                 /*
2199                  * FIXME: Do we want to invalidate secondary mmu by calling
2200                  * mmu_notifier_invalidate_range() see comments below inside
2201                  * __split_huge_pmd() ?
2202                  *
2203                  * We are going from a zero huge page write protected to zero
2204                  * small page also write protected so it does not seems useful
2205                  * to invalidate secondary mmu at this time.
2206                  */
2207                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2208         }
2209
2210         /*
2211          * Up to this point the pmd is present and huge and userland has the
2212          * whole access to the hugepage during the split (which happens in
2213          * place). If we overwrite the pmd with the not-huge version pointing
2214          * to the pte here (which of course we could if all CPUs were bug
2215          * free), userland could trigger a small page size TLB miss on the
2216          * small sized TLB while the hugepage TLB entry is still established in
2217          * the huge TLB. Some CPU doesn't like that.
2218          * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2219          * 383 on page 93. Intel should be safe but is also warns that it's
2220          * only safe if the permission and cache attributes of the two entries
2221          * loaded in the two TLB is identical (which should be the case here).
2222          * But it is generally safer to never allow small and huge TLB entries
2223          * for the same virtual address to be loaded simultaneously. So instead
2224          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2225          * current pmd notpresent (atomically because here the pmd_trans_huge
2226          * must remain set at all times on the pmd until the split is complete
2227          * for this pmd), then we flush the SMP TLB and finally we write the
2228          * non-huge version of the pmd entry with pmd_populate.
2229          */
2230         old_pmd = pmdp_invalidate(vma, haddr, pmd);
2231
2232         pmd_migration = is_pmd_migration_entry(old_pmd);
2233         if (unlikely(pmd_migration)) {
2234                 swp_entry_t entry;
2235
2236                 entry = pmd_to_swp_entry(old_pmd);
2237                 page = pfn_to_page(swp_offset(entry));
2238                 write = is_write_migration_entry(entry);
2239                 young = false;
2240                 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2241         } else {
2242                 page = pmd_page(old_pmd);
2243                 if (pmd_dirty(old_pmd))
2244                         SetPageDirty(page);
2245                 write = pmd_write(old_pmd);
2246                 young = pmd_young(old_pmd);
2247                 soft_dirty = pmd_soft_dirty(old_pmd);
2248         }
2249         VM_BUG_ON_PAGE(!page_count(page), page);
2250         page_ref_add(page, HPAGE_PMD_NR - 1);
2251
2252         /*
2253          * Withdraw the table only after we mark the pmd entry invalid.
2254          * This's critical for some architectures (Power).
2255          */
2256         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2257         pmd_populate(mm, &_pmd, pgtable);
2258
2259         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2260                 pte_t entry, *pte;
2261                 /*
2262                  * Note that NUMA hinting access restrictions are not
2263                  * transferred to avoid any possibility of altering
2264                  * permissions across VMAs.
2265                  */
2266                 if (freeze || pmd_migration) {
2267                         swp_entry_t swp_entry;
2268                         swp_entry = make_migration_entry(page + i, write);
2269                         entry = swp_entry_to_pte(swp_entry);
2270                         if (soft_dirty)
2271                                 entry = pte_swp_mksoft_dirty(entry);
2272                 } else {
2273                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2274                         entry = maybe_mkwrite(entry, vma);
2275                         if (!write)
2276                                 entry = pte_wrprotect(entry);
2277                         if (!young)
2278                                 entry = pte_mkold(entry);
2279                         if (soft_dirty)
2280                                 entry = pte_mksoft_dirty(entry);
2281                 }
2282                 pte = pte_offset_map(&_pmd, addr);
2283                 BUG_ON(!pte_none(*pte));
2284                 set_pte_at(mm, addr, pte, entry);
2285                 atomic_inc(&page[i]._mapcount);
2286                 pte_unmap(pte);
2287         }
2288
2289         /*
2290          * Set PG_double_map before dropping compound_mapcount to avoid
2291          * false-negative page_mapped().
2292          */
2293         if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2294                 for (i = 0; i < HPAGE_PMD_NR; i++)
2295                         atomic_inc(&page[i]._mapcount);
2296         }
2297
2298         if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2299                 /* Last compound_mapcount is gone. */
2300                 __dec_node_page_state(page, NR_ANON_THPS);
2301                 if (TestClearPageDoubleMap(page)) {
2302                         /* No need in mapcount reference anymore */
2303                         for (i = 0; i < HPAGE_PMD_NR; i++)
2304                                 atomic_dec(&page[i]._mapcount);
2305                 }
2306         }
2307
2308         smp_wmb(); /* make pte visible before pmd */
2309         pmd_populate(mm, pmd, pgtable);
2310
2311         if (freeze) {
2312                 for (i = 0; i < HPAGE_PMD_NR; i++) {
2313                         page_remove_rmap(page + i, false);
2314                         put_page(page + i);
2315                 }
2316         }
2317 }
2318
2319 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2320                 unsigned long address, bool freeze, struct page *page)
2321 {
2322         spinlock_t *ptl;
2323         struct mmu_notifier_range range;
2324
2325         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2326                                 address & HPAGE_PMD_MASK,
2327                                 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2328         mmu_notifier_invalidate_range_start(&range);
2329         ptl = pmd_lock(vma->vm_mm, pmd);
2330
2331         /*
2332          * If caller asks to setup a migration entries, we need a page to check
2333          * pmd against. Otherwise we can end up replacing wrong page.
2334          */
2335         VM_BUG_ON(freeze && !page);
2336         if (page && page != pmd_page(*pmd))
2337                 goto out;
2338
2339         if (pmd_trans_huge(*pmd)) {
2340                 page = pmd_page(*pmd);
2341                 if (PageMlocked(page))
2342                         clear_page_mlock(page);
2343         } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2344                 goto out;
2345         __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2346 out:
2347         spin_unlock(ptl);
2348         /*
2349          * No need to double call mmu_notifier->invalidate_range() callback.
2350          * They are 3 cases to consider inside __split_huge_pmd_locked():
2351          *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2352          *  2) __split_huge_zero_page_pmd() read only zero page and any write
2353          *    fault will trigger a flush_notify before pointing to a new page
2354          *    (it is fine if the secondary mmu keeps pointing to the old zero
2355          *    page in the meantime)
2356          *  3) Split a huge pmd into pte pointing to the same page. No need
2357          *     to invalidate secondary tlb entry they are all still valid.
2358          *     any further changes to individual pte will notify. So no need
2359          *     to call mmu_notifier->invalidate_range()
2360          */
2361         mmu_notifier_invalidate_range_only_end(&range);
2362 }
2363
2364 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2365                 bool freeze, struct page *page)
2366 {
2367         pgd_t *pgd;
2368         p4d_t *p4d;
2369         pud_t *pud;
2370         pmd_t *pmd;
2371
2372         pgd = pgd_offset(vma->vm_mm, address);
2373         if (!pgd_present(*pgd))
2374                 return;
2375
2376         p4d = p4d_offset(pgd, address);
2377         if (!p4d_present(*p4d))
2378                 return;
2379
2380         pud = pud_offset(p4d, address);
2381         if (!pud_present(*pud))
2382                 return;
2383
2384         pmd = pmd_offset(pud, address);
2385
2386         __split_huge_pmd(vma, pmd, address, freeze, page);
2387 }
2388
2389 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2390                              unsigned long start,
2391                              unsigned long end,
2392                              long adjust_next)
2393 {
2394         /*
2395          * If the new start address isn't hpage aligned and it could
2396          * previously contain an hugepage: check if we need to split
2397          * an huge pmd.
2398          */
2399         if (start & ~HPAGE_PMD_MASK &&
2400             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2401             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2402                 split_huge_pmd_address(vma, start, false, NULL);
2403
2404         /*
2405          * If the new end address isn't hpage aligned and it could
2406          * previously contain an hugepage: check if we need to split
2407          * an huge pmd.
2408          */
2409         if (end & ~HPAGE_PMD_MASK &&
2410             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2411             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2412                 split_huge_pmd_address(vma, end, false, NULL);
2413
2414         /*
2415          * If we're also updating the vma->vm_next->vm_start, if the new
2416          * vm_next->vm_start isn't page aligned and it could previously
2417          * contain an hugepage: check if we need to split an huge pmd.
2418          */
2419         if (adjust_next > 0) {
2420                 struct vm_area_struct *next = vma->vm_next;
2421                 unsigned long nstart = next->vm_start;
2422                 nstart += adjust_next << PAGE_SHIFT;
2423                 if (nstart & ~HPAGE_PMD_MASK &&
2424                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2425                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2426                         split_huge_pmd_address(next, nstart, false, NULL);
2427         }
2428 }
2429
2430 static void unmap_page(struct page *page)
2431 {
2432         enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
2433                 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2434         bool unmap_success;
2435
2436         VM_BUG_ON_PAGE(!PageHead(page), page);
2437
2438         if (PageAnon(page))
2439                 ttu_flags |= TTU_SPLIT_FREEZE;
2440
2441         unmap_success = try_to_unmap(page, ttu_flags);
2442         VM_BUG_ON_PAGE(!unmap_success, page);
2443 }
2444
2445 static void remap_page(struct page *page)
2446 {
2447         int i;
2448         if (PageTransHuge(page)) {
2449                 remove_migration_ptes(page, page, true);
2450         } else {
2451                 for (i = 0; i < HPAGE_PMD_NR; i++)
2452                         remove_migration_ptes(page + i, page + i, true);
2453         }
2454 }
2455
2456 static void __split_huge_page_tail(struct page *head, int tail,
2457                 struct lruvec *lruvec, struct list_head *list)
2458 {
2459         struct page *page_tail = head + tail;
2460
2461         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2462
2463         /*
2464          * Clone page flags before unfreezing refcount.
2465          *
2466          * After successful get_page_unless_zero() might follow flags change,
2467          * for exmaple lock_page() which set PG_waiters.
2468          */
2469         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2470         page_tail->flags |= (head->flags &
2471                         ((1L << PG_referenced) |
2472                          (1L << PG_swapbacked) |
2473                          (1L << PG_swapcache) |
2474                          (1L << PG_mlocked) |
2475                          (1L << PG_uptodate) |
2476                          (1L << PG_active) |
2477                          (1L << PG_workingset) |
2478                          (1L << PG_locked) |
2479                          (1L << PG_unevictable) |
2480                          (1L << PG_dirty)));
2481
2482         /* ->mapping in first tail page is compound_mapcount */
2483         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2484                         page_tail);
2485         page_tail->mapping = head->mapping;
2486         page_tail->index = head->index + tail;
2487
2488         /* Page flags must be visible before we make the page non-compound. */
2489         smp_wmb();
2490
2491         /*
2492          * Clear PageTail before unfreezing page refcount.
2493          *
2494          * After successful get_page_unless_zero() might follow put_page()
2495          * which needs correct compound_head().
2496          */
2497         clear_compound_head(page_tail);
2498
2499         /* Finally unfreeze refcount. Additional reference from page cache. */
2500         page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2501                                           PageSwapCache(head)));
2502
2503         if (page_is_young(head))
2504                 set_page_young(page_tail);
2505         if (page_is_idle(head))
2506                 set_page_idle(page_tail);
2507
2508         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2509
2510         /*
2511          * always add to the tail because some iterators expect new
2512          * pages to show after the currently processed elements - e.g.
2513          * migrate_pages
2514          */
2515         lru_add_page_tail(head, page_tail, lruvec, list);
2516 }
2517
2518 static void __split_huge_page(struct page *page, struct list_head *list,
2519                 pgoff_t end, unsigned long flags)
2520 {
2521         struct page *head = compound_head(page);
2522         pg_data_t *pgdat = page_pgdat(head);
2523         struct lruvec *lruvec;
2524         struct address_space *swap_cache = NULL;
2525         unsigned long offset = 0;
2526         int i;
2527
2528         lruvec = mem_cgroup_page_lruvec(head, pgdat);
2529
2530         /* complete memcg works before add pages to LRU */
2531         mem_cgroup_split_huge_fixup(head);
2532
2533         if (PageAnon(head) && PageSwapCache(head)) {
2534                 swp_entry_t entry = { .val = page_private(head) };
2535
2536                 offset = swp_offset(entry);
2537                 swap_cache = swap_address_space(entry);
2538                 xa_lock(&swap_cache->i_pages);
2539         }
2540
2541         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
2542                 __split_huge_page_tail(head, i, lruvec, list);
2543                 /* Some pages can be beyond i_size: drop them from page cache */
2544                 if (head[i].index >= end) {
2545                         ClearPageDirty(head + i);
2546                         __delete_from_page_cache(head + i, NULL);
2547                         if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2548                                 shmem_uncharge(head->mapping->host, 1);
2549                         put_page(head + i);
2550                 } else if (!PageAnon(page)) {
2551                         __xa_store(&head->mapping->i_pages, head[i].index,
2552                                         head + i, 0);
2553                 } else if (swap_cache) {
2554                         __xa_store(&swap_cache->i_pages, offset + i,
2555                                         head + i, 0);
2556                 }
2557         }
2558
2559         ClearPageCompound(head);
2560
2561         split_page_owner(head, HPAGE_PMD_ORDER);
2562
2563         /* See comment in __split_huge_page_tail() */
2564         if (PageAnon(head)) {
2565                 /* Additional pin to swap cache */
2566                 if (PageSwapCache(head)) {
2567                         page_ref_add(head, 2);
2568                         xa_unlock(&swap_cache->i_pages);
2569                 } else {
2570                         page_ref_inc(head);
2571                 }
2572         } else {
2573                 /* Additional pin to page cache */
2574                 page_ref_add(head, 2);
2575                 xa_unlock(&head->mapping->i_pages);
2576         }
2577
2578         spin_unlock_irqrestore(&pgdat->lru_lock, flags);
2579
2580         remap_page(head);
2581
2582         for (i = 0; i < HPAGE_PMD_NR; i++) {
2583                 struct page *subpage = head + i;
2584                 if (subpage == page)
2585                         continue;
2586                 unlock_page(subpage);
2587
2588                 /*
2589                  * Subpages may be freed if there wasn't any mapping
2590                  * like if add_to_swap() is running on a lru page that
2591                  * had its mapping zapped. And freeing these pages
2592                  * requires taking the lru_lock so we do the put_page
2593                  * of the tail pages after the split is complete.
2594                  */
2595                 put_page(subpage);
2596         }
2597 }
2598
2599 int total_mapcount(struct page *page)
2600 {
2601         int i, compound, ret;
2602
2603         VM_BUG_ON_PAGE(PageTail(page), page);
2604
2605         if (likely(!PageCompound(page)))
2606                 return atomic_read(&page->_mapcount) + 1;
2607
2608         compound = compound_mapcount(page);
2609         if (PageHuge(page))
2610                 return compound;
2611         ret = compound;
2612         for (i = 0; i < HPAGE_PMD_NR; i++)
2613                 ret += atomic_read(&page[i]._mapcount) + 1;
2614         /* File pages has compound_mapcount included in _mapcount */
2615         if (!PageAnon(page))
2616                 return ret - compound * HPAGE_PMD_NR;
2617         if (PageDoubleMap(page))
2618                 ret -= HPAGE_PMD_NR;
2619         return ret;
2620 }
2621
2622 /*
2623  * This calculates accurately how many mappings a transparent hugepage
2624  * has (unlike page_mapcount() which isn't fully accurate). This full
2625  * accuracy is primarily needed to know if copy-on-write faults can
2626  * reuse the page and change the mapping to read-write instead of
2627  * copying them. At the same time this returns the total_mapcount too.
2628  *
2629  * The function returns the highest mapcount any one of the subpages
2630  * has. If the return value is one, even if different processes are
2631  * mapping different subpages of the transparent hugepage, they can
2632  * all reuse it, because each process is reusing a different subpage.
2633  *
2634  * The total_mapcount is instead counting all virtual mappings of the
2635  * subpages. If the total_mapcount is equal to "one", it tells the
2636  * caller all mappings belong to the same "mm" and in turn the
2637  * anon_vma of the transparent hugepage can become the vma->anon_vma
2638  * local one as no other process may be mapping any of the subpages.
2639  *
2640  * It would be more accurate to replace page_mapcount() with
2641  * page_trans_huge_mapcount(), however we only use
2642  * page_trans_huge_mapcount() in the copy-on-write faults where we
2643  * need full accuracy to avoid breaking page pinning, because
2644  * page_trans_huge_mapcount() is slower than page_mapcount().
2645  */
2646 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2647 {
2648         int i, ret, _total_mapcount, mapcount;
2649
2650         /* hugetlbfs shouldn't call it */
2651         VM_BUG_ON_PAGE(PageHuge(page), page);
2652
2653         if (likely(!PageTransCompound(page))) {
2654                 mapcount = atomic_read(&page->_mapcount) + 1;
2655                 if (total_mapcount)
2656                         *total_mapcount = mapcount;
2657                 return mapcount;
2658         }
2659
2660         page = compound_head(page);
2661
2662         _total_mapcount = ret = 0;
2663         for (i = 0; i < HPAGE_PMD_NR; i++) {
2664                 mapcount = atomic_read(&page[i]._mapcount) + 1;
2665                 ret = max(ret, mapcount);
2666                 _total_mapcount += mapcount;
2667         }
2668         if (PageDoubleMap(page)) {
2669                 ret -= 1;
2670                 _total_mapcount -= HPAGE_PMD_NR;
2671         }
2672         mapcount = compound_mapcount(page);
2673         ret += mapcount;
2674         _total_mapcount += mapcount;
2675         if (total_mapcount)
2676                 *total_mapcount = _total_mapcount;
2677         return ret;
2678 }
2679
2680 /* Racy check whether the huge page can be split */
2681 bool can_split_huge_page(struct page *page, int *pextra_pins)
2682 {
2683         int extra_pins;
2684
2685         /* Additional pins from page cache */
2686         if (PageAnon(page))
2687                 extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0;
2688         else
2689                 extra_pins = HPAGE_PMD_NR;
2690         if (pextra_pins)
2691                 *pextra_pins = extra_pins;
2692         return total_mapcount(page) == page_count(page) - extra_pins - 1;
2693 }
2694
2695 /*
2696  * This function splits huge page into normal pages. @page can point to any
2697  * subpage of huge page to split. Split doesn't change the position of @page.
2698  *
2699  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2700  * The huge page must be locked.
2701  *
2702  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2703  *
2704  * Both head page and tail pages will inherit mapping, flags, and so on from
2705  * the hugepage.
2706  *
2707  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2708  * they are not mapped.
2709  *
2710  * Returns 0 if the hugepage is split successfully.
2711  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2712  * us.
2713  */
2714 int split_huge_page_to_list(struct page *page, struct list_head *list)
2715 {
2716         struct page *head = compound_head(page);
2717         struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2718         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2719         struct anon_vma *anon_vma = NULL;
2720         struct address_space *mapping = NULL;
2721         int count, mapcount, extra_pins, ret;
2722         bool mlocked;
2723         unsigned long flags;
2724         pgoff_t end;
2725
2726         VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
2727         VM_BUG_ON_PAGE(!PageLocked(page), page);
2728         VM_BUG_ON_PAGE(!PageCompound(page), page);
2729
2730         if (PageWriteback(page))
2731                 return -EBUSY;
2732
2733         if (PageAnon(head)) {
2734                 /*
2735                  * The caller does not necessarily hold an mmap_sem that would
2736                  * prevent the anon_vma disappearing so we first we take a
2737                  * reference to it and then lock the anon_vma for write. This
2738                  * is similar to page_lock_anon_vma_read except the write lock
2739                  * is taken to serialise against parallel split or collapse
2740                  * operations.
2741                  */
2742                 anon_vma = page_get_anon_vma(head);
2743                 if (!anon_vma) {
2744                         ret = -EBUSY;
2745                         goto out;
2746                 }
2747                 end = -1;
2748                 mapping = NULL;
2749                 anon_vma_lock_write(anon_vma);
2750         } else {
2751                 mapping = head->mapping;
2752
2753                 /* Truncated ? */
2754                 if (!mapping) {
2755                         ret = -EBUSY;
2756                         goto out;
2757                 }
2758
2759                 anon_vma = NULL;
2760                 i_mmap_lock_read(mapping);
2761
2762                 /*
2763                  *__split_huge_page() may need to trim off pages beyond EOF:
2764                  * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2765                  * which cannot be nested inside the page tree lock. So note
2766                  * end now: i_size itself may be changed at any moment, but
2767                  * head page lock is good enough to serialize the trimming.
2768                  */
2769                 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2770         }
2771
2772         /*
2773          * Racy check if we can split the page, before unmap_page() will
2774          * split PMDs
2775          */
2776         if (!can_split_huge_page(head, &extra_pins)) {
2777                 ret = -EBUSY;
2778                 goto out_unlock;
2779         }
2780
2781         mlocked = PageMlocked(page);
2782         unmap_page(head);
2783         VM_BUG_ON_PAGE(compound_mapcount(head), head);
2784
2785         /* Make sure the page is not on per-CPU pagevec as it takes pin */
2786         if (mlocked)
2787                 lru_add_drain();
2788
2789         /* prevent PageLRU to go away from under us, and freeze lru stats */
2790         spin_lock_irqsave(&pgdata->lru_lock, flags);
2791
2792         if (mapping) {
2793                 XA_STATE(xas, &mapping->i_pages, page_index(head));
2794
2795                 /*
2796                  * Check if the head page is present in page cache.
2797                  * We assume all tail are present too, if head is there.
2798                  */
2799                 xa_lock(&mapping->i_pages);
2800                 if (xas_load(&xas) != head)
2801                         goto fail;
2802         }
2803
2804         /* Prevent deferred_split_scan() touching ->_refcount */
2805         spin_lock(&ds_queue->split_queue_lock);
2806         count = page_count(head);
2807         mapcount = total_mapcount(head);
2808         if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2809                 if (!list_empty(page_deferred_list(head))) {
2810                         ds_queue->split_queue_len--;
2811                         list_del(page_deferred_list(head));
2812                 }
2813                 if (mapping) {
2814                         if (PageSwapBacked(page))
2815                                 __dec_node_page_state(page, NR_SHMEM_THPS);
2816                         else
2817                                 __dec_node_page_state(page, NR_FILE_THPS);
2818                 }
2819
2820                 spin_unlock(&ds_queue->split_queue_lock);
2821                 __split_huge_page(page, list, end, flags);
2822                 if (PageSwapCache(head)) {
2823                         swp_entry_t entry = { .val = page_private(head) };
2824
2825                         ret = split_swap_cluster(entry);
2826                 } else
2827                         ret = 0;
2828         } else {
2829                 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2830                         pr_alert("total_mapcount: %u, page_count(): %u\n",
2831                                         mapcount, count);
2832                         if (PageTail(page))
2833                                 dump_page(head, NULL);
2834                         dump_page(page, "total_mapcount(head) > 0");
2835                         BUG();
2836                 }
2837                 spin_unlock(&ds_queue->split_queue_lock);
2838 fail:           if (mapping)
2839                         xa_unlock(&mapping->i_pages);
2840                 spin_unlock_irqrestore(&pgdata->lru_lock, flags);
2841                 remap_page(head);
2842                 ret = -EBUSY;
2843         }
2844
2845 out_unlock:
2846         if (anon_vma) {
2847                 anon_vma_unlock_write(anon_vma);
2848                 put_anon_vma(anon_vma);
2849         }
2850         if (mapping)
2851                 i_mmap_unlock_read(mapping);
2852 out:
2853         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2854         return ret;
2855 }
2856
2857 void free_transhuge_page(struct page *page)
2858 {
2859         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2860         unsigned long flags;
2861
2862         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2863         if (!list_empty(page_deferred_list(page))) {
2864                 ds_queue->split_queue_len--;
2865                 list_del(page_deferred_list(page));
2866         }
2867         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2868         free_compound_page(page);
2869 }
2870
2871 void deferred_split_huge_page(struct page *page)
2872 {
2873         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2874 #ifdef CONFIG_MEMCG
2875         struct mem_cgroup *memcg = compound_head(page)->mem_cgroup;
2876 #endif
2877         unsigned long flags;
2878
2879         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2880
2881         /*
2882          * The try_to_unmap() in page reclaim path might reach here too,
2883          * this may cause a race condition to corrupt deferred split queue.
2884          * And, if page reclaim is already handling the same page, it is
2885          * unnecessary to handle it again in shrinker.
2886          *
2887          * Check PageSwapCache to determine if the page is being
2888          * handled by page reclaim since THP swap would add the page into
2889          * swap cache before calling try_to_unmap().
2890          */
2891         if (PageSwapCache(page))
2892                 return;
2893
2894         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2895         if (list_empty(page_deferred_list(page))) {
2896                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2897                 list_add_tail(page_deferred_list(page), &ds_queue->split_queue);
2898                 ds_queue->split_queue_len++;
2899 #ifdef CONFIG_MEMCG
2900                 if (memcg)
2901                         memcg_set_shrinker_bit(memcg, page_to_nid(page),
2902                                                deferred_split_shrinker.id);
2903 #endif
2904         }
2905         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2906 }
2907
2908 static unsigned long deferred_split_count(struct shrinker *shrink,
2909                 struct shrink_control *sc)
2910 {
2911         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2912         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2913
2914 #ifdef CONFIG_MEMCG
2915         if (sc->memcg)
2916                 ds_queue = &sc->memcg->deferred_split_queue;
2917 #endif
2918         return READ_ONCE(ds_queue->split_queue_len);
2919 }
2920
2921 static unsigned long deferred_split_scan(struct shrinker *shrink,
2922                 struct shrink_control *sc)
2923 {
2924         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2925         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2926         unsigned long flags;
2927         LIST_HEAD(list), *pos, *next;
2928         struct page *page;
2929         int split = 0;
2930
2931 #ifdef CONFIG_MEMCG
2932         if (sc->memcg)
2933                 ds_queue = &sc->memcg->deferred_split_queue;
2934 #endif
2935
2936         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2937         /* Take pin on all head pages to avoid freeing them under us */
2938         list_for_each_safe(pos, next, &ds_queue->split_queue) {
2939                 page = list_entry((void *)pos, struct page, mapping);
2940                 page = compound_head(page);
2941                 if (get_page_unless_zero(page)) {
2942                         list_move(page_deferred_list(page), &list);
2943                 } else {
2944                         /* We lost race with put_compound_page() */
2945                         list_del_init(page_deferred_list(page));
2946                         ds_queue->split_queue_len--;
2947                 }
2948                 if (!--sc->nr_to_scan)
2949                         break;
2950         }
2951         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2952
2953         list_for_each_safe(pos, next, &list) {
2954                 page = list_entry((void *)pos, struct page, mapping);
2955                 if (!trylock_page(page))
2956                         goto next;
2957                 /* split_huge_page() removes page from list on success */
2958                 if (!split_huge_page(page))
2959                         split++;
2960                 unlock_page(page);
2961 next:
2962                 put_page(page);
2963         }
2964
2965         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2966         list_splice_tail(&list, &ds_queue->split_queue);
2967         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2968
2969         /*
2970          * Stop shrinker if we didn't split any page, but the queue is empty.
2971          * This can happen if pages were freed under us.
2972          */
2973         if (!split && list_empty(&ds_queue->split_queue))
2974                 return SHRINK_STOP;
2975         return split;
2976 }
2977
2978 static struct shrinker deferred_split_shrinker = {
2979         .count_objects = deferred_split_count,
2980         .scan_objects = deferred_split_scan,
2981         .seeks = DEFAULT_SEEKS,
2982         .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
2983                  SHRINKER_NONSLAB,
2984 };
2985
2986 #ifdef CONFIG_DEBUG_FS
2987 static int split_huge_pages_set(void *data, u64 val)
2988 {
2989         struct zone *zone;
2990         struct page *page;
2991         unsigned long pfn, max_zone_pfn;
2992         unsigned long total = 0, split = 0;
2993
2994         if (val != 1)
2995                 return -EINVAL;
2996
2997         for_each_populated_zone(zone) {
2998                 max_zone_pfn = zone_end_pfn(zone);
2999                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
3000                         if (!pfn_valid(pfn))
3001                                 continue;
3002
3003                         page = pfn_to_page(pfn);
3004                         if (!get_page_unless_zero(page))
3005                                 continue;
3006
3007                         if (zone != page_zone(page))
3008                                 goto next;
3009
3010                         if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
3011                                 goto next;
3012
3013                         total++;
3014                         lock_page(page);
3015                         if (!split_huge_page(page))
3016                                 split++;
3017                         unlock_page(page);
3018 next:
3019                         put_page(page);
3020                 }
3021         }
3022
3023         pr_info("%lu of %lu THP split\n", split, total);
3024
3025         return 0;
3026 }
3027 DEFINE_DEBUGFS_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
3028                 "%llu\n");
3029
3030 static int __init split_huge_pages_debugfs(void)
3031 {
3032         debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3033                             &split_huge_pages_fops);
3034         return 0;
3035 }
3036 late_initcall(split_huge_pages_debugfs);
3037 #endif
3038
3039 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
3040 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
3041                 struct page *page)
3042 {
3043         struct vm_area_struct *vma = pvmw->vma;
3044         struct mm_struct *mm = vma->vm_mm;
3045         unsigned long address = pvmw->address;
3046         pmd_t pmdval;
3047         swp_entry_t entry;
3048         pmd_t pmdswp;
3049
3050         if (!(pvmw->pmd && !pvmw->pte))
3051                 return;
3052
3053         flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
3054         pmdval = *pvmw->pmd;
3055         pmdp_invalidate(vma, address, pvmw->pmd);
3056         if (pmd_dirty(pmdval))
3057                 set_page_dirty(page);
3058         entry = make_migration_entry(page, pmd_write(pmdval));
3059         pmdswp = swp_entry_to_pmd(entry);
3060         if (pmd_soft_dirty(pmdval))
3061                 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
3062         set_pmd_at(mm, address, pvmw->pmd, pmdswp);
3063         page_remove_rmap(page, true);
3064         put_page(page);
3065 }
3066
3067 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
3068 {
3069         struct vm_area_struct *vma = pvmw->vma;
3070         struct mm_struct *mm = vma->vm_mm;
3071         unsigned long address = pvmw->address;
3072         unsigned long mmun_start = address & HPAGE_PMD_MASK;
3073         pmd_t pmde;
3074         swp_entry_t entry;
3075
3076         if (!(pvmw->pmd && !pvmw->pte))
3077                 return;
3078
3079         entry = pmd_to_swp_entry(*pvmw->pmd);
3080         get_page(new);
3081         pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
3082         if (pmd_swp_soft_dirty(*pvmw->pmd))
3083                 pmde = pmd_mksoft_dirty(pmde);
3084         if (is_write_migration_entry(entry))
3085                 pmde = maybe_pmd_mkwrite(pmde, vma);
3086
3087         flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
3088         if (PageAnon(new))
3089                 page_add_anon_rmap(new, vma, mmun_start, true);
3090         else
3091                 page_add_file_rmap(new, true);
3092         set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
3093         if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new))
3094                 mlock_vma_page(new);
3095         update_mmu_cache_pmd(vma, address, pvmw->pmd);
3096 }
3097 #endif