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