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