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