mm/huge_memory.c

   1 /*
   2  *  Copyright (C) 2009  Red Hat, Inc.
   3  *
   4  *  This work is licensed under the terms of the GNU GPL, version 2. See
   5  *  the COPYING file in the top-level directory.
   6  */
   7
   8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   9
  10 #include <linux/mm.h>
  11 #include <linux/sched.h>
  12 #include <linux/sched/coredump.h>
  13 #include <linux/sched/numa_balancing.h>
  14 #include <linux/highmem.h>
  15 #include <linux/hugetlb.h>
  16 #include <linux/mmu_notifier.h>
  17 #include <linux/rmap.h>
  18 #include <linux/swap.h>
  19 #include <linux/shrinker.h>
  20 #include <linux/mm_inline.h>
  21 #include <linux/swapops.h>
  22 #include <linux/dax.h>
  23 #include <linux/khugepaged.h>
  24 #include <linux/freezer.h>
  25 #include <linux/pfn_t.h>
  26 #include <linux/mman.h>
  27 #include <linux/memremap.h>
  28 #include <linux/pagemap.h>
  29 #include <linux/debugfs.h>
  30 #include <linux/migrate.h>
  31 #include <linux/hashtable.h>
  32 #include <linux/userfaultfd_k.h>
  33 #include <linux/page_idle.h>
  34 #include <linux/shmem_fs.h>
  35 #include <linux/oom.h>
  36 #include <linux/numa.h>
  37
  38 #include <asm/tlb.h>
  39 #include <asm/pgalloc.h>
  40 #include "internal.h"
  41
  42 /*
  43  * By default, transparent hugepage support is disabled in order to avoid
  44  * risking an increased memory footprint for applications that are not
  45  * guaranteed to benefit from it. When transparent hugepage support is
  46  * enabled, it is for all mappings, and khugepaged scans all mappings.
  47  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  48  * for all hugepage allocations.
  49  */
  50 unsigned long transparent_hugepage_flags __read_mostly =
  51 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  52         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  53 #endif
  54 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  55         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  56 #endif
  57         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
  58         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  59         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  60
  61 static struct shrinker deferred_split_shrinker;
  62
  63 static atomic_t huge_zero_refcount;
  64 struct page *huge_zero_page __read_mostly;
  65
  66 bool transparent_hugepage_enabled(struct vm_area_struct *vma)
  67 {
  68         if (vma_is_anonymous(vma))
  69                 return __transparent_hugepage_enabled(vma);
  70         if (vma_is_shmem(vma) && shmem_huge_enabled(vma))
  71                 return __transparent_hugepage_enabled(vma);
  72
  73         return false;
  74 }
  75
  76 static struct page *get_huge_zero_page(void)
  77 {
  78         struct page *zero_page;
  79 retry:
  80         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
  81                 return READ_ONCE(huge_zero_page);
  82
  83         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
  84                         HPAGE_PMD_ORDER);
  85         if (!zero_page) {
  86                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
  87                 return NULL;
  88         }
  89         count_vm_event(THP_ZERO_PAGE_ALLOC);
  90         preempt_disable();
  91         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
  92                 preempt_enable();
  93                 __free_pages(zero_page, compound_order(zero_page));
  94                 goto retry;
  95         }
  96
  97         /* We take additional reference here. It will be put back by shrinker */
  98         atomic_set(&huge_zero_refcount, 2);
  99         preempt_enable();
 100         return READ_ONCE(huge_zero_page);
 101 }
 102
 103 static void put_huge_zero_page(void)
 104 {
 105         /*
 106          * Counter should never go to zero here. Only shrinker can put
 107          * last reference.
 108          */
 109         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 110 }
 111
 112 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
 113 {
 114         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 115                 return READ_ONCE(huge_zero_page);
 116
 117         if (!get_huge_zero_page())
 118                 return NULL;
 119
 120         if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 121                 put_huge_zero_page();
 122
 123         return READ_ONCE(huge_zero_page);
 124 }
 125
 126 void mm_put_huge_zero_page(struct mm_struct *mm)
 127 {
 128         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 129                 put_huge_zero_page();
 130 }
 131
 132 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 133                                         struct shrink_control *sc)
 134 {
 135         /* we can free zero page only if last reference remains */
 136         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 137 }
 138
 139 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 140                                        struct shrink_control *sc)
 141 {
 142         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 143                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 144                 BUG_ON(zero_page == NULL);
 145                 __free_pages(zero_page, compound_order(zero_page));
 146                 return HPAGE_PMD_NR;
 147         }
 148
 149         return 0;
 150 }
 151
 152 static struct shrinker huge_zero_page_shrinker = {
 153         .count_objects = shrink_huge_zero_page_count,
 154         .scan_objects = shrink_huge_zero_page_scan,
 155         .seeks = DEFAULT_SEEKS,
 156 };
 157
 158 #ifdef CONFIG_SYSFS
 159 static ssize_t enabled_show(struct kobject *kobj,
 160                             struct kobj_attribute *attr, char *buf)
 161 {
 162         if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
 163                 return sprintf(buf, "[always] madvise never\n");
 164         else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
 165                 return sprintf(buf, "always [madvise] never\n");
 166         else
 167                 return sprintf(buf, "always madvise [never]\n");
 168 }
 169
 170 static ssize_t enabled_store(struct kobject *kobj,
 171                              struct kobj_attribute *attr,
 172                              const char *buf, size_t count)
 173 {
 174         ssize_t ret = count;
 175
 176         if (!memcmp("always", buf,
 177                     min(sizeof("always")-1, count))) {
 178                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 179                 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 180         } else if (!memcmp("madvise", buf,
 181                            min(sizeof("madvise")-1, count))) {
 182                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 183                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 184         } else if (!memcmp("never", buf,
 185                            min(sizeof("never")-1, count))) {
 186                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 187                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 188         } else
 189                 ret = -EINVAL;
 190
 191         if (ret > 0) {
 192                 int err = start_stop_khugepaged();
 193                 if (err)
 194                         ret = err;
 195         }
 196         return ret;
 197 }
 198 static struct kobj_attribute enabled_attr =
 199         __ATTR(enabled, 0644, enabled_show, enabled_store);
 200
 201 ssize_t single_hugepage_flag_show(struct kobject *kobj,
 202                                 struct kobj_attribute *attr, char *buf,
 203                                 enum transparent_hugepage_flag flag)
 204 {
 205         return sprintf(buf, "%d\n",
 206                        !!test_bit(flag, &transparent_hugepage_flags));
 207 }
 208
 209 ssize_t single_hugepage_flag_store(struct kobject *kobj,
 210                                  struct kobj_attribute *attr,
 211                                  const char *buf, size_t count,
 212                                  enum transparent_hugepage_flag flag)
 213 {
 214         unsigned long value;
 215         int ret;
 216
 217         ret = kstrtoul(buf, 10, &value);
 218         if (ret < 0)
 219                 return ret;
 220         if (value > 1)
 221                 return -EINVAL;
 222
 223         if (value)
 224                 set_bit(flag, &transparent_hugepage_flags);
 225         else
 226                 clear_bit(flag, &transparent_hugepage_flags);
 227
 228         return count;
 229 }
 230
 231 static ssize_t defrag_show(struct kobject *kobj,
 232                            struct kobj_attribute *attr, char *buf)
 233 {
 234         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 235                 return sprintf(buf, "[always] defer defer+madvise madvise never\n");
 236         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 237                 return sprintf(buf, "always [defer] defer+madvise madvise never\n");
 238         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 239                 return sprintf(buf, "always defer [defer+madvise] madvise never\n");
 240         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 241                 return sprintf(buf, "always defer defer+madvise [madvise] never\n");
 242         return sprintf(buf, "always defer defer+madvise madvise [never]\n");
 243 }
 244
 245 static ssize_t defrag_store(struct kobject *kobj,
 246                             struct kobj_attribute *attr,
 247                             const char *buf, size_t count)
 248 {
 249         if (!memcmp("always", buf,
 250                     min(sizeof("always")-1, count))) {
 251                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 252                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 253                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 254                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 255         } else if (!memcmp("defer+madvise", buf,
 256                     min(sizeof("defer+madvise")-1, count))) {
 257                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 258                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 259                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 260                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 261         } else if (!memcmp("defer", buf,
 262                     min(sizeof("defer")-1, count))) {
 263                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 264                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 265                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 266                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 267         } else if (!memcmp("madvise", buf,
 268                            min(sizeof("madvise")-1, count))) {
 269                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 270                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 271                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 272                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 273         } else if (!memcmp("never", buf,
 274                            min(sizeof("never")-1, count))) {
 275                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 276                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 277                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 278                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 279         } else
 280                 return -EINVAL;
 281
 282         return count;
 283 }
 284 static struct kobj_attribute defrag_attr =
 285         __ATTR(defrag, 0644, defrag_show, defrag_store);
 286
 287 static ssize_t use_zero_page_show(struct kobject *kobj,
 288                 struct kobj_attribute *attr, char *buf)
 289 {
 290         return single_hugepage_flag_show(kobj, attr, buf,
 291                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 292 }
 293 static ssize_t use_zero_page_store(struct kobject *kobj,
 294                 struct kobj_attribute *attr, const char *buf, size_t count)
 295 {
 296         return single_hugepage_flag_store(kobj, attr, buf, count,
 297                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 298 }
 299 static struct kobj_attribute use_zero_page_attr =
 300         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 301
 302 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
 303                 struct kobj_attribute *attr, char *buf)
 304 {
 305         return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE);
 306 }
 307 static struct kobj_attribute hpage_pmd_size_attr =
 308         __ATTR_RO(hpage_pmd_size);
 309
 310 #ifdef CONFIG_DEBUG_VM
 311 static ssize_t debug_cow_show(struct kobject *kobj,
 312                                 struct kobj_attribute *attr, char *buf)
 313 {
 314         return single_hugepage_flag_show(kobj, attr, buf,
 315                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 316 }
 317 static ssize_t debug_cow_store(struct kobject *kobj,
 318                                struct kobj_attribute *attr,
 319                                const char *buf, size_t count)
 320 {
 321         return single_hugepage_flag_store(kobj, attr, buf, count,
 322                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 323 }
 324 static struct kobj_attribute debug_cow_attr =
 325         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 326 #endif /* CONFIG_DEBUG_VM */
 327
 328 static struct attribute *hugepage_attr[] = {
 329         &enabled_attr.attr,
 330         &defrag_attr.attr,
 331         &use_zero_page_attr.attr,
 332         &hpage_pmd_size_attr.attr,
 333 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
 334         &shmem_enabled_attr.attr,
 335 #endif
 336 #ifdef CONFIG_DEBUG_VM
 337         &debug_cow_attr.attr,
 338 #endif
 339         NULL,
 340 };
 341
 342 static const struct attribute_group hugepage_attr_group = {
 343         .attrs = hugepage_attr,
 344 };
 345
 346 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 347 {
 348         int err;
 349
 350         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 351         if (unlikely(!*hugepage_kobj)) {
 352                 pr_err("failed to create transparent hugepage kobject\n");
 353                 return -ENOMEM;
 354         }
 355
 356         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 357         if (err) {
 358                 pr_err("failed to register transparent hugepage group\n");
 359                 goto delete_obj;
 360         }
 361
 362         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 363         if (err) {
 364                 pr_err("failed to register transparent hugepage group\n");
 365                 goto remove_hp_group;
 366         }
 367
 368         return 0;
 369
 370 remove_hp_group:
 371         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 372 delete_obj:
 373         kobject_put(*hugepage_kobj);
 374         return err;
 375 }
 376
 377 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 378 {
 379         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 380         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 381         kobject_put(hugepage_kobj);
 382 }
 383 #else
 384 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 385 {
 386         return 0;
 387 }
 388
 389 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 390 {
 391 }
 392 #endif /* CONFIG_SYSFS */
 393
 394 static int __init hugepage_init(void)
 395 {
 396         int err;
 397         struct kobject *hugepage_kobj;
 398
 399         if (!has_transparent_hugepage()) {
 400                 transparent_hugepage_flags = 0;
 401                 return -EINVAL;
 402         }
 403
 404         /*
 405          * hugepages can't be allocated by the buddy allocator
 406          */
 407         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
 408         /*
 409          * we use page->mapping and page->index in second tail page
 410          * as list_head: assuming THP order >= 2
 411          */
 412         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 413
 414         err = hugepage_init_sysfs(&hugepage_kobj);
 415         if (err)
 416                 goto err_sysfs;
 417
 418         err = khugepaged_init();
 419         if (err)
 420                 goto err_slab;
 421
 422         err = register_shrinker(&huge_zero_page_shrinker);
 423         if (err)
 424                 goto err_hzp_shrinker;
 425         err = register_shrinker(&deferred_split_shrinker);
 426         if (err)
 427                 goto err_split_shrinker;
 428
 429         /*
 430          * By default disable transparent hugepages on smaller systems,
 431          * where the extra memory used could hurt more than TLB overhead
 432          * is likely to save.  The admin can still enable it through /sys.
 433          */
 434         if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
 435                 transparent_hugepage_flags = 0;
 436                 return 0;
 437         }
 438
 439         err = start_stop_khugepaged();
 440         if (err)
 441                 goto err_khugepaged;
 442
 443         return 0;
 444 err_khugepaged:
 445         unregister_shrinker(&deferred_split_shrinker);
 446 err_split_shrinker:
 447         unregister_shrinker(&huge_zero_page_shrinker);
 448 err_hzp_shrinker:
 449         khugepaged_destroy();
 450 err_slab:
 451         hugepage_exit_sysfs(hugepage_kobj);
 452 err_sysfs:
 453         return err;
 454 }
 455 subsys_initcall(hugepage_init);
 456
 457 static int __init setup_transparent_hugepage(char *str)
 458 {
 459         int ret = 0;
 460         if (!str)
 461                 goto out;
 462         if (!strcmp(str, "always")) {
 463                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 464                         &transparent_hugepage_flags);
 465                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 466                           &transparent_hugepage_flags);
 467                 ret = 1;
 468         } else if (!strcmp(str, "madvise")) {
 469                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 470                           &transparent_hugepage_flags);
 471                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 472                         &transparent_hugepage_flags);
 473                 ret = 1;
 474         } else if (!strcmp(str, "never")) {
 475                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 476                           &transparent_hugepage_flags);
 477                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 478                           &transparent_hugepage_flags);
 479                 ret = 1;
 480         }
 481 out:
 482         if (!ret)
 483                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 484         return ret;
 485 }
 486 __setup("transparent_hugepage=", setup_transparent_hugepage);
 487
 488 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 489 {
 490         if (likely(vma->vm_flags & VM_WRITE))
 491                 pmd = pmd_mkwrite(pmd);
 492         return pmd;
 493 }
 494
 495 static inline struct list_head *page_deferred_list(struct page *page)
 496 {
 497         /* ->lru in the tail pages is occupied by compound_head. */
 498         return &page[2].deferred_list;
 499 }
 500
 501 void prep_transhuge_page(struct page *page)
 502 {
 503         /*
 504          * we use page->mapping and page->indexlru in second tail page
 505          * as list_head: assuming THP order >= 2
 506          */
 507
 508         INIT_LIST_HEAD(page_deferred_list(page));
 509         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 510 }
 511
 512 unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len,
 513                 loff_t off, unsigned long flags, unsigned long size)
 514 {
 515         unsigned long addr;
 516         loff_t off_end = off + len;
 517         loff_t off_align = round_up(off, size);
 518         unsigned long len_pad;
 519
 520         if (off_end <= off_align || (off_end - off_align) < size)
 521                 return 0;
 522
 523         len_pad = len + size;
 524         if (len_pad < len || (off + len_pad) < off)
 525                 return 0;
 526
 527         addr = current->mm->get_unmapped_area(filp, 0, len_pad,
 528                                               off >> PAGE_SHIFT, flags);
 529         if (IS_ERR_VALUE(addr))
 530                 return 0;
 531
 532         addr += (off - addr) & (size - 1);
 533         return addr;
 534 }
 535
 536 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 537                 unsigned long len, unsigned long pgoff, unsigned long flags)
 538 {
 539         loff_t off = (loff_t)pgoff << PAGE_SHIFT;
 540
 541         if (addr)
 542                 goto out;
 543         if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
 544                 goto out;
 545
 546         addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE);
 547         if (addr)
 548                 return addr;
 549
 550  out:
 551         return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
 552 }
 553 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
 554
 555 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
 556                         struct page *page, gfp_t gfp)
 557 {
 558         struct vm_area_struct *vma = vmf->vma;
 559         struct mem_cgroup *memcg;
 560         pgtable_t pgtable;
 561         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 562         vm_fault_t ret = 0;
 563
 564         VM_BUG_ON_PAGE(!PageCompound(page), page);
 565
 566         if (mem_cgroup_try_charge_delay(page, vma->vm_mm, gfp, &memcg, true)) {
 567                 put_page(page);
 568                 count_vm_event(THP_FAULT_FALLBACK);
 569                 return VM_FAULT_FALLBACK;
 570         }
 571
 572         pgtable = pte_alloc_one(vma->vm_mm);
 573         if (unlikely(!pgtable)) {
 574                 ret = VM_FAULT_OOM;
 575                 goto release;
 576         }
 577
 578         clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
 579         /*
 580          * The memory barrier inside __SetPageUptodate makes sure that
 581          * clear_huge_page writes become visible before the set_pmd_at()
 582          * write.
 583          */
 584         __SetPageUptodate(page);
 585
 586         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 587         if (unlikely(!pmd_none(*vmf->pmd))) {
 588                 goto unlock_release;
 589         } else {
 590                 pmd_t entry;
 591
 592                 ret = check_stable_address_space(vma->vm_mm);
 593                 if (ret)
 594                         goto unlock_release;
 595
 596                 /* Deliver the page fault to userland */
 597                 if (userfaultfd_missing(vma)) {
 598                         vm_fault_t ret2;
 599
 600                         spin_unlock(vmf->ptl);
 601                         mem_cgroup_cancel_charge(page, memcg, true);
 602                         put_page(page);
 603                         pte_free(vma->vm_mm, pgtable);
 604                         ret2 = handle_userfault(vmf, VM_UFFD_MISSING);
 605                         VM_BUG_ON(ret2 & VM_FAULT_FALLBACK);
 606                         return ret2;
 607                 }
 608
 609                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 610                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 611                 page_add_new_anon_rmap(page, vma, haddr, true);
 612                 mem_cgroup_commit_charge(page, memcg, false, true);
 613                 lru_cache_add_active_or_unevictable(page, vma);
 614                 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
 615                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
 616                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 617                 mm_inc_nr_ptes(vma->vm_mm);
 618                 spin_unlock(vmf->ptl);
 619                 count_vm_event(THP_FAULT_ALLOC);
 620                 count_memcg_events(memcg, THP_FAULT_ALLOC, 1);
 621         }
 622
 623         return 0;
 624 unlock_release:
 625         spin_unlock(vmf->ptl);
 626 release:
 627         if (pgtable)
 628                 pte_free(vma->vm_mm, pgtable);
 629         mem_cgroup_cancel_charge(page, memcg, true);
 630         put_page(page);
 631         return ret;
 632
 633 }
 634
 635 /*
 636  * always: directly stall for all thp allocations
 637  * defer: wake kswapd and fail if not immediately available
 638  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
 639  *                fail if not immediately available
 640  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
 641  *          available
 642  * never: never stall for any thp allocation
 643  */
 644 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
 645 {
 646         const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
 647
 648         /* Always do synchronous compaction */
 649         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 650                 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
 651
 652         /* Kick kcompactd and fail quickly */
 653         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 654                 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
 655
 656         /* Synchronous compaction if madvised, otherwise kick kcompactd */
 657         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 658                 return GFP_TRANSHUGE_LIGHT |
 659                         (vma_madvised ? __GFP_DIRECT_RECLAIM :
 660                                         __GFP_KSWAPD_RECLAIM);
 661
 662         /* Only do synchronous compaction if madvised */
 663         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 664                 return GFP_TRANSHUGE_LIGHT |
 665                        (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
 666
 667         return GFP_TRANSHUGE_LIGHT;
 668 }
 669
 670 /* Caller must hold page table lock. */
 671 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 672                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 673                 struct page *zero_page)
 674 {
 675         pmd_t entry;
 676         if (!pmd_none(*pmd))
 677                 return false;
 678         entry = mk_pmd(zero_page, vma->vm_page_prot);
 679         entry = pmd_mkhuge(entry);
 680         if (pgtable)
 681                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 682         set_pmd_at(mm, haddr, pmd, entry);
 683         mm_inc_nr_ptes(mm);
 684         return true;
 685 }
 686
 687 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
 688 {
 689         struct vm_area_struct *vma = vmf->vma;
 690         gfp_t gfp;
 691         struct page *page;
 692         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 693
 694         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 695                 return VM_FAULT_FALLBACK;
 696         if (unlikely(anon_vma_prepare(vma)))
 697                 return VM_FAULT_OOM;
 698         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 699                 return VM_FAULT_OOM;
 700         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
 701                         !mm_forbids_zeropage(vma->vm_mm) &&
 702                         transparent_hugepage_use_zero_page()) {
 703                 pgtable_t pgtable;
 704                 struct page *zero_page;
 705                 bool set;
 706                 vm_fault_t ret;
 707                 pgtable = pte_alloc_one(vma->vm_mm);
 708                 if (unlikely(!pgtable))
 709                         return VM_FAULT_OOM;
 710                 zero_page = mm_get_huge_zero_page(vma->vm_mm);
 711                 if (unlikely(!zero_page)) {
 712                         pte_free(vma->vm_mm, pgtable);
 713                         count_vm_event(THP_FAULT_FALLBACK);
 714                         return VM_FAULT_FALLBACK;
 715                 }
 716                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 717                 ret = 0;
 718                 set = false;
 719                 if (pmd_none(*vmf->pmd)) {
 720                         ret = check_stable_address_space(vma->vm_mm);
 721                         if (ret) {
 722                                 spin_unlock(vmf->ptl);
 723                         } else if (userfaultfd_missing(vma)) {
 724                                 spin_unlock(vmf->ptl);
 725                                 ret = handle_userfault(vmf, VM_UFFD_MISSING);
 726                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 727                         } else {
 728                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
 729                                                    haddr, vmf->pmd, zero_page);
 730                                 spin_unlock(vmf->ptl);
 731                                 set = true;
 732                         }
 733                 } else
 734                         spin_unlock(vmf->ptl);
 735                 if (!set)
 736                         pte_free(vma->vm_mm, pgtable);
 737                 return ret;
 738         }
 739         gfp = alloc_hugepage_direct_gfpmask(vma);
 740         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 741         if (unlikely(!page)) {
 742                 count_vm_event(THP_FAULT_FALLBACK);
 743                 return VM_FAULT_FALLBACK;
 744         }
 745         prep_transhuge_page(page);
 746         return __do_huge_pmd_anonymous_page(vmf, page, gfp);
 747 }
 748
 749 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 750                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
 751                 pgtable_t pgtable)
 752 {
 753         struct mm_struct *mm = vma->vm_mm;
 754         pmd_t entry;
 755         spinlock_t *ptl;
 756
 757         ptl = pmd_lock(mm, pmd);
 758         if (!pmd_none(*pmd)) {
 759                 if (write) {
 760                         if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
 761                                 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
 762                                 goto out_unlock;
 763                         }
 764                         entry = pmd_mkyoung(*pmd);
 765                         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 766                         if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
 767                                 update_mmu_cache_pmd(vma, addr, pmd);
 768                 }
 769
 770                 goto out_unlock;
 771         }
 772
 773         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 774         if (pfn_t_devmap(pfn))
 775                 entry = pmd_mkdevmap(entry);
 776         if (write) {
 777                 entry = pmd_mkyoung(pmd_mkdirty(entry));
 778                 entry = maybe_pmd_mkwrite(entry, vma);
 779         }
 780
 781         if (pgtable) {
 782                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 783                 mm_inc_nr_ptes(mm);
 784                 pgtable = NULL;
 785         }
 786
 787         set_pmd_at(mm, addr, pmd, entry);
 788         update_mmu_cache_pmd(vma, addr, pmd);
 789
 790 out_unlock:
 791         spin_unlock(ptl);
 792         if (pgtable)
 793                 pte_free(mm, pgtable);
 794 }
 795
 796 vm_fault_t vmf_insert_pfn_pmd(struct vm_fault *vmf, pfn_t pfn, bool write)
 797 {
 798         unsigned long addr = vmf->address & PMD_MASK;
 799         struct vm_area_struct *vma = vmf->vma;
 800         pgprot_t pgprot = vma->vm_page_prot;
 801         pgtable_t pgtable = NULL;
 802
 803         /*
 804          * If we had pmd_special, we could avoid all these restrictions,
 805          * but we need to be consistent with PTEs and architectures that
 806          * can't support a 'special' bit.
 807          */
 808         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 809                         !pfn_t_devmap(pfn));
 810         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 811                                                 (VM_PFNMAP|VM_MIXEDMAP));
 812         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 813
 814         if (addr < vma->vm_start || addr >= vma->vm_end)
 815                 return VM_FAULT_SIGBUS;
 816
 817         if (arch_needs_pgtable_deposit()) {
 818                 pgtable = pte_alloc_one(vma->vm_mm);
 819                 if (!pgtable)
 820                         return VM_FAULT_OOM;
 821         }
 822
 823         track_pfn_insert(vma, &pgprot, pfn);
 824
 825         insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
 826         return VM_FAULT_NOPAGE;
 827 }
 828 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
 829
 830 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 831 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
 832 {
 833         if (likely(vma->vm_flags & VM_WRITE))
 834                 pud = pud_mkwrite(pud);
 835         return pud;
 836 }
 837
 838 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 839                 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
 840 {
 841         struct mm_struct *mm = vma->vm_mm;
 842         pud_t entry;
 843         spinlock_t *ptl;
 844
 845         ptl = pud_lock(mm, pud);
 846         if (!pud_none(*pud)) {
 847                 if (write) {
 848                         if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
 849                                 WARN_ON_ONCE(!is_huge_zero_pud(*pud));
 850                                 goto out_unlock;
 851                         }
 852                         entry = pud_mkyoung(*pud);
 853                         entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
 854                         if (pudp_set_access_flags(vma, addr, pud, entry, 1))
 855                                 update_mmu_cache_pud(vma, addr, pud);
 856                 }
 857                 goto out_unlock;
 858         }
 859
 860         entry = pud_mkhuge(pfn_t_pud(pfn, prot));
 861         if (pfn_t_devmap(pfn))
 862                 entry = pud_mkdevmap(entry);
 863         if (write) {
 864                 entry = pud_mkyoung(pud_mkdirty(entry));
 865                 entry = maybe_pud_mkwrite(entry, vma);
 866         }
 867         set_pud_at(mm, addr, pud, entry);
 868         update_mmu_cache_pud(vma, addr, pud);
 869
 870 out_unlock:
 871         spin_unlock(ptl);
 872 }
 873
 874 vm_fault_t vmf_insert_pfn_pud(struct vm_fault *vmf, pfn_t pfn, bool write)
 875 {
 876         unsigned long addr = vmf->address & PUD_MASK;
 877         struct vm_area_struct *vma = vmf->vma;
 878         pgprot_t pgprot = vma->vm_page_prot;
 879
 880         /*
 881          * If we had pud_special, we could avoid all these restrictions,
 882          * but we need to be consistent with PTEs and architectures that
 883          * can't support a 'special' bit.
 884          */
 885         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 886                         !pfn_t_devmap(pfn));
 887         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 888                                                 (VM_PFNMAP|VM_MIXEDMAP));
 889         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 890
 891         if (addr < vma->vm_start || addr >= vma->vm_end)
 892                 return VM_FAULT_SIGBUS;
 893
 894         track_pfn_insert(vma, &pgprot, pfn);
 895
 896         insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
 897         return VM_FAULT_NOPAGE;
 898 }
 899 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
 900 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
 901
 902 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
 903                 pmd_t *pmd, int flags)
 904 {
 905         pmd_t _pmd;
 906
 907         _pmd = pmd_mkyoung(*pmd);
 908         if (flags & FOLL_WRITE)
 909                 _pmd = pmd_mkdirty(_pmd);
 910         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
 911                                 pmd, _pmd, flags & FOLL_WRITE))
 912                 update_mmu_cache_pmd(vma, addr, pmd);
 913 }
 914
 915 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 916                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
 917 {
 918         unsigned long pfn = pmd_pfn(*pmd);
 919         struct mm_struct *mm = vma->vm_mm;
 920         struct page *page;
 921
 922         assert_spin_locked(pmd_lockptr(mm, pmd));
 923
 924         /*
 925          * When we COW a devmap PMD entry, we split it into PTEs, so we should
 926          * not be in this function with `flags & FOLL_COW` set.
 927          */
 928         WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
 929
 930         if (flags & FOLL_WRITE && !pmd_write(*pmd))
 931                 return NULL;
 932
 933         if (pmd_present(*pmd) && pmd_devmap(*pmd))
 934                 /* pass */;
 935         else
 936                 return NULL;
 937
 938         if (flags & FOLL_TOUCH)
 939                 touch_pmd(vma, addr, pmd, flags);
 940
 941         /*
 942          * device mapped pages can only be returned if the
 943          * caller will manage the page reference count.
 944          */
 945         if (!(flags & FOLL_GET))
 946                 return ERR_PTR(-EEXIST);
 947
 948         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
 949         *pgmap = get_dev_pagemap(pfn, *pgmap);
 950         if (!*pgmap)
 951                 return ERR_PTR(-EFAULT);
 952         page = pfn_to_page(pfn);
 953         get_page(page);
 954
 955         return page;
 956 }
 957
 958 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 959                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 960                   struct vm_area_struct *vma)
 961 {
 962         spinlock_t *dst_ptl, *src_ptl;
 963         struct page *src_page;
 964         pmd_t pmd;
 965         pgtable_t pgtable = NULL;
 966         int ret = -ENOMEM;
 967
 968         /* Skip if can be re-fill on fault */
 969         if (!vma_is_anonymous(vma))
 970                 return 0;
 971
 972         pgtable = pte_alloc_one(dst_mm);
 973         if (unlikely(!pgtable))
 974                 goto out;
 975
 976         dst_ptl = pmd_lock(dst_mm, dst_pmd);
 977         src_ptl = pmd_lockptr(src_mm, src_pmd);
 978         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 979
 980         ret = -EAGAIN;
 981         pmd = *src_pmd;
 982
 983 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
 984         if (unlikely(is_swap_pmd(pmd))) {
 985                 swp_entry_t entry = pmd_to_swp_entry(pmd);
 986
 987                 VM_BUG_ON(!is_pmd_migration_entry(pmd));
 988                 if (is_write_migration_entry(entry)) {
 989                         make_migration_entry_read(&entry);
 990                         pmd = swp_entry_to_pmd(entry);
 991                         if (pmd_swp_soft_dirty(*src_pmd))
 992                                 pmd = pmd_swp_mksoft_dirty(pmd);
 993                         set_pmd_at(src_mm, addr, src_pmd, pmd);
 994                 }
 995                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 996                 mm_inc_nr_ptes(dst_mm);
 997                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
 998                 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 999                 ret = 0;
1000                 goto out_unlock;
1001         }
1002 #endif
1003
1004         if (unlikely(!pmd_trans_huge(pmd))) {
1005                 pte_free(dst_mm, pgtable);
1006                 goto out_unlock;
1007         }
1008         /*
1009          * When page table lock is held, the huge zero pmd should not be
1010          * under splitting since we don't split the page itself, only pmd to
1011          * a page table.
1012          */
1013         if (is_huge_zero_pmd(pmd)) {
1014                 struct page *zero_page;
1015                 /*
1016                  * get_huge_zero_page() will never allocate a new page here,
1017                  * since we already have a zero page to copy. It just takes a
1018                  * reference.
1019                  */
1020                 zero_page = mm_get_huge_zero_page(dst_mm);
1021                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1022                                 zero_page);
1023                 ret = 0;
1024                 goto out_unlock;
1025         }
1026
1027         src_page = pmd_page(pmd);
1028         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1029         get_page(src_page);
1030         page_dup_rmap(src_page, true);
1031         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1032         mm_inc_nr_ptes(dst_mm);
1033         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1034
1035         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1036         pmd = pmd_mkold(pmd_wrprotect(pmd));
1037         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1038
1039         ret = 0;
1040 out_unlock:
1041         spin_unlock(src_ptl);
1042         spin_unlock(dst_ptl);
1043 out:
1044         return ret;
1045 }
1046
1047 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1048 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1049                 pud_t *pud, int flags)
1050 {
1051         pud_t _pud;
1052
1053         _pud = pud_mkyoung(*pud);
1054         if (flags & FOLL_WRITE)
1055                 _pud = pud_mkdirty(_pud);
1056         if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1057                                 pud, _pud, flags & FOLL_WRITE))
1058                 update_mmu_cache_pud(vma, addr, pud);
1059 }
1060
1061 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1062                 pud_t *pud, int flags, struct dev_pagemap **pgmap)
1063 {
1064         unsigned long pfn = pud_pfn(*pud);
1065         struct mm_struct *mm = vma->vm_mm;
1066         struct page *page;
1067
1068         assert_spin_locked(pud_lockptr(mm, pud));
1069
1070         if (flags & FOLL_WRITE && !pud_write(*pud))
1071                 return NULL;
1072
1073         if (pud_present(*pud) && pud_devmap(*pud))
1074                 /* pass */;
1075         else
1076                 return NULL;
1077
1078         if (flags & FOLL_TOUCH)
1079                 touch_pud(vma, addr, pud, flags);
1080
1081         /*
1082          * device mapped pages can only be returned if the
1083          * caller will manage the page reference count.
1084          */
1085         if (!(flags & FOLL_GET))
1086                 return ERR_PTR(-EEXIST);
1087
1088         pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1089         *pgmap = get_dev_pagemap(pfn, *pgmap);
1090         if (!*pgmap)
1091                 return ERR_PTR(-EFAULT);
1092         page = pfn_to_page(pfn);
1093         get_page(page);
1094
1095         return page;
1096 }
1097
1098 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1099                   pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1100                   struct vm_area_struct *vma)
1101 {
1102         spinlock_t *dst_ptl, *src_ptl;
1103         pud_t pud;
1104         int ret;
1105
1106         dst_ptl = pud_lock(dst_mm, dst_pud);
1107         src_ptl = pud_lockptr(src_mm, src_pud);
1108         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1109
1110         ret = -EAGAIN;
1111         pud = *src_pud;
1112         if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1113                 goto out_unlock;
1114
1115         /*
1116          * When page table lock is held, the huge zero pud should not be
1117          * under splitting since we don't split the page itself, only pud to
1118          * a page table.
1119          */
1120         if (is_huge_zero_pud(pud)) {
1121                 /* No huge zero pud yet */
1122         }
1123
1124         pudp_set_wrprotect(src_mm, addr, src_pud);
1125         pud = pud_mkold(pud_wrprotect(pud));
1126         set_pud_at(dst_mm, addr, dst_pud, pud);
1127
1128         ret = 0;
1129 out_unlock:
1130         spin_unlock(src_ptl);
1131         spin_unlock(dst_ptl);
1132         return ret;
1133 }
1134
1135 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1136 {
1137         pud_t entry;
1138         unsigned long haddr;
1139         bool write = vmf->flags & FAULT_FLAG_WRITE;
1140
1141         vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1142         if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1143                 goto unlock;
1144
1145         entry = pud_mkyoung(orig_pud);
1146         if (write)
1147                 entry = pud_mkdirty(entry);
1148         haddr = vmf->address & HPAGE_PUD_MASK;
1149         if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1150                 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1151
1152 unlock:
1153         spin_unlock(vmf->ptl);
1154 }
1155 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1156
1157 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1158 {
1159         pmd_t entry;
1160         unsigned long haddr;
1161         bool write = vmf->flags & FAULT_FLAG_WRITE;
1162
1163         vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1164         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1165                 goto unlock;
1166
1167         entry = pmd_mkyoung(orig_pmd);
1168         if (write)
1169                 entry = pmd_mkdirty(entry);
1170         haddr = vmf->address & HPAGE_PMD_MASK;
1171         if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1172                 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1173
1174 unlock:
1175         spin_unlock(vmf->ptl);
1176 }
1177
1178 static vm_fault_t do_huge_pmd_wp_page_fallback(struct vm_fault *vmf,
1179                         pmd_t orig_pmd, struct page *page)
1180 {
1181         struct vm_area_struct *vma = vmf->vma;
1182         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1183         struct mem_cgroup *memcg;
1184         pgtable_t pgtable;
1185         pmd_t _pmd;
1186         int i;
1187         vm_fault_t ret = 0;
1188         struct page **pages;
1189         struct mmu_notifier_range range;
1190
1191         pages = kmalloc_array(HPAGE_PMD_NR, sizeof(struct page *),
1192                               GFP_KERNEL);
1193         if (unlikely(!pages)) {
1194                 ret |= VM_FAULT_OOM;
1195                 goto out;
1196         }
1197
1198         for (i = 0; i < HPAGE_PMD_NR; i++) {
1199                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
1200                                                vmf->address, page_to_nid(page));
1201                 if (unlikely(!pages[i] ||
1202                              mem_cgroup_try_charge_delay(pages[i], vma->vm_mm,
1203                                      GFP_KERNEL, &memcg, false))) {
1204                         if (pages[i])
1205                                 put_page(pages[i]);
1206                         while (--i >= 0) {
1207                                 memcg = (void *)page_private(pages[i]);
1208                                 set_page_private(pages[i], 0);
1209                                 mem_cgroup_cancel_charge(pages[i], memcg,
1210                                                 false);
1211                                 put_page(pages[i]);
1212                         }
1213                         kfree(pages);
1214                         ret |= VM_FAULT_OOM;
1215                         goto out;
1216                 }
1217                 set_page_private(pages[i], (unsigned long)memcg);
1218         }
1219
1220         for (i = 0; i < HPAGE_PMD_NR; i++) {
1221                 copy_user_highpage(pages[i], page + i,
1222                                    haddr + PAGE_SIZE * i, vma);
1223                 __SetPageUptodate(pages[i]);
1224                 cond_resched();
1225         }
1226
1227         mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
1228                                 haddr,
1229                                 haddr + HPAGE_PMD_SIZE);
1230         mmu_notifier_invalidate_range_start(&range);
1231
1232         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1233         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1234                 goto out_free_pages;
1235         VM_BUG_ON_PAGE(!PageHead(page), page);
1236
1237         /*
1238          * Leave pmd empty until pte is filled note we must notify here as
1239          * concurrent CPU thread might write to new page before the call to
1240          * mmu_notifier_invalidate_range_end() happens which can lead to a
1241          * device seeing memory write in different order than CPU.
1242          *
1243          * See Documentation/vm/mmu_notifier.rst
1244          */
1245         pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1246
1247         pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
1248         pmd_populate(vma->vm_mm, &_pmd, pgtable);
1249
1250         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1251                 pte_t entry;
1252                 entry = mk_pte(pages[i], vma->vm_page_prot);
1253                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1254                 memcg = (void *)page_private(pages[i]);
1255                 set_page_private(pages[i], 0);
1256                 page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
1257                 mem_cgroup_commit_charge(pages[i], memcg, false, false);
1258                 lru_cache_add_active_or_unevictable(pages[i], vma);
1259                 vmf->pte = pte_offset_map(&_pmd, haddr);
1260                 VM_BUG_ON(!pte_none(*vmf->pte));
1261                 set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
1262                 pte_unmap(vmf->pte);
1263         }
1264         kfree(pages);
1265
1266         smp_wmb(); /* make pte visible before pmd */
1267         pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
1268         page_remove_rmap(page, true);
1269         spin_unlock(vmf->ptl);
1270
1271         /*
1272          * No need to double call mmu_notifier->invalidate_range() callback as
1273          * the above pmdp_huge_clear_flush_notify() did already call it.
1274          */
1275         mmu_notifier_invalidate_range_only_end(&range);
1276
1277         ret |= VM_FAULT_WRITE;
1278         put_page(page);
1279
1280 out:
1281         return ret;
1282
1283 out_free_pages:
1284         spin_unlock(vmf->ptl);
1285         mmu_notifier_invalidate_range_end(&range);
1286         for (i = 0; i < HPAGE_PMD_NR; i++) {
1287                 memcg = (void *)page_private(pages[i]);
1288                 set_page_private(pages[i], 0);
1289                 mem_cgroup_cancel_charge(pages[i], memcg, false);
1290                 put_page(pages[i]);
1291         }
1292         kfree(pages);
1293         goto out;
1294 }
1295
1296 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1297 {
1298         struct vm_area_struct *vma = vmf->vma;
1299         struct page *page = NULL, *new_page;
1300         struct mem_cgroup *memcg;
1301         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1302         struct mmu_notifier_range range;
1303         gfp_t huge_gfp;                 /* for allocation and charge */
1304         vm_fault_t ret = 0;
1305
1306         vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1307         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1308         if (is_huge_zero_pmd(orig_pmd))
1309                 goto alloc;
1310         spin_lock(vmf->ptl);
1311         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1312                 goto out_unlock;
1313
1314         page = pmd_page(orig_pmd);
1315         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1316         /*
1317          * We can only reuse the page if nobody else maps the huge page or it's
1318          * part.
1319          */
1320         if (!trylock_page(page)) {
1321                 get_page(page);
1322                 spin_unlock(vmf->ptl);
1323                 lock_page(page);
1324                 spin_lock(vmf->ptl);
1325                 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1326                         unlock_page(page);
1327                         put_page(page);
1328                         goto out_unlock;
1329                 }
1330                 put_page(page);
1331         }
1332         if (reuse_swap_page(page, NULL)) {
1333                 pmd_t entry;
1334                 entry = pmd_mkyoung(orig_pmd);
1335                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1336                 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry,  1))
1337                         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1338                 ret |= VM_FAULT_WRITE;
1339                 unlock_page(page);
1340                 goto out_unlock;
1341         }
1342         unlock_page(page);
1343         get_page(page);
1344         spin_unlock(vmf->ptl);
1345 alloc:
1346         if (__transparent_hugepage_enabled(vma) &&
1347             !transparent_hugepage_debug_cow()) {
1348                 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1349                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1350         } else
1351                 new_page = NULL;
1352
1353         if (likely(new_page)) {
1354                 prep_transhuge_page(new_page);
1355         } else {
1356                 if (!page) {
1357                         split_huge_pmd(vma, vmf->pmd, vmf->address);
1358                         ret |= VM_FAULT_FALLBACK;
1359                 } else {
1360                         ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
1361                         if (ret & VM_FAULT_OOM) {
1362                                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1363                                 ret |= VM_FAULT_FALLBACK;
1364                         }
1365                         put_page(page);
1366                 }
1367                 count_vm_event(THP_FAULT_FALLBACK);
1368                 goto out;
1369         }
1370
1371         if (unlikely(mem_cgroup_try_charge_delay(new_page, vma->vm_mm,
1372                                         huge_gfp, &memcg, true))) {
1373                 put_page(new_page);
1374                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1375                 if (page)
1376                         put_page(page);
1377                 ret |= VM_FAULT_FALLBACK;
1378                 count_vm_event(THP_FAULT_FALLBACK);
1379                 goto out;
1380         }
1381
1382         count_vm_event(THP_FAULT_ALLOC);
1383         count_memcg_events(memcg, THP_FAULT_ALLOC, 1);
1384
1385         if (!page)
1386                 clear_huge_page(new_page, vmf->address, HPAGE_PMD_NR);
1387         else
1388                 copy_user_huge_page(new_page, page, vmf->address,
1389                                     vma, HPAGE_PMD_NR);
1390         __SetPageUptodate(new_page);
1391
1392         mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
1393                                 haddr,
1394                                 haddr + HPAGE_PMD_SIZE);
1395         mmu_notifier_invalidate_range_start(&range);
1396
1397         spin_lock(vmf->ptl);
1398         if (page)
1399                 put_page(page);
1400         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1401                 spin_unlock(vmf->ptl);
1402                 mem_cgroup_cancel_charge(new_page, memcg, true);
1403                 put_page(new_page);
1404                 goto out_mn;
1405         } else {
1406                 pmd_t entry;
1407                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1408                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1409                 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1410                 page_add_new_anon_rmap(new_page, vma, haddr, true);
1411                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1412                 lru_cache_add_active_or_unevictable(new_page, vma);
1413                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1414                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1415                 if (!page) {
1416                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1417                 } else {
1418                         VM_BUG_ON_PAGE(!PageHead(page), page);
1419                         page_remove_rmap(page, true);
1420                         put_page(page);
1421                 }
1422                 ret |= VM_FAULT_WRITE;
1423         }
1424         spin_unlock(vmf->ptl);
1425 out_mn:
1426         /*
1427          * No need to double call mmu_notifier->invalidate_range() callback as
1428          * the above pmdp_huge_clear_flush_notify() did already call it.
1429          */
1430         mmu_notifier_invalidate_range_only_end(&range);
1431 out:
1432         return ret;
1433 out_unlock:
1434         spin_unlock(vmf->ptl);
1435         return ret;
1436 }
1437
1438 /*
1439  * FOLL_FORCE can write to even unwritable pmd's, but only
1440  * after we've gone through a COW cycle and they are dirty.
1441  */
1442 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1443 {
1444         return pmd_write(pmd) ||
1445                ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1446 }
1447
1448 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1449                                    unsigned long addr,
1450                                    pmd_t *pmd,
1451                                    unsigned int flags)
1452 {
1453         struct mm_struct *mm = vma->vm_mm;
1454         struct page *page = NULL;
1455
1456         assert_spin_locked(pmd_lockptr(mm, pmd));
1457
1458         if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1459                 goto out;
1460
1461         /* Avoid dumping huge zero page */
1462         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1463                 return ERR_PTR(-EFAULT);
1464
1465         /* Full NUMA hinting faults to serialise migration in fault paths */
1466         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1467                 goto out;
1468
1469         page = pmd_page(*pmd);
1470         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1471         if (flags & FOLL_TOUCH)
1472                 touch_pmd(vma, addr, pmd, flags);
1473         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1474                 /*
1475                  * We don't mlock() pte-mapped THPs. This way we can avoid
1476                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1477                  *
1478                  * For anon THP:
1479                  *
1480                  * In most cases the pmd is the only mapping of the page as we
1481                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1482                  * writable private mappings in populate_vma_page_range().
1483                  *
1484                  * The only scenario when we have the page shared here is if we
1485                  * mlocking read-only mapping shared over fork(). We skip
1486                  * mlocking such pages.
1487                  *
1488                  * For file THP:
1489                  *
1490                  * We can expect PageDoubleMap() to be stable under page lock:
1491                  * for file pages we set it in page_add_file_rmap(), which
1492                  * requires page to be locked.
1493                  */
1494
1495                 if (PageAnon(page) && compound_mapcount(page) != 1)
1496                         goto skip_mlock;
1497                 if (PageDoubleMap(page) || !page->mapping)
1498                         goto skip_mlock;
1499                 if (!trylock_page(page))
1500                         goto skip_mlock;
1501                 lru_add_drain();
1502                 if (page->mapping && !PageDoubleMap(page))
1503                         mlock_vma_page(page);
1504                 unlock_page(page);
1505         }
1506 skip_mlock:
1507         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1508         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1509         if (flags & FOLL_GET)
1510                 get_page(page);
1511
1512 out:
1513         return page;
1514 }
1515
1516 /* NUMA hinting page fault entry point for trans huge pmds */
1517 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1518 {
1519         struct vm_area_struct *vma = vmf->vma;
1520         struct anon_vma *anon_vma = NULL;
1521         struct page *page;
1522         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1523         int page_nid = NUMA_NO_NODE, this_nid = numa_node_id();
1524         int target_nid, last_cpupid = -1;
1525         bool page_locked;
1526         bool migrated = false;
1527         bool was_writable;
1528         int flags = 0;
1529
1530         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1531         if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1532                 goto out_unlock;
1533
1534         /*
1535          * If there are potential migrations, wait for completion and retry
1536          * without disrupting NUMA hinting information. Do not relock and
1537          * check_same as the page may no longer be mapped.
1538          */
1539         if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1540                 page = pmd_page(*vmf->pmd);
1541                 if (!get_page_unless_zero(page))
1542                         goto out_unlock;
1543                 spin_unlock(vmf->ptl);
1544                 put_and_wait_on_page_locked(page);
1545                 goto out;
1546         }
1547
1548         page = pmd_page(pmd);
1549         BUG_ON(is_huge_zero_page(page));
1550         page_nid = page_to_nid(page);
1551         last_cpupid = page_cpupid_last(page);
1552         count_vm_numa_event(NUMA_HINT_FAULTS);
1553         if (page_nid == this_nid) {
1554                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1555                 flags |= TNF_FAULT_LOCAL;
1556         }
1557
1558         /* See similar comment in do_numa_page for explanation */
1559         if (!pmd_savedwrite(pmd))
1560                 flags |= TNF_NO_GROUP;
1561
1562         /*
1563          * Acquire the page lock to serialise THP migrations but avoid dropping
1564          * page_table_lock if at all possible
1565          */
1566         page_locked = trylock_page(page);
1567         target_nid = mpol_misplaced(page, vma, haddr);
1568         if (target_nid == NUMA_NO_NODE) {
1569                 /* If the page was locked, there are no parallel migrations */
1570                 if (page_locked)
1571                         goto clear_pmdnuma;
1572         }
1573
1574         /* Migration could have started since the pmd_trans_migrating check */
1575         if (!page_locked) {
1576                 page_nid = NUMA_NO_NODE;
1577                 if (!get_page_unless_zero(page))
1578                         goto out_unlock;
1579                 spin_unlock(vmf->ptl);
1580                 put_and_wait_on_page_locked(page);
1581                 goto out;
1582         }
1583
1584         /*
1585          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1586          * to serialises splits
1587          */
1588         get_page(page);
1589         spin_unlock(vmf->ptl);
1590         anon_vma = page_lock_anon_vma_read(page);
1591
1592         /* Confirm the PMD did not change while page_table_lock was released */
1593         spin_lock(vmf->ptl);
1594         if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1595                 unlock_page(page);
1596                 put_page(page);
1597                 page_nid = NUMA_NO_NODE;
1598                 goto out_unlock;
1599         }
1600
1601         /* Bail if we fail to protect against THP splits for any reason */
1602         if (unlikely(!anon_vma)) {
1603                 put_page(page);
1604                 page_nid = NUMA_NO_NODE;
1605                 goto clear_pmdnuma;
1606         }
1607
1608         /*
1609          * Since we took the NUMA fault, we must have observed the !accessible
1610          * bit. Make sure all other CPUs agree with that, to avoid them
1611          * modifying the page we're about to migrate.
1612          *
1613          * Must be done under PTL such that we'll observe the relevant
1614          * inc_tlb_flush_pending().
1615          *
1616          * We are not sure a pending tlb flush here is for a huge page
1617          * mapping or not. Hence use the tlb range variant
1618          */
1619         if (mm_tlb_flush_pending(vma->vm_mm)) {
1620                 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1621                 /*
1622                  * change_huge_pmd() released the pmd lock before
1623                  * invalidating the secondary MMUs sharing the primary
1624                  * MMU pagetables (with ->invalidate_range()). The
1625                  * mmu_notifier_invalidate_range_end() (which
1626                  * internally calls ->invalidate_range()) in
1627                  * change_pmd_range() will run after us, so we can't
1628                  * rely on it here and we need an explicit invalidate.
1629                  */
1630                 mmu_notifier_invalidate_range(vma->vm_mm, haddr,
1631                                               haddr + HPAGE_PMD_SIZE);
1632         }
1633
1634         /*
1635          * Migrate the THP to the requested node, returns with page unlocked
1636          * and access rights restored.
1637          */
1638         spin_unlock(vmf->ptl);
1639
1640         migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1641                                 vmf->pmd, pmd, vmf->address, page, target_nid);
1642         if (migrated) {
1643                 flags |= TNF_MIGRATED;
1644                 page_nid = target_nid;
1645         } else
1646                 flags |= TNF_MIGRATE_FAIL;
1647
1648         goto out;
1649 clear_pmdnuma:
1650         BUG_ON(!PageLocked(page));
1651         was_writable = pmd_savedwrite(pmd);
1652         pmd = pmd_modify(pmd, vma->vm_page_prot);
1653         pmd = pmd_mkyoung(pmd);
1654         if (was_writable)
1655                 pmd = pmd_mkwrite(pmd);
1656         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1657         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1658         unlock_page(page);
1659 out_unlock:
1660         spin_unlock(vmf->ptl);
1661
1662 out:
1663         if (anon_vma)
1664                 page_unlock_anon_vma_read(anon_vma);
1665
1666         if (page_nid != NUMA_NO_NODE)
1667                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1668                                 flags);
1669
1670         return 0;
1671 }
1672
1673 /*
1674  * Return true if we do MADV_FREE successfully on entire pmd page.
1675  * Otherwise, return false.
1676  */
1677 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1678                 pmd_t *pmd, unsigned long addr, unsigned long next)
1679 {
1680         spinlock_t *ptl;
1681         pmd_t orig_pmd;
1682         struct page *page;
1683         struct mm_struct *mm = tlb->mm;
1684         bool ret = false;
1685
1686         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1687
1688         ptl = pmd_trans_huge_lock(pmd, vma);
1689         if (!ptl)
1690                 goto out_unlocked;
1691
1692         orig_pmd = *pmd;
1693         if (is_huge_zero_pmd(orig_pmd))
1694                 goto out;
1695
1696         if (unlikely(!pmd_present(orig_pmd))) {
1697                 VM_BUG_ON(thp_migration_supported() &&
1698                                   !is_pmd_migration_entry(orig_pmd));
1699                 goto out;
1700         }
1701
1702         page = pmd_page(orig_pmd);
1703         /*
1704          * If other processes are mapping this page, we couldn't discard
1705          * the page unless they all do MADV_FREE so let's skip the page.
1706          */
1707         if (page_mapcount(page) != 1)
1708                 goto out;
1709
1710         if (!trylock_page(page))
1711                 goto out;
1712
1713         /*
1714          * If user want to discard part-pages of THP, split it so MADV_FREE
1715          * will deactivate only them.
1716          */
1717         if (next - addr != HPAGE_PMD_SIZE) {
1718                 get_page(page);
1719                 spin_unlock(ptl);
1720                 split_huge_page(page);
1721                 unlock_page(page);
1722                 put_page(page);
1723                 goto out_unlocked;
1724         }
1725
1726         if (PageDirty(page))
1727                 ClearPageDirty(page);
1728         unlock_page(page);
1729
1730         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1731                 pmdp_invalidate(vma, addr, pmd);
1732                 orig_pmd = pmd_mkold(orig_pmd);
1733                 orig_pmd = pmd_mkclean(orig_pmd);
1734
1735                 set_pmd_at(mm, addr, pmd, orig_pmd);
1736                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1737         }
1738
1739         mark_page_lazyfree(page);
1740         ret = true;
1741 out:
1742         spin_unlock(ptl);
1743 out_unlocked:
1744         return ret;
1745 }
1746
1747 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1748 {
1749         pgtable_t pgtable;
1750
1751         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1752         pte_free(mm, pgtable);
1753         mm_dec_nr_ptes(mm);
1754 }
1755
1756 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1757                  pmd_t *pmd, unsigned long addr)
1758 {
1759         pmd_t orig_pmd;
1760         spinlock_t *ptl;
1761
1762         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1763
1764         ptl = __pmd_trans_huge_lock(pmd, vma);
1765         if (!ptl)
1766                 return 0;
1767         /*
1768          * For architectures like ppc64 we look at deposited pgtable
1769          * when calling pmdp_huge_get_and_clear. So do the
1770          * pgtable_trans_huge_withdraw after finishing pmdp related
1771          * operations.
1772          */
1773         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1774                         tlb->fullmm);
1775         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1776         if (vma_is_dax(vma)) {
1777                 if (arch_needs_pgtable_deposit())
1778                         zap_deposited_table(tlb->mm, pmd);
1779                 spin_unlock(ptl);
1780                 if (is_huge_zero_pmd(orig_pmd))
1781                         tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1782         } else if (is_huge_zero_pmd(orig_pmd)) {
1783                 zap_deposited_table(tlb->mm, pmd);
1784                 spin_unlock(ptl);
1785                 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1786         } else {
1787                 struct page *page = NULL;
1788                 int flush_needed = 1;
1789
1790                 if (pmd_present(orig_pmd)) {
1791                         page = pmd_page(orig_pmd);
1792                         page_remove_rmap(page, true);
1793                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1794                         VM_BUG_ON_PAGE(!PageHead(page), page);
1795                 } else if (thp_migration_supported()) {
1796                         swp_entry_t entry;
1797
1798                         VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1799                         entry = pmd_to_swp_entry(orig_pmd);
1800                         page = pfn_to_page(swp_offset(entry));
1801                         flush_needed = 0;
1802                 } else
1803                         WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1804
1805                 if (PageAnon(page)) {
1806                         zap_deposited_table(tlb->mm, pmd);
1807                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1808                 } else {
1809                         if (arch_needs_pgtable_deposit())
1810                                 zap_deposited_table(tlb->mm, pmd);
1811                         add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1812                 }
1813
1814                 spin_unlock(ptl);
1815                 if (flush_needed)
1816                         tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1817         }
1818         return 1;
1819 }
1820
1821 #ifndef pmd_move_must_withdraw
1822 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1823                                          spinlock_t *old_pmd_ptl,
1824                                          struct vm_area_struct *vma)
1825 {
1826         /*
1827          * With split pmd lock we also need to move preallocated
1828          * PTE page table if new_pmd is on different PMD page table.
1829          *
1830          * We also don't deposit and withdraw tables for file pages.
1831          */
1832         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1833 }
1834 #endif
1835
1836 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1837 {
1838 #ifdef CONFIG_MEM_SOFT_DIRTY
1839         if (unlikely(is_pmd_migration_entry(pmd)))
1840                 pmd = pmd_swp_mksoft_dirty(pmd);
1841         else if (pmd_present(pmd))
1842                 pmd = pmd_mksoft_dirty(pmd);
1843 #endif
1844         return pmd;
1845 }
1846
1847 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1848                   unsigned long new_addr, unsigned long old_end,
1849                   pmd_t *old_pmd, pmd_t *new_pmd)
1850 {
1851         spinlock_t *old_ptl, *new_ptl;
1852         pmd_t pmd;
1853         struct mm_struct *mm = vma->vm_mm;
1854         bool force_flush = false;
1855
1856         if ((old_addr & ~HPAGE_PMD_MASK) ||
1857             (new_addr & ~HPAGE_PMD_MASK) ||
1858             old_end - old_addr < HPAGE_PMD_SIZE)
1859                 return false;
1860
1861         /*
1862          * The destination pmd shouldn't be established, free_pgtables()
1863          * should have release it.
1864          */
1865         if (WARN_ON(!pmd_none(*new_pmd))) {
1866                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1867                 return false;
1868         }
1869
1870         /*
1871          * We don't have to worry about the ordering of src and dst
1872          * ptlocks because exclusive mmap_sem prevents deadlock.
1873          */
1874         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1875         if (old_ptl) {
1876                 new_ptl = pmd_lockptr(mm, new_pmd);
1877                 if (new_ptl != old_ptl)
1878                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1879                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1880                 if (pmd_present(pmd))
1881                         force_flush = true;
1882                 VM_BUG_ON(!pmd_none(*new_pmd));
1883
1884                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1885                         pgtable_t pgtable;
1886                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1887                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1888                 }
1889                 pmd = move_soft_dirty_pmd(pmd);
1890                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1891                 if (force_flush)
1892                         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1893                 if (new_ptl != old_ptl)
1894                         spin_unlock(new_ptl);
1895                 spin_unlock(old_ptl);
1896                 return true;
1897         }
1898         return false;
1899 }
1900
1901 /*
1902  * Returns
1903  *  - 0 if PMD could not be locked
1904  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1905  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1906  */
1907 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1908                 unsigned long addr, pgprot_t newprot, int prot_numa)
1909 {
1910         struct mm_struct *mm = vma->vm_mm;
1911         spinlock_t *ptl;
1912         pmd_t entry;
1913         bool preserve_write;
1914         int ret;
1915
1916         ptl = __pmd_trans_huge_lock(pmd, vma);
1917         if (!ptl)
1918                 return 0;
1919
1920         preserve_write = prot_numa && pmd_write(*pmd);
1921         ret = 1;
1922
1923 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1924         if (is_swap_pmd(*pmd)) {
1925                 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1926
1927                 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1928                 if (is_write_migration_entry(entry)) {
1929                         pmd_t newpmd;
1930                         /*
1931                          * A protection check is difficult so
1932                          * just be safe and disable write
1933                          */
1934                         make_migration_entry_read(&entry);
1935                         newpmd = swp_entry_to_pmd(entry);
1936                         if (pmd_swp_soft_dirty(*pmd))
1937                                 newpmd = pmd_swp_mksoft_dirty(newpmd);
1938                         set_pmd_at(mm, addr, pmd, newpmd);
1939                 }
1940                 goto unlock;
1941         }
1942 #endif
1943
1944         /*
1945          * Avoid trapping faults against the zero page. The read-only
1946          * data is likely to be read-cached on the local CPU and
1947          * local/remote hits to the zero page are not interesting.
1948          */
1949         if (prot_numa && is_huge_zero_pmd(*pmd))
1950                 goto unlock;
1951
1952         if (prot_numa && pmd_protnone(*pmd))
1953                 goto unlock;
1954
1955         /*
1956          * In case prot_numa, we are under down_read(mmap_sem). It's critical
1957          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1958          * which is also under down_read(mmap_sem):
1959          *
1960          *      CPU0:                           CPU1:
1961          *                              change_huge_pmd(prot_numa=1)
1962          *                               pmdp_huge_get_and_clear_notify()
1963          * madvise_dontneed()
1964          *  zap_pmd_range()
1965          *   pmd_trans_huge(*pmd) == 0 (without ptl)
1966          *   // skip the pmd
1967          *                               set_pmd_at();
1968          *                               // pmd is re-established
1969          *
1970          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1971          * which may break userspace.
1972          *
1973          * pmdp_invalidate() is required to make sure we don't miss
1974          * dirty/young flags set by hardware.
1975          */
1976         entry = pmdp_invalidate(vma, addr, pmd);
1977
1978         entry = pmd_modify(entry, newprot);
1979         if (preserve_write)
1980                 entry = pmd_mk_savedwrite(entry);
1981         ret = HPAGE_PMD_NR;
1982         set_pmd_at(mm, addr, pmd, entry);
1983         BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1984 unlock:
1985         spin_unlock(ptl);
1986         return ret;
1987 }
1988
1989 /*
1990  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1991  *
1992  * Note that if it returns page table lock pointer, this routine returns without
1993  * unlocking page table lock. So callers must unlock it.
1994  */
1995 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1996 {
1997         spinlock_t *ptl;
1998         ptl = pmd_lock(vma->vm_mm, pmd);
1999         if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
2000                         pmd_devmap(*pmd)))
2001                 return ptl;
2002         spin_unlock(ptl);
2003         return NULL;
2004 }
2005
2006 /*
2007  * Returns true if a given pud maps a thp, false otherwise.
2008  *
2009  * Note that if it returns true, this routine returns without unlocking page
2010  * table lock. So callers must unlock it.
2011  */
2012 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
2013 {
2014         spinlock_t *ptl;
2015
2016         ptl = pud_lock(vma->vm_mm, pud);
2017         if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
2018                 return ptl;
2019         spin_unlock(ptl);
2020         return NULL;
2021 }
2022
2023 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
2024 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
2025                  pud_t *pud, unsigned long addr)
2026 {
2027         spinlock_t *ptl;
2028
2029         ptl = __pud_trans_huge_lock(pud, vma);
2030         if (!ptl)
2031                 return 0;
2032         /*
2033          * For architectures like ppc64 we look at deposited pgtable
2034          * when calling pudp_huge_get_and_clear. So do the
2035          * pgtable_trans_huge_withdraw after finishing pudp related
2036          * operations.
2037          */
2038         pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
2039         tlb_remove_pud_tlb_entry(tlb, pud, addr);
2040         if (vma_is_dax(vma)) {
2041                 spin_unlock(ptl);
2042                 /* No zero page support yet */
2043         } else {
2044                 /* No support for anonymous PUD pages yet */
2045                 BUG();
2046         }
2047         return 1;
2048 }
2049
2050 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
2051                 unsigned long haddr)
2052 {
2053         VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
2054         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2055         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
2056         VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
2057
2058         count_vm_event(THP_SPLIT_PUD);
2059
2060         pudp_huge_clear_flush_notify(vma, haddr, pud);
2061 }
2062
2063 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2064                 unsigned long address)
2065 {
2066         spinlock_t *ptl;
2067         struct mmu_notifier_range range;
2068
2069         mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
2070                                 address & HPAGE_PUD_MASK,
2071                                 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
2072         mmu_notifier_invalidate_range_start(&range);
2073         ptl = pud_lock(vma->vm_mm, pud);
2074         if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2075                 goto out;
2076         __split_huge_pud_locked(vma, pud, range.start);
2077
2078 out:
2079         spin_unlock(ptl);
2080         /*
2081          * No need to double call mmu_notifier->invalidate_range() callback as
2082          * the above pudp_huge_clear_flush_notify() did already call it.
2083          */
2084         mmu_notifier_invalidate_range_only_end(&range);
2085 }
2086 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2087
2088 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2089                 unsigned long haddr, pmd_t *pmd)
2090 {
2091         struct mm_struct *mm = vma->vm_mm;
2092         pgtable_t pgtable;
2093         pmd_t _pmd;
2094         int i;
2095
2096         /*
2097          * Leave pmd empty until pte is filled note that it is fine to delay
2098          * notification until mmu_notifier_invalidate_range_end() as we are
2099          * replacing a zero pmd write protected page with a zero pte write
2100          * protected page.
2101          *
2102          * See Documentation/vm/mmu_notifier.rst
2103          */
2104         pmdp_huge_clear_flush(vma, haddr, pmd);
2105
2106         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2107         pmd_populate(mm, &_pmd, pgtable);
2108
2109         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2110                 pte_t *pte, entry;
2111                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2112                 entry = pte_mkspecial(entry);
2113                 pte = pte_offset_map(&_pmd, haddr);
2114                 VM_BUG_ON(!pte_none(*pte));
2115                 set_pte_at(mm, haddr, pte, entry);
2116                 pte_unmap(pte);
2117         }
2118         smp_wmb(); /* make pte visible before pmd */
2119         pmd_populate(mm, pmd, pgtable);
2120 }
2121
2122 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2123                 unsigned long haddr, bool freeze)
2124 {
2125         struct mm_struct *mm = vma->vm_mm;
2126         struct page *page;
2127         pgtable_t pgtable;
2128         pmd_t old_pmd, _pmd;
2129         bool young, write, soft_dirty, pmd_migration = false;
2130         unsigned long addr;
2131         int i;
2132
2133         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2134         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2135         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2136         VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2137                                 && !pmd_devmap(*pmd));
2138
2139         count_vm_event(THP_SPLIT_PMD);
2140
2141         if (!vma_is_anonymous(vma)) {
2142                 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2143                 /*
2144                  * We are going to unmap this huge page. So
2145                  * just go ahead and zap it
2146                  */
2147                 if (arch_needs_pgtable_deposit())
2148                         zap_deposited_table(mm, pmd);
2149                 if (vma_is_dax(vma))
2150                         return;
2151                 page = pmd_page(_pmd);
2152                 if (!PageDirty(page) && pmd_dirty(_pmd))
2153                         set_page_dirty(page);
2154                 if (!PageReferenced(page) && pmd_young(_pmd))
2155                         SetPageReferenced(page);
2156                 page_remove_rmap(page, true);
2157                 put_page(page);
2158                 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2159                 return;
2160         } else if (is_huge_zero_pmd(*pmd)) {
2161                 /*
2162                  * FIXME: Do we want to invalidate secondary mmu by calling
2163                  * mmu_notifier_invalidate_range() see comments below inside
2164                  * __split_huge_pmd() ?
2165                  *
2166                  * We are going from a zero huge page write protected to zero
2167                  * small page also write protected so it does not seems useful
2168                  * to invalidate secondary mmu at this time.
2169                  */
2170                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2171         }
2172
2173         /*
2174          * Up to this point the pmd is present and huge and userland has the
2175          * whole access to the hugepage during the split (which happens in
2176          * place). If we overwrite the pmd with the not-huge version pointing
2177          * to the pte here (which of course we could if all CPUs were bug
2178          * free), userland could trigger a small page size TLB miss on the
2179          * small sized TLB while the hugepage TLB entry is still established in
2180          * the huge TLB. Some CPU doesn't like that.
2181          * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2182          * 383 on page 93. Intel should be safe but is also warns that it's
2183          * only safe if the permission and cache attributes of the two entries
2184          * loaded in the two TLB is identical (which should be the case here).
2185          * But it is generally safer to never allow small and huge TLB entries
2186          * for the same virtual address to be loaded simultaneously. So instead
2187          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2188          * current pmd notpresent (atomically because here the pmd_trans_huge
2189          * must remain set at all times on the pmd until the split is complete
2190          * for this pmd), then we flush the SMP TLB and finally we write the
2191          * non-huge version of the pmd entry with pmd_populate.
2192          */
2193         old_pmd = pmdp_invalidate(vma, haddr, pmd);
2194
2195         pmd_migration = is_pmd_migration_entry(old_pmd);
2196         if (unlikely(pmd_migration)) {
2197                 swp_entry_t entry;
2198
2199                 entry = pmd_to_swp_entry(old_pmd);
2200                 page = pfn_to_page(swp_offset(entry));
2201                 write = is_write_migration_entry(entry);
2202                 young = false;
2203                 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2204         } else {
2205                 page = pmd_page(old_pmd);
2206                 if (pmd_dirty(old_pmd))
2207                         SetPageDirty(page);
2208                 write = pmd_write(old_pmd);
2209                 young = pmd_young(old_pmd);
2210                 soft_dirty = pmd_soft_dirty(old_pmd);
2211         }
2212         VM_BUG_ON_PAGE(!page_count(page), page);
2213         page_ref_add(page, HPAGE_PMD_NR - 1);
2214
2215         /*
2216          * Withdraw the table only after we mark the pmd entry invalid.
2217          * This's critical for some architectures (Power).
2218          */
2219         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2220         pmd_populate(mm, &_pmd, pgtable);
2221
2222         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2223                 pte_t entry, *pte;
2224                 /*
2225                  * Note that NUMA hinting access restrictions are not
2226                  * transferred to avoid any possibility of altering
2227                  * permissions across VMAs.
2228                  */
2229                 if (freeze || pmd_migration) {
2230                         swp_entry_t swp_entry;
2231                         swp_entry = make_migration_entry(page + i, write);
2232                         entry = swp_entry_to_pte(swp_entry);
2233                         if (soft_dirty)
2234                                 entry = pte_swp_mksoft_dirty(entry);
2235                 } else {
2236                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2237                         entry = maybe_mkwrite(entry, vma);
2238                         if (!write)
2239                                 entry = pte_wrprotect(entry);
2240                         if (!young)
2241                                 entry = pte_mkold(entry);
2242                         if (soft_dirty)
2243                                 entry = pte_mksoft_dirty(entry);
2244                 }
2245                 pte = pte_offset_map(&_pmd, addr);
2246                 BUG_ON(!pte_none(*pte));
2247                 set_pte_at(mm, addr, pte, entry);
2248                 atomic_inc(&page[i]._mapcount);
2249                 pte_unmap(pte);
2250         }
2251
2252         /*
2253          * Set PG_double_map before dropping compound_mapcount to avoid
2254          * false-negative page_mapped().
2255          */
2256         if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2257                 for (i = 0; i < HPAGE_PMD_NR; i++)
2258                         atomic_inc(&page[i]._mapcount);
2259         }
2260
2261         if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2262                 /* Last compound_mapcount is gone. */
2263                 __dec_node_page_state(page, NR_ANON_THPS);
2264                 if (TestClearPageDoubleMap(page)) {
2265                         /* No need in mapcount reference anymore */
2266                         for (i = 0; i < HPAGE_PMD_NR; i++)
2267                                 atomic_dec(&page[i]._mapcount);
2268                 }
2269         }
2270
2271         smp_wmb(); /* make pte visible before pmd */
2272         pmd_populate(mm, pmd, pgtable);
2273
2274         if (freeze) {
2275                 for (i = 0; i < HPAGE_PMD_NR; i++) {
2276                         page_remove_rmap(page + i, false);
2277                         put_page(page + i);
2278                 }
2279         }
2280 }
2281
2282 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2283                 unsigned long address, bool freeze, struct page *page)
2284 {
2285         spinlock_t *ptl;
2286         struct mmu_notifier_range range;
2287
2288         mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
2289                                 address & HPAGE_PMD_MASK,
2290                                 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2291         mmu_notifier_invalidate_range_start(&range);
2292         ptl = pmd_lock(vma->vm_mm, pmd);
2293
2294         /*
2295          * If caller asks to setup a migration entries, we need a page to check
2296          * pmd against. Otherwise we can end up replacing wrong page.
2297          */
2298         VM_BUG_ON(freeze && !page);
2299         if (page && page != pmd_page(*pmd))
2300                 goto out;
2301
2302         if (pmd_trans_huge(*pmd)) {
2303                 page = pmd_page(*pmd);
2304                 if (PageMlocked(page))
2305                         clear_page_mlock(page);
2306         } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2307                 goto out;
2308         __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2309 out:
2310         spin_unlock(ptl);
2311         /*
2312          * No need to double call mmu_notifier->invalidate_range() callback.
2313          * They are 3 cases to consider inside __split_huge_pmd_locked():
2314          *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2315          *  2) __split_huge_zero_page_pmd() read only zero page and any write
2316          *    fault will trigger a flush_notify before pointing to a new page
2317          *    (it is fine if the secondary mmu keeps pointing to the old zero
2318          *    page in the meantime)
2319          *  3) Split a huge pmd into pte pointing to the same page. No need
2320          *     to invalidate secondary tlb entry they are all still valid.
2321          *     any further changes to individual pte will notify. So no need
2322          *     to call mmu_notifier->invalidate_range()
2323          */
2324         mmu_notifier_invalidate_range_only_end(&range);
2325 }
2326
2327 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2328                 bool freeze, struct page *page)
2329 {
2330         pgd_t *pgd;
2331         p4d_t *p4d;
2332         pud_t *pud;
2333         pmd_t *pmd;
2334
2335         pgd = pgd_offset(vma->vm_mm, address);
2336         if (!pgd_present(*pgd))
2337                 return;
2338
2339         p4d = p4d_offset(pgd, address);
2340         if (!p4d_present(*p4d))
2341                 return;
2342
2343         pud = pud_offset(p4d, address);
2344         if (!pud_present(*pud))
2345                 return;
2346
2347         pmd = pmd_offset(pud, address);
2348
2349         __split_huge_pmd(vma, pmd, address, freeze, page);
2350 }
2351
2352 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2353                              unsigned long start,
2354                              unsigned long end,
2355                              long adjust_next)
2356 {
2357         /*
2358          * If the new start address isn't hpage aligned and it could
2359          * previously contain an hugepage: check if we need to split
2360          * an huge pmd.
2361          */
2362         if (start & ~HPAGE_PMD_MASK &&
2363             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2364             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2365                 split_huge_pmd_address(vma, start, false, NULL);
2366
2367         /*
2368          * If the new end address isn't hpage aligned and it could
2369          * previously contain an hugepage: check if we need to split
2370          * an huge pmd.
2371          */
2372         if (end & ~HPAGE_PMD_MASK &&
2373             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2374             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2375                 split_huge_pmd_address(vma, end, false, NULL);
2376
2377         /*
2378          * If we're also updating the vma->vm_next->vm_start, if the new
2379          * vm_next->vm_start isn't page aligned and it could previously
2380          * contain an hugepage: check if we need to split an huge pmd.
2381          */
2382         if (adjust_next > 0) {
2383                 struct vm_area_struct *next = vma->vm_next;
2384                 unsigned long nstart = next->vm_start;
2385                 nstart += adjust_next << PAGE_SHIFT;
2386                 if (nstart & ~HPAGE_PMD_MASK &&
2387                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2388                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2389                         split_huge_pmd_address(next, nstart, false, NULL);
2390         }
2391 }
2392
2393 static void unmap_page(struct page *page)
2394 {
2395         enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
2396                 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2397         bool unmap_success;
2398
2399         VM_BUG_ON_PAGE(!PageHead(page), page);
2400
2401         if (PageAnon(page))
2402                 ttu_flags |= TTU_SPLIT_FREEZE;
2403
2404         unmap_success = try_to_unmap(page, ttu_flags);
2405         VM_BUG_ON_PAGE(!unmap_success, page);
2406 }
2407
2408 static void remap_page(struct page *page)
2409 {
2410         int i;
2411         if (PageTransHuge(page)) {
2412                 remove_migration_ptes(page, page, true);
2413         } else {
2414                 for (i = 0; i < HPAGE_PMD_NR; i++)
2415                         remove_migration_ptes(page + i, page + i, true);
2416         }
2417 }
2418
2419 static void __split_huge_page_tail(struct page *head, int tail,
2420                 struct lruvec *lruvec, struct list_head *list)
2421 {
2422         struct page *page_tail = head + tail;
2423
2424         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2425
2426         /*
2427          * Clone page flags before unfreezing refcount.
2428          *
2429          * After successful get_page_unless_zero() might follow flags change,
2430          * for exmaple lock_page() which set PG_waiters.
2431          */
2432         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2433         page_tail->flags |= (head->flags &
2434                         ((1L << PG_referenced) |
2435                          (1L << PG_swapbacked) |
2436                          (1L << PG_swapcache) |
2437                          (1L << PG_mlocked) |
2438                          (1L << PG_uptodate) |
2439                          (1L << PG_active) |
2440                          (1L << PG_workingset) |
2441                          (1L << PG_locked) |
2442                          (1L << PG_unevictable) |
2443                          (1L << PG_dirty)));
2444
2445         /* ->mapping in first tail page is compound_mapcount */
2446         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2447                         page_tail);
2448         page_tail->mapping = head->mapping;
2449         page_tail->index = head->index + tail;
2450
2451         /* Page flags must be visible before we make the page non-compound. */
2452         smp_wmb();
2453
2454         /*
2455          * Clear PageTail before unfreezing page refcount.
2456          *
2457          * After successful get_page_unless_zero() might follow put_page()
2458          * which needs correct compound_head().
2459          */
2460         clear_compound_head(page_tail);
2461
2462         /* Finally unfreeze refcount. Additional reference from page cache. */
2463         page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2464                                           PageSwapCache(head)));
2465
2466         if (page_is_young(head))
2467                 set_page_young(page_tail);
2468         if (page_is_idle(head))
2469                 set_page_idle(page_tail);
2470
2471         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2472
2473         /*
2474          * always add to the tail because some iterators expect new
2475          * pages to show after the currently processed elements - e.g.
2476          * migrate_pages
2477          */
2478         lru_add_page_tail(head, page_tail, lruvec, list);
2479 }
2480
2481 static void __split_huge_page(struct page *page, struct list_head *list,
2482                 pgoff_t end, unsigned long flags)
2483 {
2484         struct page *head = compound_head(page);
2485         pg_data_t *pgdat = page_pgdat(head);
2486         struct lruvec *lruvec;
2487         int i;
2488
2489         lruvec = mem_cgroup_page_lruvec(head, pgdat);
2490
2491         /* complete memcg works before add pages to LRU */
2492         mem_cgroup_split_huge_fixup(head);
2493
2494         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
2495                 __split_huge_page_tail(head, i, lruvec, list);
2496                 /* Some pages can be beyond i_size: drop them from page cache */
2497                 if (head[i].index >= end) {
2498                         ClearPageDirty(head + i);
2499                         __delete_from_page_cache(head + i, NULL);
2500                         if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2501                                 shmem_uncharge(head->mapping->host, 1);
2502                         put_page(head + i);
2503                 } else if (!PageAnon(page)) {
2504                         __xa_store(&head->mapping->i_pages, head[i].index,
2505                                         head + i, 0);
2506                 }
2507         }
2508
2509         ClearPageCompound(head);
2510         /* See comment in __split_huge_page_tail() */
2511         if (PageAnon(head)) {
2512                 /* Additional pin to swap cache */
2513                 if (PageSwapCache(head))
2514                         page_ref_add(head, 2);
2515                 else
2516                         page_ref_inc(head);
2517         } else {
2518                 /* Additional pin to page cache */
2519                 page_ref_add(head, 2);
2520                 xa_unlock(&head->mapping->i_pages);
2521         }
2522
2523         spin_unlock_irqrestore(&pgdat->lru_lock, flags);
2524
2525         remap_page(head);
2526
2527         for (i = 0; i < HPAGE_PMD_NR; i++) {
2528                 struct page *subpage = head + i;
2529                 if (subpage == page)
2530                         continue;
2531                 unlock_page(subpage);
2532
2533                 /*
2534                  * Subpages may be freed if there wasn't any mapping
2535                  * like if add_to_swap() is running on a lru page that
2536                  * had its mapping zapped. And freeing these pages
2537                  * requires taking the lru_lock so we do the put_page
2538                  * of the tail pages after the split is complete.
2539                  */
2540                 put_page(subpage);
2541         }
2542 }
2543
2544 int total_mapcount(struct page *page)
2545 {
2546         int i, compound, ret;
2547
2548         VM_BUG_ON_PAGE(PageTail(page), page);
2549
2550         if (likely(!PageCompound(page)))
2551                 return atomic_read(&page->_mapcount) + 1;
2552
2553         compound = compound_mapcount(page);
2554         if (PageHuge(page))
2555                 return compound;
2556         ret = compound;
2557         for (i = 0; i < HPAGE_PMD_NR; i++)
2558                 ret += atomic_read(&page[i]._mapcount) + 1;
2559         /* File pages has compound_mapcount included in _mapcount */
2560         if (!PageAnon(page))
2561                 return ret - compound * HPAGE_PMD_NR;
2562         if (PageDoubleMap(page))
2563                 ret -= HPAGE_PMD_NR;
2564         return ret;
2565 }
2566
2567 /*
2568  * This calculates accurately how many mappings a transparent hugepage
2569  * has (unlike page_mapcount() which isn't fully accurate). This full
2570  * accuracy is primarily needed to know if copy-on-write faults can
2571  * reuse the page and change the mapping to read-write instead of
2572  * copying them. At the same time this returns the total_mapcount too.
2573  *
2574  * The function returns the highest mapcount any one of the subpages
2575  * has. If the return value is one, even if different processes are
2576  * mapping different subpages of the transparent hugepage, they can
2577  * all reuse it, because each process is reusing a different subpage.
2578  *
2579  * The total_mapcount is instead counting all virtual mappings of the
2580  * subpages. If the total_mapcount is equal to "one", it tells the
2581  * caller all mappings belong to the same "mm" and in turn the
2582  * anon_vma of the transparent hugepage can become the vma->anon_vma
2583  * local one as no other process may be mapping any of the subpages.
2584  *
2585  * It would be more accurate to replace page_mapcount() with
2586  * page_trans_huge_mapcount(), however we only use
2587  * page_trans_huge_mapcount() in the copy-on-write faults where we
2588  * need full accuracy to avoid breaking page pinning, because
2589  * page_trans_huge_mapcount() is slower than page_mapcount().
2590  */
2591 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2592 {
2593         int i, ret, _total_mapcount, mapcount;
2594
2595         /* hugetlbfs shouldn't call it */
2596         VM_BUG_ON_PAGE(PageHuge(page), page);
2597
2598         if (likely(!PageTransCompound(page))) {
2599                 mapcount = atomic_read(&page->_mapcount) + 1;
2600                 if (total_mapcount)
2601                         *total_mapcount = mapcount;
2602                 return mapcount;
2603         }
2604
2605         page = compound_head(page);
2606
2607         _total_mapcount = ret = 0;
2608         for (i = 0; i < HPAGE_PMD_NR; i++) {
2609                 mapcount = atomic_read(&page[i]._mapcount) + 1;
2610                 ret = max(ret, mapcount);
2611                 _total_mapcount += mapcount;
2612         }
2613         if (PageDoubleMap(page)) {
2614                 ret -= 1;
2615                 _total_mapcount -= HPAGE_PMD_NR;
2616         }
2617         mapcount = compound_mapcount(page);
2618         ret += mapcount;
2619         _total_mapcount += mapcount;
2620         if (total_mapcount)
2621                 *total_mapcount = _total_mapcount;
2622         return ret;
2623 }
2624
2625 /* Racy check whether the huge page can be split */
2626 bool can_split_huge_page(struct page *page, int *pextra_pins)
2627 {
2628         int extra_pins;
2629
2630         /* Additional pins from page cache */
2631         if (PageAnon(page))
2632                 extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0;
2633         else
2634                 extra_pins = HPAGE_PMD_NR;
2635         if (pextra_pins)
2636                 *pextra_pins = extra_pins;
2637         return total_mapcount(page) == page_count(page) - extra_pins - 1;
2638 }
2639
2640 /*
2641  * This function splits huge page into normal pages. @page can point to any
2642  * subpage of huge page to split. Split doesn't change the position of @page.
2643  *
2644  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2645  * The huge page must be locked.
2646  *
2647  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2648  *
2649  * Both head page and tail pages will inherit mapping, flags, and so on from
2650  * the hugepage.
2651  *
2652  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2653  * they are not mapped.
2654  *
2655  * Returns 0 if the hugepage is split successfully.
2656  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2657  * us.
2658  */
2659 int split_huge_page_to_list(struct page *page, struct list_head *list)
2660 {
2661         struct page *head = compound_head(page);
2662         struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2663         struct anon_vma *anon_vma = NULL;
2664         struct address_space *mapping = NULL;
2665         int count, mapcount, extra_pins, ret;
2666         bool mlocked;
2667         unsigned long flags;
2668         pgoff_t end;
2669
2670         VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
2671         VM_BUG_ON_PAGE(!PageLocked(page), page);
2672         VM_BUG_ON_PAGE(!PageCompound(page), page);
2673
2674         if (PageWriteback(page))
2675                 return -EBUSY;
2676
2677         if (PageAnon(head)) {
2678                 /*
2679                  * The caller does not necessarily hold an mmap_sem that would
2680                  * prevent the anon_vma disappearing so we first we take a
2681                  * reference to it and then lock the anon_vma for write. This
2682                  * is similar to page_lock_anon_vma_read except the write lock
2683                  * is taken to serialise against parallel split or collapse
2684                  * operations.
2685                  */
2686                 anon_vma = page_get_anon_vma(head);
2687                 if (!anon_vma) {
2688                         ret = -EBUSY;
2689                         goto out;
2690                 }
2691                 end = -1;
2692                 mapping = NULL;
2693                 anon_vma_lock_write(anon_vma);
2694         } else {
2695                 mapping = head->mapping;
2696
2697                 /* Truncated ? */
2698                 if (!mapping) {
2699                         ret = -EBUSY;
2700                         goto out;
2701                 }
2702
2703                 anon_vma = NULL;
2704                 i_mmap_lock_read(mapping);
2705
2706                 /*
2707                  *__split_huge_page() may need to trim off pages beyond EOF:
2708                  * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2709                  * which cannot be nested inside the page tree lock. So note
2710                  * end now: i_size itself may be changed at any moment, but
2711                  * head page lock is good enough to serialize the trimming.
2712                  */
2713                 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2714         }
2715
2716         /*
2717          * Racy check if we can split the page, before unmap_page() will
2718          * split PMDs
2719          */
2720         if (!can_split_huge_page(head, &extra_pins)) {
2721                 ret = -EBUSY;
2722                 goto out_unlock;
2723         }
2724
2725         mlocked = PageMlocked(page);
2726         unmap_page(head);
2727         VM_BUG_ON_PAGE(compound_mapcount(head), head);
2728
2729         /* Make sure the page is not on per-CPU pagevec as it takes pin */
2730         if (mlocked)
2731                 lru_add_drain();
2732
2733         /* prevent PageLRU to go away from under us, and freeze lru stats */
2734         spin_lock_irqsave(&pgdata->lru_lock, flags);
2735
2736         if (mapping) {
2737                 XA_STATE(xas, &mapping->i_pages, page_index(head));
2738
2739                 /*
2740                  * Check if the head page is present in page cache.
2741                  * We assume all tail are present too, if head is there.
2742                  */
2743                 xa_lock(&mapping->i_pages);
2744                 if (xas_load(&xas) != head)
2745                         goto fail;
2746         }
2747
2748         /* Prevent deferred_split_scan() touching ->_refcount */
2749         spin_lock(&pgdata->split_queue_lock);
2750         count = page_count(head);
2751         mapcount = total_mapcount(head);
2752         if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2753                 if (!list_empty(page_deferred_list(head))) {
2754                         pgdata->split_queue_len--;
2755                         list_del(page_deferred_list(head));
2756                 }
2757                 if (mapping)
2758                         __dec_node_page_state(page, NR_SHMEM_THPS);
2759                 spin_unlock(&pgdata->split_queue_lock);
2760                 __split_huge_page(page, list, end, flags);
2761                 if (PageSwapCache(head)) {
2762                         swp_entry_t entry = { .val = page_private(head) };
2763
2764                         ret = split_swap_cluster(entry);
2765                 } else
2766                         ret = 0;
2767         } else {
2768                 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2769                         pr_alert("total_mapcount: %u, page_count(): %u\n",
2770                                         mapcount, count);
2771                         if (PageTail(page))
2772                                 dump_page(head, NULL);
2773                         dump_page(page, "total_mapcount(head) > 0");
2774                         BUG();
2775                 }
2776                 spin_unlock(&pgdata->split_queue_lock);
2777 fail:           if (mapping)
2778                         xa_unlock(&mapping->i_pages);
2779                 spin_unlock_irqrestore(&pgdata->lru_lock, flags);
2780                 remap_page(head);
2781                 ret = -EBUSY;
2782         }
2783
2784 out_unlock:
2785         if (anon_vma) {
2786                 anon_vma_unlock_write(anon_vma);
2787                 put_anon_vma(anon_vma);
2788         }
2789         if (mapping)
2790                 i_mmap_unlock_read(mapping);
2791 out:
2792         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2793         return ret;
2794 }
2795
2796 void free_transhuge_page(struct page *page)
2797 {
2798         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2799         unsigned long flags;
2800
2801         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2802         if (!list_empty(page_deferred_list(page))) {
2803                 pgdata->split_queue_len--;
2804                 list_del(page_deferred_list(page));
2805         }
2806         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2807         free_compound_page(page);
2808 }
2809
2810 void deferred_split_huge_page(struct page *page)
2811 {
2812         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2813         unsigned long flags;
2814
2815         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2816
2817         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2818         if (list_empty(page_deferred_list(page))) {
2819                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2820                 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2821                 pgdata->split_queue_len++;
2822         }
2823         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2824 }
2825
2826 static unsigned long deferred_split_count(struct shrinker *shrink,
2827                 struct shrink_control *sc)
2828 {
2829         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2830         return READ_ONCE(pgdata->split_queue_len);
2831 }
2832
2833 static unsigned long deferred_split_scan(struct shrinker *shrink,
2834                 struct shrink_control *sc)
2835 {
2836         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2837         unsigned long flags;
2838         LIST_HEAD(list), *pos, *next;
2839         struct page *page;
2840         int split = 0;
2841
2842         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2843         /* Take pin on all head pages to avoid freeing them under us */
2844         list_for_each_safe(pos, next, &pgdata->split_queue) {
2845                 page = list_entry((void *)pos, struct page, mapping);
2846                 page = compound_head(page);
2847                 if (get_page_unless_zero(page)) {
2848                         list_move(page_deferred_list(page), &list);
2849                 } else {
2850                         /* We lost race with put_compound_page() */
2851                         list_del_init(page_deferred_list(page));
2852                         pgdata->split_queue_len--;
2853                 }
2854                 if (!--sc->nr_to_scan)
2855                         break;
2856         }
2857         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2858
2859         list_for_each_safe(pos, next, &list) {
2860                 page = list_entry((void *)pos, struct page, mapping);
2861                 if (!trylock_page(page))
2862                         goto next;
2863                 /* split_huge_page() removes page from list on success */
2864                 if (!split_huge_page(page))
2865                         split++;
2866                 unlock_page(page);
2867 next:
2868                 put_page(page);
2869         }
2870
2871         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2872         list_splice_tail(&list, &pgdata->split_queue);
2873         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2874
2875         /*
2876          * Stop shrinker if we didn't split any page, but the queue is empty.
2877          * This can happen if pages were freed under us.
2878          */
2879         if (!split && list_empty(&pgdata->split_queue))
2880                 return SHRINK_STOP;
2881         return split;
2882 }
2883
2884 static struct shrinker deferred_split_shrinker = {
2885         .count_objects = deferred_split_count,
2886         .scan_objects = deferred_split_scan,
2887         .seeks = DEFAULT_SEEKS,
2888         .flags = SHRINKER_NUMA_AWARE,
2889 };
2890
2891 #ifdef CONFIG_DEBUG_FS
2892 static int split_huge_pages_set(void *data, u64 val)
2893 {
2894         struct zone *zone;
2895         struct page *page;
2896         unsigned long pfn, max_zone_pfn;
2897         unsigned long total = 0, split = 0;
2898
2899         if (val != 1)
2900                 return -EINVAL;
2901
2902         for_each_populated_zone(zone) {
2903                 max_zone_pfn = zone_end_pfn(zone);
2904                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2905                         if (!pfn_valid(pfn))
2906                                 continue;
2907
2908                         page = pfn_to_page(pfn);
2909                         if (!get_page_unless_zero(page))
2910                                 continue;
2911
2912                         if (zone != page_zone(page))
2913                                 goto next;
2914
2915                         if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2916                                 goto next;
2917
2918                         total++;
2919                         lock_page(page);
2920                         if (!split_huge_page(page))
2921                                 split++;
2922                         unlock_page(page);
2923 next:
2924                         put_page(page);
2925                 }
2926         }
2927
2928         pr_info("%lu of %lu THP split\n", split, total);
2929
2930         return 0;
2931 }
2932 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2933                 "%llu\n");
2934
2935 static int __init split_huge_pages_debugfs(void)
2936 {
2937         debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2938                             &split_huge_pages_fops);
2939         return 0;
2940 }
2941 late_initcall(split_huge_pages_debugfs);
2942 #endif
2943
2944 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2945 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
2946                 struct page *page)
2947 {
2948         struct vm_area_struct *vma = pvmw->vma;
2949         struct mm_struct *mm = vma->vm_mm;
2950         unsigned long address = pvmw->address;
2951         pmd_t pmdval;
2952         swp_entry_t entry;
2953         pmd_t pmdswp;
2954
2955         if (!(pvmw->pmd && !pvmw->pte))
2956                 return;
2957
2958         flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
2959         pmdval = *pvmw->pmd;
2960         pmdp_invalidate(vma, address, pvmw->pmd);
2961         if (pmd_dirty(pmdval))
2962                 set_page_dirty(page);
2963         entry = make_migration_entry(page, pmd_write(pmdval));
2964         pmdswp = swp_entry_to_pmd(entry);
2965         if (pmd_soft_dirty(pmdval))
2966                 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
2967         set_pmd_at(mm, address, pvmw->pmd, pmdswp);
2968         page_remove_rmap(page, true);
2969         put_page(page);
2970 }
2971
2972 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
2973 {
2974         struct vm_area_struct *vma = pvmw->vma;
2975         struct mm_struct *mm = vma->vm_mm;
2976         unsigned long address = pvmw->address;
2977         unsigned long mmun_start = address & HPAGE_PMD_MASK;
2978         pmd_t pmde;
2979         swp_entry_t entry;
2980
2981         if (!(pvmw->pmd && !pvmw->pte))
2982                 return;
2983
2984         entry = pmd_to_swp_entry(*pvmw->pmd);
2985         get_page(new);
2986         pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
2987         if (pmd_swp_soft_dirty(*pvmw->pmd))
2988                 pmde = pmd_mksoft_dirty(pmde);
2989         if (is_write_migration_entry(entry))
2990                 pmde = maybe_pmd_mkwrite(pmde, vma);
2991
2992         flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
2993         if (PageAnon(new))
2994                 page_add_anon_rmap(new, vma, mmun_start, true);
2995         else
2996                 page_add_file_rmap(new, true);
2997         set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
2998         if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new))
2999                 mlock_vma_page(new);
3000         update_mmu_cache_pmd(vma, address, pvmw->pmd);
3001 }
3002 #endif