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