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/highmem.h>
  13 #include <linux/hugetlb.h>
  14 #include <linux/mmu_notifier.h>
  15 #include <linux/rmap.h>
  16 #include <linux/swap.h>
  17 #include <linux/shrinker.h>
  18 #include <linux/mm_inline.h>
  19 #include <linux/swapops.h>
  20 #include <linux/dax.h>
  21 #include <linux/khugepaged.h>
  22 #include <linux/freezer.h>
  23 #include <linux/pfn_t.h>
  24 #include <linux/mman.h>
  25 #include <linux/memremap.h>
  26 #include <linux/pagemap.h>
  27 #include <linux/debugfs.h>
  28 #include <linux/migrate.h>
  29 #include <linux/hashtable.h>
  30 #include <linux/userfaultfd_k.h>
  31 #include <linux/page_idle.h>
  32 #include <linux/shmem_fs.h>
  33
  34 #include <asm/tlb.h>
  35 #include <asm/pgalloc.h>
  36 #include "internal.h"
  37
  38 /*
  39  * By default transparent hugepage support is disabled in order that avoid
  40  * to risk increase the memory footprint of applications without a guaranteed
  41  * benefit. When transparent hugepage support is enabled, is for all mappings,
  42  * and khugepaged scans all mappings.
  43  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  44  * for all hugepage allocations.
  45  */
  46 unsigned long transparent_hugepage_flags __read_mostly =
  47 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  48         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  49 #endif
  50 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  51         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  52 #endif
  53         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
  54         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  55         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  56
  57 static struct shrinker deferred_split_shrinker;
  58
  59 static atomic_t huge_zero_refcount;
  60 struct page *huge_zero_page __read_mostly;
  61
  62 static struct page *get_huge_zero_page(void)
  63 {
  64         struct page *zero_page;
  65 retry:
  66         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
  67                 return READ_ONCE(huge_zero_page);
  68
  69         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
  70                         HPAGE_PMD_ORDER);
  71         if (!zero_page) {
  72                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
  73                 return NULL;
  74         }
  75         count_vm_event(THP_ZERO_PAGE_ALLOC);
  76         preempt_disable();
  77         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
  78                 preempt_enable();
  79                 __free_pages(zero_page, compound_order(zero_page));
  80                 goto retry;
  81         }
  82
  83         /* We take additional reference here. It will be put back by shrinker */
  84         atomic_set(&huge_zero_refcount, 2);
  85         preempt_enable();
  86         return READ_ONCE(huge_zero_page);
  87 }
  88
  89 static void put_huge_zero_page(void)
  90 {
  91         /*
  92          * Counter should never go to zero here. Only shrinker can put
  93          * last reference.
  94          */
  95         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
  96 }
  97
  98 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
  99 {
 100         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 101                 return READ_ONCE(huge_zero_page);
 102
 103         if (!get_huge_zero_page())
 104                 return NULL;
 105
 106         if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 107                 put_huge_zero_page();
 108
 109         return READ_ONCE(huge_zero_page);
 110 }
 111
 112 void mm_put_huge_zero_page(struct mm_struct *mm)
 113 {
 114         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 115                 put_huge_zero_page();
 116 }
 117
 118 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 119                                         struct shrink_control *sc)
 120 {
 121         /* we can free zero page only if last reference remains */
 122         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 123 }
 124
 125 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 126                                        struct shrink_control *sc)
 127 {
 128         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 129                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 130                 BUG_ON(zero_page == NULL);
 131                 __free_pages(zero_page, compound_order(zero_page));
 132                 return HPAGE_PMD_NR;
 133         }
 134
 135         return 0;
 136 }
 137
 138 static struct shrinker huge_zero_page_shrinker = {
 139         .count_objects = shrink_huge_zero_page_count,
 140         .scan_objects = shrink_huge_zero_page_scan,
 141         .seeks = DEFAULT_SEEKS,
 142 };
 143
 144 #ifdef CONFIG_SYSFS
 145 static ssize_t enabled_show(struct kobject *kobj,
 146                             struct kobj_attribute *attr, char *buf)
 147 {
 148         if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
 149                 return sprintf(buf, "[always] madvise never\n");
 150         else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
 151                 return sprintf(buf, "always [madvise] never\n");
 152         else
 153                 return sprintf(buf, "always madvise [never]\n");
 154 }
 155
 156 static ssize_t enabled_store(struct kobject *kobj,
 157                              struct kobj_attribute *attr,
 158                              const char *buf, size_t count)
 159 {
 160         ssize_t ret = count;
 161
 162         if (!memcmp("always", buf,
 163                     min(sizeof("always")-1, count))) {
 164                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 165                 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 166         } else if (!memcmp("madvise", buf,
 167                            min(sizeof("madvise")-1, count))) {
 168                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 169                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 170         } else if (!memcmp("never", buf,
 171                            min(sizeof("never")-1, count))) {
 172                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 173                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 174         } else
 175                 ret = -EINVAL;
 176
 177         if (ret > 0) {
 178                 int err = start_stop_khugepaged();
 179                 if (err)
 180                         ret = err;
 181         }
 182         return ret;
 183 }
 184 static struct kobj_attribute enabled_attr =
 185         __ATTR(enabled, 0644, enabled_show, enabled_store);
 186
 187 ssize_t single_hugepage_flag_show(struct kobject *kobj,
 188                                 struct kobj_attribute *attr, char *buf,
 189                                 enum transparent_hugepage_flag flag)
 190 {
 191         return sprintf(buf, "%d\n",
 192                        !!test_bit(flag, &transparent_hugepage_flags));
 193 }
 194
 195 ssize_t single_hugepage_flag_store(struct kobject *kobj,
 196                                  struct kobj_attribute *attr,
 197                                  const char *buf, size_t count,
 198                                  enum transparent_hugepage_flag flag)
 199 {
 200         unsigned long value;
 201         int ret;
 202
 203         ret = kstrtoul(buf, 10, &value);
 204         if (ret < 0)
 205                 return ret;
 206         if (value > 1)
 207                 return -EINVAL;
 208
 209         if (value)
 210                 set_bit(flag, &transparent_hugepage_flags);
 211         else
 212                 clear_bit(flag, &transparent_hugepage_flags);
 213
 214         return count;
 215 }
 216
 217 static ssize_t defrag_show(struct kobject *kobj,
 218                            struct kobj_attribute *attr, char *buf)
 219 {
 220         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 221                 return sprintf(buf, "[always] defer defer+madvise madvise never\n");
 222         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 223                 return sprintf(buf, "always [defer] defer+madvise madvise never\n");
 224         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 225                 return sprintf(buf, "always defer [defer+madvise] madvise never\n");
 226         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 227                 return sprintf(buf, "always defer defer+madvise [madvise] never\n");
 228         return sprintf(buf, "always defer defer+madvise madvise [never]\n");
 229 }
 230
 231 static ssize_t defrag_store(struct kobject *kobj,
 232                             struct kobj_attribute *attr,
 233                             const char *buf, size_t count)
 234 {
 235         if (!memcmp("always", buf,
 236                     min(sizeof("always")-1, count))) {
 237                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 238                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 239                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 240                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 241         } else if (!memcmp("defer", buf,
 242                     min(sizeof("defer")-1, count))) {
 243                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 244                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 245                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 246                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 247         } else if (!memcmp("defer+madvise", buf,
 248                     min(sizeof("defer+madvise")-1, count))) {
 249                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 250                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 251                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 252                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 253         } else if (!memcmp("madvise", buf,
 254                            min(sizeof("madvise")-1, count))) {
 255                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 256                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 257                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 258                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 259         } else if (!memcmp("never", buf,
 260                            min(sizeof("never")-1, count))) {
 261                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 262                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 263                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 264                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 265         } else
 266                 return -EINVAL;
 267
 268         return count;
 269 }
 270 static struct kobj_attribute defrag_attr =
 271         __ATTR(defrag, 0644, defrag_show, defrag_store);
 272
 273 static ssize_t use_zero_page_show(struct kobject *kobj,
 274                 struct kobj_attribute *attr, char *buf)
 275 {
 276         return single_hugepage_flag_show(kobj, attr, buf,
 277                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 278 }
 279 static ssize_t use_zero_page_store(struct kobject *kobj,
 280                 struct kobj_attribute *attr, const char *buf, size_t count)
 281 {
 282         return single_hugepage_flag_store(kobj, attr, buf, count,
 283                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 284 }
 285 static struct kobj_attribute use_zero_page_attr =
 286         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 287
 288 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
 289                 struct kobj_attribute *attr, char *buf)
 290 {
 291         return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE);
 292 }
 293 static struct kobj_attribute hpage_pmd_size_attr =
 294         __ATTR_RO(hpage_pmd_size);
 295
 296 #ifdef CONFIG_DEBUG_VM
 297 static ssize_t debug_cow_show(struct kobject *kobj,
 298                                 struct kobj_attribute *attr, char *buf)
 299 {
 300         return single_hugepage_flag_show(kobj, attr, buf,
 301                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 302 }
 303 static ssize_t debug_cow_store(struct kobject *kobj,
 304                                struct kobj_attribute *attr,
 305                                const char *buf, size_t count)
 306 {
 307         return single_hugepage_flag_store(kobj, attr, buf, count,
 308                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 309 }
 310 static struct kobj_attribute debug_cow_attr =
 311         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 312 #endif /* CONFIG_DEBUG_VM */
 313
 314 static struct attribute *hugepage_attr[] = {
 315         &enabled_attr.attr,
 316         &defrag_attr.attr,
 317         &use_zero_page_attr.attr,
 318         &hpage_pmd_size_attr.attr,
 319 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
 320         &shmem_enabled_attr.attr,
 321 #endif
 322 #ifdef CONFIG_DEBUG_VM
 323         &debug_cow_attr.attr,
 324 #endif
 325         NULL,
 326 };
 327
 328 static struct attribute_group hugepage_attr_group = {
 329         .attrs = hugepage_attr,
 330 };
 331
 332 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 333 {
 334         int err;
 335
 336         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 337         if (unlikely(!*hugepage_kobj)) {
 338                 pr_err("failed to create transparent hugepage kobject\n");
 339                 return -ENOMEM;
 340         }
 341
 342         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 343         if (err) {
 344                 pr_err("failed to register transparent hugepage group\n");
 345                 goto delete_obj;
 346         }
 347
 348         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 349         if (err) {
 350                 pr_err("failed to register transparent hugepage group\n");
 351                 goto remove_hp_group;
 352         }
 353
 354         return 0;
 355
 356 remove_hp_group:
 357         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 358 delete_obj:
 359         kobject_put(*hugepage_kobj);
 360         return err;
 361 }
 362
 363 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 364 {
 365         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 366         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 367         kobject_put(hugepage_kobj);
 368 }
 369 #else
 370 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 371 {
 372         return 0;
 373 }
 374
 375 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 376 {
 377 }
 378 #endif /* CONFIG_SYSFS */
 379
 380 static int __init hugepage_init(void)
 381 {
 382         int err;
 383         struct kobject *hugepage_kobj;
 384
 385         if (!has_transparent_hugepage()) {
 386                 transparent_hugepage_flags = 0;
 387                 return -EINVAL;
 388         }
 389
 390         /*
 391          * hugepages can't be allocated by the buddy allocator
 392          */
 393         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
 394         /*
 395          * we use page->mapping and page->index in second tail page
 396          * as list_head: assuming THP order >= 2
 397          */
 398         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 399
 400         err = hugepage_init_sysfs(&hugepage_kobj);
 401         if (err)
 402                 goto err_sysfs;
 403
 404         err = khugepaged_init();
 405         if (err)
 406                 goto err_slab;
 407
 408         err = register_shrinker(&huge_zero_page_shrinker);
 409         if (err)
 410                 goto err_hzp_shrinker;
 411         err = register_shrinker(&deferred_split_shrinker);
 412         if (err)
 413                 goto err_split_shrinker;
 414
 415         /*
 416          * By default disable transparent hugepages on smaller systems,
 417          * where the extra memory used could hurt more than TLB overhead
 418          * is likely to save.  The admin can still enable it through /sys.
 419          */
 420         if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
 421                 transparent_hugepage_flags = 0;
 422                 return 0;
 423         }
 424
 425         err = start_stop_khugepaged();
 426         if (err)
 427                 goto err_khugepaged;
 428
 429         return 0;
 430 err_khugepaged:
 431         unregister_shrinker(&deferred_split_shrinker);
 432 err_split_shrinker:
 433         unregister_shrinker(&huge_zero_page_shrinker);
 434 err_hzp_shrinker:
 435         khugepaged_destroy();
 436 err_slab:
 437         hugepage_exit_sysfs(hugepage_kobj);
 438 err_sysfs:
 439         return err;
 440 }
 441 subsys_initcall(hugepage_init);
 442
 443 static int __init setup_transparent_hugepage(char *str)
 444 {
 445         int ret = 0;
 446         if (!str)
 447                 goto out;
 448         if (!strcmp(str, "always")) {
 449                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 450                         &transparent_hugepage_flags);
 451                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 452                           &transparent_hugepage_flags);
 453                 ret = 1;
 454         } else if (!strcmp(str, "madvise")) {
 455                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 456                           &transparent_hugepage_flags);
 457                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 458                         &transparent_hugepage_flags);
 459                 ret = 1;
 460         } else if (!strcmp(str, "never")) {
 461                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 462                           &transparent_hugepage_flags);
 463                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 464                           &transparent_hugepage_flags);
 465                 ret = 1;
 466         }
 467 out:
 468         if (!ret)
 469                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 470         return ret;
 471 }
 472 __setup("transparent_hugepage=", setup_transparent_hugepage);
 473
 474 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 475 {
 476         if (likely(vma->vm_flags & VM_WRITE))
 477                 pmd = pmd_mkwrite(pmd);
 478         return pmd;
 479 }
 480
 481 static inline struct list_head *page_deferred_list(struct page *page)
 482 {
 483         /*
 484          * ->lru in the tail pages is occupied by compound_head.
 485          * Let's use ->mapping + ->index in the second tail page as list_head.
 486          */
 487         return (struct list_head *)&page[2].mapping;
 488 }
 489
 490 void prep_transhuge_page(struct page *page)
 491 {
 492         /*
 493          * we use page->mapping and page->indexlru in second tail page
 494          * as list_head: assuming THP order >= 2
 495          */
 496
 497         INIT_LIST_HEAD(page_deferred_list(page));
 498         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 499 }
 500
 501 unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len,
 502                 loff_t off, unsigned long flags, unsigned long size)
 503 {
 504         unsigned long addr;
 505         loff_t off_end = off + len;
 506         loff_t off_align = round_up(off, size);
 507         unsigned long len_pad;
 508
 509         if (off_end <= off_align || (off_end - off_align) < size)
 510                 return 0;
 511
 512         len_pad = len + size;
 513         if (len_pad < len || (off + len_pad) < off)
 514                 return 0;
 515
 516         addr = current->mm->get_unmapped_area(filp, 0, len_pad,
 517                                               off >> PAGE_SHIFT, flags);
 518         if (IS_ERR_VALUE(addr))
 519                 return 0;
 520
 521         addr += (off - addr) & (size - 1);
 522         return addr;
 523 }
 524
 525 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 526                 unsigned long len, unsigned long pgoff, unsigned long flags)
 527 {
 528         loff_t off = (loff_t)pgoff << PAGE_SHIFT;
 529
 530         if (addr)
 531                 goto out;
 532         if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
 533                 goto out;
 534
 535         addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE);
 536         if (addr)
 537                 return addr;
 538
 539  out:
 540         return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
 541 }
 542 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
 543
 544 static int __do_huge_pmd_anonymous_page(struct vm_fault *vmf, struct page *page,
 545                 gfp_t gfp)
 546 {
 547         struct vm_area_struct *vma = vmf->vma;
 548         struct mem_cgroup *memcg;
 549         pgtable_t pgtable;
 550         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 551
 552         VM_BUG_ON_PAGE(!PageCompound(page), page);
 553
 554         if (mem_cgroup_try_charge(page, vma->vm_mm, gfp, &memcg, true)) {
 555                 put_page(page);
 556                 count_vm_event(THP_FAULT_FALLBACK);
 557                 return VM_FAULT_FALLBACK;
 558         }
 559
 560         pgtable = pte_alloc_one(vma->vm_mm, haddr);
 561         if (unlikely(!pgtable)) {
 562                 mem_cgroup_cancel_charge(page, memcg, true);
 563                 put_page(page);
 564                 return VM_FAULT_OOM;
 565         }
 566
 567         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 568         /*
 569          * The memory barrier inside __SetPageUptodate makes sure that
 570          * clear_huge_page writes become visible before the set_pmd_at()
 571          * write.
 572          */
 573         __SetPageUptodate(page);
 574
 575         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 576         if (unlikely(!pmd_none(*vmf->pmd))) {
 577                 spin_unlock(vmf->ptl);
 578                 mem_cgroup_cancel_charge(page, memcg, true);
 579                 put_page(page);
 580                 pte_free(vma->vm_mm, pgtable);
 581         } else {
 582                 pmd_t entry;
 583
 584                 /* Deliver the page fault to userland */
 585                 if (userfaultfd_missing(vma)) {
 586                         int ret;
 587
 588                         spin_unlock(vmf->ptl);
 589                         mem_cgroup_cancel_charge(page, memcg, true);
 590                         put_page(page);
 591                         pte_free(vma->vm_mm, pgtable);
 592                         ret = handle_userfault(vmf, VM_UFFD_MISSING);
 593                         VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 594                         return ret;
 595                 }
 596
 597                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 598                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 599                 page_add_new_anon_rmap(page, vma, haddr, true);
 600                 mem_cgroup_commit_charge(page, memcg, false, true);
 601                 lru_cache_add_active_or_unevictable(page, vma);
 602                 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
 603                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
 604                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 605                 atomic_long_inc(&vma->vm_mm->nr_ptes);
 606                 spin_unlock(vmf->ptl);
 607                 count_vm_event(THP_FAULT_ALLOC);
 608         }
 609
 610         return 0;
 611 }
 612
 613 /*
 614  * always: directly stall for all thp allocations
 615  * defer: wake kswapd and fail if not immediately available
 616  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
 617  *                fail if not immediately available
 618  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
 619  *          available
 620  * never: never stall for any thp allocation
 621  */
 622 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
 623 {
 624         const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
 625
 626         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 627                 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
 628         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 629                 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
 630         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 631                 return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM :
 632                                                              __GFP_KSWAPD_RECLAIM);
 633         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 634                 return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM :
 635                                                              0);
 636         return GFP_TRANSHUGE_LIGHT;
 637 }
 638
 639 /* Caller must hold page table lock. */
 640 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 641                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 642                 struct page *zero_page)
 643 {
 644         pmd_t entry;
 645         if (!pmd_none(*pmd))
 646                 return false;
 647         entry = mk_pmd(zero_page, vma->vm_page_prot);
 648         entry = pmd_mkhuge(entry);
 649         if (pgtable)
 650                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 651         set_pmd_at(mm, haddr, pmd, entry);
 652         atomic_long_inc(&mm->nr_ptes);
 653         return true;
 654 }
 655
 656 int do_huge_pmd_anonymous_page(struct vm_fault *vmf)
 657 {
 658         struct vm_area_struct *vma = vmf->vma;
 659         gfp_t gfp;
 660         struct page *page;
 661         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 662
 663         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 664                 return VM_FAULT_FALLBACK;
 665         if (unlikely(anon_vma_prepare(vma)))
 666                 return VM_FAULT_OOM;
 667         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 668                 return VM_FAULT_OOM;
 669         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
 670                         !mm_forbids_zeropage(vma->vm_mm) &&
 671                         transparent_hugepage_use_zero_page()) {
 672                 pgtable_t pgtable;
 673                 struct page *zero_page;
 674                 bool set;
 675                 int ret;
 676                 pgtable = pte_alloc_one(vma->vm_mm, haddr);
 677                 if (unlikely(!pgtable))
 678                         return VM_FAULT_OOM;
 679                 zero_page = mm_get_huge_zero_page(vma->vm_mm);
 680                 if (unlikely(!zero_page)) {
 681                         pte_free(vma->vm_mm, pgtable);
 682                         count_vm_event(THP_FAULT_FALLBACK);
 683                         return VM_FAULT_FALLBACK;
 684                 }
 685                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 686                 ret = 0;
 687                 set = false;
 688                 if (pmd_none(*vmf->pmd)) {
 689                         if (userfaultfd_missing(vma)) {
 690                                 spin_unlock(vmf->ptl);
 691                                 ret = handle_userfault(vmf, VM_UFFD_MISSING);
 692                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 693                         } else {
 694                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
 695                                                    haddr, vmf->pmd, zero_page);
 696                                 spin_unlock(vmf->ptl);
 697                                 set = true;
 698                         }
 699                 } else
 700                         spin_unlock(vmf->ptl);
 701                 if (!set)
 702                         pte_free(vma->vm_mm, pgtable);
 703                 return ret;
 704         }
 705         gfp = alloc_hugepage_direct_gfpmask(vma);
 706         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 707         if (unlikely(!page)) {
 708                 count_vm_event(THP_FAULT_FALLBACK);
 709                 return VM_FAULT_FALLBACK;
 710         }
 711         prep_transhuge_page(page);
 712         return __do_huge_pmd_anonymous_page(vmf, page, gfp);
 713 }
 714
 715 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 716                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write)
 717 {
 718         struct mm_struct *mm = vma->vm_mm;
 719         pmd_t entry;
 720         spinlock_t *ptl;
 721
 722         ptl = pmd_lock(mm, pmd);
 723         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 724         if (pfn_t_devmap(pfn))
 725                 entry = pmd_mkdevmap(entry);
 726         if (write) {
 727                 entry = pmd_mkyoung(pmd_mkdirty(entry));
 728                 entry = maybe_pmd_mkwrite(entry, vma);
 729         }
 730         set_pmd_at(mm, addr, pmd, entry);
 731         update_mmu_cache_pmd(vma, addr, pmd);
 732         spin_unlock(ptl);
 733 }
 734
 735 int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 736                         pmd_t *pmd, pfn_t pfn, bool write)
 737 {
 738         pgprot_t pgprot = vma->vm_page_prot;
 739         /*
 740          * If we had pmd_special, we could avoid all these restrictions,
 741          * but we need to be consistent with PTEs and architectures that
 742          * can't support a 'special' bit.
 743          */
 744         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
 745         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 746                                                 (VM_PFNMAP|VM_MIXEDMAP));
 747         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 748         BUG_ON(!pfn_t_devmap(pfn));
 749
 750         if (addr < vma->vm_start || addr >= vma->vm_end)
 751                 return VM_FAULT_SIGBUS;
 752
 753         track_pfn_insert(vma, &pgprot, pfn);
 754
 755         insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
 756         return VM_FAULT_NOPAGE;
 757 }
 758 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
 759
 760 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
 761                 pmd_t *pmd)
 762 {
 763         pmd_t _pmd;
 764
 765         /*
 766          * We should set the dirty bit only for FOLL_WRITE but for now
 767          * the dirty bit in the pmd is meaningless.  And if the dirty
 768          * bit will become meaningful and we'll only set it with
 769          * FOLL_WRITE, an atomic set_bit will be required on the pmd to
 770          * set the young bit, instead of the current set_pmd_at.
 771          */
 772         _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
 773         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
 774                                 pmd, _pmd,  1))
 775                 update_mmu_cache_pmd(vma, addr, pmd);
 776 }
 777
 778 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 779                 pmd_t *pmd, int flags)
 780 {
 781         unsigned long pfn = pmd_pfn(*pmd);
 782         struct mm_struct *mm = vma->vm_mm;
 783         struct dev_pagemap *pgmap;
 784         struct page *page;
 785
 786         assert_spin_locked(pmd_lockptr(mm, pmd));
 787
 788         /*
 789          * When we COW a devmap PMD entry, we split it into PTEs, so we should
 790          * not be in this function with `flags & FOLL_COW` set.
 791          */
 792         WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
 793
 794         if (flags & FOLL_WRITE && !pmd_write(*pmd))
 795                 return NULL;
 796
 797         if (pmd_present(*pmd) && pmd_devmap(*pmd))
 798                 /* pass */;
 799         else
 800                 return NULL;
 801
 802         if (flags & FOLL_TOUCH)
 803                 touch_pmd(vma, addr, pmd);
 804
 805         /*
 806          * device mapped pages can only be returned if the
 807          * caller will manage the page reference count.
 808          */
 809         if (!(flags & FOLL_GET))
 810                 return ERR_PTR(-EEXIST);
 811
 812         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
 813         pgmap = get_dev_pagemap(pfn, NULL);
 814         if (!pgmap)
 815                 return ERR_PTR(-EFAULT);
 816         page = pfn_to_page(pfn);
 817         get_page(page);
 818         put_dev_pagemap(pgmap);
 819
 820         return page;
 821 }
 822
 823 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 824                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 825                   struct vm_area_struct *vma)
 826 {
 827         spinlock_t *dst_ptl, *src_ptl;
 828         struct page *src_page;
 829         pmd_t pmd;
 830         pgtable_t pgtable = NULL;
 831         int ret = -ENOMEM;
 832
 833         /* Skip if can be re-fill on fault */
 834         if (!vma_is_anonymous(vma))
 835                 return 0;
 836
 837         pgtable = pte_alloc_one(dst_mm, addr);
 838         if (unlikely(!pgtable))
 839                 goto out;
 840
 841         dst_ptl = pmd_lock(dst_mm, dst_pmd);
 842         src_ptl = pmd_lockptr(src_mm, src_pmd);
 843         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 844
 845         ret = -EAGAIN;
 846         pmd = *src_pmd;
 847         if (unlikely(!pmd_trans_huge(pmd))) {
 848                 pte_free(dst_mm, pgtable);
 849                 goto out_unlock;
 850         }
 851         /*
 852          * When page table lock is held, the huge zero pmd should not be
 853          * under splitting since we don't split the page itself, only pmd to
 854          * a page table.
 855          */
 856         if (is_huge_zero_pmd(pmd)) {
 857                 struct page *zero_page;
 858                 /*
 859                  * get_huge_zero_page() will never allocate a new page here,
 860                  * since we already have a zero page to copy. It just takes a
 861                  * reference.
 862                  */
 863                 zero_page = mm_get_huge_zero_page(dst_mm);
 864                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
 865                                 zero_page);
 866                 ret = 0;
 867                 goto out_unlock;
 868         }
 869
 870         src_page = pmd_page(pmd);
 871         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
 872         get_page(src_page);
 873         page_dup_rmap(src_page, true);
 874         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 875         atomic_long_inc(&dst_mm->nr_ptes);
 876         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
 877
 878         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 879         pmd = pmd_mkold(pmd_wrprotect(pmd));
 880         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 881
 882         ret = 0;
 883 out_unlock:
 884         spin_unlock(src_ptl);
 885         spin_unlock(dst_ptl);
 886 out:
 887         return ret;
 888 }
 889
 890 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
 891 {
 892         pmd_t entry;
 893         unsigned long haddr;
 894         bool write = vmf->flags & FAULT_FLAG_WRITE;
 895
 896         vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
 897         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
 898                 goto unlock;
 899
 900         entry = pmd_mkyoung(orig_pmd);
 901         if (write)
 902                 entry = pmd_mkdirty(entry);
 903         haddr = vmf->address & HPAGE_PMD_MASK;
 904         if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
 905                 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
 906
 907 unlock:
 908         spin_unlock(vmf->ptl);
 909 }
 910
 911 static int do_huge_pmd_wp_page_fallback(struct vm_fault *vmf, pmd_t orig_pmd,
 912                 struct page *page)
 913 {
 914         struct vm_area_struct *vma = vmf->vma;
 915         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 916         struct mem_cgroup *memcg;
 917         pgtable_t pgtable;
 918         pmd_t _pmd;
 919         int ret = 0, i;
 920         struct page **pages;
 921         unsigned long mmun_start;       /* For mmu_notifiers */
 922         unsigned long mmun_end;         /* For mmu_notifiers */
 923
 924         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
 925                         GFP_KERNEL);
 926         if (unlikely(!pages)) {
 927                 ret |= VM_FAULT_OOM;
 928                 goto out;
 929         }
 930
 931         for (i = 0; i < HPAGE_PMD_NR; i++) {
 932                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
 933                                                vmf->address, page_to_nid(page));
 934                 if (unlikely(!pages[i] ||
 935                              mem_cgroup_try_charge(pages[i], vma->vm_mm,
 936                                      GFP_KERNEL, &memcg, false))) {
 937                         if (pages[i])
 938                                 put_page(pages[i]);
 939                         while (--i >= 0) {
 940                                 memcg = (void *)page_private(pages[i]);
 941                                 set_page_private(pages[i], 0);
 942                                 mem_cgroup_cancel_charge(pages[i], memcg,
 943                                                 false);
 944                                 put_page(pages[i]);
 945                         }
 946                         kfree(pages);
 947                         ret |= VM_FAULT_OOM;
 948                         goto out;
 949                 }
 950                 set_page_private(pages[i], (unsigned long)memcg);
 951         }
 952
 953         for (i = 0; i < HPAGE_PMD_NR; i++) {
 954                 copy_user_highpage(pages[i], page + i,
 955                                    haddr + PAGE_SIZE * i, vma);
 956                 __SetPageUptodate(pages[i]);
 957                 cond_resched();
 958         }
 959
 960         mmun_start = haddr;
 961         mmun_end   = haddr + HPAGE_PMD_SIZE;
 962         mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
 963
 964         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 965         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
 966                 goto out_free_pages;
 967         VM_BUG_ON_PAGE(!PageHead(page), page);
 968
 969         pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
 970         /* leave pmd empty until pte is filled */
 971
 972         pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
 973         pmd_populate(vma->vm_mm, &_pmd, pgtable);
 974
 975         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
 976                 pte_t entry;
 977                 entry = mk_pte(pages[i], vma->vm_page_prot);
 978                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
 979                 memcg = (void *)page_private(pages[i]);
 980                 set_page_private(pages[i], 0);
 981                 page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
 982                 mem_cgroup_commit_charge(pages[i], memcg, false, false);
 983                 lru_cache_add_active_or_unevictable(pages[i], vma);
 984                 vmf->pte = pte_offset_map(&_pmd, haddr);
 985                 VM_BUG_ON(!pte_none(*vmf->pte));
 986                 set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
 987                 pte_unmap(vmf->pte);
 988         }
 989         kfree(pages);
 990
 991         smp_wmb(); /* make pte visible before pmd */
 992         pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
 993         page_remove_rmap(page, true);
 994         spin_unlock(vmf->ptl);
 995
 996         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
 997
 998         ret |= VM_FAULT_WRITE;
 999         put_page(page);
1000
1001 out:
1002         return ret;
1003
1004 out_free_pages:
1005         spin_unlock(vmf->ptl);
1006         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1007         for (i = 0; i < HPAGE_PMD_NR; i++) {
1008                 memcg = (void *)page_private(pages[i]);
1009                 set_page_private(pages[i], 0);
1010                 mem_cgroup_cancel_charge(pages[i], memcg, false);
1011                 put_page(pages[i]);
1012         }
1013         kfree(pages);
1014         goto out;
1015 }
1016
1017 int do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1018 {
1019         struct vm_area_struct *vma = vmf->vma;
1020         struct page *page = NULL, *new_page;
1021         struct mem_cgroup *memcg;
1022         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1023         unsigned long mmun_start;       /* For mmu_notifiers */
1024         unsigned long mmun_end;         /* For mmu_notifiers */
1025         gfp_t huge_gfp;                 /* for allocation and charge */
1026         int ret = 0;
1027
1028         vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1029         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1030         if (is_huge_zero_pmd(orig_pmd))
1031                 goto alloc;
1032         spin_lock(vmf->ptl);
1033         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1034                 goto out_unlock;
1035
1036         page = pmd_page(orig_pmd);
1037         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1038         /*
1039          * We can only reuse the page if nobody else maps the huge page or it's
1040          * part.
1041          */
1042         if (page_trans_huge_mapcount(page, NULL) == 1) {
1043                 pmd_t entry;
1044                 entry = pmd_mkyoung(orig_pmd);
1045                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1046                 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry,  1))
1047                         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1048                 ret |= VM_FAULT_WRITE;
1049                 goto out_unlock;
1050         }
1051         get_page(page);
1052         spin_unlock(vmf->ptl);
1053 alloc:
1054         if (transparent_hugepage_enabled(vma) &&
1055             !transparent_hugepage_debug_cow()) {
1056                 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1057                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1058         } else
1059                 new_page = NULL;
1060
1061         if (likely(new_page)) {
1062                 prep_transhuge_page(new_page);
1063         } else {
1064                 if (!page) {
1065                         split_huge_pmd(vma, vmf->pmd, vmf->address);
1066                         ret |= VM_FAULT_FALLBACK;
1067                 } else {
1068                         ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
1069                         if (ret & VM_FAULT_OOM) {
1070                                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1071                                 ret |= VM_FAULT_FALLBACK;
1072                         }
1073                         put_page(page);
1074                 }
1075                 count_vm_event(THP_FAULT_FALLBACK);
1076                 goto out;
1077         }
1078
1079         if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm,
1080                                         huge_gfp, &memcg, true))) {
1081                 put_page(new_page);
1082                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1083                 if (page)
1084                         put_page(page);
1085                 ret |= VM_FAULT_FALLBACK;
1086                 count_vm_event(THP_FAULT_FALLBACK);
1087                 goto out;
1088         }
1089
1090         count_vm_event(THP_FAULT_ALLOC);
1091
1092         if (!page)
1093                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1094         else
1095                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1096         __SetPageUptodate(new_page);
1097
1098         mmun_start = haddr;
1099         mmun_end   = haddr + HPAGE_PMD_SIZE;
1100         mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
1101
1102         spin_lock(vmf->ptl);
1103         if (page)
1104                 put_page(page);
1105         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1106                 spin_unlock(vmf->ptl);
1107                 mem_cgroup_cancel_charge(new_page, memcg, true);
1108                 put_page(new_page);
1109                 goto out_mn;
1110         } else {
1111                 pmd_t entry;
1112                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1113                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1114                 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1115                 page_add_new_anon_rmap(new_page, vma, haddr, true);
1116                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1117                 lru_cache_add_active_or_unevictable(new_page, vma);
1118                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1119                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1120                 if (!page) {
1121                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1122                 } else {
1123                         VM_BUG_ON_PAGE(!PageHead(page), page);
1124                         page_remove_rmap(page, true);
1125                         put_page(page);
1126                 }
1127                 ret |= VM_FAULT_WRITE;
1128         }
1129         spin_unlock(vmf->ptl);
1130 out_mn:
1131         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1132 out:
1133         return ret;
1134 out_unlock:
1135         spin_unlock(vmf->ptl);
1136         return ret;
1137 }
1138
1139 /*
1140  * FOLL_FORCE can write to even unwritable pmd's, but only
1141  * after we've gone through a COW cycle and they are dirty.
1142  */
1143 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1144 {
1145         return pmd_write(pmd) ||
1146                ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1147 }
1148
1149 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1150                                    unsigned long addr,
1151                                    pmd_t *pmd,
1152                                    unsigned int flags)
1153 {
1154         struct mm_struct *mm = vma->vm_mm;
1155         struct page *page = NULL;
1156
1157         assert_spin_locked(pmd_lockptr(mm, pmd));
1158
1159         if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1160                 goto out;
1161
1162         /* Avoid dumping huge zero page */
1163         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1164                 return ERR_PTR(-EFAULT);
1165
1166         /* Full NUMA hinting faults to serialise migration in fault paths */
1167         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1168                 goto out;
1169
1170         page = pmd_page(*pmd);
1171         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1172         if (flags & FOLL_TOUCH)
1173                 touch_pmd(vma, addr, pmd);
1174         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1175                 /*
1176                  * We don't mlock() pte-mapped THPs. This way we can avoid
1177                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1178                  *
1179                  * For anon THP:
1180                  *
1181                  * In most cases the pmd is the only mapping of the page as we
1182                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1183                  * writable private mappings in populate_vma_page_range().
1184                  *
1185                  * The only scenario when we have the page shared here is if we
1186                  * mlocking read-only mapping shared over fork(). We skip
1187                  * mlocking such pages.
1188                  *
1189                  * For file THP:
1190                  *
1191                  * We can expect PageDoubleMap() to be stable under page lock:
1192                  * for file pages we set it in page_add_file_rmap(), which
1193                  * requires page to be locked.
1194                  */
1195
1196                 if (PageAnon(page) && compound_mapcount(page) != 1)
1197                         goto skip_mlock;
1198                 if (PageDoubleMap(page) || !page->mapping)
1199                         goto skip_mlock;
1200                 if (!trylock_page(page))
1201                         goto skip_mlock;
1202                 lru_add_drain();
1203                 if (page->mapping && !PageDoubleMap(page))
1204                         mlock_vma_page(page);
1205                 unlock_page(page);
1206         }
1207 skip_mlock:
1208         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1209         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1210         if (flags & FOLL_GET)
1211                 get_page(page);
1212
1213 out:
1214         return page;
1215 }
1216
1217 /* NUMA hinting page fault entry point for trans huge pmds */
1218 int do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1219 {
1220         struct vm_area_struct *vma = vmf->vma;
1221         struct anon_vma *anon_vma = NULL;
1222         struct page *page;
1223         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1224         int page_nid = -1, this_nid = numa_node_id();
1225         int target_nid, last_cpupid = -1;
1226         bool page_locked;
1227         bool migrated = false;
1228         bool was_writable;
1229         int flags = 0;
1230
1231         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1232         if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1233                 goto out_unlock;
1234
1235         /*
1236          * If there are potential migrations, wait for completion and retry
1237          * without disrupting NUMA hinting information. Do not relock and
1238          * check_same as the page may no longer be mapped.
1239          */
1240         if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1241                 page = pmd_page(*vmf->pmd);
1242                 spin_unlock(vmf->ptl);
1243                 wait_on_page_locked(page);
1244                 goto out;
1245         }
1246
1247         page = pmd_page(pmd);
1248         BUG_ON(is_huge_zero_page(page));
1249         page_nid = page_to_nid(page);
1250         last_cpupid = page_cpupid_last(page);
1251         count_vm_numa_event(NUMA_HINT_FAULTS);
1252         if (page_nid == this_nid) {
1253                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1254                 flags |= TNF_FAULT_LOCAL;
1255         }
1256
1257         /* See similar comment in do_numa_page for explanation */
1258         if (!pmd_write(pmd))
1259                 flags |= TNF_NO_GROUP;
1260
1261         /*
1262          * Acquire the page lock to serialise THP migrations but avoid dropping
1263          * page_table_lock if at all possible
1264          */
1265         page_locked = trylock_page(page);
1266         target_nid = mpol_misplaced(page, vma, haddr);
1267         if (target_nid == -1) {
1268                 /* If the page was locked, there are no parallel migrations */
1269                 if (page_locked)
1270                         goto clear_pmdnuma;
1271         }
1272
1273         /* Migration could have started since the pmd_trans_migrating check */
1274         if (!page_locked) {
1275                 spin_unlock(vmf->ptl);
1276                 wait_on_page_locked(page);
1277                 page_nid = -1;
1278                 goto out;
1279         }
1280
1281         /*
1282          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1283          * to serialises splits
1284          */
1285         get_page(page);
1286         spin_unlock(vmf->ptl);
1287         anon_vma = page_lock_anon_vma_read(page);
1288
1289         /* Confirm the PMD did not change while page_table_lock was released */
1290         spin_lock(vmf->ptl);
1291         if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1292                 unlock_page(page);
1293                 put_page(page);
1294                 page_nid = -1;
1295                 goto out_unlock;
1296         }
1297
1298         /* Bail if we fail to protect against THP splits for any reason */
1299         if (unlikely(!anon_vma)) {
1300                 put_page(page);
1301                 page_nid = -1;
1302                 goto clear_pmdnuma;
1303         }
1304
1305         /*
1306          * Migrate the THP to the requested node, returns with page unlocked
1307          * and access rights restored.
1308          */
1309         spin_unlock(vmf->ptl);
1310         migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1311                                 vmf->pmd, pmd, vmf->address, page, target_nid);
1312         if (migrated) {
1313                 flags |= TNF_MIGRATED;
1314                 page_nid = target_nid;
1315         } else
1316                 flags |= TNF_MIGRATE_FAIL;
1317
1318         goto out;
1319 clear_pmdnuma:
1320         BUG_ON(!PageLocked(page));
1321         was_writable = pmd_write(pmd);
1322         pmd = pmd_modify(pmd, vma->vm_page_prot);
1323         pmd = pmd_mkyoung(pmd);
1324         if (was_writable)
1325                 pmd = pmd_mkwrite(pmd);
1326         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1327         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1328         unlock_page(page);
1329 out_unlock:
1330         spin_unlock(vmf->ptl);
1331
1332 out:
1333         if (anon_vma)
1334                 page_unlock_anon_vma_read(anon_vma);
1335
1336         if (page_nid != -1)
1337                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1338                                 vmf->flags);
1339
1340         return 0;
1341 }
1342
1343 /*
1344  * Return true if we do MADV_FREE successfully on entire pmd page.
1345  * Otherwise, return false.
1346  */
1347 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1348                 pmd_t *pmd, unsigned long addr, unsigned long next)
1349 {
1350         spinlock_t *ptl;
1351         pmd_t orig_pmd;
1352         struct page *page;
1353         struct mm_struct *mm = tlb->mm;
1354         bool ret = false;
1355
1356         tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1357
1358         ptl = pmd_trans_huge_lock(pmd, vma);
1359         if (!ptl)
1360                 goto out_unlocked;
1361
1362         orig_pmd = *pmd;
1363         if (is_huge_zero_pmd(orig_pmd))
1364                 goto out;
1365
1366         page = pmd_page(orig_pmd);
1367         /*
1368          * If other processes are mapping this page, we couldn't discard
1369          * the page unless they all do MADV_FREE so let's skip the page.
1370          */
1371         if (page_mapcount(page) != 1)
1372                 goto out;
1373
1374         if (!trylock_page(page))
1375                 goto out;
1376
1377         /*
1378          * If user want to discard part-pages of THP, split it so MADV_FREE
1379          * will deactivate only them.
1380          */
1381         if (next - addr != HPAGE_PMD_SIZE) {
1382                 get_page(page);
1383                 spin_unlock(ptl);
1384                 split_huge_page(page);
1385                 put_page(page);
1386                 unlock_page(page);
1387                 goto out_unlocked;
1388         }
1389
1390         if (PageDirty(page))
1391                 ClearPageDirty(page);
1392         unlock_page(page);
1393
1394         if (PageActive(page))
1395                 deactivate_page(page);
1396
1397         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1398                 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1399                         tlb->fullmm);
1400                 orig_pmd = pmd_mkold(orig_pmd);
1401                 orig_pmd = pmd_mkclean(orig_pmd);
1402
1403                 set_pmd_at(mm, addr, pmd, orig_pmd);
1404                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1405         }
1406         ret = true;
1407 out:
1408         spin_unlock(ptl);
1409 out_unlocked:
1410         return ret;
1411 }
1412
1413 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1414 {
1415         pgtable_t pgtable;
1416
1417         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1418         pte_free(mm, pgtable);
1419         atomic_long_dec(&mm->nr_ptes);
1420 }
1421
1422 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1423                  pmd_t *pmd, unsigned long addr)
1424 {
1425         pmd_t orig_pmd;
1426         spinlock_t *ptl;
1427
1428         tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1429
1430         ptl = __pmd_trans_huge_lock(pmd, vma);
1431         if (!ptl)
1432                 return 0;
1433         /*
1434          * For architectures like ppc64 we look at deposited pgtable
1435          * when calling pmdp_huge_get_and_clear. So do the
1436          * pgtable_trans_huge_withdraw after finishing pmdp related
1437          * operations.
1438          */
1439         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1440                         tlb->fullmm);
1441         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1442         if (vma_is_dax(vma)) {
1443                 spin_unlock(ptl);
1444                 if (is_huge_zero_pmd(orig_pmd))
1445                         tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1446         } else if (is_huge_zero_pmd(orig_pmd)) {
1447                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1448                 atomic_long_dec(&tlb->mm->nr_ptes);
1449                 spin_unlock(ptl);
1450                 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1451         } else {
1452                 struct page *page = pmd_page(orig_pmd);
1453                 page_remove_rmap(page, true);
1454                 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1455                 VM_BUG_ON_PAGE(!PageHead(page), page);
1456                 if (PageAnon(page)) {
1457                         pgtable_t pgtable;
1458                         pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1459                         pte_free(tlb->mm, pgtable);
1460                         atomic_long_dec(&tlb->mm->nr_ptes);
1461                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1462                 } else {
1463                         if (arch_needs_pgtable_deposit())
1464                                 zap_deposited_table(tlb->mm, pmd);
1465                         add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1466                 }
1467                 spin_unlock(ptl);
1468                 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1469         }
1470         return 1;
1471 }
1472
1473 #ifndef pmd_move_must_withdraw
1474 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1475                                          spinlock_t *old_pmd_ptl,
1476                                          struct vm_area_struct *vma)
1477 {
1478         /*
1479          * With split pmd lock we also need to move preallocated
1480          * PTE page table if new_pmd is on different PMD page table.
1481          *
1482          * We also don't deposit and withdraw tables for file pages.
1483          */
1484         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1485 }
1486 #endif
1487
1488 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1489                   unsigned long new_addr, unsigned long old_end,
1490                   pmd_t *old_pmd, pmd_t *new_pmd, bool *need_flush)
1491 {
1492         spinlock_t *old_ptl, *new_ptl;
1493         pmd_t pmd;
1494         struct mm_struct *mm = vma->vm_mm;
1495         bool force_flush = false;
1496
1497         if ((old_addr & ~HPAGE_PMD_MASK) ||
1498             (new_addr & ~HPAGE_PMD_MASK) ||
1499             old_end - old_addr < HPAGE_PMD_SIZE)
1500                 return false;
1501
1502         /*
1503          * The destination pmd shouldn't be established, free_pgtables()
1504          * should have release it.
1505          */
1506         if (WARN_ON(!pmd_none(*new_pmd))) {
1507                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1508                 return false;
1509         }
1510
1511         /*
1512          * We don't have to worry about the ordering of src and dst
1513          * ptlocks because exclusive mmap_sem prevents deadlock.
1514          */
1515         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1516         if (old_ptl) {
1517                 new_ptl = pmd_lockptr(mm, new_pmd);
1518                 if (new_ptl != old_ptl)
1519                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1520                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1521                 if (pmd_present(pmd) && pmd_dirty(pmd))
1522                         force_flush = true;
1523                 VM_BUG_ON(!pmd_none(*new_pmd));
1524
1525                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1526                         pgtable_t pgtable;
1527                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1528                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1529                 }
1530                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1531                 if (new_ptl != old_ptl)
1532                         spin_unlock(new_ptl);
1533                 if (force_flush)
1534                         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1535                 else
1536                         *need_flush = true;
1537                 spin_unlock(old_ptl);
1538                 return true;
1539         }
1540         return false;
1541 }
1542
1543 /*
1544  * Returns
1545  *  - 0 if PMD could not be locked
1546  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1547  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1548  */
1549 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1550                 unsigned long addr, pgprot_t newprot, int prot_numa)
1551 {
1552         struct mm_struct *mm = vma->vm_mm;
1553         spinlock_t *ptl;
1554         int ret = 0;
1555
1556         ptl = __pmd_trans_huge_lock(pmd, vma);
1557         if (ptl) {
1558                 pmd_t entry;
1559                 bool preserve_write = prot_numa && pmd_write(*pmd);
1560                 ret = 1;
1561
1562                 /*
1563                  * Avoid trapping faults against the zero page. The read-only
1564                  * data is likely to be read-cached on the local CPU and
1565                  * local/remote hits to the zero page are not interesting.
1566                  */
1567                 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1568                         spin_unlock(ptl);
1569                         return ret;
1570                 }
1571
1572                 if (!prot_numa || !pmd_protnone(*pmd)) {
1573                         entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
1574                         entry = pmd_modify(entry, newprot);
1575                         if (preserve_write)
1576                                 entry = pmd_mkwrite(entry);
1577                         ret = HPAGE_PMD_NR;
1578                         set_pmd_at(mm, addr, pmd, entry);
1579                         BUG_ON(vma_is_anonymous(vma) && !preserve_write &&
1580                                         pmd_write(entry));
1581                 }
1582                 spin_unlock(ptl);
1583         }
1584
1585         return ret;
1586 }
1587
1588 /*
1589  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1590  *
1591  * Note that if it returns page table lock pointer, this routine returns without
1592  * unlocking page table lock. So callers must unlock it.
1593  */
1594 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1595 {
1596         spinlock_t *ptl;
1597         ptl = pmd_lock(vma->vm_mm, pmd);
1598         if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd)))
1599                 return ptl;
1600         spin_unlock(ptl);
1601         return NULL;
1602 }
1603
1604 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1605                 unsigned long haddr, pmd_t *pmd)
1606 {
1607         struct mm_struct *mm = vma->vm_mm;
1608         pgtable_t pgtable;
1609         pmd_t _pmd;
1610         int i;
1611
1612         /* leave pmd empty until pte is filled */
1613         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1614
1615         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1616         pmd_populate(mm, &_pmd, pgtable);
1617
1618         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1619                 pte_t *pte, entry;
1620                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
1621                 entry = pte_mkspecial(entry);
1622                 pte = pte_offset_map(&_pmd, haddr);
1623                 VM_BUG_ON(!pte_none(*pte));
1624                 set_pte_at(mm, haddr, pte, entry);
1625                 pte_unmap(pte);
1626         }
1627         smp_wmb(); /* make pte visible before pmd */
1628         pmd_populate(mm, pmd, pgtable);
1629 }
1630
1631 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
1632                 unsigned long haddr, bool freeze)
1633 {
1634         struct mm_struct *mm = vma->vm_mm;
1635         struct page *page;
1636         pgtable_t pgtable;
1637         pmd_t _pmd;
1638         bool young, write, dirty, soft_dirty;
1639         unsigned long addr;
1640         int i;
1641
1642         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
1643         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1644         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
1645         VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd));
1646
1647         count_vm_event(THP_SPLIT_PMD);
1648
1649         if (!vma_is_anonymous(vma)) {
1650                 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1651                 /*
1652                  * We are going to unmap this huge page. So
1653                  * just go ahead and zap it
1654                  */
1655                 if (arch_needs_pgtable_deposit())
1656                         zap_deposited_table(mm, pmd);
1657                 if (vma_is_dax(vma))
1658                         return;
1659                 page = pmd_page(_pmd);
1660                 if (!PageReferenced(page) && pmd_young(_pmd))
1661                         SetPageReferenced(page);
1662                 page_remove_rmap(page, true);
1663                 put_page(page);
1664                 add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1665                 return;
1666         } else if (is_huge_zero_pmd(*pmd)) {
1667                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
1668         }
1669
1670         page = pmd_page(*pmd);
1671         VM_BUG_ON_PAGE(!page_count(page), page);
1672         page_ref_add(page, HPAGE_PMD_NR - 1);
1673         write = pmd_write(*pmd);
1674         young = pmd_young(*pmd);
1675         dirty = pmd_dirty(*pmd);
1676         soft_dirty = pmd_soft_dirty(*pmd);
1677
1678         pmdp_huge_split_prepare(vma, haddr, pmd);
1679         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1680         pmd_populate(mm, &_pmd, pgtable);
1681
1682         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
1683                 pte_t entry, *pte;
1684                 /*
1685                  * Note that NUMA hinting access restrictions are not
1686                  * transferred to avoid any possibility of altering
1687                  * permissions across VMAs.
1688                  */
1689                 if (freeze) {
1690                         swp_entry_t swp_entry;
1691                         swp_entry = make_migration_entry(page + i, write);
1692                         entry = swp_entry_to_pte(swp_entry);
1693                         if (soft_dirty)
1694                                 entry = pte_swp_mksoft_dirty(entry);
1695                 } else {
1696                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
1697                         entry = maybe_mkwrite(entry, vma);
1698                         if (!write)
1699                                 entry = pte_wrprotect(entry);
1700                         if (!young)
1701                                 entry = pte_mkold(entry);
1702                         if (soft_dirty)
1703                                 entry = pte_mksoft_dirty(entry);
1704                 }
1705                 if (dirty)
1706                         SetPageDirty(page + i);
1707                 pte = pte_offset_map(&_pmd, addr);
1708                 BUG_ON(!pte_none(*pte));
1709                 set_pte_at(mm, addr, pte, entry);
1710                 atomic_inc(&page[i]._mapcount);
1711                 pte_unmap(pte);
1712         }
1713
1714         /*
1715          * Set PG_double_map before dropping compound_mapcount to avoid
1716          * false-negative page_mapped().
1717          */
1718         if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
1719                 for (i = 0; i < HPAGE_PMD_NR; i++)
1720                         atomic_inc(&page[i]._mapcount);
1721         }
1722
1723         if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
1724                 /* Last compound_mapcount is gone. */
1725                 __dec_node_page_state(page, NR_ANON_THPS);
1726                 if (TestClearPageDoubleMap(page)) {
1727                         /* No need in mapcount reference anymore */
1728                         for (i = 0; i < HPAGE_PMD_NR; i++)
1729                                 atomic_dec(&page[i]._mapcount);
1730                 }
1731         }
1732
1733         smp_wmb(); /* make pte visible before pmd */
1734         /*
1735          * Up to this point the pmd is present and huge and userland has the
1736          * whole access to the hugepage during the split (which happens in
1737          * place). If we overwrite the pmd with the not-huge version pointing
1738          * to the pte here (which of course we could if all CPUs were bug
1739          * free), userland could trigger a small page size TLB miss on the
1740          * small sized TLB while the hugepage TLB entry is still established in
1741          * the huge TLB. Some CPU doesn't like that.
1742          * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
1743          * 383 on page 93. Intel should be safe but is also warns that it's
1744          * only safe if the permission and cache attributes of the two entries
1745          * loaded in the two TLB is identical (which should be the case here).
1746          * But it is generally safer to never allow small and huge TLB entries
1747          * for the same virtual address to be loaded simultaneously. So instead
1748          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
1749          * current pmd notpresent (atomically because here the pmd_trans_huge
1750          * and pmd_trans_splitting must remain set at all times on the pmd
1751          * until the split is complete for this pmd), then we flush the SMP TLB
1752          * and finally we write the non-huge version of the pmd entry with
1753          * pmd_populate.
1754          */
1755         pmdp_invalidate(vma, haddr, pmd);
1756         pmd_populate(mm, pmd, pgtable);
1757
1758         if (freeze) {
1759                 for (i = 0; i < HPAGE_PMD_NR; i++) {
1760                         page_remove_rmap(page + i, false);
1761                         put_page(page + i);
1762                 }
1763         }
1764 }
1765
1766 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1767                 unsigned long address, bool freeze, struct page *page)
1768 {
1769         spinlock_t *ptl;
1770         struct mm_struct *mm = vma->vm_mm;
1771         unsigned long haddr = address & HPAGE_PMD_MASK;
1772
1773         mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
1774         ptl = pmd_lock(mm, pmd);
1775
1776         /*
1777          * If caller asks to setup a migration entries, we need a page to check
1778          * pmd against. Otherwise we can end up replacing wrong page.
1779          */
1780         VM_BUG_ON(freeze && !page);
1781         if (page && page != pmd_page(*pmd))
1782                 goto out;
1783
1784         if (pmd_trans_huge(*pmd)) {
1785                 page = pmd_page(*pmd);
1786                 if (PageMlocked(page))
1787                         clear_page_mlock(page);
1788         } else if (!pmd_devmap(*pmd))
1789                 goto out;
1790         __split_huge_pmd_locked(vma, pmd, haddr, freeze);
1791 out:
1792         spin_unlock(ptl);
1793         mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
1794 }
1795
1796 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
1797                 bool freeze, struct page *page)
1798 {
1799         pgd_t *pgd;
1800         pud_t *pud;
1801         pmd_t *pmd;
1802
1803         pgd = pgd_offset(vma->vm_mm, address);
1804         if (!pgd_present(*pgd))
1805                 return;
1806
1807         pud = pud_offset(pgd, address);
1808         if (!pud_present(*pud))
1809                 return;
1810
1811         pmd = pmd_offset(pud, address);
1812
1813         __split_huge_pmd(vma, pmd, address, freeze, page);
1814 }
1815
1816 void vma_adjust_trans_huge(struct vm_area_struct *vma,
1817                              unsigned long start,
1818                              unsigned long end,
1819                              long adjust_next)
1820 {
1821         /*
1822          * If the new start address isn't hpage aligned and it could
1823          * previously contain an hugepage: check if we need to split
1824          * an huge pmd.
1825          */
1826         if (start & ~HPAGE_PMD_MASK &&
1827             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
1828             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
1829                 split_huge_pmd_address(vma, start, false, NULL);
1830
1831         /*
1832          * If the new end address isn't hpage aligned and it could
1833          * previously contain an hugepage: check if we need to split
1834          * an huge pmd.
1835          */
1836         if (end & ~HPAGE_PMD_MASK &&
1837             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
1838             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
1839                 split_huge_pmd_address(vma, end, false, NULL);
1840
1841         /*
1842          * If we're also updating the vma->vm_next->vm_start, if the new
1843          * vm_next->vm_start isn't page aligned and it could previously
1844          * contain an hugepage: check if we need to split an huge pmd.
1845          */
1846         if (adjust_next > 0) {
1847                 struct vm_area_struct *next = vma->vm_next;
1848                 unsigned long nstart = next->vm_start;
1849                 nstart += adjust_next << PAGE_SHIFT;
1850                 if (nstart & ~HPAGE_PMD_MASK &&
1851                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
1852                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
1853                         split_huge_pmd_address(next, nstart, false, NULL);
1854         }
1855 }
1856
1857 static void freeze_page(struct page *page)
1858 {
1859         enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
1860                 TTU_RMAP_LOCKED;
1861         int i, ret;
1862
1863         VM_BUG_ON_PAGE(!PageHead(page), page);
1864
1865         if (PageAnon(page))
1866                 ttu_flags |= TTU_MIGRATION;
1867
1868         /* We only need TTU_SPLIT_HUGE_PMD once */
1869         ret = try_to_unmap(page, ttu_flags | TTU_SPLIT_HUGE_PMD);
1870         for (i = 1; !ret && i < HPAGE_PMD_NR; i++) {
1871                 /* Cut short if the page is unmapped */
1872                 if (page_count(page) == 1)
1873                         return;
1874
1875                 ret = try_to_unmap(page + i, ttu_flags);
1876         }
1877         VM_BUG_ON_PAGE(ret, page + i - 1);
1878 }
1879
1880 static void unfreeze_page(struct page *page)
1881 {
1882         int i;
1883
1884         for (i = 0; i < HPAGE_PMD_NR; i++)
1885                 remove_migration_ptes(page + i, page + i, true);
1886 }
1887
1888 static void __split_huge_page_tail(struct page *head, int tail,
1889                 struct lruvec *lruvec, struct list_head *list)
1890 {
1891         struct page *page_tail = head + tail;
1892
1893         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
1894         VM_BUG_ON_PAGE(page_ref_count(page_tail) != 0, page_tail);
1895
1896         /*
1897          * tail_page->_refcount is zero and not changing from under us. But
1898          * get_page_unless_zero() may be running from under us on the
1899          * tail_page. If we used atomic_set() below instead of atomic_inc() or
1900          * atomic_add(), we would then run atomic_set() concurrently with
1901          * get_page_unless_zero(), and atomic_set() is implemented in C not
1902          * using locked ops. spin_unlock on x86 sometime uses locked ops
1903          * because of PPro errata 66, 92, so unless somebody can guarantee
1904          * atomic_set() here would be safe on all archs (and not only on x86),
1905          * it's safer to use atomic_inc()/atomic_add().
1906          */
1907         if (PageAnon(head)) {
1908                 page_ref_inc(page_tail);
1909         } else {
1910                 /* Additional pin to radix tree */
1911                 page_ref_add(page_tail, 2);
1912         }
1913
1914         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1915         page_tail->flags |= (head->flags &
1916                         ((1L << PG_referenced) |
1917                          (1L << PG_swapbacked) |
1918                          (1L << PG_mlocked) |
1919                          (1L << PG_uptodate) |
1920                          (1L << PG_active) |
1921                          (1L << PG_locked) |
1922                          (1L << PG_unevictable) |
1923                          (1L << PG_dirty)));
1924
1925         /*
1926          * After clearing PageTail the gup refcount can be released.
1927          * Page flags also must be visible before we make the page non-compound.
1928          */
1929         smp_wmb();
1930
1931         clear_compound_head(page_tail);
1932
1933         if (page_is_young(head))
1934                 set_page_young(page_tail);
1935         if (page_is_idle(head))
1936                 set_page_idle(page_tail);
1937
1938         /* ->mapping in first tail page is compound_mapcount */
1939         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
1940                         page_tail);
1941         page_tail->mapping = head->mapping;
1942
1943         page_tail->index = head->index + tail;
1944         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
1945         lru_add_page_tail(head, page_tail, lruvec, list);
1946 }
1947
1948 static void __split_huge_page(struct page *page, struct list_head *list,
1949                 unsigned long flags)
1950 {
1951         struct page *head = compound_head(page);
1952         struct zone *zone = page_zone(head);
1953         struct lruvec *lruvec;
1954         pgoff_t end = -1;
1955         int i;
1956
1957         lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat);
1958
1959         /* complete memcg works before add pages to LRU */
1960         mem_cgroup_split_huge_fixup(head);
1961
1962         if (!PageAnon(page))
1963                 end = DIV_ROUND_UP(i_size_read(head->mapping->host), PAGE_SIZE);
1964
1965         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1966                 __split_huge_page_tail(head, i, lruvec, list);
1967                 /* Some pages can be beyond i_size: drop them from page cache */
1968                 if (head[i].index >= end) {
1969                         __ClearPageDirty(head + i);
1970                         __delete_from_page_cache(head + i, NULL);
1971                         if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
1972                                 shmem_uncharge(head->mapping->host, 1);
1973                         put_page(head + i);
1974                 }
1975         }
1976
1977         ClearPageCompound(head);
1978         /* See comment in __split_huge_page_tail() */
1979         if (PageAnon(head)) {
1980                 page_ref_inc(head);
1981         } else {
1982                 /* Additional pin to radix tree */
1983                 page_ref_add(head, 2);
1984                 spin_unlock(&head->mapping->tree_lock);
1985         }
1986
1987         spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
1988
1989         unfreeze_page(head);
1990
1991         for (i = 0; i < HPAGE_PMD_NR; i++) {
1992                 struct page *subpage = head + i;
1993                 if (subpage == page)
1994                         continue;
1995                 unlock_page(subpage);
1996
1997                 /*
1998                  * Subpages may be freed if there wasn't any mapping
1999                  * like if add_to_swap() is running on a lru page that
2000                  * had its mapping zapped. And freeing these pages
2001                  * requires taking the lru_lock so we do the put_page
2002                  * of the tail pages after the split is complete.
2003                  */
2004                 put_page(subpage);
2005         }
2006 }
2007
2008 int total_mapcount(struct page *page)
2009 {
2010         int i, compound, ret;
2011
2012         VM_BUG_ON_PAGE(PageTail(page), page);
2013
2014         if (likely(!PageCompound(page)))
2015                 return atomic_read(&page->_mapcount) + 1;
2016
2017         compound = compound_mapcount(page);
2018         if (PageHuge(page))
2019                 return compound;
2020         ret = compound;
2021         for (i = 0; i < HPAGE_PMD_NR; i++)
2022                 ret += atomic_read(&page[i]._mapcount) + 1;
2023         /* File pages has compound_mapcount included in _mapcount */
2024         if (!PageAnon(page))
2025                 return ret - compound * HPAGE_PMD_NR;
2026         if (PageDoubleMap(page))
2027                 ret -= HPAGE_PMD_NR;
2028         return ret;
2029 }
2030
2031 /*
2032  * This calculates accurately how many mappings a transparent hugepage
2033  * has (unlike page_mapcount() which isn't fully accurate). This full
2034  * accuracy is primarily needed to know if copy-on-write faults can
2035  * reuse the page and change the mapping to read-write instead of
2036  * copying them. At the same time this returns the total_mapcount too.
2037  *
2038  * The function returns the highest mapcount any one of the subpages
2039  * has. If the return value is one, even if different processes are
2040  * mapping different subpages of the transparent hugepage, they can
2041  * all reuse it, because each process is reusing a different subpage.
2042  *
2043  * The total_mapcount is instead counting all virtual mappings of the
2044  * subpages. If the total_mapcount is equal to "one", it tells the
2045  * caller all mappings belong to the same "mm" and in turn the
2046  * anon_vma of the transparent hugepage can become the vma->anon_vma
2047  * local one as no other process may be mapping any of the subpages.
2048  *
2049  * It would be more accurate to replace page_mapcount() with
2050  * page_trans_huge_mapcount(), however we only use
2051  * page_trans_huge_mapcount() in the copy-on-write faults where we
2052  * need full accuracy to avoid breaking page pinning, because
2053  * page_trans_huge_mapcount() is slower than page_mapcount().
2054  */
2055 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2056 {
2057         int i, ret, _total_mapcount, mapcount;
2058
2059         /* hugetlbfs shouldn't call it */
2060         VM_BUG_ON_PAGE(PageHuge(page), page);
2061
2062         if (likely(!PageTransCompound(page))) {
2063                 mapcount = atomic_read(&page->_mapcount) + 1;
2064                 if (total_mapcount)
2065                         *total_mapcount = mapcount;
2066                 return mapcount;
2067         }
2068
2069         page = compound_head(page);
2070
2071         _total_mapcount = ret = 0;
2072         for (i = 0; i < HPAGE_PMD_NR; i++) {
2073                 mapcount = atomic_read(&page[i]._mapcount) + 1;
2074                 ret = max(ret, mapcount);
2075                 _total_mapcount += mapcount;
2076         }
2077         if (PageDoubleMap(page)) {
2078                 ret -= 1;
2079                 _total_mapcount -= HPAGE_PMD_NR;
2080         }
2081         mapcount = compound_mapcount(page);
2082         ret += mapcount;
2083         _total_mapcount += mapcount;
2084         if (total_mapcount)
2085                 *total_mapcount = _total_mapcount;
2086         return ret;
2087 }
2088
2089 /*
2090  * This function splits huge page into normal pages. @page can point to any
2091  * subpage of huge page to split. Split doesn't change the position of @page.
2092  *
2093  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2094  * The huge page must be locked.
2095  *
2096  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2097  *
2098  * Both head page and tail pages will inherit mapping, flags, and so on from
2099  * the hugepage.
2100  *
2101  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2102  * they are not mapped.
2103  *
2104  * Returns 0 if the hugepage is split successfully.
2105  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2106  * us.
2107  */
2108 int split_huge_page_to_list(struct page *page, struct list_head *list)
2109 {
2110         struct page *head = compound_head(page);
2111         struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2112         struct anon_vma *anon_vma = NULL;
2113         struct address_space *mapping = NULL;
2114         int count, mapcount, extra_pins, ret;
2115         bool mlocked;
2116         unsigned long flags;
2117
2118         VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
2119         VM_BUG_ON_PAGE(!PageLocked(page), page);
2120         VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2121         VM_BUG_ON_PAGE(!PageCompound(page), page);
2122
2123         if (PageAnon(head)) {
2124                 /*
2125                  * The caller does not necessarily hold an mmap_sem that would
2126                  * prevent the anon_vma disappearing so we first we take a
2127                  * reference to it and then lock the anon_vma for write. This
2128                  * is similar to page_lock_anon_vma_read except the write lock
2129                  * is taken to serialise against parallel split or collapse
2130                  * operations.
2131                  */
2132                 anon_vma = page_get_anon_vma(head);
2133                 if (!anon_vma) {
2134                         ret = -EBUSY;
2135                         goto out;
2136                 }
2137                 extra_pins = 0;
2138                 mapping = NULL;
2139                 anon_vma_lock_write(anon_vma);
2140         } else {
2141                 mapping = head->mapping;
2142
2143                 /* Truncated ? */
2144                 if (!mapping) {
2145                         ret = -EBUSY;
2146                         goto out;
2147                 }
2148
2149                 /* Addidional pins from radix tree */
2150                 extra_pins = HPAGE_PMD_NR;
2151                 anon_vma = NULL;
2152                 i_mmap_lock_read(mapping);
2153         }
2154
2155         /*
2156          * Racy check if we can split the page, before freeze_page() will
2157          * split PMDs
2158          */
2159         if (total_mapcount(head) != page_count(head) - extra_pins - 1) {
2160                 ret = -EBUSY;
2161                 goto out_unlock;
2162         }
2163
2164         mlocked = PageMlocked(page);
2165         freeze_page(head);
2166         VM_BUG_ON_PAGE(compound_mapcount(head), head);
2167
2168         /* Make sure the page is not on per-CPU pagevec as it takes pin */
2169         if (mlocked)
2170                 lru_add_drain();
2171
2172         /* prevent PageLRU to go away from under us, and freeze lru stats */
2173         spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags);
2174
2175         if (mapping) {
2176                 void **pslot;
2177
2178                 spin_lock(&mapping->tree_lock);
2179                 pslot = radix_tree_lookup_slot(&mapping->page_tree,
2180                                 page_index(head));
2181                 /*
2182                  * Check if the head page is present in radix tree.
2183                  * We assume all tail are present too, if head is there.
2184                  */
2185                 if (radix_tree_deref_slot_protected(pslot,
2186                                         &mapping->tree_lock) != head)
2187                         goto fail;
2188         }
2189
2190         /* Prevent deferred_split_scan() touching ->_refcount */
2191         spin_lock(&pgdata->split_queue_lock);
2192         count = page_count(head);
2193         mapcount = total_mapcount(head);
2194         if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2195                 if (!list_empty(page_deferred_list(head))) {
2196                         pgdata->split_queue_len--;
2197                         list_del(page_deferred_list(head));
2198                 }
2199                 if (mapping)
2200                         __dec_node_page_state(page, NR_SHMEM_THPS);
2201                 spin_unlock(&pgdata->split_queue_lock);
2202                 __split_huge_page(page, list, flags);
2203                 ret = 0;
2204         } else {
2205                 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2206                         pr_alert("total_mapcount: %u, page_count(): %u\n",
2207                                         mapcount, count);
2208                         if (PageTail(page))
2209                                 dump_page(head, NULL);
2210                         dump_page(page, "total_mapcount(head) > 0");
2211                         BUG();
2212                 }
2213                 spin_unlock(&pgdata->split_queue_lock);
2214 fail:           if (mapping)
2215                         spin_unlock(&mapping->tree_lock);
2216                 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
2217                 unfreeze_page(head);
2218                 ret = -EBUSY;
2219         }
2220
2221 out_unlock:
2222         if (anon_vma) {
2223                 anon_vma_unlock_write(anon_vma);
2224                 put_anon_vma(anon_vma);
2225         }
2226         if (mapping)
2227                 i_mmap_unlock_read(mapping);
2228 out:
2229         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2230         return ret;
2231 }
2232
2233 void free_transhuge_page(struct page *page)
2234 {
2235         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2236         unsigned long flags;
2237
2238         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2239         if (!list_empty(page_deferred_list(page))) {
2240                 pgdata->split_queue_len--;
2241                 list_del(page_deferred_list(page));
2242         }
2243         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2244         free_compound_page(page);
2245 }
2246
2247 void deferred_split_huge_page(struct page *page)
2248 {
2249         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2250         unsigned long flags;
2251
2252         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2253
2254         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2255         if (list_empty(page_deferred_list(page))) {
2256                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2257                 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2258                 pgdata->split_queue_len++;
2259         }
2260         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2261 }
2262
2263 static unsigned long deferred_split_count(struct shrinker *shrink,
2264                 struct shrink_control *sc)
2265 {
2266         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2267         return ACCESS_ONCE(pgdata->split_queue_len);
2268 }
2269
2270 static unsigned long deferred_split_scan(struct shrinker *shrink,
2271                 struct shrink_control *sc)
2272 {
2273         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2274         unsigned long flags;
2275         LIST_HEAD(list), *pos, *next;
2276         struct page *page;
2277         int split = 0;
2278
2279         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2280         /* Take pin on all head pages to avoid freeing them under us */
2281         list_for_each_safe(pos, next, &pgdata->split_queue) {
2282                 page = list_entry((void *)pos, struct page, mapping);
2283                 page = compound_head(page);
2284                 if (get_page_unless_zero(page)) {
2285                         list_move(page_deferred_list(page), &list);
2286                 } else {
2287                         /* We lost race with put_compound_page() */
2288                         list_del_init(page_deferred_list(page));
2289                         pgdata->split_queue_len--;
2290                 }
2291                 if (!--sc->nr_to_scan)
2292                         break;
2293         }
2294         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2295
2296         list_for_each_safe(pos, next, &list) {
2297                 page = list_entry((void *)pos, struct page, mapping);
2298                 lock_page(page);
2299                 /* split_huge_page() removes page from list on success */
2300                 if (!split_huge_page(page))
2301                         split++;
2302                 unlock_page(page);
2303                 put_page(page);
2304         }
2305
2306         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2307         list_splice_tail(&list, &pgdata->split_queue);
2308         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2309
2310         /*
2311          * Stop shrinker if we didn't split any page, but the queue is empty.
2312          * This can happen if pages were freed under us.
2313          */
2314         if (!split && list_empty(&pgdata->split_queue))
2315                 return SHRINK_STOP;
2316         return split;
2317 }
2318
2319 static struct shrinker deferred_split_shrinker = {
2320         .count_objects = deferred_split_count,
2321         .scan_objects = deferred_split_scan,
2322         .seeks = DEFAULT_SEEKS,
2323         .flags = SHRINKER_NUMA_AWARE,
2324 };
2325
2326 #ifdef CONFIG_DEBUG_FS
2327 static int split_huge_pages_set(void *data, u64 val)
2328 {
2329         struct zone *zone;
2330         struct page *page;
2331         unsigned long pfn, max_zone_pfn;
2332         unsigned long total = 0, split = 0;
2333
2334         if (val != 1)
2335                 return -EINVAL;
2336
2337         for_each_populated_zone(zone) {
2338                 max_zone_pfn = zone_end_pfn(zone);
2339                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2340                         if (!pfn_valid(pfn))
2341                                 continue;
2342
2343                         page = pfn_to_page(pfn);
2344                         if (!get_page_unless_zero(page))
2345                                 continue;
2346
2347                         if (zone != page_zone(page))
2348                                 goto next;
2349
2350                         if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2351                                 goto next;
2352
2353                         total++;
2354                         lock_page(page);
2355                         if (!split_huge_page(page))
2356                                 split++;
2357                         unlock_page(page);
2358 next:
2359                         put_page(page);
2360                 }
2361         }
2362
2363         pr_info("%lu of %lu THP split\n", split, total);
2364
2365         return 0;
2366 }
2367 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2368                 "%llu\n");
2369
2370 static int __init split_huge_pages_debugfs(void)
2371 {
2372         void *ret;
2373
2374         ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2375                         &split_huge_pages_fops);
2376         if (!ret)
2377                 pr_warn("Failed to create split_huge_pages in debugfs");
2378         return 0;
2379 }
2380 late_initcall(split_huge_pages_debugfs);
2381 #endif