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 #include <linux/page_owner.h>
  34
  35 #include <asm/tlb.h>
  36 #include <asm/pgalloc.h>
  37 #include "internal.h"
  38
  39 /*
  40  * By default transparent hugepage support is disabled in order that avoid
  41  * to risk increase the memory footprint of applications without a guaranteed
  42  * benefit. When transparent hugepage support is enabled, is for all mappings,
  43  * and khugepaged scans all mappings.
  44  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  45  * for all hugepage allocations.
  46  */
  47 unsigned long transparent_hugepage_flags __read_mostly =
  48 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  49         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  50 #endif
  51 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  52         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  53 #endif
  54         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
  55         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  56         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  57
  58 static struct shrinker deferred_split_shrinker;
  59
  60 static atomic_t huge_zero_refcount;
  61 struct page *huge_zero_page __read_mostly;
  62
  63 static struct page *get_huge_zero_page(void)
  64 {
  65         struct page *zero_page;
  66 retry:
  67         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
  68                 return READ_ONCE(huge_zero_page);
  69
  70         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
  71                         HPAGE_PMD_ORDER);
  72         if (!zero_page) {
  73                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
  74                 return NULL;
  75         }
  76         count_vm_event(THP_ZERO_PAGE_ALLOC);
  77         preempt_disable();
  78         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
  79                 preempt_enable();
  80                 __free_pages(zero_page, compound_order(zero_page));
  81                 goto retry;
  82         }
  83
  84         /* We take additional reference here. It will be put back by shrinker */
  85         atomic_set(&huge_zero_refcount, 2);
  86         preempt_enable();
  87         return READ_ONCE(huge_zero_page);
  88 }
  89
  90 static void put_huge_zero_page(void)
  91 {
  92         /*
  93          * Counter should never go to zero here. Only shrinker can put
  94          * last reference.
  95          */
  96         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
  97 }
  98
  99 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
 100 {
 101         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 102                 return READ_ONCE(huge_zero_page);
 103
 104         if (!get_huge_zero_page())
 105                 return NULL;
 106
 107         if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 108                 put_huge_zero_page();
 109
 110         return READ_ONCE(huge_zero_page);
 111 }
 112
 113 void mm_put_huge_zero_page(struct mm_struct *mm)
 114 {
 115         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 116                 put_huge_zero_page();
 117 }
 118
 119 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 120                                         struct shrink_control *sc)
 121 {
 122         /* we can free zero page only if last reference remains */
 123         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 124 }
 125
 126 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 127                                        struct shrink_control *sc)
 128 {
 129         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 130                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 131                 BUG_ON(zero_page == NULL);
 132                 __free_pages(zero_page, compound_order(zero_page));
 133                 return HPAGE_PMD_NR;
 134         }
 135
 136         return 0;
 137 }
 138
 139 static struct shrinker huge_zero_page_shrinker = {
 140         .count_objects = shrink_huge_zero_page_count,
 141         .scan_objects = shrink_huge_zero_page_scan,
 142         .seeks = DEFAULT_SEEKS,
 143 };
 144
 145 #ifdef CONFIG_SYSFS
 146
 147 static ssize_t triple_flag_store(struct kobject *kobj,
 148                                  struct kobj_attribute *attr,
 149                                  const char *buf, size_t count,
 150                                  enum transparent_hugepage_flag enabled,
 151                                  enum transparent_hugepage_flag deferred,
 152                                  enum transparent_hugepage_flag req_madv)
 153 {
 154         if (!memcmp("defer", buf,
 155                     min(sizeof("defer")-1, count))) {
 156                 if (enabled == deferred)
 157                         return -EINVAL;
 158                 clear_bit(enabled, &transparent_hugepage_flags);
 159                 clear_bit(req_madv, &transparent_hugepage_flags);
 160                 set_bit(deferred, &transparent_hugepage_flags);
 161         } else if (!memcmp("always", buf,
 162                     min(sizeof("always")-1, count))) {
 163                 clear_bit(deferred, &transparent_hugepage_flags);
 164                 clear_bit(req_madv, &transparent_hugepage_flags);
 165                 set_bit(enabled, &transparent_hugepage_flags);
 166         } else if (!memcmp("madvise", buf,
 167                            min(sizeof("madvise")-1, count))) {
 168                 clear_bit(enabled, &transparent_hugepage_flags);
 169                 clear_bit(deferred, &transparent_hugepage_flags);
 170                 set_bit(req_madv, &transparent_hugepage_flags);
 171         } else if (!memcmp("never", buf,
 172                            min(sizeof("never")-1, count))) {
 173                 clear_bit(enabled, &transparent_hugepage_flags);
 174                 clear_bit(req_madv, &transparent_hugepage_flags);
 175                 clear_bit(deferred, &transparent_hugepage_flags);
 176         } else
 177                 return -EINVAL;
 178
 179         return count;
 180 }
 181
 182 static ssize_t enabled_show(struct kobject *kobj,
 183                             struct kobj_attribute *attr, char *buf)
 184 {
 185         if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
 186                 return sprintf(buf, "[always] madvise never\n");
 187         else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
 188                 return sprintf(buf, "always [madvise] never\n");
 189         else
 190                 return sprintf(buf, "always madvise [never]\n");
 191 }
 192
 193 static ssize_t enabled_store(struct kobject *kobj,
 194                              struct kobj_attribute *attr,
 195                              const char *buf, size_t count)
 196 {
 197         ssize_t ret;
 198
 199         ret = triple_flag_store(kobj, attr, buf, count,
 200                                 TRANSPARENT_HUGEPAGE_FLAG,
 201                                 TRANSPARENT_HUGEPAGE_FLAG,
 202                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 203
 204         if (ret > 0) {
 205                 int err = start_stop_khugepaged();
 206                 if (err)
 207                         ret = err;
 208         }
 209
 210         return ret;
 211 }
 212 static struct kobj_attribute enabled_attr =
 213         __ATTR(enabled, 0644, enabled_show, enabled_store);
 214
 215 ssize_t single_hugepage_flag_show(struct kobject *kobj,
 216                                 struct kobj_attribute *attr, char *buf,
 217                                 enum transparent_hugepage_flag flag)
 218 {
 219         return sprintf(buf, "%d\n",
 220                        !!test_bit(flag, &transparent_hugepage_flags));
 221 }
 222
 223 ssize_t single_hugepage_flag_store(struct kobject *kobj,
 224                                  struct kobj_attribute *attr,
 225                                  const char *buf, size_t count,
 226                                  enum transparent_hugepage_flag flag)
 227 {
 228         unsigned long value;
 229         int ret;
 230
 231         ret = kstrtoul(buf, 10, &value);
 232         if (ret < 0)
 233                 return ret;
 234         if (value > 1)
 235                 return -EINVAL;
 236
 237         if (value)
 238                 set_bit(flag, &transparent_hugepage_flags);
 239         else
 240                 clear_bit(flag, &transparent_hugepage_flags);
 241
 242         return count;
 243 }
 244
 245 /*
 246  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 247  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 248  * memory just to allocate one more hugepage.
 249  */
 250 static ssize_t defrag_show(struct kobject *kobj,
 251                            struct kobj_attribute *attr, char *buf)
 252 {
 253         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 254                 return sprintf(buf, "[always] defer madvise never\n");
 255         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 256                 return sprintf(buf, "always [defer] madvise never\n");
 257         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 258                 return sprintf(buf, "always defer [madvise] never\n");
 259         else
 260                 return sprintf(buf, "always defer madvise [never]\n");
 261
 262 }
 263 static ssize_t defrag_store(struct kobject *kobj,
 264                             struct kobj_attribute *attr,
 265                             const char *buf, size_t count)
 266 {
 267         return triple_flag_store(kobj, attr, buf, count,
 268                                  TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
 269                                  TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
 270                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 271 }
 272 static struct kobj_attribute defrag_attr =
 273         __ATTR(defrag, 0644, defrag_show, defrag_store);
 274
 275 static ssize_t use_zero_page_show(struct kobject *kobj,
 276                 struct kobj_attribute *attr, char *buf)
 277 {
 278         return single_hugepage_flag_show(kobj, attr, buf,
 279                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 280 }
 281 static ssize_t use_zero_page_store(struct kobject *kobj,
 282                 struct kobj_attribute *attr, const char *buf, size_t count)
 283 {
 284         return single_hugepage_flag_store(kobj, attr, buf, count,
 285                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 286 }
 287 static struct kobj_attribute use_zero_page_attr =
 288         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 289 #ifdef CONFIG_DEBUG_VM
 290 static ssize_t debug_cow_show(struct kobject *kobj,
 291                                 struct kobj_attribute *attr, char *buf)
 292 {
 293         return single_hugepage_flag_show(kobj, attr, buf,
 294                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 295 }
 296 static ssize_t debug_cow_store(struct kobject *kobj,
 297                                struct kobj_attribute *attr,
 298                                const char *buf, size_t count)
 299 {
 300         return single_hugepage_flag_store(kobj, attr, buf, count,
 301                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 302 }
 303 static struct kobj_attribute debug_cow_attr =
 304         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 305 #endif /* CONFIG_DEBUG_VM */
 306
 307 static struct attribute *hugepage_attr[] = {
 308         &enabled_attr.attr,
 309         &defrag_attr.attr,
 310         &use_zero_page_attr.attr,
 311 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
 312         &shmem_enabled_attr.attr,
 313 #endif
 314 #ifdef CONFIG_DEBUG_VM
 315         &debug_cow_attr.attr,
 316 #endif
 317         NULL,
 318 };
 319
 320 static struct attribute_group hugepage_attr_group = {
 321         .attrs = hugepage_attr,
 322 };
 323
 324 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 325 {
 326         int err;
 327
 328         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 329         if (unlikely(!*hugepage_kobj)) {
 330                 pr_err("failed to create transparent hugepage kobject\n");
 331                 return -ENOMEM;
 332         }
 333
 334         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 335         if (err) {
 336                 pr_err("failed to register transparent hugepage group\n");
 337                 goto delete_obj;
 338         }
 339
 340         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 341         if (err) {
 342                 pr_err("failed to register transparent hugepage group\n");
 343                 goto remove_hp_group;
 344         }
 345
 346         return 0;
 347
 348 remove_hp_group:
 349         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 350 delete_obj:
 351         kobject_put(*hugepage_kobj);
 352         return err;
 353 }
 354
 355 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 356 {
 357         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 358         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 359         kobject_put(hugepage_kobj);
 360 }
 361 #else
 362 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 363 {
 364         return 0;
 365 }
 366
 367 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 368 {
 369 }
 370 #endif /* CONFIG_SYSFS */
 371
 372 static int __init hugepage_init(void)
 373 {
 374         int err;
 375         struct kobject *hugepage_kobj;
 376
 377         if (!has_transparent_hugepage()) {
 378                 transparent_hugepage_flags = 0;
 379                 return -EINVAL;
 380         }
 381
 382         /*
 383          * hugepages can't be allocated by the buddy allocator
 384          */
 385         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
 386         /*
 387          * we use page->mapping and page->index in second tail page
 388          * as list_head: assuming THP order >= 2
 389          */
 390         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 391
 392         err = hugepage_init_sysfs(&hugepage_kobj);
 393         if (err)
 394                 goto err_sysfs;
 395
 396         err = khugepaged_init();
 397         if (err)
 398                 goto err_slab;
 399
 400         err = register_shrinker(&huge_zero_page_shrinker);
 401         if (err)
 402                 goto err_hzp_shrinker;
 403         err = register_shrinker(&deferred_split_shrinker);
 404         if (err)
 405                 goto err_split_shrinker;
 406
 407         /*
 408          * By default disable transparent hugepages on smaller systems,
 409          * where the extra memory used could hurt more than TLB overhead
 410          * is likely to save.  The admin can still enable it through /sys.
 411          */
 412         if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
 413                 transparent_hugepage_flags = 0;
 414                 return 0;
 415         }
 416
 417         err = start_stop_khugepaged();
 418         if (err)
 419                 goto err_khugepaged;
 420
 421         return 0;
 422 err_khugepaged:
 423         unregister_shrinker(&deferred_split_shrinker);
 424 err_split_shrinker:
 425         unregister_shrinker(&huge_zero_page_shrinker);
 426 err_hzp_shrinker:
 427         khugepaged_destroy();
 428 err_slab:
 429         hugepage_exit_sysfs(hugepage_kobj);
 430 err_sysfs:
 431         return err;
 432 }
 433 subsys_initcall(hugepage_init);
 434
 435 static int __init setup_transparent_hugepage(char *str)
 436 {
 437         int ret = 0;
 438         if (!str)
 439                 goto out;
 440         if (!strcmp(str, "always")) {
 441                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 442                         &transparent_hugepage_flags);
 443                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 444                           &transparent_hugepage_flags);
 445                 ret = 1;
 446         } else if (!strcmp(str, "madvise")) {
 447                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 448                           &transparent_hugepage_flags);
 449                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 450                         &transparent_hugepage_flags);
 451                 ret = 1;
 452         } else if (!strcmp(str, "never")) {
 453                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 454                           &transparent_hugepage_flags);
 455                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 456                           &transparent_hugepage_flags);
 457                 ret = 1;
 458         }
 459 out:
 460         if (!ret)
 461                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 462         return ret;
 463 }
 464 __setup("transparent_hugepage=", setup_transparent_hugepage);
 465
 466 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 467 {
 468         if (likely(vma->vm_flags & VM_WRITE))
 469                 pmd = pmd_mkwrite(pmd);
 470         return pmd;
 471 }
 472
 473 static inline struct list_head *page_deferred_list(struct page *page)
 474 {
 475         /*
 476          * ->lru in the tail pages is occupied by compound_head.
 477          * Let's use ->mapping + ->index in the second tail page as list_head.
 478          */
 479         return (struct list_head *)&page[2].mapping;
 480 }
 481
 482 void prep_transhuge_page(struct page *page)
 483 {
 484         /*
 485          * we use page->mapping and page->indexlru in second tail page
 486          * as list_head: assuming THP order >= 2
 487          */
 488
 489         INIT_LIST_HEAD(page_deferred_list(page));
 490         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 491 }
 492
 493 unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len,
 494                 loff_t off, unsigned long flags, unsigned long size)
 495 {
 496         unsigned long addr;
 497         loff_t off_end = off + len;
 498         loff_t off_align = round_up(off, size);
 499         unsigned long len_pad;
 500
 501         if (off_end <= off_align || (off_end - off_align) < size)
 502                 return 0;
 503
 504         len_pad = len + size;
 505         if (len_pad < len || (off + len_pad) < off)
 506                 return 0;
 507
 508         addr = current->mm->get_unmapped_area(filp, 0, len_pad,
 509                                               off >> PAGE_SHIFT, flags);
 510         if (IS_ERR_VALUE(addr))
 511                 return 0;
 512
 513         addr += (off - addr) & (size - 1);
 514         return addr;
 515 }
 516
 517 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 518                 unsigned long len, unsigned long pgoff, unsigned long flags)
 519 {
 520         loff_t off = (loff_t)pgoff << PAGE_SHIFT;
 521
 522         if (addr)
 523                 goto out;
 524         if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
 525                 goto out;
 526
 527         addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE);
 528         if (addr)
 529                 return addr;
 530
 531  out:
 532         return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
 533 }
 534 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
 535
 536 static int __do_huge_pmd_anonymous_page(struct fault_env *fe, struct page *page,
 537                 gfp_t gfp)
 538 {
 539         struct vm_area_struct *vma = fe->vma;
 540         struct mem_cgroup *memcg;
 541         pgtable_t pgtable;
 542         unsigned long haddr = fe->address & HPAGE_PMD_MASK;
 543
 544         VM_BUG_ON_PAGE(!PageCompound(page), page);
 545
 546         if (mem_cgroup_try_charge(page, vma->vm_mm, gfp | __GFP_NORETRY, &memcg,
 547                                   true)) {
 548                 put_page(page);
 549                 count_vm_event(THP_FAULT_FALLBACK);
 550                 return VM_FAULT_FALLBACK;
 551         }
 552
 553         pgtable = pte_alloc_one(vma->vm_mm, haddr);
 554         if (unlikely(!pgtable)) {
 555                 mem_cgroup_cancel_charge(page, memcg, true);
 556                 put_page(page);
 557                 return VM_FAULT_OOM;
 558         }
 559
 560         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 561         /*
 562          * The memory barrier inside __SetPageUptodate makes sure that
 563          * clear_huge_page writes become visible before the set_pmd_at()
 564          * write.
 565          */
 566         __SetPageUptodate(page);
 567
 568         fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
 569         if (unlikely(!pmd_none(*fe->pmd))) {
 570                 spin_unlock(fe->ptl);
 571                 mem_cgroup_cancel_charge(page, memcg, true);
 572                 put_page(page);
 573                 pte_free(vma->vm_mm, pgtable);
 574         } else {
 575                 pmd_t entry;
 576
 577                 /* Deliver the page fault to userland */
 578                 if (userfaultfd_missing(vma)) {
 579                         int ret;
 580
 581                         spin_unlock(fe->ptl);
 582                         mem_cgroup_cancel_charge(page, memcg, true);
 583                         put_page(page);
 584                         pte_free(vma->vm_mm, pgtable);
 585                         ret = handle_userfault(fe, VM_UFFD_MISSING);
 586                         VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 587                         return ret;
 588                 }
 589
 590                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 591                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 592                 page_add_new_anon_rmap(page, vma, haddr, true);
 593                 mem_cgroup_commit_charge(page, memcg, false, true);
 594                 lru_cache_add_active_or_unevictable(page, vma);
 595                 pgtable_trans_huge_deposit(vma->vm_mm, fe->pmd, pgtable);
 596                 set_pmd_at(vma->vm_mm, haddr, fe->pmd, entry);
 597                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 598                 atomic_long_inc(&vma->vm_mm->nr_ptes);
 599                 spin_unlock(fe->ptl);
 600                 count_vm_event(THP_FAULT_ALLOC);
 601         }
 602
 603         return 0;
 604 }
 605
 606 /*
 607  * If THP defrag is set to always then directly reclaim/compact as necessary
 608  * If set to defer then do only background reclaim/compact and defer to khugepaged
 609  * If set to madvise and the VMA is flagged then directly reclaim/compact
 610  * When direct reclaim/compact is allowed, don't retry except for flagged VMA's
 611  */
 612 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
 613 {
 614         bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
 615
 616         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
 617                                 &transparent_hugepage_flags) && vma_madvised)
 618                 return GFP_TRANSHUGE;
 619         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
 620                                                 &transparent_hugepage_flags))
 621                 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
 622         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
 623                                                 &transparent_hugepage_flags))
 624                 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
 625
 626         return GFP_TRANSHUGE_LIGHT;
 627 }
 628
 629 /* Caller must hold page table lock. */
 630 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 631                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 632                 struct page *zero_page)
 633 {
 634         pmd_t entry;
 635         if (!pmd_none(*pmd))
 636                 return false;
 637         entry = mk_pmd(zero_page, vma->vm_page_prot);
 638         entry = pmd_mkhuge(entry);
 639         if (pgtable)
 640                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 641         set_pmd_at(mm, haddr, pmd, entry);
 642         atomic_long_inc(&mm->nr_ptes);
 643         return true;
 644 }
 645
 646 int do_huge_pmd_anonymous_page(struct fault_env *fe)
 647 {
 648         struct vm_area_struct *vma = fe->vma;
 649         gfp_t gfp;
 650         struct page *page;
 651         unsigned long haddr = fe->address & HPAGE_PMD_MASK;
 652
 653         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 654                 return VM_FAULT_FALLBACK;
 655         if (unlikely(anon_vma_prepare(vma)))
 656                 return VM_FAULT_OOM;
 657         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 658                 return VM_FAULT_OOM;
 659         if (!(fe->flags & FAULT_FLAG_WRITE) &&
 660                         !mm_forbids_zeropage(vma->vm_mm) &&
 661                         transparent_hugepage_use_zero_page()) {
 662                 pgtable_t pgtable;
 663                 struct page *zero_page;
 664                 int ret;
 665                 pgtable = pte_alloc_one(vma->vm_mm, haddr);
 666                 if (unlikely(!pgtable))
 667                         return VM_FAULT_OOM;
 668                 zero_page = mm_get_huge_zero_page(vma->vm_mm);
 669                 if (unlikely(!zero_page)) {
 670                         pte_free(vma->vm_mm, pgtable);
 671                         count_vm_event(THP_FAULT_FALLBACK);
 672                         return VM_FAULT_FALLBACK;
 673                 }
 674                 fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
 675                 ret = 0;
 676                 if (pmd_none(*fe->pmd)) {
 677                         if (userfaultfd_missing(vma)) {
 678                                 spin_unlock(fe->ptl);
 679                                 pte_free(vma->vm_mm, pgtable);
 680                                 ret = handle_userfault(fe, VM_UFFD_MISSING);
 681                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 682                         } else {
 683                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
 684                                                    haddr, fe->pmd, zero_page);
 685                                 spin_unlock(fe->ptl);
 686                         }
 687                 } else {
 688                         spin_unlock(fe->ptl);
 689                         pte_free(vma->vm_mm, pgtable);
 690                 }
 691                 return ret;
 692         }
 693         gfp = alloc_hugepage_direct_gfpmask(vma);
 694         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 695         if (unlikely(!page)) {
 696                 count_vm_event(THP_FAULT_FALLBACK);
 697                 return VM_FAULT_FALLBACK;
 698         }
 699         prep_transhuge_page(page);
 700         return __do_huge_pmd_anonymous_page(fe, page, gfp);
 701 }
 702
 703 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 704                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write)
 705 {
 706         struct mm_struct *mm = vma->vm_mm;
 707         pmd_t entry;
 708         spinlock_t *ptl;
 709
 710         ptl = pmd_lock(mm, pmd);
 711         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 712         if (pfn_t_devmap(pfn))
 713                 entry = pmd_mkdevmap(entry);
 714         if (write) {
 715                 entry = pmd_mkyoung(pmd_mkdirty(entry));
 716                 entry = maybe_pmd_mkwrite(entry, vma);
 717         }
 718         set_pmd_at(mm, addr, pmd, entry);
 719         update_mmu_cache_pmd(vma, addr, pmd);
 720         spin_unlock(ptl);
 721 }
 722
 723 int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 724                         pmd_t *pmd, pfn_t pfn, bool write)
 725 {
 726         pgprot_t pgprot = vma->vm_page_prot;
 727         /*
 728          * If we had pmd_special, we could avoid all these restrictions,
 729          * but we need to be consistent with PTEs and architectures that
 730          * can't support a 'special' bit.
 731          */
 732         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
 733         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 734                                                 (VM_PFNMAP|VM_MIXEDMAP));
 735         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 736         BUG_ON(!pfn_t_devmap(pfn));
 737
 738         if (addr < vma->vm_start || addr >= vma->vm_end)
 739                 return VM_FAULT_SIGBUS;
 740         if (track_pfn_insert(vma, &pgprot, pfn))
 741                 return VM_FAULT_SIGBUS;
 742         insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
 743         return VM_FAULT_NOPAGE;
 744 }
 745 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
 746
 747 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
 748                 pmd_t *pmd, int flags)
 749 {
 750         pmd_t _pmd;
 751
 752         _pmd = pmd_mkyoung(*pmd);
 753         if (flags & FOLL_WRITE)
 754                 _pmd = pmd_mkdirty(_pmd);
 755         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
 756                                 pmd, _pmd, flags & FOLL_WRITE))
 757                 update_mmu_cache_pmd(vma, addr, pmd);
 758 }
 759
 760 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 761                 pmd_t *pmd, int flags)
 762 {
 763         unsigned long pfn = pmd_pfn(*pmd);
 764         struct mm_struct *mm = vma->vm_mm;
 765         struct dev_pagemap *pgmap;
 766         struct page *page;
 767
 768         assert_spin_locked(pmd_lockptr(mm, pmd));
 769
 770         /*
 771          * When we COW a devmap PMD entry, we split it into PTEs, so we should
 772          * not be in this function with `flags & FOLL_COW` set.
 773          */
 774         WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
 775
 776         if (flags & FOLL_WRITE && !pmd_write(*pmd))
 777                 return NULL;
 778
 779         if (pmd_present(*pmd) && pmd_devmap(*pmd))
 780                 /* pass */;
 781         else
 782                 return NULL;
 783
 784         if (flags & FOLL_TOUCH)
 785                 touch_pmd(vma, addr, pmd, flags);
 786
 787         /*
 788          * device mapped pages can only be returned if the
 789          * caller will manage the page reference count.
 790          */
 791         if (!(flags & FOLL_GET))
 792                 return ERR_PTR(-EEXIST);
 793
 794         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
 795         pgmap = get_dev_pagemap(pfn, NULL);
 796         if (!pgmap)
 797                 return ERR_PTR(-EFAULT);
 798         page = pfn_to_page(pfn);
 799         get_page(page);
 800         put_dev_pagemap(pgmap);
 801
 802         return page;
 803 }
 804
 805 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 806                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 807                   struct vm_area_struct *vma)
 808 {
 809         spinlock_t *dst_ptl, *src_ptl;
 810         struct page *src_page;
 811         pmd_t pmd;
 812         pgtable_t pgtable = NULL;
 813         int ret = -ENOMEM;
 814
 815         /* Skip if can be re-fill on fault */
 816         if (!vma_is_anonymous(vma))
 817                 return 0;
 818
 819         pgtable = pte_alloc_one(dst_mm, addr);
 820         if (unlikely(!pgtable))
 821                 goto out;
 822
 823         dst_ptl = pmd_lock(dst_mm, dst_pmd);
 824         src_ptl = pmd_lockptr(src_mm, src_pmd);
 825         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 826
 827         ret = -EAGAIN;
 828         pmd = *src_pmd;
 829         if (unlikely(!pmd_trans_huge(pmd))) {
 830                 pte_free(dst_mm, pgtable);
 831                 goto out_unlock;
 832         }
 833         /*
 834          * When page table lock is held, the huge zero pmd should not be
 835          * under splitting since we don't split the page itself, only pmd to
 836          * a page table.
 837          */
 838         if (is_huge_zero_pmd(pmd)) {
 839                 struct page *zero_page;
 840                 /*
 841                  * get_huge_zero_page() will never allocate a new page here,
 842                  * since we already have a zero page to copy. It just takes a
 843                  * reference.
 844                  */
 845                 zero_page = mm_get_huge_zero_page(dst_mm);
 846                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
 847                                 zero_page);
 848                 ret = 0;
 849                 goto out_unlock;
 850         }
 851
 852         src_page = pmd_page(pmd);
 853         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
 854         get_page(src_page);
 855         page_dup_rmap(src_page, true);
 856         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 857         atomic_long_inc(&dst_mm->nr_ptes);
 858         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
 859
 860         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 861         pmd = pmd_mkold(pmd_wrprotect(pmd));
 862         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 863
 864         ret = 0;
 865 out_unlock:
 866         spin_unlock(src_ptl);
 867         spin_unlock(dst_ptl);
 868 out:
 869         return ret;
 870 }
 871
 872 void huge_pmd_set_accessed(struct fault_env *fe, pmd_t orig_pmd)
 873 {
 874         pmd_t entry;
 875         unsigned long haddr;
 876         bool write = fe->flags & FAULT_FLAG_WRITE;
 877
 878         fe->ptl = pmd_lock(fe->vma->vm_mm, fe->pmd);
 879         if (unlikely(!pmd_same(*fe->pmd, orig_pmd)))
 880                 goto unlock;
 881
 882         entry = pmd_mkyoung(orig_pmd);
 883         if (write)
 884                 entry = pmd_mkdirty(entry);
 885         haddr = fe->address & HPAGE_PMD_MASK;
 886         if (pmdp_set_access_flags(fe->vma, haddr, fe->pmd, entry, write))
 887                 update_mmu_cache_pmd(fe->vma, fe->address, fe->pmd);
 888
 889 unlock:
 890         spin_unlock(fe->ptl);
 891 }
 892
 893 static int do_huge_pmd_wp_page_fallback(struct fault_env *fe, pmd_t orig_pmd,
 894                 struct page *page)
 895 {
 896         struct vm_area_struct *vma = fe->vma;
 897         unsigned long haddr = fe->address & HPAGE_PMD_MASK;
 898         struct mem_cgroup *memcg;
 899         pgtable_t pgtable;
 900         pmd_t _pmd;
 901         int ret = 0, i;
 902         struct page **pages;
 903         unsigned long mmun_start;       /* For mmu_notifiers */
 904         unsigned long mmun_end;         /* For mmu_notifiers */
 905
 906         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
 907                         GFP_KERNEL);
 908         if (unlikely(!pages)) {
 909                 ret |= VM_FAULT_OOM;
 910                 goto out;
 911         }
 912
 913         for (i = 0; i < HPAGE_PMD_NR; i++) {
 914                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
 915                                                __GFP_OTHER_NODE, vma,
 916                                                fe->address, page_to_nid(page));
 917                 if (unlikely(!pages[i] ||
 918                              mem_cgroup_try_charge(pages[i], vma->vm_mm,
 919                                      GFP_KERNEL, &memcg, false))) {
 920                         if (pages[i])
 921                                 put_page(pages[i]);
 922                         while (--i >= 0) {
 923                                 memcg = (void *)page_private(pages[i]);
 924                                 set_page_private(pages[i], 0);
 925                                 mem_cgroup_cancel_charge(pages[i], memcg,
 926                                                 false);
 927                                 put_page(pages[i]);
 928                         }
 929                         kfree(pages);
 930                         ret |= VM_FAULT_OOM;
 931                         goto out;
 932                 }
 933                 set_page_private(pages[i], (unsigned long)memcg);
 934         }
 935
 936         for (i = 0; i < HPAGE_PMD_NR; i++) {
 937                 copy_user_highpage(pages[i], page + i,
 938                                    haddr + PAGE_SIZE * i, vma);
 939                 __SetPageUptodate(pages[i]);
 940                 cond_resched();
 941         }
 942
 943         mmun_start = haddr;
 944         mmun_end   = haddr + HPAGE_PMD_SIZE;
 945         mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
 946
 947         fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
 948         if (unlikely(!pmd_same(*fe->pmd, orig_pmd)))
 949                 goto out_free_pages;
 950         VM_BUG_ON_PAGE(!PageHead(page), page);
 951
 952         pmdp_huge_clear_flush_notify(vma, haddr, fe->pmd);
 953         /* leave pmd empty until pte is filled */
 954
 955         pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, fe->pmd);
 956         pmd_populate(vma->vm_mm, &_pmd, pgtable);
 957
 958         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
 959                 pte_t entry;
 960                 entry = mk_pte(pages[i], vma->vm_page_prot);
 961                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
 962                 memcg = (void *)page_private(pages[i]);
 963                 set_page_private(pages[i], 0);
 964                 page_add_new_anon_rmap(pages[i], fe->vma, haddr, false);
 965                 mem_cgroup_commit_charge(pages[i], memcg, false, false);
 966                 lru_cache_add_active_or_unevictable(pages[i], vma);
 967                 fe->pte = pte_offset_map(&_pmd, haddr);
 968                 VM_BUG_ON(!pte_none(*fe->pte));
 969                 set_pte_at(vma->vm_mm, haddr, fe->pte, entry);
 970                 pte_unmap(fe->pte);
 971         }
 972         kfree(pages);
 973
 974         smp_wmb(); /* make pte visible before pmd */
 975         pmd_populate(vma->vm_mm, fe->pmd, pgtable);
 976         page_remove_rmap(page, true);
 977         spin_unlock(fe->ptl);
 978
 979         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
 980
 981         ret |= VM_FAULT_WRITE;
 982         put_page(page);
 983
 984 out:
 985         return ret;
 986
 987 out_free_pages:
 988         spin_unlock(fe->ptl);
 989         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
 990         for (i = 0; i < HPAGE_PMD_NR; i++) {
 991                 memcg = (void *)page_private(pages[i]);
 992                 set_page_private(pages[i], 0);
 993                 mem_cgroup_cancel_charge(pages[i], memcg, false);
 994                 put_page(pages[i]);
 995         }
 996         kfree(pages);
 997         goto out;
 998 }
 999
1000 int do_huge_pmd_wp_page(struct fault_env *fe, pmd_t orig_pmd)
1001 {
1002         struct vm_area_struct *vma = fe->vma;
1003         struct page *page = NULL, *new_page;
1004         struct mem_cgroup *memcg;
1005         unsigned long haddr = fe->address & HPAGE_PMD_MASK;
1006         unsigned long mmun_start;       /* For mmu_notifiers */
1007         unsigned long mmun_end;         /* For mmu_notifiers */
1008         gfp_t huge_gfp;                 /* for allocation and charge */
1009         int ret = 0;
1010
1011         fe->ptl = pmd_lockptr(vma->vm_mm, fe->pmd);
1012         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1013         if (is_huge_zero_pmd(orig_pmd))
1014                 goto alloc;
1015         spin_lock(fe->ptl);
1016         if (unlikely(!pmd_same(*fe->pmd, orig_pmd)))
1017                 goto out_unlock;
1018
1019         page = pmd_page(orig_pmd);
1020         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1021         /*
1022          * We can only reuse the page if nobody else maps the huge page or it's
1023          * part.
1024          */
1025         if (!trylock_page(page)) {
1026                 get_page(page);
1027                 spin_unlock(fe->ptl);
1028                 lock_page(page);
1029                 spin_lock(fe->ptl);
1030                 if (unlikely(!pmd_same(*fe->pmd, orig_pmd))) {
1031                         unlock_page(page);
1032                         put_page(page);
1033                         goto out_unlock;
1034                 }
1035                 put_page(page);
1036         }
1037
1038         if (page_trans_huge_mapcount(page, NULL) == 1) {
1039                 pmd_t entry;
1040                 entry = pmd_mkyoung(orig_pmd);
1041                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1042                 if (pmdp_set_access_flags(vma, haddr, fe->pmd, entry,  1))
1043                         update_mmu_cache_pmd(vma, fe->address, fe->pmd);
1044                 ret |= VM_FAULT_WRITE;
1045                 unlock_page(page);
1046                 goto out_unlock;
1047         }
1048         unlock_page(page);
1049         get_page(page);
1050         spin_unlock(fe->ptl);
1051 alloc:
1052         if (transparent_hugepage_enabled(vma) &&
1053             !transparent_hugepage_debug_cow()) {
1054                 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1055                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1056         } else
1057                 new_page = NULL;
1058
1059         if (likely(new_page)) {
1060                 prep_transhuge_page(new_page);
1061         } else {
1062                 if (!page) {
1063                         split_huge_pmd(vma, fe->pmd, fe->address);
1064                         ret |= VM_FAULT_FALLBACK;
1065                 } else {
1066                         ret = do_huge_pmd_wp_page_fallback(fe, orig_pmd, page);
1067                         if (ret & VM_FAULT_OOM) {
1068                                 split_huge_pmd(vma, fe->pmd, fe->address);
1069                                 ret |= VM_FAULT_FALLBACK;
1070                         }
1071                         put_page(page);
1072                 }
1073                 count_vm_event(THP_FAULT_FALLBACK);
1074                 goto out;
1075         }
1076
1077         if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm,
1078                                 huge_gfp | __GFP_NORETRY, &memcg, true))) {
1079                 put_page(new_page);
1080                 split_huge_pmd(vma, fe->pmd, fe->address);
1081                 if (page)
1082                         put_page(page);
1083                 ret |= VM_FAULT_FALLBACK;
1084                 count_vm_event(THP_FAULT_FALLBACK);
1085                 goto out;
1086         }
1087
1088         count_vm_event(THP_FAULT_ALLOC);
1089
1090         if (!page)
1091                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1092         else
1093                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1094         __SetPageUptodate(new_page);
1095
1096         mmun_start = haddr;
1097         mmun_end   = haddr + HPAGE_PMD_SIZE;
1098         mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
1099
1100         spin_lock(fe->ptl);
1101         if (page)
1102                 put_page(page);
1103         if (unlikely(!pmd_same(*fe->pmd, orig_pmd))) {
1104                 spin_unlock(fe->ptl);
1105                 mem_cgroup_cancel_charge(new_page, memcg, true);
1106                 put_page(new_page);
1107                 goto out_mn;
1108         } else {
1109                 pmd_t entry;
1110                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1111                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1112                 pmdp_huge_clear_flush_notify(vma, haddr, fe->pmd);
1113                 page_add_new_anon_rmap(new_page, vma, haddr, true);
1114                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1115                 lru_cache_add_active_or_unevictable(new_page, vma);
1116                 set_pmd_at(vma->vm_mm, haddr, fe->pmd, entry);
1117                 update_mmu_cache_pmd(vma, fe->address, fe->pmd);
1118                 if (!page) {
1119                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1120                 } else {
1121                         VM_BUG_ON_PAGE(!PageHead(page), page);
1122                         page_remove_rmap(page, true);
1123                         put_page(page);
1124                 }
1125                 ret |= VM_FAULT_WRITE;
1126         }
1127         spin_unlock(fe->ptl);
1128 out_mn:
1129         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1130 out:
1131         return ret;
1132 out_unlock:
1133         spin_unlock(fe->ptl);
1134         return ret;
1135 }
1136
1137 /*
1138  * FOLL_FORCE or a forced COW break can write even to unwritable pmd's,
1139  * but only after we've gone through a COW cycle and they are dirty.
1140  */
1141 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1142 {
1143         return pmd_write(pmd) || ((flags & FOLL_COW) && pmd_dirty(pmd));
1144 }
1145
1146 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1147                                    unsigned long addr,
1148                                    pmd_t *pmd,
1149                                    unsigned int flags)
1150 {
1151         struct mm_struct *mm = vma->vm_mm;
1152         struct page *page = NULL;
1153
1154         assert_spin_locked(pmd_lockptr(mm, pmd));
1155
1156         if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1157                 goto out;
1158
1159         /* Avoid dumping huge zero page */
1160         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1161                 return ERR_PTR(-EFAULT);
1162
1163         /* Full NUMA hinting faults to serialise migration in fault paths */
1164         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1165                 goto out;
1166
1167         page = pmd_page(*pmd);
1168         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1169         if (flags & FOLL_TOUCH)
1170                 touch_pmd(vma, addr, pmd, flags);
1171         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1172                 /*
1173                  * We don't mlock() pte-mapped THPs. This way we can avoid
1174                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1175                  *
1176                  * For anon THP:
1177                  *
1178                  * In most cases the pmd is the only mapping of the page as we
1179                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1180                  * writable private mappings in populate_vma_page_range().
1181                  *
1182                  * The only scenario when we have the page shared here is if we
1183                  * mlocking read-only mapping shared over fork(). We skip
1184                  * mlocking such pages.
1185                  *
1186                  * For file THP:
1187                  *
1188                  * We can expect PageDoubleMap() to be stable under page lock:
1189                  * for file pages we set it in page_add_file_rmap(), which
1190                  * requires page to be locked.
1191                  */
1192
1193                 if (PageAnon(page) && compound_mapcount(page) != 1)
1194                         goto skip_mlock;
1195                 if (PageDoubleMap(page) || !page->mapping)
1196                         goto skip_mlock;
1197                 if (!trylock_page(page))
1198                         goto skip_mlock;
1199                 lru_add_drain();
1200                 if (page->mapping && !PageDoubleMap(page))
1201                         mlock_vma_page(page);
1202                 unlock_page(page);
1203         }
1204 skip_mlock:
1205         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1206         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1207         if (flags & FOLL_GET)
1208                 get_page(page);
1209
1210 out:
1211         return page;
1212 }
1213
1214 /* NUMA hinting page fault entry point for trans huge pmds */
1215 int do_huge_pmd_numa_page(struct fault_env *fe, pmd_t pmd)
1216 {
1217         struct vm_area_struct *vma = fe->vma;
1218         struct anon_vma *anon_vma = NULL;
1219         struct page *page;
1220         unsigned long haddr = fe->address & HPAGE_PMD_MASK;
1221         int page_nid = -1, this_nid = numa_node_id();
1222         int target_nid, last_cpupid = -1;
1223         bool page_locked;
1224         bool migrated = false;
1225         bool was_writable;
1226         int flags = 0;
1227
1228         fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
1229         if (unlikely(!pmd_same(pmd, *fe->pmd)))
1230                 goto out_unlock;
1231
1232         /*
1233          * If there are potential migrations, wait for completion and retry
1234          * without disrupting NUMA hinting information. Do not relock and
1235          * check_same as the page may no longer be mapped.
1236          */
1237         if (unlikely(pmd_trans_migrating(*fe->pmd))) {
1238                 page = pmd_page(*fe->pmd);
1239                 if (!get_page_unless_zero(page))
1240                         goto out_unlock;
1241                 spin_unlock(fe->ptl);
1242                 wait_on_page_locked(page);
1243                 put_page(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                 page_nid = -1;
1276                 if (!get_page_unless_zero(page))
1277                         goto out_unlock;
1278                 spin_unlock(fe->ptl);
1279                 wait_on_page_locked(page);
1280                 put_page(page);
1281                 goto out;
1282         }
1283
1284         /*
1285          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1286          * to serialises splits
1287          */
1288         get_page(page);
1289         spin_unlock(fe->ptl);
1290         anon_vma = page_lock_anon_vma_read(page);
1291
1292         /* Confirm the PMD did not change while page_table_lock was released */
1293         spin_lock(fe->ptl);
1294         if (unlikely(!pmd_same(pmd, *fe->pmd))) {
1295                 unlock_page(page);
1296                 put_page(page);
1297                 page_nid = -1;
1298                 goto out_unlock;
1299         }
1300
1301         /* Bail if we fail to protect against THP splits for any reason */
1302         if (unlikely(!anon_vma)) {
1303                 put_page(page);
1304                 page_nid = -1;
1305                 goto clear_pmdnuma;
1306         }
1307
1308         /*
1309          * Migrate the THP to the requested node, returns with page unlocked
1310          * and access rights restored.
1311          */
1312         spin_unlock(fe->ptl);
1313         migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1314                                 fe->pmd, pmd, fe->address, page, target_nid);
1315         if (migrated) {
1316                 flags |= TNF_MIGRATED;
1317                 page_nid = target_nid;
1318         } else
1319                 flags |= TNF_MIGRATE_FAIL;
1320
1321         goto out;
1322 clear_pmdnuma:
1323         BUG_ON(!PageLocked(page));
1324         was_writable = pmd_write(pmd);
1325         pmd = pmd_modify(pmd, vma->vm_page_prot);
1326         pmd = pmd_mkyoung(pmd);
1327         if (was_writable)
1328                 pmd = pmd_mkwrite(pmd);
1329         set_pmd_at(vma->vm_mm, haddr, fe->pmd, pmd);
1330         update_mmu_cache_pmd(vma, fe->address, fe->pmd);
1331         unlock_page(page);
1332 out_unlock:
1333         spin_unlock(fe->ptl);
1334
1335 out:
1336         if (anon_vma)
1337                 page_unlock_anon_vma_read(anon_vma);
1338
1339         if (page_nid != -1)
1340                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, fe->flags);
1341
1342         return 0;
1343 }
1344
1345 /*
1346  * Return true if we do MADV_FREE successfully on entire pmd page.
1347  * Otherwise, return false.
1348  */
1349 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1350                 pmd_t *pmd, unsigned long addr, unsigned long next)
1351 {
1352         spinlock_t *ptl;
1353         pmd_t orig_pmd;
1354         struct page *page;
1355         struct mm_struct *mm = tlb->mm;
1356         bool ret = false;
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 (total_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                 unlock_page(page);
1386                 put_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                 pmdp_invalidate(vma, addr, pmd);
1399                 orig_pmd = pmd_mkold(orig_pmd);
1400                 orig_pmd = pmd_mkclean(orig_pmd);
1401
1402                 set_pmd_at(mm, addr, pmd, orig_pmd);
1403                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1404         }
1405         ret = true;
1406 out:
1407         spin_unlock(ptl);
1408 out_unlocked:
1409         return ret;
1410 }
1411
1412 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1413                  pmd_t *pmd, unsigned long addr)
1414 {
1415         pmd_t orig_pmd;
1416         spinlock_t *ptl;
1417
1418         ptl = __pmd_trans_huge_lock(pmd, vma);
1419         if (!ptl)
1420                 return 0;
1421         /*
1422          * For architectures like ppc64 we look at deposited pgtable
1423          * when calling pmdp_huge_get_and_clear. So do the
1424          * pgtable_trans_huge_withdraw after finishing pmdp related
1425          * operations.
1426          */
1427         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1428                         tlb->fullmm);
1429         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1430         if (vma_is_dax(vma)) {
1431                 spin_unlock(ptl);
1432                 if (is_huge_zero_pmd(orig_pmd))
1433                         tlb_remove_page(tlb, pmd_page(orig_pmd));
1434         } else if (is_huge_zero_pmd(orig_pmd)) {
1435                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1436                 atomic_long_dec(&tlb->mm->nr_ptes);
1437                 spin_unlock(ptl);
1438                 tlb_remove_page(tlb, pmd_page(orig_pmd));
1439         } else {
1440                 struct page *page = pmd_page(orig_pmd);
1441                 page_remove_rmap(page, true);
1442                 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1443                 VM_BUG_ON_PAGE(!PageHead(page), page);
1444                 if (PageAnon(page)) {
1445                         pgtable_t pgtable;
1446                         pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1447                         pte_free(tlb->mm, pgtable);
1448                         atomic_long_dec(&tlb->mm->nr_ptes);
1449                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1450                 } else {
1451                         add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1452                 }
1453                 spin_unlock(ptl);
1454                 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1455         }
1456         return 1;
1457 }
1458
1459 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1460                   unsigned long new_addr, unsigned long old_end,
1461                   pmd_t *old_pmd, pmd_t *new_pmd)
1462 {
1463         spinlock_t *old_ptl, *new_ptl;
1464         pmd_t pmd;
1465         struct mm_struct *mm = vma->vm_mm;
1466         bool force_flush = false;
1467
1468         if ((old_addr & ~HPAGE_PMD_MASK) ||
1469             (new_addr & ~HPAGE_PMD_MASK) ||
1470             old_end - old_addr < HPAGE_PMD_SIZE)
1471                 return false;
1472
1473         /*
1474          * The destination pmd shouldn't be established, free_pgtables()
1475          * should have release it.
1476          */
1477         if (WARN_ON(!pmd_none(*new_pmd))) {
1478                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1479                 return false;
1480         }
1481
1482         /*
1483          * We don't have to worry about the ordering of src and dst
1484          * ptlocks because exclusive mmap_sem prevents deadlock.
1485          */
1486         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1487         if (old_ptl) {
1488                 new_ptl = pmd_lockptr(mm, new_pmd);
1489                 if (new_ptl != old_ptl)
1490                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1491                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1492                 if (pmd_present(pmd))
1493                         force_flush = true;
1494                 VM_BUG_ON(!pmd_none(*new_pmd));
1495
1496                 if (pmd_move_must_withdraw(new_ptl, old_ptl) &&
1497                                 vma_is_anonymous(vma)) {
1498                         pgtable_t pgtable;
1499                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1500                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1501                 }
1502                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1503                 if (force_flush)
1504                         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1505                 if (new_ptl != old_ptl)
1506                         spin_unlock(new_ptl);
1507                 spin_unlock(old_ptl);
1508                 return true;
1509         }
1510         return false;
1511 }
1512
1513 /*
1514  * Returns
1515  *  - 0 if PMD could not be locked
1516  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1517  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1518  */
1519 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1520                 unsigned long addr, pgprot_t newprot, int prot_numa)
1521 {
1522         struct mm_struct *mm = vma->vm_mm;
1523         spinlock_t *ptl;
1524         pmd_t entry;
1525         bool preserve_write;
1526         int ret;
1527
1528         ptl = __pmd_trans_huge_lock(pmd, vma);
1529         if (!ptl)
1530                 return 0;
1531
1532         preserve_write = prot_numa && pmd_write(*pmd);
1533         ret = 1;
1534
1535         /*
1536          * Avoid trapping faults against the zero page. The read-only
1537          * data is likely to be read-cached on the local CPU and
1538          * local/remote hits to the zero page are not interesting.
1539          */
1540         if (prot_numa && is_huge_zero_pmd(*pmd))
1541                 goto unlock;
1542
1543         if (prot_numa && pmd_protnone(*pmd))
1544                 goto unlock;
1545
1546         /*
1547          * In case prot_numa, we are under down_read(mmap_sem). It's critical
1548          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1549          * which is also under down_read(mmap_sem):
1550          *
1551          *      CPU0:                           CPU1:
1552          *                              change_huge_pmd(prot_numa=1)
1553          *                               pmdp_huge_get_and_clear_notify()
1554          * madvise_dontneed()
1555          *  zap_pmd_range()
1556          *   pmd_trans_huge(*pmd) == 0 (without ptl)
1557          *   // skip the pmd
1558          *                               set_pmd_at();
1559          *                               // pmd is re-established
1560          *
1561          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1562          * which may break userspace.
1563          *
1564          * pmdp_invalidate() is required to make sure we don't miss
1565          * dirty/young flags set by hardware.
1566          */
1567         entry = *pmd;
1568         pmdp_invalidate(vma, addr, pmd);
1569
1570         /*
1571          * Recover dirty/young flags.  It relies on pmdp_invalidate to not
1572          * corrupt them.
1573          */
1574         if (pmd_dirty(*pmd))
1575                 entry = pmd_mkdirty(entry);
1576         if (pmd_young(*pmd))
1577                 entry = pmd_mkyoung(entry);
1578
1579         entry = pmd_modify(entry, newprot);
1580         if (preserve_write)
1581                 entry = pmd_mkwrite(entry);
1582         ret = HPAGE_PMD_NR;
1583         set_pmd_at(mm, addr, pmd, entry);
1584         BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1585 unlock:
1586         spin_unlock(ptl);
1587         return ret;
1588 }
1589
1590 /*
1591  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1592  *
1593  * Note that if it returns page table lock pointer, this routine returns without
1594  * unlocking page table lock. So callers must unlock it.
1595  */
1596 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1597 {
1598         spinlock_t *ptl;
1599         ptl = pmd_lock(vma->vm_mm, pmd);
1600         if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd)))
1601                 return ptl;
1602         spin_unlock(ptl);
1603         return NULL;
1604 }
1605
1606 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1607                 unsigned long haddr, pmd_t *pmd)
1608 {
1609         struct mm_struct *mm = vma->vm_mm;
1610         pgtable_t pgtable;
1611         pmd_t _pmd;
1612         int i;
1613
1614         /* leave pmd empty until pte is filled */
1615         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1616
1617         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1618         pmd_populate(mm, &_pmd, pgtable);
1619
1620         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1621                 pte_t *pte, entry;
1622                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
1623                 entry = pte_mkspecial(entry);
1624                 pte = pte_offset_map(&_pmd, haddr);
1625                 VM_BUG_ON(!pte_none(*pte));
1626                 set_pte_at(mm, haddr, pte, entry);
1627                 pte_unmap(pte);
1628         }
1629         smp_wmb(); /* make pte visible before pmd */
1630         pmd_populate(mm, pmd, pgtable);
1631 }
1632
1633 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
1634                 unsigned long haddr, bool freeze)
1635 {
1636         struct mm_struct *mm = vma->vm_mm;
1637         struct page *page;
1638         pgtable_t pgtable;
1639         pmd_t _pmd;
1640         bool young, write, dirty, soft_dirty;
1641         unsigned long addr;
1642         int i;
1643
1644         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
1645         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1646         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
1647         VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd));
1648
1649         count_vm_event(THP_SPLIT_PMD);
1650
1651         if (!vma_is_anonymous(vma)) {
1652                 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1653                 if (vma_is_dax(vma))
1654                         return;
1655                 page = pmd_page(_pmd);
1656                 if (!PageDirty(page) && pmd_dirty(_pmd))
1657                         set_page_dirty(page);
1658                 if (!PageReferenced(page) && pmd_young(_pmd))
1659                         SetPageReferenced(page);
1660                 page_remove_rmap(page, true);
1661                 put_page(page);
1662                 add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1663                 return;
1664         } else if (is_huge_zero_pmd(*pmd)) {
1665                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
1666         }
1667
1668         page = pmd_page(*pmd);
1669         VM_BUG_ON_PAGE(!page_count(page), page);
1670         page_ref_add(page, HPAGE_PMD_NR - 1);
1671         write = pmd_write(*pmd);
1672         young = pmd_young(*pmd);
1673         dirty = pmd_dirty(*pmd);
1674         soft_dirty = pmd_soft_dirty(*pmd);
1675
1676         pmdp_huge_split_prepare(vma, haddr, pmd);
1677         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1678         pmd_populate(mm, &_pmd, pgtable);
1679
1680         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
1681                 pte_t entry, *pte;
1682                 /*
1683                  * Note that NUMA hinting access restrictions are not
1684                  * transferred to avoid any possibility of altering
1685                  * permissions across VMAs.
1686                  */
1687                 if (freeze) {
1688                         swp_entry_t swp_entry;
1689                         swp_entry = make_migration_entry(page + i, write);
1690                         entry = swp_entry_to_pte(swp_entry);
1691                         if (soft_dirty)
1692                                 entry = pte_swp_mksoft_dirty(entry);
1693                 } else {
1694                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
1695                         entry = maybe_mkwrite(entry, vma);
1696                         if (!write)
1697                                 entry = pte_wrprotect(entry);
1698                         if (!young)
1699                                 entry = pte_mkold(entry);
1700                         if (soft_dirty)
1701                                 entry = pte_mksoft_dirty(entry);
1702                 }
1703                 if (dirty)
1704                         SetPageDirty(page + i);
1705                 pte = pte_offset_map(&_pmd, addr);
1706                 BUG_ON(!pte_none(*pte));
1707                 set_pte_at(mm, addr, pte, entry);
1708                 atomic_inc(&page[i]._mapcount);
1709                 pte_unmap(pte);
1710         }
1711
1712         /*
1713          * Set PG_double_map before dropping compound_mapcount to avoid
1714          * false-negative page_mapped().
1715          */
1716         if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
1717                 for (i = 0; i < HPAGE_PMD_NR; i++)
1718                         atomic_inc(&page[i]._mapcount);
1719         }
1720
1721         if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
1722                 /* Last compound_mapcount is gone. */
1723                 __dec_node_page_state(page, NR_ANON_THPS);
1724                 if (TestClearPageDoubleMap(page)) {
1725                         /* No need in mapcount reference anymore */
1726                         for (i = 0; i < HPAGE_PMD_NR; i++)
1727                                 atomic_dec(&page[i]._mapcount);
1728                 }
1729         }
1730
1731         smp_wmb(); /* make pte visible before pmd */
1732         /*
1733          * Up to this point the pmd is present and huge and userland has the
1734          * whole access to the hugepage during the split (which happens in
1735          * place). If we overwrite the pmd with the not-huge version pointing
1736          * to the pte here (which of course we could if all CPUs were bug
1737          * free), userland could trigger a small page size TLB miss on the
1738          * small sized TLB while the hugepage TLB entry is still established in
1739          * the huge TLB. Some CPU doesn't like that.
1740          * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
1741          * 383 on page 93. Intel should be safe but is also warns that it's
1742          * only safe if the permission and cache attributes of the two entries
1743          * loaded in the two TLB is identical (which should be the case here).
1744          * But it is generally safer to never allow small and huge TLB entries
1745          * for the same virtual address to be loaded simultaneously. So instead
1746          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
1747          * current pmd notpresent (atomically because here the pmd_trans_huge
1748          * and pmd_trans_splitting must remain set at all times on the pmd
1749          * until the split is complete for this pmd), then we flush the SMP TLB
1750          * and finally we write the non-huge version of the pmd entry with
1751          * pmd_populate.
1752          */
1753         pmdp_invalidate(vma, haddr, pmd);
1754         pmd_populate(mm, pmd, pgtable);
1755
1756         if (freeze) {
1757                 for (i = 0; i < HPAGE_PMD_NR; i++) {
1758                         page_remove_rmap(page + i, false);
1759                         put_page(page + i);
1760                 }
1761         }
1762 }
1763
1764 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1765                 unsigned long address, bool freeze, struct page *page)
1766 {
1767         spinlock_t *ptl;
1768         struct mm_struct *mm = vma->vm_mm;
1769         unsigned long haddr = address & HPAGE_PMD_MASK;
1770         bool do_unlock_page = false;
1771         pmd_t _pmd;
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) {
1782                 VM_WARN_ON_ONCE(!PageLocked(page));
1783                 if (page != pmd_page(*pmd))
1784                         goto out;
1785         }
1786
1787 repeat:
1788         if (pmd_trans_huge(*pmd)) {
1789                 if (!page) {
1790                         page = pmd_page(*pmd);
1791                         /*
1792                          * An anonymous page must be locked, to ensure that a
1793                          * concurrent reuse_swap_page() sees stable mapcount;
1794                          * but reuse_swap_page() is not used on shmem or file,
1795                          * and page lock must not be taken when zap_pmd_range()
1796                          * calls __split_huge_pmd() while i_mmap_lock is held.
1797                          */
1798                         if (PageAnon(page)) {
1799                                 if (unlikely(!trylock_page(page))) {
1800                                         get_page(page);
1801                                         _pmd = *pmd;
1802                                         spin_unlock(ptl);
1803                                         lock_page(page);
1804                                         spin_lock(ptl);
1805                                         if (unlikely(!pmd_same(*pmd, _pmd))) {
1806                                                 unlock_page(page);
1807                                                 put_page(page);
1808                                                 page = NULL;
1809                                                 goto repeat;
1810                                         }
1811                                         put_page(page);
1812                                 }
1813                                 do_unlock_page = true;
1814                         }
1815                 }
1816                 if (PageMlocked(page))
1817                         clear_page_mlock(page);
1818         } else if (!pmd_devmap(*pmd))
1819                 goto out;
1820         __split_huge_pmd_locked(vma, pmd, haddr, freeze);
1821 out:
1822         spin_unlock(ptl);
1823         if (do_unlock_page)
1824                 unlock_page(page);
1825         mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
1826 }
1827
1828 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
1829                 bool freeze, struct page *page)
1830 {
1831         pgd_t *pgd;
1832         pud_t *pud;
1833         pmd_t *pmd;
1834
1835         pgd = pgd_offset(vma->vm_mm, address);
1836         if (!pgd_present(*pgd))
1837                 return;
1838
1839         pud = pud_offset(pgd, address);
1840         if (!pud_present(*pud))
1841                 return;
1842
1843         pmd = pmd_offset(pud, address);
1844
1845         __split_huge_pmd(vma, pmd, address, freeze, page);
1846 }
1847
1848 void vma_adjust_trans_huge(struct vm_area_struct *vma,
1849                              unsigned long start,
1850                              unsigned long end,
1851                              long adjust_next)
1852 {
1853         /*
1854          * If the new start address isn't hpage aligned and it could
1855          * previously contain an hugepage: check if we need to split
1856          * an huge pmd.
1857          */
1858         if (start & ~HPAGE_PMD_MASK &&
1859             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
1860             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
1861                 split_huge_pmd_address(vma, start, false, NULL);
1862
1863         /*
1864          * If the new end address isn't hpage aligned and it could
1865          * previously contain an hugepage: check if we need to split
1866          * an huge pmd.
1867          */
1868         if (end & ~HPAGE_PMD_MASK &&
1869             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
1870             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
1871                 split_huge_pmd_address(vma, end, false, NULL);
1872
1873         /*
1874          * If we're also updating the vma->vm_next->vm_start, if the new
1875          * vm_next->vm_start isn't page aligned and it could previously
1876          * contain an hugepage: check if we need to split an huge pmd.
1877          */
1878         if (adjust_next > 0) {
1879                 struct vm_area_struct *next = vma->vm_next;
1880                 unsigned long nstart = next->vm_start;
1881                 nstart += adjust_next << PAGE_SHIFT;
1882                 if (nstart & ~HPAGE_PMD_MASK &&
1883                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
1884                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
1885                         split_huge_pmd_address(next, nstart, false, NULL);
1886         }
1887 }
1888
1889 static void unmap_page(struct page *page)
1890 {
1891         enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
1892                 TTU_RMAP_LOCKED;
1893         int i;
1894
1895         VM_BUG_ON_PAGE(!PageHead(page), page);
1896
1897         if (PageAnon(page))
1898                 ttu_flags |= TTU_MIGRATION;
1899
1900         /* We only need TTU_SPLIT_HUGE_PMD once */
1901         try_to_unmap(page, ttu_flags | TTU_SPLIT_HUGE_PMD);
1902         for (i = 1; i < HPAGE_PMD_NR; i++) {
1903                 /* Cut short if the page is unmapped */
1904                 if (page_count(page) == 1)
1905                         return;
1906
1907                 try_to_unmap(page + i, ttu_flags);
1908         }
1909
1910         VM_WARN_ON_ONCE_PAGE(page_mapped(page), page);
1911 }
1912
1913 static void remap_page(struct page *page)
1914 {
1915         int i;
1916
1917         for (i = 0; i < HPAGE_PMD_NR; i++)
1918                 remove_migration_ptes(page + i, page + i, true);
1919 }
1920
1921 static void __split_huge_page_tail(struct page *head, int tail,
1922                 struct lruvec *lruvec, struct list_head *list)
1923 {
1924         struct page *page_tail = head + tail;
1925
1926         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
1927
1928         /*
1929          * Clone page flags before unfreezing refcount.
1930          *
1931          * After successful get_page_unless_zero() might follow flags change,
1932          * for exmaple lock_page() which set PG_waiters.
1933          */
1934         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1935         page_tail->flags |= (head->flags &
1936                         ((1L << PG_referenced) |
1937                          (1L << PG_swapbacked) |
1938                          (1L << PG_mlocked) |
1939                          (1L << PG_uptodate) |
1940                          (1L << PG_active) |
1941                          (1L << PG_locked) |
1942                          (1L << PG_unevictable) |
1943                          (1L << PG_dirty)));
1944
1945         /* ->mapping in first tail page is compound_mapcount */
1946         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
1947                         page_tail);
1948         page_tail->mapping = head->mapping;
1949         page_tail->index = head->index + tail;
1950
1951         /* Page flags must be visible before we make the page non-compound. */
1952         smp_wmb();
1953
1954         /*
1955          * Clear PageTail before unfreezing page refcount.
1956          *
1957          * After successful get_page_unless_zero() might follow put_page()
1958          * which needs correct compound_head().
1959          */
1960         clear_compound_head(page_tail);
1961
1962         /* Finally unfreeze refcount. Additional reference from page cache. */
1963         page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
1964                                           PageSwapCache(head)));
1965
1966         if (page_is_young(head))
1967                 set_page_young(page_tail);
1968         if (page_is_idle(head))
1969                 set_page_idle(page_tail);
1970
1971         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
1972         lru_add_page_tail(head, page_tail, lruvec, list);
1973 }
1974
1975 static void __split_huge_page(struct page *page, struct list_head *list,
1976                 pgoff_t end, unsigned long flags)
1977 {
1978         struct page *head = compound_head(page);
1979         struct zone *zone = page_zone(head);
1980         struct lruvec *lruvec;
1981         int i;
1982
1983         lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat);
1984
1985         /* complete memcg works before add pages to LRU */
1986         mem_cgroup_split_huge_fixup(head);
1987
1988         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1989                 __split_huge_page_tail(head, i, lruvec, list);
1990                 /* Some pages can be beyond i_size: drop them from page cache */
1991                 if (head[i].index >= end) {
1992                         __ClearPageDirty(head + i);
1993                         __delete_from_page_cache(head + i, NULL);
1994                         if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
1995                                 shmem_uncharge(head->mapping->host, 1);
1996                         put_page(head + i);
1997                 }
1998         }
1999
2000         ClearPageCompound(head);
2001
2002         split_page_owner(head, HPAGE_PMD_ORDER);
2003
2004         /* See comment in __split_huge_page_tail() */
2005         if (PageAnon(head)) {
2006                 page_ref_inc(head);
2007         } else {
2008                 /* Additional pin to radix tree */
2009                 page_ref_add(head, 2);
2010                 spin_unlock(&head->mapping->tree_lock);
2011         }
2012
2013         spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
2014
2015         remap_page(head);
2016
2017         for (i = 0; i < HPAGE_PMD_NR; i++) {
2018                 struct page *subpage = head + i;
2019                 if (subpage == page)
2020                         continue;
2021                 unlock_page(subpage);
2022
2023                 /*
2024                  * Subpages may be freed if there wasn't any mapping
2025                  * like if add_to_swap() is running on a lru page that
2026                  * had its mapping zapped. And freeing these pages
2027                  * requires taking the lru_lock so we do the put_page
2028                  * of the tail pages after the split is complete.
2029                  */
2030                 put_page(subpage);
2031         }
2032 }
2033
2034 int total_mapcount(struct page *page)
2035 {
2036         int i, compound, ret;
2037
2038         VM_BUG_ON_PAGE(PageTail(page), page);
2039
2040         if (likely(!PageCompound(page)))
2041                 return atomic_read(&page->_mapcount) + 1;
2042
2043         compound = compound_mapcount(page);
2044         if (PageHuge(page))
2045                 return compound;
2046         ret = compound;
2047         for (i = 0; i < HPAGE_PMD_NR; i++)
2048                 ret += atomic_read(&page[i]._mapcount) + 1;
2049         /* File pages has compound_mapcount included in _mapcount */
2050         if (!PageAnon(page))
2051                 return ret - compound * HPAGE_PMD_NR;
2052         if (PageDoubleMap(page))
2053                 ret -= HPAGE_PMD_NR;
2054         return ret;
2055 }
2056
2057 /*
2058  * This calculates accurately how many mappings a transparent hugepage
2059  * has (unlike page_mapcount() which isn't fully accurate). This full
2060  * accuracy is primarily needed to know if copy-on-write faults can
2061  * reuse the page and change the mapping to read-write instead of
2062  * copying them. At the same time this returns the total_mapcount too.
2063  *
2064  * The function returns the highest mapcount any one of the subpages
2065  * has. If the return value is one, even if different processes are
2066  * mapping different subpages of the transparent hugepage, they can
2067  * all reuse it, because each process is reusing a different subpage.
2068  *
2069  * The total_mapcount is instead counting all virtual mappings of the
2070  * subpages. If the total_mapcount is equal to "one", it tells the
2071  * caller all mappings belong to the same "mm" and in turn the
2072  * anon_vma of the transparent hugepage can become the vma->anon_vma
2073  * local one as no other process may be mapping any of the subpages.
2074  *
2075  * It would be more accurate to replace page_mapcount() with
2076  * page_trans_huge_mapcount(), however we only use
2077  * page_trans_huge_mapcount() in the copy-on-write faults where we
2078  * need full accuracy to avoid breaking page pinning, because
2079  * page_trans_huge_mapcount() is slower than page_mapcount().
2080  */
2081 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2082 {
2083         int i, ret, _total_mapcount, mapcount;
2084
2085         /* hugetlbfs shouldn't call it */
2086         VM_BUG_ON_PAGE(PageHuge(page), page);
2087
2088         if (likely(!PageTransCompound(page))) {
2089                 mapcount = atomic_read(&page->_mapcount) + 1;
2090                 if (total_mapcount)
2091                         *total_mapcount = mapcount;
2092                 return mapcount;
2093         }
2094
2095         page = compound_head(page);
2096
2097         _total_mapcount = ret = 0;
2098         for (i = 0; i < HPAGE_PMD_NR; i++) {
2099                 mapcount = atomic_read(&page[i]._mapcount) + 1;
2100                 ret = max(ret, mapcount);
2101                 _total_mapcount += mapcount;
2102         }
2103         if (PageDoubleMap(page)) {
2104                 ret -= 1;
2105                 _total_mapcount -= HPAGE_PMD_NR;
2106         }
2107         mapcount = compound_mapcount(page);
2108         ret += mapcount;
2109         _total_mapcount += mapcount;
2110         if (total_mapcount)
2111                 *total_mapcount = _total_mapcount;
2112         return ret;
2113 }
2114
2115 /*
2116  * This function splits huge page into normal pages. @page can point to any
2117  * subpage of huge page to split. Split doesn't change the position of @page.
2118  *
2119  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2120  * The huge page must be locked.
2121  *
2122  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2123  *
2124  * Both head page and tail pages will inherit mapping, flags, and so on from
2125  * the hugepage.
2126  *
2127  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2128  * they are not mapped.
2129  *
2130  * Returns 0 if the hugepage is split successfully.
2131  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2132  * us.
2133  */
2134 int split_huge_page_to_list(struct page *page, struct list_head *list)
2135 {
2136         struct page *head = compound_head(page);
2137         struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2138         struct anon_vma *anon_vma = NULL;
2139         struct address_space *mapping = NULL;
2140         int extra_pins, ret;
2141         bool mlocked;
2142         unsigned long flags;
2143         pgoff_t end;
2144
2145         VM_BUG_ON_PAGE(is_huge_zero_page(head), head);
2146         VM_BUG_ON_PAGE(!PageLocked(page), page);
2147         VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2148         VM_BUG_ON_PAGE(!PageCompound(page), page);
2149
2150         if (PageAnon(head)) {
2151                 /*
2152                  * The caller does not necessarily hold an mmap_sem that would
2153                  * prevent the anon_vma disappearing so we first we take a
2154                  * reference to it and then lock the anon_vma for write. This
2155                  * is similar to page_lock_anon_vma_read except the write lock
2156                  * is taken to serialise against parallel split or collapse
2157                  * operations.
2158                  */
2159                 anon_vma = page_get_anon_vma(head);
2160                 if (!anon_vma) {
2161                         ret = -EBUSY;
2162                         goto out;
2163                 }
2164                 extra_pins = 0;
2165                 end = -1;
2166                 mapping = NULL;
2167                 anon_vma_lock_write(anon_vma);
2168         } else {
2169                 mapping = head->mapping;
2170
2171                 /* Truncated ? */
2172                 if (!mapping) {
2173                         ret = -EBUSY;
2174                         goto out;
2175                 }
2176
2177                 /* Addidional pins from radix tree */
2178                 extra_pins = HPAGE_PMD_NR;
2179                 anon_vma = NULL;
2180                 i_mmap_lock_read(mapping);
2181
2182                 /*
2183                  *__split_huge_page() may need to trim off pages beyond EOF:
2184                  * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2185                  * which cannot be nested inside the page tree lock. So note
2186                  * end now: i_size itself may be changed at any moment, but
2187                  * head page lock is good enough to serialize the trimming.
2188                  */
2189                 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2190         }
2191
2192         /*
2193          * Racy check if we can split the page, before unmap_page() will
2194          * split PMDs
2195          */
2196         if (total_mapcount(head) != page_count(head) - extra_pins - 1) {
2197                 ret = -EBUSY;
2198                 goto out_unlock;
2199         }
2200
2201         mlocked = PageMlocked(page);
2202         unmap_page(head);
2203
2204         /* Make sure the page is not on per-CPU pagevec as it takes pin */
2205         if (mlocked)
2206                 lru_add_drain();
2207
2208         /* prevent PageLRU to go away from under us, and freeze lru stats */
2209         spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags);
2210
2211         if (mapping) {
2212                 void **pslot;
2213
2214                 spin_lock(&mapping->tree_lock);
2215                 pslot = radix_tree_lookup_slot(&mapping->page_tree,
2216                                 page_index(head));
2217                 /*
2218                  * Check if the head page is present in radix tree.
2219                  * We assume all tail are present too, if head is there.
2220                  */
2221                 if (radix_tree_deref_slot_protected(pslot,
2222                                         &mapping->tree_lock) != head)
2223                         goto fail;
2224         }
2225
2226         /* Prevent deferred_split_scan() touching ->_refcount */
2227         spin_lock(&pgdata->split_queue_lock);
2228         if (page_ref_freeze(head, 1 + extra_pins)) {
2229                 if (!list_empty(page_deferred_list(head))) {
2230                         pgdata->split_queue_len--;
2231                         list_del(page_deferred_list(head));
2232                 }
2233                 if (mapping)
2234                         __dec_node_page_state(page, NR_SHMEM_THPS);
2235                 spin_unlock(&pgdata->split_queue_lock);
2236                 __split_huge_page(page, list, end, flags);
2237                 ret = 0;
2238         } else {
2239                 spin_unlock(&pgdata->split_queue_lock);
2240 fail:
2241                 if (mapping)
2242                         spin_unlock(&mapping->tree_lock);
2243                 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
2244                 remap_page(head);
2245                 ret = -EBUSY;
2246         }
2247
2248 out_unlock:
2249         if (anon_vma) {
2250                 anon_vma_unlock_write(anon_vma);
2251                 put_anon_vma(anon_vma);
2252         }
2253         if (mapping)
2254                 i_mmap_unlock_read(mapping);
2255 out:
2256         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2257         return ret;
2258 }
2259
2260 void free_transhuge_page(struct page *page)
2261 {
2262         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2263         unsigned long flags;
2264
2265         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2266         if (!list_empty(page_deferred_list(page))) {
2267                 pgdata->split_queue_len--;
2268                 list_del(page_deferred_list(page));
2269         }
2270         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2271         free_compound_page(page);
2272 }
2273
2274 void deferred_split_huge_page(struct page *page)
2275 {
2276         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2277         unsigned long flags;
2278
2279         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2280
2281         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2282         if (list_empty(page_deferred_list(page))) {
2283                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2284                 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2285                 pgdata->split_queue_len++;
2286         }
2287         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2288 }
2289
2290 static unsigned long deferred_split_count(struct shrinker *shrink,
2291                 struct shrink_control *sc)
2292 {
2293         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2294         return ACCESS_ONCE(pgdata->split_queue_len);
2295 }
2296
2297 static unsigned long deferred_split_scan(struct shrinker *shrink,
2298                 struct shrink_control *sc)
2299 {
2300         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2301         unsigned long flags;
2302         LIST_HEAD(list), *pos, *next;
2303         struct page *page;
2304         int split = 0;
2305
2306         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2307         /* Take pin on all head pages to avoid freeing them under us */
2308         list_for_each_safe(pos, next, &pgdata->split_queue) {
2309                 page = list_entry((void *)pos, struct page, mapping);
2310                 page = compound_head(page);
2311                 if (get_page_unless_zero(page)) {
2312                         list_move(page_deferred_list(page), &list);
2313                 } else {
2314                         /* We lost race with put_compound_page() */
2315                         list_del_init(page_deferred_list(page));
2316                         pgdata->split_queue_len--;
2317                 }
2318                 if (!--sc->nr_to_scan)
2319                         break;
2320         }
2321         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2322
2323         list_for_each_safe(pos, next, &list) {
2324                 page = list_entry((void *)pos, struct page, mapping);
2325                 if (!trylock_page(page))
2326                         goto next;
2327                 /* split_huge_page() removes page from list on success */
2328                 if (!split_huge_page(page))
2329                         split++;
2330                 unlock_page(page);
2331 next:
2332                 put_page(page);
2333         }
2334
2335         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2336         list_splice_tail(&list, &pgdata->split_queue);
2337         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2338
2339         /*
2340          * Stop shrinker if we didn't split any page, but the queue is empty.
2341          * This can happen if pages were freed under us.
2342          */
2343         if (!split && list_empty(&pgdata->split_queue))
2344                 return SHRINK_STOP;
2345         return split;
2346 }
2347
2348 static struct shrinker deferred_split_shrinker = {
2349         .count_objects = deferred_split_count,
2350         .scan_objects = deferred_split_scan,
2351         .seeks = DEFAULT_SEEKS,
2352         .flags = SHRINKER_NUMA_AWARE,
2353 };
2354
2355 #ifdef CONFIG_DEBUG_FS
2356 static int split_huge_pages_set(void *data, u64 val)
2357 {
2358         struct zone *zone;
2359         struct page *page;
2360         unsigned long pfn, max_zone_pfn;
2361         unsigned long total = 0, split = 0;
2362
2363         if (val != 1)
2364                 return -EINVAL;
2365
2366         for_each_populated_zone(zone) {
2367                 max_zone_pfn = zone_end_pfn(zone);
2368                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2369                         if (!pfn_valid(pfn))
2370                                 continue;
2371
2372                         page = pfn_to_page(pfn);
2373                         if (!get_page_unless_zero(page))
2374                                 continue;
2375
2376                         if (zone != page_zone(page))
2377                                 goto next;
2378
2379                         if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2380                                 goto next;
2381
2382                         total++;
2383                         lock_page(page);
2384                         if (!split_huge_page(page))
2385                                 split++;
2386                         unlock_page(page);
2387 next:
2388                         put_page(page);
2389                 }
2390         }
2391
2392         pr_info("%lu of %lu THP split\n", split, total);
2393
2394         return 0;
2395 }
2396 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2397                 "%llu\n");
2398
2399 static int __init split_huge_pages_debugfs(void)
2400 {
2401         void *ret;
2402
2403         ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2404                         &split_huge_pages_fops);
2405         if (!ret)
2406                 pr_warn("Failed to create split_huge_pages in debugfs");
2407         return 0;
2408 }
2409 late_initcall(split_huge_pages_debugfs);
2410 #endif