mm/huge_memory.c

   1 /*
   2  *  Copyright (C) 2009  Red Hat, Inc.
   3  *
   4  *  This work is licensed under the terms of the GNU GPL, version 2. See
   5  *  the COPYING file in the top-level directory.
   6  */
   7
   8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   9
  10 #include <linux/mm.h>
  11 #include <linux/sched.h>
  12 #include <linux/highmem.h>
  13 #include <linux/hugetlb.h>
  14 #include <linux/mmu_notifier.h>
  15 #include <linux/rmap.h>
  16 #include <linux/swap.h>
  17 #include <linux/shrinker.h>
  18 #include <linux/mm_inline.h>
  19 #include <linux/swapops.h>
  20 #include <linux/dax.h>
  21 #include <linux/khugepaged.h>
  22 #include <linux/freezer.h>
  23 #include <linux/pfn_t.h>
  24 #include <linux/mman.h>
  25 #include <linux/memremap.h>
  26 #include <linux/pagemap.h>
  27 #include <linux/debugfs.h>
  28 #include <linux/migrate.h>
  29 #include <linux/hashtable.h>
  30 #include <linux/userfaultfd_k.h>
  31 #include <linux/page_idle.h>
  32 #include <linux/shmem_fs.h>
  33
  34 #include <asm/tlb.h>
  35 #include <asm/pgalloc.h>
  36 #include "internal.h"
  37
  38 /*
  39  * By default transparent hugepage support is disabled in order that avoid
  40  * to risk increase the memory footprint of applications without a guaranteed
  41  * benefit. When transparent hugepage support is enabled, is for all mappings,
  42  * and khugepaged scans all mappings.
  43  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  44  * for all hugepage allocations.
  45  */
  46 unsigned long transparent_hugepage_flags __read_mostly =
  47 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  48         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  49 #endif
  50 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  51         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  52 #endif
  53         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
  54         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  55         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  56
  57 static struct shrinker deferred_split_shrinker;
  58
  59 static atomic_t huge_zero_refcount;
  60 struct page *huge_zero_page __read_mostly;
  61
  62 static struct page *get_huge_zero_page(void)
  63 {
  64         struct page *zero_page;
  65 retry:
  66         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
  67                 return READ_ONCE(huge_zero_page);
  68
  69         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
  70                         HPAGE_PMD_ORDER);
  71         if (!zero_page) {
  72                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
  73                 return NULL;
  74         }
  75         count_vm_event(THP_ZERO_PAGE_ALLOC);
  76         preempt_disable();
  77         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
  78                 preempt_enable();
  79                 __free_pages(zero_page, compound_order(zero_page));
  80                 goto retry;
  81         }
  82
  83         /* We take additional reference here. It will be put back by shrinker */
  84         atomic_set(&huge_zero_refcount, 2);
  85         preempt_enable();
  86         return READ_ONCE(huge_zero_page);
  87 }
  88
  89 static void put_huge_zero_page(void)
  90 {
  91         /*
  92          * Counter should never go to zero here. Only shrinker can put
  93          * last reference.
  94          */
  95         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
  96 }
  97
  98 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
  99 {
 100         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 101                 return READ_ONCE(huge_zero_page);
 102
 103         if (!get_huge_zero_page())
 104                 return NULL;
 105
 106         if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 107                 put_huge_zero_page();
 108
 109         return READ_ONCE(huge_zero_page);
 110 }
 111
 112 void mm_put_huge_zero_page(struct mm_struct *mm)
 113 {
 114         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 115                 put_huge_zero_page();
 116 }
 117
 118 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 119                                         struct shrink_control *sc)
 120 {
 121         /* we can free zero page only if last reference remains */
 122         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 123 }
 124
 125 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 126                                        struct shrink_control *sc)
 127 {
 128         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 129                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 130                 BUG_ON(zero_page == NULL);
 131                 __free_pages(zero_page, compound_order(zero_page));
 132                 return HPAGE_PMD_NR;
 133         }
 134
 135         return 0;
 136 }
 137
 138 static struct shrinker huge_zero_page_shrinker = {
 139         .count_objects = shrink_huge_zero_page_count,
 140         .scan_objects = shrink_huge_zero_page_scan,
 141         .seeks = DEFAULT_SEEKS,
 142 };
 143
 144 #ifdef CONFIG_SYSFS
 145
 146 static ssize_t triple_flag_store(struct kobject *kobj,
 147                                  struct kobj_attribute *attr,
 148                                  const char *buf, size_t count,
 149                                  enum transparent_hugepage_flag enabled,
 150                                  enum transparent_hugepage_flag deferred,
 151                                  enum transparent_hugepage_flag req_madv)
 152 {
 153         if (!memcmp("defer", buf,
 154                     min(sizeof("defer")-1, count))) {
 155                 if (enabled == deferred)
 156                         return -EINVAL;
 157                 clear_bit(enabled, &transparent_hugepage_flags);
 158                 clear_bit(req_madv, &transparent_hugepage_flags);
 159                 set_bit(deferred, &transparent_hugepage_flags);
 160         } else if (!memcmp("always", buf,
 161                     min(sizeof("always")-1, count))) {
 162                 clear_bit(deferred, &transparent_hugepage_flags);
 163                 clear_bit(req_madv, &transparent_hugepage_flags);
 164                 set_bit(enabled, &transparent_hugepage_flags);
 165         } else if (!memcmp("madvise", buf,
 166                            min(sizeof("madvise")-1, count))) {
 167                 clear_bit(enabled, &transparent_hugepage_flags);
 168                 clear_bit(deferred, &transparent_hugepage_flags);
 169                 set_bit(req_madv, &transparent_hugepage_flags);
 170         } else if (!memcmp("never", buf,
 171                            min(sizeof("never")-1, count))) {
 172                 clear_bit(enabled, &transparent_hugepage_flags);
 173                 clear_bit(req_madv, &transparent_hugepage_flags);
 174                 clear_bit(deferred, &transparent_hugepage_flags);
 175         } else
 176                 return -EINVAL;
 177
 178         return count;
 179 }
 180
 181 static ssize_t enabled_show(struct kobject *kobj,
 182                             struct kobj_attribute *attr, char *buf)
 183 {
 184         if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
 185                 return sprintf(buf, "[always] madvise never\n");
 186         else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
 187                 return sprintf(buf, "always [madvise] never\n");
 188         else
 189                 return sprintf(buf, "always madvise [never]\n");
 190 }
 191
 192 static ssize_t enabled_store(struct kobject *kobj,
 193                              struct kobj_attribute *attr,
 194                              const char *buf, size_t count)
 195 {
 196         ssize_t ret;
 197
 198         ret = triple_flag_store(kobj, attr, buf, count,
 199                                 TRANSPARENT_HUGEPAGE_FLAG,
 200                                 TRANSPARENT_HUGEPAGE_FLAG,
 201                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 202
 203         if (ret > 0) {
 204                 int err = start_stop_khugepaged();
 205                 if (err)
 206                         ret = err;
 207         }
 208
 209         return ret;
 210 }
 211 static struct kobj_attribute enabled_attr =
 212         __ATTR(enabled, 0644, enabled_show, enabled_store);
 213
 214 ssize_t single_hugepage_flag_show(struct kobject *kobj,
 215                                 struct kobj_attribute *attr, char *buf,
 216                                 enum transparent_hugepage_flag flag)
 217 {
 218         return sprintf(buf, "%d\n",
 219                        !!test_bit(flag, &transparent_hugepage_flags));
 220 }
 221
 222 ssize_t single_hugepage_flag_store(struct kobject *kobj,
 223                                  struct kobj_attribute *attr,
 224                                  const char *buf, size_t count,
 225                                  enum transparent_hugepage_flag flag)
 226 {
 227         unsigned long value;
 228         int ret;
 229
 230         ret = kstrtoul(buf, 10, &value);
 231         if (ret < 0)
 232                 return ret;
 233         if (value > 1)
 234                 return -EINVAL;
 235
 236         if (value)
 237                 set_bit(flag, &transparent_hugepage_flags);
 238         else
 239                 clear_bit(flag, &transparent_hugepage_flags);
 240
 241         return count;
 242 }
 243
 244 /*
 245  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 246  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 247  * memory just to allocate one more hugepage.
 248  */
 249 static ssize_t defrag_show(struct kobject *kobj,
 250                            struct kobj_attribute *attr, char *buf)
 251 {
 252         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 253                 return sprintf(buf, "[always] defer madvise never\n");
 254         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 255                 return sprintf(buf, "always [defer] madvise never\n");
 256         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 257                 return sprintf(buf, "always defer [madvise] never\n");
 258         else
 259                 return sprintf(buf, "always defer madvise [never]\n");
 260
 261 }
 262 static ssize_t defrag_store(struct kobject *kobj,
 263                             struct kobj_attribute *attr,
 264                             const char *buf, size_t count)
 265 {
 266         return triple_flag_store(kobj, attr, buf, count,
 267                                  TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
 268                                  TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
 269                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 270 }
 271 static struct kobj_attribute defrag_attr =
 272         __ATTR(defrag, 0644, defrag_show, defrag_store);
 273
 274 static ssize_t use_zero_page_show(struct kobject *kobj,
 275                 struct kobj_attribute *attr, char *buf)
 276 {
 277         return single_hugepage_flag_show(kobj, attr, buf,
 278                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 279 }
 280 static ssize_t use_zero_page_store(struct kobject *kobj,
 281                 struct kobj_attribute *attr, const char *buf, size_t count)
 282 {
 283         return single_hugepage_flag_store(kobj, attr, buf, count,
 284                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 285 }
 286 static struct kobj_attribute use_zero_page_attr =
 287         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 288 #ifdef CONFIG_DEBUG_VM
 289 static ssize_t debug_cow_show(struct kobject *kobj,
 290                                 struct kobj_attribute *attr, char *buf)
 291 {
 292         return single_hugepage_flag_show(kobj, attr, buf,
 293                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 294 }
 295 static ssize_t debug_cow_store(struct kobject *kobj,
 296                                struct kobj_attribute *attr,
 297                                const char *buf, size_t count)
 298 {
 299         return single_hugepage_flag_store(kobj, attr, buf, count,
 300                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 301 }
 302 static struct kobj_attribute debug_cow_attr =
 303         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 304 #endif /* CONFIG_DEBUG_VM */
 305
 306 static struct attribute *hugepage_attr[] = {
 307         &enabled_attr.attr,
 308         &defrag_attr.attr,
 309         &use_zero_page_attr.attr,
 310 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
 311         &shmem_enabled_attr.attr,
 312 #endif
 313 #ifdef CONFIG_DEBUG_VM
 314         &debug_cow_attr.attr,
 315 #endif
 316         NULL,
 317 };
 318
 319 static struct attribute_group hugepage_attr_group = {
 320         .attrs = hugepage_attr,
 321 };
 322
 323 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 324 {
 325         int err;
 326
 327         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 328         if (unlikely(!*hugepage_kobj)) {
 329                 pr_err("failed to create transparent hugepage kobject\n");
 330                 return -ENOMEM;
 331         }
 332
 333         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 334         if (err) {
 335                 pr_err("failed to register transparent hugepage group\n");
 336                 goto delete_obj;
 337         }
 338
 339         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 340         if (err) {
 341                 pr_err("failed to register transparent hugepage group\n");
 342                 goto remove_hp_group;
 343         }
 344
 345         return 0;
 346
 347 remove_hp_group:
 348         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 349 delete_obj:
 350         kobject_put(*hugepage_kobj);
 351         return err;
 352 }
 353
 354 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 355 {
 356         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 357         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 358         kobject_put(hugepage_kobj);
 359 }
 360 #else
 361 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 362 {
 363         return 0;
 364 }
 365
 366 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 367 {
 368 }
 369 #endif /* CONFIG_SYSFS */
 370
 371 static int __init hugepage_init(void)
 372 {
 373         int err;
 374         struct kobject *hugepage_kobj;
 375
 376         if (!has_transparent_hugepage()) {
 377                 transparent_hugepage_flags = 0;
 378                 return -EINVAL;
 379         }
 380
 381         /*
 382          * hugepages can't be allocated by the buddy allocator
 383          */
 384         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
 385         /*
 386          * we use page->mapping and page->index in second tail page
 387          * as list_head: assuming THP order >= 2
 388          */
 389         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 390
 391         err = hugepage_init_sysfs(&hugepage_kobj);
 392         if (err)
 393                 goto err_sysfs;
 394
 395         err = khugepaged_init();
 396         if (err)
 397                 goto err_slab;
 398
 399         err = register_shrinker(&huge_zero_page_shrinker);
 400         if (err)
 401                 goto err_hzp_shrinker;
 402         err = register_shrinker(&deferred_split_shrinker);
 403         if (err)
 404                 goto err_split_shrinker;
 405
 406         /*
 407          * By default disable transparent hugepages on smaller systems,
 408          * where the extra memory used could hurt more than TLB overhead
 409          * is likely to save.  The admin can still enable it through /sys.
 410          */
 411         if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
 412                 transparent_hugepage_flags = 0;
 413                 return 0;
 414         }
 415
 416         err = start_stop_khugepaged();
 417         if (err)
 418                 goto err_khugepaged;
 419
 420         return 0;
 421 err_khugepaged:
 422         unregister_shrinker(&deferred_split_shrinker);
 423 err_split_shrinker:
 424         unregister_shrinker(&huge_zero_page_shrinker);
 425 err_hzp_shrinker:
 426         khugepaged_destroy();
 427 err_slab:
 428         hugepage_exit_sysfs(hugepage_kobj);
 429 err_sysfs:
 430         return err;
 431 }
 432 subsys_initcall(hugepage_init);
 433
 434 static int __init setup_transparent_hugepage(char *str)
 435 {
 436         int ret = 0;
 437         if (!str)
 438                 goto out;
 439         if (!strcmp(str, "always")) {
 440                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 441                         &transparent_hugepage_flags);
 442                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 443                           &transparent_hugepage_flags);
 444                 ret = 1;
 445         } else if (!strcmp(str, "madvise")) {
 446                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 447                           &transparent_hugepage_flags);
 448                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 449                         &transparent_hugepage_flags);
 450                 ret = 1;
 451         } else if (!strcmp(str, "never")) {
 452                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 453                           &transparent_hugepage_flags);
 454                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 455                           &transparent_hugepage_flags);
 456                 ret = 1;
 457         }
 458 out:
 459         if (!ret)
 460                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 461         return ret;
 462 }
 463 __setup("transparent_hugepage=", setup_transparent_hugepage);
 464
 465 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 466 {
 467         if (likely(vma->vm_flags & VM_WRITE))
 468                 pmd = pmd_mkwrite(pmd);
 469         return pmd;
 470 }
 471
 472 static inline struct list_head *page_deferred_list(struct page *page)
 473 {
 474         /*
 475          * ->lru in the tail pages is occupied by compound_head.
 476          * Let's use ->mapping + ->index in the second tail page as list_head.
 477          */
 478         return (struct list_head *)&page[2].mapping;
 479 }
 480
 481 void prep_transhuge_page(struct page *page)
 482 {
 483         /*
 484          * we use page->mapping and page->indexlru in second tail page
 485          * as list_head: assuming THP order >= 2
 486          */
 487
 488         INIT_LIST_HEAD(page_deferred_list(page));
 489         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 490 }
 491
 492 unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len,
 493                 loff_t off, unsigned long flags, unsigned long size)
 494 {
 495         unsigned long addr;
 496         loff_t off_end = off + len;
 497         loff_t off_align = round_up(off, size);
 498         unsigned long len_pad;
 499
 500         if (off_end <= off_align || (off_end - off_align) < size)
 501                 return 0;
 502
 503         len_pad = len + size;
 504         if (len_pad < len || (off + len_pad) < off)
 505                 return 0;
 506
 507         addr = current->mm->get_unmapped_area(filp, 0, len_pad,
 508                                               off >> PAGE_SHIFT, flags);
 509         if (IS_ERR_VALUE(addr))
 510                 return 0;
 511
 512         addr += (off - addr) & (size - 1);
 513         return addr;
 514 }
 515
 516 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 517                 unsigned long len, unsigned long pgoff, unsigned long flags)
 518 {
 519         loff_t off = (loff_t)pgoff << PAGE_SHIFT;
 520
 521         if (addr)
 522                 goto out;
 523         if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
 524                 goto out;
 525
 526         addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE);
 527         if (addr)
 528                 return addr;
 529
 530  out:
 531         return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
 532 }
 533 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
 534
 535 static int __do_huge_pmd_anonymous_page(struct fault_env *fe, struct page *page,
 536                 gfp_t gfp)
 537 {
 538         struct vm_area_struct *vma = fe->vma;
 539         struct mem_cgroup *memcg;
 540         pgtable_t pgtable;
 541         unsigned long haddr = fe->address & HPAGE_PMD_MASK;
 542
 543         VM_BUG_ON_PAGE(!PageCompound(page), page);
 544
 545         if (mem_cgroup_try_charge(page, vma->vm_mm, gfp, &memcg, true)) {
 546                 put_page(page);
 547                 count_vm_event(THP_FAULT_FALLBACK);
 548                 return VM_FAULT_FALLBACK;
 549         }
 550
 551         pgtable = pte_alloc_one(vma->vm_mm, haddr);
 552         if (unlikely(!pgtable)) {
 553                 mem_cgroup_cancel_charge(page, memcg, true);
 554                 put_page(page);
 555                 return VM_FAULT_OOM;
 556         }
 557
 558         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 559         /*
 560          * The memory barrier inside __SetPageUptodate makes sure that
 561          * clear_huge_page writes become visible before the set_pmd_at()
 562          * write.
 563          */
 564         __SetPageUptodate(page);
 565
 566         fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
 567         if (unlikely(!pmd_none(*fe->pmd))) {
 568                 spin_unlock(fe->ptl);
 569                 mem_cgroup_cancel_charge(page, memcg, true);
 570                 put_page(page);
 571                 pte_free(vma->vm_mm, pgtable);
 572         } else {
 573                 pmd_t entry;
 574
 575                 /* Deliver the page fault to userland */
 576                 if (userfaultfd_missing(vma)) {
 577                         int ret;
 578
 579                         spin_unlock(fe->ptl);
 580                         mem_cgroup_cancel_charge(page, memcg, true);
 581                         put_page(page);
 582                         pte_free(vma->vm_mm, pgtable);
 583                         ret = handle_userfault(fe, VM_UFFD_MISSING);
 584                         VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 585                         return ret;
 586                 }
 587
 588                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 589                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 590                 page_add_new_anon_rmap(page, vma, haddr, true);
 591                 mem_cgroup_commit_charge(page, memcg, false, true);
 592                 lru_cache_add_active_or_unevictable(page, vma);
 593                 pgtable_trans_huge_deposit(vma->vm_mm, fe->pmd, pgtable);
 594                 set_pmd_at(vma->vm_mm, haddr, fe->pmd, entry);
 595                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 596                 atomic_long_inc(&vma->vm_mm->nr_ptes);
 597                 spin_unlock(fe->ptl);
 598                 count_vm_event(THP_FAULT_ALLOC);
 599         }
 600
 601         return 0;
 602 }
 603
 604 /*
 605  * If THP defrag is set to always then directly reclaim/compact as necessary
 606  * If set to defer then do only background reclaim/compact and defer to khugepaged
 607  * If set to madvise and the VMA is flagged then directly reclaim/compact
 608  * When direct reclaim/compact is allowed, don't retry except for flagged VMA's
 609  */
 610 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
 611 {
 612         bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
 613
 614         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
 615                                 &transparent_hugepage_flags) && vma_madvised)
 616                 return GFP_TRANSHUGE;
 617         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
 618                                                 &transparent_hugepage_flags))
 619                 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
 620         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
 621                                                 &transparent_hugepage_flags))
 622                 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
 623
 624         return GFP_TRANSHUGE_LIGHT;
 625 }
 626
 627 /* Caller must hold page table lock. */
 628 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 629                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 630                 struct page *zero_page)
 631 {
 632         pmd_t entry;
 633         if (!pmd_none(*pmd))
 634                 return false;
 635         entry = mk_pmd(zero_page, vma->vm_page_prot);
 636         entry = pmd_mkhuge(entry);
 637         if (pgtable)
 638                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 639         set_pmd_at(mm, haddr, pmd, entry);
 640         atomic_long_inc(&mm->nr_ptes);
 641         return true;
 642 }
 643
 644 int do_huge_pmd_anonymous_page(struct fault_env *fe)
 645 {
 646         struct vm_area_struct *vma = fe->vma;
 647         gfp_t gfp;
 648         struct page *page;
 649         unsigned long haddr = fe->address & HPAGE_PMD_MASK;
 650
 651         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 652                 return VM_FAULT_FALLBACK;
 653         if (unlikely(anon_vma_prepare(vma)))
 654                 return VM_FAULT_OOM;
 655         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 656                 return VM_FAULT_OOM;
 657         if (!(fe->flags & FAULT_FLAG_WRITE) &&
 658                         !mm_forbids_zeropage(vma->vm_mm) &&
 659                         transparent_hugepage_use_zero_page()) {
 660                 pgtable_t pgtable;
 661                 struct page *zero_page;
 662                 bool set;
 663                 int ret;
 664                 pgtable = pte_alloc_one(vma->vm_mm, haddr);
 665                 if (unlikely(!pgtable))
 666                         return VM_FAULT_OOM;
 667                 zero_page = mm_get_huge_zero_page(vma->vm_mm);
 668                 if (unlikely(!zero_page)) {
 669                         pte_free(vma->vm_mm, pgtable);
 670                         count_vm_event(THP_FAULT_FALLBACK);
 671                         return VM_FAULT_FALLBACK;
 672                 }
 673                 fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
 674                 ret = 0;
 675                 set = false;
 676                 if (pmd_none(*fe->pmd)) {
 677                         if (userfaultfd_missing(vma)) {
 678                                 spin_unlock(fe->ptl);
 679                                 ret = handle_userfault(fe, VM_UFFD_MISSING);
 680                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 681                         } else {
 682                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
 683                                                    haddr, fe->pmd, zero_page);
 684                                 spin_unlock(fe->ptl);
 685                                 set = true;
 686                         }
 687                 } else
 688                         spin_unlock(fe->ptl);
 689                 if (!set)
 690                         pte_free(vma->vm_mm, pgtable);
 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)
 749 {
 750         pmd_t _pmd;
 751
 752         /*
 753          * We should set the dirty bit only for FOLL_WRITE but for now
 754          * the dirty bit in the pmd is meaningless.  And if the dirty
 755          * bit will become meaningful and we'll only set it with
 756          * FOLL_WRITE, an atomic set_bit will be required on the pmd to
 757          * set the young bit, instead of the current set_pmd_at.
 758          */
 759         _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
 760         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
 761                                 pmd, _pmd,  1))
 762                 update_mmu_cache_pmd(vma, addr, pmd);
 763 }
 764
 765 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 766                 pmd_t *pmd, int flags)
 767 {
 768         unsigned long pfn = pmd_pfn(*pmd);
 769         struct mm_struct *mm = vma->vm_mm;
 770         struct dev_pagemap *pgmap;
 771         struct page *page;
 772
 773         assert_spin_locked(pmd_lockptr(mm, pmd));
 774
 775         if (flags & FOLL_WRITE && !pmd_write(*pmd))
 776                 return NULL;
 777
 778         if (pmd_present(*pmd) && pmd_devmap(*pmd))
 779                 /* pass */;
 780         else
 781                 return NULL;
 782
 783         if (flags & FOLL_TOUCH)
 784                 touch_pmd(vma, addr, pmd);
 785
 786         /*
 787          * device mapped pages can only be returned if the
 788          * caller will manage the page reference count.
 789          */
 790         if (!(flags & FOLL_GET))
 791                 return ERR_PTR(-EEXIST);
 792
 793         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
 794         pgmap = get_dev_pagemap(pfn, NULL);
 795         if (!pgmap)
 796                 return ERR_PTR(-EFAULT);
 797         page = pfn_to_page(pfn);
 798         get_page(page);
 799         put_dev_pagemap(pgmap);
 800
 801         return page;
 802 }
 803
 804 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 805                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 806                   struct vm_area_struct *vma)
 807 {
 808         spinlock_t *dst_ptl, *src_ptl;
 809         struct page *src_page;
 810         pmd_t pmd;
 811         pgtable_t pgtable = NULL;
 812         int ret = -ENOMEM;
 813
 814         /* Skip if can be re-fill on fault */
 815         if (!vma_is_anonymous(vma))
 816                 return 0;
 817
 818         pgtable = pte_alloc_one(dst_mm, addr);
 819         if (unlikely(!pgtable))
 820                 goto out;
 821
 822         dst_ptl = pmd_lock(dst_mm, dst_pmd);
 823         src_ptl = pmd_lockptr(src_mm, src_pmd);
 824         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 825
 826         ret = -EAGAIN;
 827         pmd = *src_pmd;
 828         if (unlikely(!pmd_trans_huge(pmd))) {
 829                 pte_free(dst_mm, pgtable);
 830                 goto out_unlock;
 831         }
 832         /*
 833          * When page table lock is held, the huge zero pmd should not be
 834          * under splitting since we don't split the page itself, only pmd to
 835          * a page table.
 836          */
 837         if (is_huge_zero_pmd(pmd)) {
 838                 struct page *zero_page;
 839                 /*
 840                  * get_huge_zero_page() will never allocate a new page here,
 841                  * since we already have a zero page to copy. It just takes a
 842                  * reference.
 843                  */
 844                 zero_page = mm_get_huge_zero_page(dst_mm);
 845                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
 846                                 zero_page);
 847                 ret = 0;
 848                 goto out_unlock;
 849         }
 850
 851         src_page = pmd_page(pmd);
 852         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
 853         get_page(src_page);
 854         page_dup_rmap(src_page, true);
 855         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 856         atomic_long_inc(&dst_mm->nr_ptes);
 857         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
 858
 859         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 860         pmd = pmd_mkold(pmd_wrprotect(pmd));
 861         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 862
 863         ret = 0;
 864 out_unlock:
 865         spin_unlock(src_ptl);
 866         spin_unlock(dst_ptl);
 867 out:
 868         return ret;
 869 }
 870
 871 void huge_pmd_set_accessed(struct fault_env *fe, pmd_t orig_pmd)
 872 {
 873         pmd_t entry;
 874         unsigned long haddr;
 875
 876         fe->ptl = pmd_lock(fe->vma->vm_mm, fe->pmd);
 877         if (unlikely(!pmd_same(*fe->pmd, orig_pmd)))
 878                 goto unlock;
 879
 880         entry = pmd_mkyoung(orig_pmd);
 881         haddr = fe->address & HPAGE_PMD_MASK;
 882         if (pmdp_set_access_flags(fe->vma, haddr, fe->pmd, entry,
 883                                 fe->flags & FAULT_FLAG_WRITE))
 884                 update_mmu_cache_pmd(fe->vma, fe->address, fe->pmd);
 885
 886 unlock:
 887         spin_unlock(fe->ptl);
 888 }
 889
 890 static int do_huge_pmd_wp_page_fallback(struct fault_env *fe, pmd_t orig_pmd,
 891                 struct page *page)
 892 {
 893         struct vm_area_struct *vma = fe->vma;
 894         unsigned long haddr = fe->address & HPAGE_PMD_MASK;
 895         struct mem_cgroup *memcg;
 896         pgtable_t pgtable;
 897         pmd_t _pmd;
 898         int ret = 0, i;
 899         struct page **pages;
 900         unsigned long mmun_start;       /* For mmu_notifiers */
 901         unsigned long mmun_end;         /* For mmu_notifiers */
 902
 903         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
 904                         GFP_KERNEL);
 905         if (unlikely(!pages)) {
 906                 ret |= VM_FAULT_OOM;
 907                 goto out;
 908         }
 909
 910         for (i = 0; i < HPAGE_PMD_NR; i++) {
 911                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
 912                                                __GFP_OTHER_NODE, vma,
 913                                                fe->address, page_to_nid(page));
 914                 if (unlikely(!pages[i] ||
 915                              mem_cgroup_try_charge(pages[i], vma->vm_mm,
 916                                      GFP_KERNEL, &memcg, false))) {
 917                         if (pages[i])
 918                                 put_page(pages[i]);
 919                         while (--i >= 0) {
 920                                 memcg = (void *)page_private(pages[i]);
 921                                 set_page_private(pages[i], 0);
 922                                 mem_cgroup_cancel_charge(pages[i], memcg,
 923                                                 false);
 924                                 put_page(pages[i]);
 925                         }
 926                         kfree(pages);
 927                         ret |= VM_FAULT_OOM;
 928                         goto out;
 929                 }
 930                 set_page_private(pages[i], (unsigned long)memcg);
 931         }
 932
 933         for (i = 0; i < HPAGE_PMD_NR; i++) {
 934                 copy_user_highpage(pages[i], page + i,
 935                                    haddr + PAGE_SIZE * i, vma);
 936                 __SetPageUptodate(pages[i]);
 937                 cond_resched();
 938         }
 939
 940         mmun_start = haddr;
 941         mmun_end   = haddr + HPAGE_PMD_SIZE;
 942         mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
 943
 944         fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
 945         if (unlikely(!pmd_same(*fe->pmd, orig_pmd)))
 946                 goto out_free_pages;
 947         VM_BUG_ON_PAGE(!PageHead(page), page);
 948
 949         pmdp_huge_clear_flush_notify(vma, haddr, fe->pmd);
 950         /* leave pmd empty until pte is filled */
 951
 952         pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, fe->pmd);
 953         pmd_populate(vma->vm_mm, &_pmd, pgtable);
 954
 955         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
 956                 pte_t entry;
 957                 entry = mk_pte(pages[i], vma->vm_page_prot);
 958                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
 959                 memcg = (void *)page_private(pages[i]);
 960                 set_page_private(pages[i], 0);
 961                 page_add_new_anon_rmap(pages[i], fe->vma, haddr, false);
 962                 mem_cgroup_commit_charge(pages[i], memcg, false, false);
 963                 lru_cache_add_active_or_unevictable(pages[i], vma);
 964                 fe->pte = pte_offset_map(&_pmd, haddr);
 965                 VM_BUG_ON(!pte_none(*fe->pte));
 966                 set_pte_at(vma->vm_mm, haddr, fe->pte, entry);
 967                 pte_unmap(fe->pte);
 968         }
 969         kfree(pages);
 970
 971         smp_wmb(); /* make pte visible before pmd */
 972         pmd_populate(vma->vm_mm, fe->pmd, pgtable);
 973         page_remove_rmap(page, true);
 974         spin_unlock(fe->ptl);
 975
 976         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
 977
 978         ret |= VM_FAULT_WRITE;
 979         put_page(page);
 980
 981 out:
 982         return ret;
 983
 984 out_free_pages:
 985         spin_unlock(fe->ptl);
 986         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
 987         for (i = 0; i < HPAGE_PMD_NR; i++) {
 988                 memcg = (void *)page_private(pages[i]);
 989                 set_page_private(pages[i], 0);
 990                 mem_cgroup_cancel_charge(pages[i], memcg, false);
 991                 put_page(pages[i]);
 992         }
 993         kfree(pages);
 994         goto out;
 995 }
 996
 997 int do_huge_pmd_wp_page(struct fault_env *fe, pmd_t orig_pmd)
 998 {
 999         struct vm_area_struct *vma = fe->vma;
1000         struct page *page = NULL, *new_page;
1001         struct mem_cgroup *memcg;
1002         unsigned long haddr = fe->address & HPAGE_PMD_MASK;
1003         unsigned long mmun_start;       /* For mmu_notifiers */
1004         unsigned long mmun_end;         /* For mmu_notifiers */
1005         gfp_t huge_gfp;                 /* for allocation and charge */
1006         int ret = 0;
1007
1008         fe->ptl = pmd_lockptr(vma->vm_mm, fe->pmd);
1009         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1010         if (is_huge_zero_pmd(orig_pmd))
1011                 goto alloc;
1012         spin_lock(fe->ptl);
1013         if (unlikely(!pmd_same(*fe->pmd, orig_pmd)))
1014                 goto out_unlock;
1015
1016         page = pmd_page(orig_pmd);
1017         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1018         /*
1019          * We can only reuse the page if nobody else maps the huge page or it's
1020          * part.
1021          */
1022         if (page_trans_huge_mapcount(page, NULL) == 1) {
1023                 pmd_t entry;
1024                 entry = pmd_mkyoung(orig_pmd);
1025                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1026                 if (pmdp_set_access_flags(vma, haddr, fe->pmd, entry,  1))
1027                         update_mmu_cache_pmd(vma, fe->address, fe->pmd);
1028                 ret |= VM_FAULT_WRITE;
1029                 goto out_unlock;
1030         }
1031         get_page(page);
1032         spin_unlock(fe->ptl);
1033 alloc:
1034         if (transparent_hugepage_enabled(vma) &&
1035             !transparent_hugepage_debug_cow()) {
1036                 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1037                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1038         } else
1039                 new_page = NULL;
1040
1041         if (likely(new_page)) {
1042                 prep_transhuge_page(new_page);
1043         } else {
1044                 if (!page) {
1045                         split_huge_pmd(vma, fe->pmd, fe->address);
1046                         ret |= VM_FAULT_FALLBACK;
1047                 } else {
1048                         ret = do_huge_pmd_wp_page_fallback(fe, orig_pmd, page);
1049                         if (ret & VM_FAULT_OOM) {
1050                                 split_huge_pmd(vma, fe->pmd, fe->address);
1051                                 ret |= VM_FAULT_FALLBACK;
1052                         }
1053                         put_page(page);
1054                 }
1055                 count_vm_event(THP_FAULT_FALLBACK);
1056                 goto out;
1057         }
1058
1059         if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm,
1060                                         huge_gfp, &memcg, true))) {
1061                 put_page(new_page);
1062                 split_huge_pmd(vma, fe->pmd, fe->address);
1063                 if (page)
1064                         put_page(page);
1065                 ret |= VM_FAULT_FALLBACK;
1066                 count_vm_event(THP_FAULT_FALLBACK);
1067                 goto out;
1068         }
1069
1070         count_vm_event(THP_FAULT_ALLOC);
1071
1072         if (!page)
1073                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1074         else
1075                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1076         __SetPageUptodate(new_page);
1077
1078         mmun_start = haddr;
1079         mmun_end   = haddr + HPAGE_PMD_SIZE;
1080         mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
1081
1082         spin_lock(fe->ptl);
1083         if (page)
1084                 put_page(page);
1085         if (unlikely(!pmd_same(*fe->pmd, orig_pmd))) {
1086                 spin_unlock(fe->ptl);
1087                 mem_cgroup_cancel_charge(new_page, memcg, true);
1088                 put_page(new_page);
1089                 goto out_mn;
1090         } else {
1091                 pmd_t entry;
1092                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1093                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1094                 pmdp_huge_clear_flush_notify(vma, haddr, fe->pmd);
1095                 page_add_new_anon_rmap(new_page, vma, haddr, true);
1096                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1097                 lru_cache_add_active_or_unevictable(new_page, vma);
1098                 set_pmd_at(vma->vm_mm, haddr, fe->pmd, entry);
1099                 update_mmu_cache_pmd(vma, fe->address, fe->pmd);
1100                 if (!page) {
1101                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1102                 } else {
1103                         VM_BUG_ON_PAGE(!PageHead(page), page);
1104                         page_remove_rmap(page, true);
1105                         put_page(page);
1106                 }
1107                 ret |= VM_FAULT_WRITE;
1108         }
1109         spin_unlock(fe->ptl);
1110 out_mn:
1111         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1112 out:
1113         return ret;
1114 out_unlock:
1115         spin_unlock(fe->ptl);
1116         return ret;
1117 }
1118
1119 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1120                                    unsigned long addr,
1121                                    pmd_t *pmd,
1122                                    unsigned int flags)
1123 {
1124         struct mm_struct *mm = vma->vm_mm;
1125         struct page *page = NULL;
1126
1127         assert_spin_locked(pmd_lockptr(mm, pmd));
1128
1129         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1130                 goto out;
1131
1132         /* Avoid dumping huge zero page */
1133         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1134                 return ERR_PTR(-EFAULT);
1135
1136         /* Full NUMA hinting faults to serialise migration in fault paths */
1137         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1138                 goto out;
1139
1140         page = pmd_page(*pmd);
1141         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1142         if (flags & FOLL_TOUCH)
1143                 touch_pmd(vma, addr, pmd);
1144         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1145                 /*
1146                  * We don't mlock() pte-mapped THPs. This way we can avoid
1147                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1148                  *
1149                  * For anon THP:
1150                  *
1151                  * In most cases the pmd is the only mapping of the page as we
1152                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1153                  * writable private mappings in populate_vma_page_range().
1154                  *
1155                  * The only scenario when we have the page shared here is if we
1156                  * mlocking read-only mapping shared over fork(). We skip
1157                  * mlocking such pages.
1158                  *
1159                  * For file THP:
1160                  *
1161                  * We can expect PageDoubleMap() to be stable under page lock:
1162                  * for file pages we set it in page_add_file_rmap(), which
1163                  * requires page to be locked.
1164                  */
1165
1166                 if (PageAnon(page) && compound_mapcount(page) != 1)
1167                         goto skip_mlock;
1168                 if (PageDoubleMap(page) || !page->mapping)
1169                         goto skip_mlock;
1170                 if (!trylock_page(page))
1171                         goto skip_mlock;
1172                 lru_add_drain();
1173                 if (page->mapping && !PageDoubleMap(page))
1174                         mlock_vma_page(page);
1175                 unlock_page(page);
1176         }
1177 skip_mlock:
1178         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1179         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1180         if (flags & FOLL_GET)
1181                 get_page(page);
1182
1183 out:
1184         return page;
1185 }
1186
1187 /* NUMA hinting page fault entry point for trans huge pmds */
1188 int do_huge_pmd_numa_page(struct fault_env *fe, pmd_t pmd)
1189 {
1190         struct vm_area_struct *vma = fe->vma;
1191         struct anon_vma *anon_vma = NULL;
1192         struct page *page;
1193         unsigned long haddr = fe->address & HPAGE_PMD_MASK;
1194         int page_nid = -1, this_nid = numa_node_id();
1195         int target_nid, last_cpupid = -1;
1196         bool page_locked;
1197         bool migrated = false;
1198         bool was_writable;
1199         int flags = 0;
1200
1201         fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
1202         if (unlikely(!pmd_same(pmd, *fe->pmd)))
1203                 goto out_unlock;
1204
1205         /*
1206          * If there are potential migrations, wait for completion and retry
1207          * without disrupting NUMA hinting information. Do not relock and
1208          * check_same as the page may no longer be mapped.
1209          */
1210         if (unlikely(pmd_trans_migrating(*fe->pmd))) {
1211                 page = pmd_page(*fe->pmd);
1212                 spin_unlock(fe->ptl);
1213                 wait_on_page_locked(page);
1214                 goto out;
1215         }
1216
1217         page = pmd_page(pmd);
1218         BUG_ON(is_huge_zero_page(page));
1219         page_nid = page_to_nid(page);
1220         last_cpupid = page_cpupid_last(page);
1221         count_vm_numa_event(NUMA_HINT_FAULTS);
1222         if (page_nid == this_nid) {
1223                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1224                 flags |= TNF_FAULT_LOCAL;
1225         }
1226
1227         /* See similar comment in do_numa_page for explanation */
1228         if (!pmd_write(pmd))
1229                 flags |= TNF_NO_GROUP;
1230
1231         /*
1232          * Acquire the page lock to serialise THP migrations but avoid dropping
1233          * page_table_lock if at all possible
1234          */
1235         page_locked = trylock_page(page);
1236         target_nid = mpol_misplaced(page, vma, haddr);
1237         if (target_nid == -1) {
1238                 /* If the page was locked, there are no parallel migrations */
1239                 if (page_locked)
1240                         goto clear_pmdnuma;
1241         }
1242
1243         /* Migration could have started since the pmd_trans_migrating check */
1244         if (!page_locked) {
1245                 spin_unlock(fe->ptl);
1246                 wait_on_page_locked(page);
1247                 page_nid = -1;
1248                 goto out;
1249         }
1250
1251         /*
1252          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1253          * to serialises splits
1254          */
1255         get_page(page);
1256         spin_unlock(fe->ptl);
1257         anon_vma = page_lock_anon_vma_read(page);
1258
1259         /* Confirm the PMD did not change while page_table_lock was released */
1260         spin_lock(fe->ptl);
1261         if (unlikely(!pmd_same(pmd, *fe->pmd))) {
1262                 unlock_page(page);
1263                 put_page(page);
1264                 page_nid = -1;
1265                 goto out_unlock;
1266         }
1267
1268         /* Bail if we fail to protect against THP splits for any reason */
1269         if (unlikely(!anon_vma)) {
1270                 put_page(page);
1271                 page_nid = -1;
1272                 goto clear_pmdnuma;
1273         }
1274
1275         /*
1276          * Migrate the THP to the requested node, returns with page unlocked
1277          * and access rights restored.
1278          */
1279         spin_unlock(fe->ptl);
1280         migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1281                                 fe->pmd, pmd, fe->address, page, target_nid);
1282         if (migrated) {
1283                 flags |= TNF_MIGRATED;
1284                 page_nid = target_nid;
1285         } else
1286                 flags |= TNF_MIGRATE_FAIL;
1287
1288         goto out;
1289 clear_pmdnuma:
1290         BUG_ON(!PageLocked(page));
1291         was_writable = pmd_write(pmd);
1292         pmd = pmd_modify(pmd, vma->vm_page_prot);
1293         pmd = pmd_mkyoung(pmd);
1294         if (was_writable)
1295                 pmd = pmd_mkwrite(pmd);
1296         set_pmd_at(vma->vm_mm, haddr, fe->pmd, pmd);
1297         update_mmu_cache_pmd(vma, fe->address, fe->pmd);
1298         unlock_page(page);
1299 out_unlock:
1300         spin_unlock(fe->ptl);
1301
1302 out:
1303         if (anon_vma)
1304                 page_unlock_anon_vma_read(anon_vma);
1305
1306         if (page_nid != -1)
1307                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, fe->flags);
1308
1309         return 0;
1310 }
1311
1312 /*
1313  * Return true if we do MADV_FREE successfully on entire pmd page.
1314  * Otherwise, return false.
1315  */
1316 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1317                 pmd_t *pmd, unsigned long addr, unsigned long next)
1318 {
1319         spinlock_t *ptl;
1320         pmd_t orig_pmd;
1321         struct page *page;
1322         struct mm_struct *mm = tlb->mm;
1323         bool ret = false;
1324
1325         ptl = pmd_trans_huge_lock(pmd, vma);
1326         if (!ptl)
1327                 goto out_unlocked;
1328
1329         orig_pmd = *pmd;
1330         if (is_huge_zero_pmd(orig_pmd))
1331                 goto out;
1332
1333         page = pmd_page(orig_pmd);
1334         /*
1335          * If other processes are mapping this page, we couldn't discard
1336          * the page unless they all do MADV_FREE so let's skip the page.
1337          */
1338         if (page_mapcount(page) != 1)
1339                 goto out;
1340
1341         if (!trylock_page(page))
1342                 goto out;
1343
1344         /*
1345          * If user want to discard part-pages of THP, split it so MADV_FREE
1346          * will deactivate only them.
1347          */
1348         if (next - addr != HPAGE_PMD_SIZE) {
1349                 get_page(page);
1350                 spin_unlock(ptl);
1351                 split_huge_page(page);
1352                 put_page(page);
1353                 unlock_page(page);
1354                 goto out_unlocked;
1355         }
1356
1357         if (PageDirty(page))
1358                 ClearPageDirty(page);
1359         unlock_page(page);
1360
1361         if (PageActive(page))
1362                 deactivate_page(page);
1363
1364         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1365                 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1366                         tlb->fullmm);
1367                 orig_pmd = pmd_mkold(orig_pmd);
1368                 orig_pmd = pmd_mkclean(orig_pmd);
1369
1370                 set_pmd_at(mm, addr, pmd, orig_pmd);
1371                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1372         }
1373         ret = true;
1374 out:
1375         spin_unlock(ptl);
1376 out_unlocked:
1377         return ret;
1378 }
1379
1380 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1381                  pmd_t *pmd, unsigned long addr)
1382 {
1383         pmd_t orig_pmd;
1384         spinlock_t *ptl;
1385
1386         ptl = __pmd_trans_huge_lock(pmd, vma);
1387         if (!ptl)
1388                 return 0;
1389         /*
1390          * For architectures like ppc64 we look at deposited pgtable
1391          * when calling pmdp_huge_get_and_clear. So do the
1392          * pgtable_trans_huge_withdraw after finishing pmdp related
1393          * operations.
1394          */
1395         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1396                         tlb->fullmm);
1397         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1398         if (vma_is_dax(vma)) {
1399                 spin_unlock(ptl);
1400                 if (is_huge_zero_pmd(orig_pmd))
1401                         tlb_remove_page(tlb, pmd_page(orig_pmd));
1402         } else if (is_huge_zero_pmd(orig_pmd)) {
1403                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1404                 atomic_long_dec(&tlb->mm->nr_ptes);
1405                 spin_unlock(ptl);
1406                 tlb_remove_page(tlb, pmd_page(orig_pmd));
1407         } else {
1408                 struct page *page = pmd_page(orig_pmd);
1409                 page_remove_rmap(page, true);
1410                 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1411                 VM_BUG_ON_PAGE(!PageHead(page), page);
1412                 if (PageAnon(page)) {
1413                         pgtable_t pgtable;
1414                         pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1415                         pte_free(tlb->mm, pgtable);
1416                         atomic_long_dec(&tlb->mm->nr_ptes);
1417                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1418                 } else {
1419                         add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1420                 }
1421                 spin_unlock(ptl);
1422                 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1423         }
1424         return 1;
1425 }
1426
1427 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1428                   unsigned long new_addr, unsigned long old_end,
1429                   pmd_t *old_pmd, pmd_t *new_pmd, bool *need_flush)
1430 {
1431         spinlock_t *old_ptl, *new_ptl;
1432         pmd_t pmd;
1433         struct mm_struct *mm = vma->vm_mm;
1434         bool force_flush = false;
1435
1436         if ((old_addr & ~HPAGE_PMD_MASK) ||
1437             (new_addr & ~HPAGE_PMD_MASK) ||
1438             old_end - old_addr < HPAGE_PMD_SIZE)
1439                 return false;
1440
1441         /*
1442          * The destination pmd shouldn't be established, free_pgtables()
1443          * should have release it.
1444          */
1445         if (WARN_ON(!pmd_none(*new_pmd))) {
1446                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1447                 return false;
1448         }
1449
1450         /*
1451          * We don't have to worry about the ordering of src and dst
1452          * ptlocks because exclusive mmap_sem prevents deadlock.
1453          */
1454         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1455         if (old_ptl) {
1456                 new_ptl = pmd_lockptr(mm, new_pmd);
1457                 if (new_ptl != old_ptl)
1458                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1459                 if (pmd_present(*old_pmd) && pmd_dirty(*old_pmd))
1460                         force_flush = true;
1461                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1462                 VM_BUG_ON(!pmd_none(*new_pmd));
1463
1464                 if (pmd_move_must_withdraw(new_ptl, old_ptl) &&
1465                                 vma_is_anonymous(vma)) {
1466                         pgtable_t pgtable;
1467                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1468                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1469                 }
1470                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1471                 if (new_ptl != old_ptl)
1472                         spin_unlock(new_ptl);
1473                 if (force_flush)
1474                         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1475                 else
1476                         *need_flush = true;
1477                 spin_unlock(old_ptl);
1478                 return true;
1479         }
1480         return false;
1481 }
1482
1483 /*
1484  * Returns
1485  *  - 0 if PMD could not be locked
1486  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1487  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1488  */
1489 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1490                 unsigned long addr, pgprot_t newprot, int prot_numa)
1491 {
1492         struct mm_struct *mm = vma->vm_mm;
1493         spinlock_t *ptl;
1494         int ret = 0;
1495
1496         ptl = __pmd_trans_huge_lock(pmd, vma);
1497         if (ptl) {
1498                 pmd_t entry;
1499                 bool preserve_write = prot_numa && pmd_write(*pmd);
1500                 ret = 1;
1501
1502                 /*
1503                  * Avoid trapping faults against the zero page. The read-only
1504                  * data is likely to be read-cached on the local CPU and
1505                  * local/remote hits to the zero page are not interesting.
1506                  */
1507                 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1508                         spin_unlock(ptl);
1509                         return ret;
1510                 }
1511
1512                 if (!prot_numa || !pmd_protnone(*pmd)) {
1513                         entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
1514                         entry = pmd_modify(entry, newprot);
1515                         if (preserve_write)
1516                                 entry = pmd_mkwrite(entry);
1517                         ret = HPAGE_PMD_NR;
1518                         set_pmd_at(mm, addr, pmd, entry);
1519                         BUG_ON(vma_is_anonymous(vma) && !preserve_write &&
1520                                         pmd_write(entry));
1521                 }
1522                 spin_unlock(ptl);
1523         }
1524
1525         return ret;
1526 }
1527
1528 /*
1529  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1530  *
1531  * Note that if it returns page table lock pointer, this routine returns without
1532  * unlocking page table lock. So callers must unlock it.
1533  */
1534 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1535 {
1536         spinlock_t *ptl;
1537         ptl = pmd_lock(vma->vm_mm, pmd);
1538         if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd)))
1539                 return ptl;
1540         spin_unlock(ptl);
1541         return NULL;
1542 }
1543
1544 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1545                 unsigned long haddr, pmd_t *pmd)
1546 {
1547         struct mm_struct *mm = vma->vm_mm;
1548         pgtable_t pgtable;
1549         pmd_t _pmd;
1550         int i;
1551
1552         /* leave pmd empty until pte is filled */
1553         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1554
1555         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1556         pmd_populate(mm, &_pmd, pgtable);
1557
1558         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1559                 pte_t *pte, entry;
1560                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
1561                 entry = pte_mkspecial(entry);
1562                 pte = pte_offset_map(&_pmd, haddr);
1563                 VM_BUG_ON(!pte_none(*pte));
1564                 set_pte_at(mm, haddr, pte, entry);
1565                 pte_unmap(pte);
1566         }
1567         smp_wmb(); /* make pte visible before pmd */
1568         pmd_populate(mm, pmd, pgtable);
1569 }
1570
1571 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
1572                 unsigned long haddr, bool freeze)
1573 {
1574         struct mm_struct *mm = vma->vm_mm;
1575         struct page *page;
1576         pgtable_t pgtable;
1577         pmd_t _pmd;
1578         bool young, write, dirty, soft_dirty;
1579         unsigned long addr;
1580         int i;
1581
1582         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
1583         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1584         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
1585         VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd));
1586
1587         count_vm_event(THP_SPLIT_PMD);
1588
1589         if (!vma_is_anonymous(vma)) {
1590                 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1591                 if (vma_is_dax(vma))
1592                         return;
1593                 page = pmd_page(_pmd);
1594                 if (!PageReferenced(page) && pmd_young(_pmd))
1595                         SetPageReferenced(page);
1596                 page_remove_rmap(page, true);
1597                 put_page(page);
1598                 add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1599                 return;
1600         } else if (is_huge_zero_pmd(*pmd)) {
1601                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
1602         }
1603
1604         page = pmd_page(*pmd);
1605         VM_BUG_ON_PAGE(!page_count(page), page);
1606         page_ref_add(page, HPAGE_PMD_NR - 1);
1607         write = pmd_write(*pmd);
1608         young = pmd_young(*pmd);
1609         dirty = pmd_dirty(*pmd);
1610         soft_dirty = pmd_soft_dirty(*pmd);
1611
1612         pmdp_huge_split_prepare(vma, haddr, pmd);
1613         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1614         pmd_populate(mm, &_pmd, pgtable);
1615
1616         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
1617                 pte_t entry, *pte;
1618                 /*
1619                  * Note that NUMA hinting access restrictions are not
1620                  * transferred to avoid any possibility of altering
1621                  * permissions across VMAs.
1622                  */
1623                 if (freeze) {
1624                         swp_entry_t swp_entry;
1625                         swp_entry = make_migration_entry(page + i, write);
1626                         entry = swp_entry_to_pte(swp_entry);
1627                         if (soft_dirty)
1628                                 entry = pte_swp_mksoft_dirty(entry);
1629                 } else {
1630                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
1631                         entry = maybe_mkwrite(entry, vma);
1632                         if (!write)
1633                                 entry = pte_wrprotect(entry);
1634                         if (!young)
1635                                 entry = pte_mkold(entry);
1636                         if (soft_dirty)
1637                                 entry = pte_mksoft_dirty(entry);
1638                 }
1639                 if (dirty)
1640                         SetPageDirty(page + i);
1641                 pte = pte_offset_map(&_pmd, addr);
1642                 BUG_ON(!pte_none(*pte));
1643                 set_pte_at(mm, addr, pte, entry);
1644                 atomic_inc(&page[i]._mapcount);
1645                 pte_unmap(pte);
1646         }
1647
1648         /*
1649          * Set PG_double_map before dropping compound_mapcount to avoid
1650          * false-negative page_mapped().
1651          */
1652         if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
1653                 for (i = 0; i < HPAGE_PMD_NR; i++)
1654                         atomic_inc(&page[i]._mapcount);
1655         }
1656
1657         if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
1658                 /* Last compound_mapcount is gone. */
1659                 __dec_node_page_state(page, NR_ANON_THPS);
1660                 if (TestClearPageDoubleMap(page)) {
1661                         /* No need in mapcount reference anymore */
1662                         for (i = 0; i < HPAGE_PMD_NR; i++)
1663                                 atomic_dec(&page[i]._mapcount);
1664                 }
1665         }
1666
1667         smp_wmb(); /* make pte visible before pmd */
1668         /*
1669          * Up to this point the pmd is present and huge and userland has the
1670          * whole access to the hugepage during the split (which happens in
1671          * place). If we overwrite the pmd with the not-huge version pointing
1672          * to the pte here (which of course we could if all CPUs were bug
1673          * free), userland could trigger a small page size TLB miss on the
1674          * small sized TLB while the hugepage TLB entry is still established in
1675          * the huge TLB. Some CPU doesn't like that.
1676          * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
1677          * 383 on page 93. Intel should be safe but is also warns that it's
1678          * only safe if the permission and cache attributes of the two entries
1679          * loaded in the two TLB is identical (which should be the case here).
1680          * But it is generally safer to never allow small and huge TLB entries
1681          * for the same virtual address to be loaded simultaneously. So instead
1682          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
1683          * current pmd notpresent (atomically because here the pmd_trans_huge
1684          * and pmd_trans_splitting must remain set at all times on the pmd
1685          * until the split is complete for this pmd), then we flush the SMP TLB
1686          * and finally we write the non-huge version of the pmd entry with
1687          * pmd_populate.
1688          */
1689         pmdp_invalidate(vma, haddr, pmd);
1690         pmd_populate(mm, pmd, pgtable);
1691
1692         if (freeze) {
1693                 for (i = 0; i < HPAGE_PMD_NR; i++) {
1694                         page_remove_rmap(page + i, false);
1695                         put_page(page + i);
1696                 }
1697         }
1698 }
1699
1700 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1701                 unsigned long address, bool freeze, struct page *page)
1702 {
1703         spinlock_t *ptl;
1704         struct mm_struct *mm = vma->vm_mm;
1705         unsigned long haddr = address & HPAGE_PMD_MASK;
1706
1707         mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
1708         ptl = pmd_lock(mm, pmd);
1709
1710         /*
1711          * If caller asks to setup a migration entries, we need a page to check
1712          * pmd against. Otherwise we can end up replacing wrong page.
1713          */
1714         VM_BUG_ON(freeze && !page);
1715         if (page && page != pmd_page(*pmd))
1716                 goto out;
1717
1718         if (pmd_trans_huge(*pmd)) {
1719                 page = pmd_page(*pmd);
1720                 if (PageMlocked(page))
1721                         clear_page_mlock(page);
1722         } else if (!pmd_devmap(*pmd))
1723                 goto out;
1724         __split_huge_pmd_locked(vma, pmd, haddr, freeze);
1725 out:
1726         spin_unlock(ptl);
1727         mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
1728 }
1729
1730 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
1731                 bool freeze, struct page *page)
1732 {
1733         pgd_t *pgd;
1734         pud_t *pud;
1735         pmd_t *pmd;
1736
1737         pgd = pgd_offset(vma->vm_mm, address);
1738         if (!pgd_present(*pgd))
1739                 return;
1740
1741         pud = pud_offset(pgd, address);
1742         if (!pud_present(*pud))
1743                 return;
1744
1745         pmd = pmd_offset(pud, address);
1746
1747         __split_huge_pmd(vma, pmd, address, freeze, page);
1748 }
1749
1750 void vma_adjust_trans_huge(struct vm_area_struct *vma,
1751                              unsigned long start,
1752                              unsigned long end,
1753                              long adjust_next)
1754 {
1755         /*
1756          * If the new start address isn't hpage aligned and it could
1757          * previously contain an hugepage: check if we need to split
1758          * an huge pmd.
1759          */
1760         if (start & ~HPAGE_PMD_MASK &&
1761             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
1762             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
1763                 split_huge_pmd_address(vma, start, false, NULL);
1764
1765         /*
1766          * If the new end address isn't hpage aligned and it could
1767          * previously contain an hugepage: check if we need to split
1768          * an huge pmd.
1769          */
1770         if (end & ~HPAGE_PMD_MASK &&
1771             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
1772             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
1773                 split_huge_pmd_address(vma, end, false, NULL);
1774
1775         /*
1776          * If we're also updating the vma->vm_next->vm_start, if the new
1777          * vm_next->vm_start isn't page aligned and it could previously
1778          * contain an hugepage: check if we need to split an huge pmd.
1779          */
1780         if (adjust_next > 0) {
1781                 struct vm_area_struct *next = vma->vm_next;
1782                 unsigned long nstart = next->vm_start;
1783                 nstart += adjust_next << PAGE_SHIFT;
1784                 if (nstart & ~HPAGE_PMD_MASK &&
1785                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
1786                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
1787                         split_huge_pmd_address(next, nstart, false, NULL);
1788         }
1789 }
1790
1791 static void freeze_page(struct page *page)
1792 {
1793         enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
1794                 TTU_RMAP_LOCKED;
1795         int i, ret;
1796
1797         VM_BUG_ON_PAGE(!PageHead(page), page);
1798
1799         if (PageAnon(page))
1800                 ttu_flags |= TTU_MIGRATION;
1801
1802         /* We only need TTU_SPLIT_HUGE_PMD once */
1803         ret = try_to_unmap(page, ttu_flags | TTU_SPLIT_HUGE_PMD);
1804         for (i = 1; !ret && i < HPAGE_PMD_NR; i++) {
1805                 /* Cut short if the page is unmapped */
1806                 if (page_count(page) == 1)
1807                         return;
1808
1809                 ret = try_to_unmap(page + i, ttu_flags);
1810         }
1811         VM_BUG_ON_PAGE(ret, page + i - 1);
1812 }
1813
1814 static void unfreeze_page(struct page *page)
1815 {
1816         int i;
1817
1818         for (i = 0; i < HPAGE_PMD_NR; i++)
1819                 remove_migration_ptes(page + i, page + i, true);
1820 }
1821
1822 static void __split_huge_page_tail(struct page *head, int tail,
1823                 struct lruvec *lruvec, struct list_head *list)
1824 {
1825         struct page *page_tail = head + tail;
1826
1827         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
1828         VM_BUG_ON_PAGE(page_ref_count(page_tail) != 0, page_tail);
1829
1830         /*
1831          * tail_page->_refcount is zero and not changing from under us. But
1832          * get_page_unless_zero() may be running from under us on the
1833          * tail_page. If we used atomic_set() below instead of atomic_inc() or
1834          * atomic_add(), we would then run atomic_set() concurrently with
1835          * get_page_unless_zero(), and atomic_set() is implemented in C not
1836          * using locked ops. spin_unlock on x86 sometime uses locked ops
1837          * because of PPro errata 66, 92, so unless somebody can guarantee
1838          * atomic_set() here would be safe on all archs (and not only on x86),
1839          * it's safer to use atomic_inc()/atomic_add().
1840          */
1841         if (PageAnon(head)) {
1842                 page_ref_inc(page_tail);
1843         } else {
1844                 /* Additional pin to radix tree */
1845                 page_ref_add(page_tail, 2);
1846         }
1847
1848         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1849         page_tail->flags |= (head->flags &
1850                         ((1L << PG_referenced) |
1851                          (1L << PG_swapbacked) |
1852                          (1L << PG_mlocked) |
1853                          (1L << PG_uptodate) |
1854                          (1L << PG_active) |
1855                          (1L << PG_locked) |
1856                          (1L << PG_unevictable) |
1857                          (1L << PG_dirty)));
1858
1859         /*
1860          * After clearing PageTail the gup refcount can be released.
1861          * Page flags also must be visible before we make the page non-compound.
1862          */
1863         smp_wmb();
1864
1865         clear_compound_head(page_tail);
1866
1867         if (page_is_young(head))
1868                 set_page_young(page_tail);
1869         if (page_is_idle(head))
1870                 set_page_idle(page_tail);
1871
1872         /* ->mapping in first tail page is compound_mapcount */
1873         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
1874                         page_tail);
1875         page_tail->mapping = head->mapping;
1876
1877         page_tail->index = head->index + tail;
1878         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
1879         lru_add_page_tail(head, page_tail, lruvec, list);
1880 }
1881
1882 static void __split_huge_page(struct page *page, struct list_head *list,
1883                 unsigned long flags)
1884 {
1885         struct page *head = compound_head(page);
1886         struct zone *zone = page_zone(head);
1887         struct lruvec *lruvec;
1888         pgoff_t end = -1;
1889         int i;
1890
1891         lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat);
1892
1893         /* complete memcg works before add pages to LRU */
1894         mem_cgroup_split_huge_fixup(head);
1895
1896         if (!PageAnon(page))
1897                 end = DIV_ROUND_UP(i_size_read(head->mapping->host), PAGE_SIZE);
1898
1899         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1900                 __split_huge_page_tail(head, i, lruvec, list);
1901                 /* Some pages can be beyond i_size: drop them from page cache */
1902                 if (head[i].index >= end) {
1903                         __ClearPageDirty(head + i);
1904                         __delete_from_page_cache(head + i, NULL);
1905                         if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
1906                                 shmem_uncharge(head->mapping->host, 1);
1907                         put_page(head + i);
1908                 }
1909         }
1910
1911         ClearPageCompound(head);
1912         /* See comment in __split_huge_page_tail() */
1913         if (PageAnon(head)) {
1914                 page_ref_inc(head);
1915         } else {
1916                 /* Additional pin to radix tree */
1917                 page_ref_add(head, 2);
1918                 spin_unlock(&head->mapping->tree_lock);
1919         }
1920
1921         spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
1922
1923         unfreeze_page(head);
1924
1925         for (i = 0; i < HPAGE_PMD_NR; i++) {
1926                 struct page *subpage = head + i;
1927                 if (subpage == page)
1928                         continue;
1929                 unlock_page(subpage);
1930
1931                 /*
1932                  * Subpages may be freed if there wasn't any mapping
1933                  * like if add_to_swap() is running on a lru page that
1934                  * had its mapping zapped. And freeing these pages
1935                  * requires taking the lru_lock so we do the put_page
1936                  * of the tail pages after the split is complete.
1937                  */
1938                 put_page(subpage);
1939         }
1940 }
1941
1942 int total_mapcount(struct page *page)
1943 {
1944         int i, compound, ret;
1945
1946         VM_BUG_ON_PAGE(PageTail(page), page);
1947
1948         if (likely(!PageCompound(page)))
1949                 return atomic_read(&page->_mapcount) + 1;
1950
1951         compound = compound_mapcount(page);
1952         if (PageHuge(page))
1953                 return compound;
1954         ret = compound;
1955         for (i = 0; i < HPAGE_PMD_NR; i++)
1956                 ret += atomic_read(&page[i]._mapcount) + 1;
1957         /* File pages has compound_mapcount included in _mapcount */
1958         if (!PageAnon(page))
1959                 return ret - compound * HPAGE_PMD_NR;
1960         if (PageDoubleMap(page))
1961                 ret -= HPAGE_PMD_NR;
1962         return ret;
1963 }
1964
1965 /*
1966  * This calculates accurately how many mappings a transparent hugepage
1967  * has (unlike page_mapcount() which isn't fully accurate). This full
1968  * accuracy is primarily needed to know if copy-on-write faults can
1969  * reuse the page and change the mapping to read-write instead of
1970  * copying them. At the same time this returns the total_mapcount too.
1971  *
1972  * The function returns the highest mapcount any one of the subpages
1973  * has. If the return value is one, even if different processes are
1974  * mapping different subpages of the transparent hugepage, they can
1975  * all reuse it, because each process is reusing a different subpage.
1976  *
1977  * The total_mapcount is instead counting all virtual mappings of the
1978  * subpages. If the total_mapcount is equal to "one", it tells the
1979  * caller all mappings belong to the same "mm" and in turn the
1980  * anon_vma of the transparent hugepage can become the vma->anon_vma
1981  * local one as no other process may be mapping any of the subpages.
1982  *
1983  * It would be more accurate to replace page_mapcount() with
1984  * page_trans_huge_mapcount(), however we only use
1985  * page_trans_huge_mapcount() in the copy-on-write faults where we
1986  * need full accuracy to avoid breaking page pinning, because
1987  * page_trans_huge_mapcount() is slower than page_mapcount().
1988  */
1989 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
1990 {
1991         int i, ret, _total_mapcount, mapcount;
1992
1993         /* hugetlbfs shouldn't call it */
1994         VM_BUG_ON_PAGE(PageHuge(page), page);
1995
1996         if (likely(!PageTransCompound(page))) {
1997                 mapcount = atomic_read(&page->_mapcount) + 1;
1998                 if (total_mapcount)
1999                         *total_mapcount = mapcount;
2000                 return mapcount;
2001         }
2002
2003         page = compound_head(page);
2004
2005         _total_mapcount = ret = 0;
2006         for (i = 0; i < HPAGE_PMD_NR; i++) {
2007                 mapcount = atomic_read(&page[i]._mapcount) + 1;
2008                 ret = max(ret, mapcount);
2009                 _total_mapcount += mapcount;
2010         }
2011         if (PageDoubleMap(page)) {
2012                 ret -= 1;
2013                 _total_mapcount -= HPAGE_PMD_NR;
2014         }
2015         mapcount = compound_mapcount(page);
2016         ret += mapcount;
2017         _total_mapcount += mapcount;
2018         if (total_mapcount)
2019                 *total_mapcount = _total_mapcount;
2020         return ret;
2021 }
2022
2023 /*
2024  * This function splits huge page into normal pages. @page can point to any
2025  * subpage of huge page to split. Split doesn't change the position of @page.
2026  *
2027  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2028  * The huge page must be locked.
2029  *
2030  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2031  *
2032  * Both head page and tail pages will inherit mapping, flags, and so on from
2033  * the hugepage.
2034  *
2035  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2036  * they are not mapped.
2037  *
2038  * Returns 0 if the hugepage is split successfully.
2039  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2040  * us.
2041  */
2042 int split_huge_page_to_list(struct page *page, struct list_head *list)
2043 {
2044         struct page *head = compound_head(page);
2045         struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2046         struct anon_vma *anon_vma = NULL;
2047         struct address_space *mapping = NULL;
2048         int count, mapcount, extra_pins, ret;
2049         bool mlocked;
2050         unsigned long flags;
2051
2052         VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
2053         VM_BUG_ON_PAGE(!PageLocked(page), page);
2054         VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2055         VM_BUG_ON_PAGE(!PageCompound(page), page);
2056
2057         if (PageAnon(head)) {
2058                 /*
2059                  * The caller does not necessarily hold an mmap_sem that would
2060                  * prevent the anon_vma disappearing so we first we take a
2061                  * reference to it and then lock the anon_vma for write. This
2062                  * is similar to page_lock_anon_vma_read except the write lock
2063                  * is taken to serialise against parallel split or collapse
2064                  * operations.
2065                  */
2066                 anon_vma = page_get_anon_vma(head);
2067                 if (!anon_vma) {
2068                         ret = -EBUSY;
2069                         goto out;
2070                 }
2071                 extra_pins = 0;
2072                 mapping = NULL;
2073                 anon_vma_lock_write(anon_vma);
2074         } else {
2075                 mapping = head->mapping;
2076
2077                 /* Truncated ? */
2078                 if (!mapping) {
2079                         ret = -EBUSY;
2080                         goto out;
2081                 }
2082
2083                 /* Addidional pins from radix tree */
2084                 extra_pins = HPAGE_PMD_NR;
2085                 anon_vma = NULL;
2086                 i_mmap_lock_read(mapping);
2087         }
2088
2089         /*
2090          * Racy check if we can split the page, before freeze_page() will
2091          * split PMDs
2092          */
2093         if (total_mapcount(head) != page_count(head) - extra_pins - 1) {
2094                 ret = -EBUSY;
2095                 goto out_unlock;
2096         }
2097
2098         mlocked = PageMlocked(page);
2099         freeze_page(head);
2100         VM_BUG_ON_PAGE(compound_mapcount(head), head);
2101
2102         /* Make sure the page is not on per-CPU pagevec as it takes pin */
2103         if (mlocked)
2104                 lru_add_drain();
2105
2106         /* prevent PageLRU to go away from under us, and freeze lru stats */
2107         spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags);
2108
2109         if (mapping) {
2110                 void **pslot;
2111
2112                 spin_lock(&mapping->tree_lock);
2113                 pslot = radix_tree_lookup_slot(&mapping->page_tree,
2114                                 page_index(head));
2115                 /*
2116                  * Check if the head page is present in radix tree.
2117                  * We assume all tail are present too, if head is there.
2118                  */
2119                 if (radix_tree_deref_slot_protected(pslot,
2120                                         &mapping->tree_lock) != head)
2121                         goto fail;
2122         }
2123
2124         /* Prevent deferred_split_scan() touching ->_refcount */
2125         spin_lock(&pgdata->split_queue_lock);
2126         count = page_count(head);
2127         mapcount = total_mapcount(head);
2128         if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2129                 if (!list_empty(page_deferred_list(head))) {
2130                         pgdata->split_queue_len--;
2131                         list_del(page_deferred_list(head));
2132                 }
2133                 if (mapping)
2134                         __dec_node_page_state(page, NR_SHMEM_THPS);
2135                 spin_unlock(&pgdata->split_queue_lock);
2136                 __split_huge_page(page, list, flags);
2137                 ret = 0;
2138         } else {
2139                 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2140                         pr_alert("total_mapcount: %u, page_count(): %u\n",
2141                                         mapcount, count);
2142                         if (PageTail(page))
2143                                 dump_page(head, NULL);
2144                         dump_page(page, "total_mapcount(head) > 0");
2145                         BUG();
2146                 }
2147                 spin_unlock(&pgdata->split_queue_lock);
2148 fail:           if (mapping)
2149                         spin_unlock(&mapping->tree_lock);
2150                 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
2151                 unfreeze_page(head);
2152                 ret = -EBUSY;
2153         }
2154
2155 out_unlock:
2156         if (anon_vma) {
2157                 anon_vma_unlock_write(anon_vma);
2158                 put_anon_vma(anon_vma);
2159         }
2160         if (mapping)
2161                 i_mmap_unlock_read(mapping);
2162 out:
2163         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2164         return ret;
2165 }
2166
2167 void free_transhuge_page(struct page *page)
2168 {
2169         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2170         unsigned long flags;
2171
2172         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2173         if (!list_empty(page_deferred_list(page))) {
2174                 pgdata->split_queue_len--;
2175                 list_del(page_deferred_list(page));
2176         }
2177         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2178         free_compound_page(page);
2179 }
2180
2181 void deferred_split_huge_page(struct page *page)
2182 {
2183         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2184         unsigned long flags;
2185
2186         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2187
2188         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2189         if (list_empty(page_deferred_list(page))) {
2190                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2191                 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2192                 pgdata->split_queue_len++;
2193         }
2194         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2195 }
2196
2197 static unsigned long deferred_split_count(struct shrinker *shrink,
2198                 struct shrink_control *sc)
2199 {
2200         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2201         return ACCESS_ONCE(pgdata->split_queue_len);
2202 }
2203
2204 static unsigned long deferred_split_scan(struct shrinker *shrink,
2205                 struct shrink_control *sc)
2206 {
2207         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2208         unsigned long flags;
2209         LIST_HEAD(list), *pos, *next;
2210         struct page *page;
2211         int split = 0;
2212
2213         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2214         /* Take pin on all head pages to avoid freeing them under us */
2215         list_for_each_safe(pos, next, &pgdata->split_queue) {
2216                 page = list_entry((void *)pos, struct page, mapping);
2217                 page = compound_head(page);
2218                 if (get_page_unless_zero(page)) {
2219                         list_move(page_deferred_list(page), &list);
2220                 } else {
2221                         /* We lost race with put_compound_page() */
2222                         list_del_init(page_deferred_list(page));
2223                         pgdata->split_queue_len--;
2224                 }
2225                 if (!--sc->nr_to_scan)
2226                         break;
2227         }
2228         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2229
2230         list_for_each_safe(pos, next, &list) {
2231                 page = list_entry((void *)pos, struct page, mapping);
2232                 lock_page(page);
2233                 /* split_huge_page() removes page from list on success */
2234                 if (!split_huge_page(page))
2235                         split++;
2236                 unlock_page(page);
2237                 put_page(page);
2238         }
2239
2240         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2241         list_splice_tail(&list, &pgdata->split_queue);
2242         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2243
2244         /*
2245          * Stop shrinker if we didn't split any page, but the queue is empty.
2246          * This can happen if pages were freed under us.
2247          */
2248         if (!split && list_empty(&pgdata->split_queue))
2249                 return SHRINK_STOP;
2250         return split;
2251 }
2252
2253 static struct shrinker deferred_split_shrinker = {
2254         .count_objects = deferred_split_count,
2255         .scan_objects = deferred_split_scan,
2256         .seeks = DEFAULT_SEEKS,
2257         .flags = SHRINKER_NUMA_AWARE,
2258 };
2259
2260 #ifdef CONFIG_DEBUG_FS
2261 static int split_huge_pages_set(void *data, u64 val)
2262 {
2263         struct zone *zone;
2264         struct page *page;
2265         unsigned long pfn, max_zone_pfn;
2266         unsigned long total = 0, split = 0;
2267
2268         if (val != 1)
2269                 return -EINVAL;
2270
2271         for_each_populated_zone(zone) {
2272                 max_zone_pfn = zone_end_pfn(zone);
2273                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2274                         if (!pfn_valid(pfn))
2275                                 continue;
2276
2277                         page = pfn_to_page(pfn);
2278                         if (!get_page_unless_zero(page))
2279                                 continue;
2280
2281                         if (zone != page_zone(page))
2282                                 goto next;
2283
2284                         if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2285                                 goto next;
2286
2287                         total++;
2288                         lock_page(page);
2289                         if (!split_huge_page(page))
2290                                 split++;
2291                         unlock_page(page);
2292 next:
2293                         put_page(page);
2294                 }
2295         }
2296
2297         pr_info("%lu of %lu THP split\n", split, total);
2298
2299         return 0;
2300 }
2301 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2302                 "%llu\n");
2303
2304 static int __init split_huge_pages_debugfs(void)
2305 {
2306         void *ret;
2307
2308         ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2309                         &split_huge_pages_fops);
2310         if (!ret)
2311                 pr_warn("Failed to create split_huge_pages in debugfs");
2312         return 0;
2313 }
2314 late_initcall(split_huge_pages_debugfs);
2315 #endif