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
 741         track_pfn_insert(vma, &pgprot, pfn);
 742
 743         insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
 744         return VM_FAULT_NOPAGE;
 745 }
 746 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
 747
 748 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
 749                 pmd_t *pmd)
 750 {
 751         pmd_t _pmd;
 752
 753         /*
 754          * We should set the dirty bit only for FOLL_WRITE but for now
 755          * the dirty bit in the pmd is meaningless.  And if the dirty
 756          * bit will become meaningful and we'll only set it with
 757          * FOLL_WRITE, an atomic set_bit will be required on the pmd to
 758          * set the young bit, instead of the current set_pmd_at.
 759          */
 760         _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
 761         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
 762                                 pmd, _pmd,  1))
 763                 update_mmu_cache_pmd(vma, addr, pmd);
 764 }
 765
 766 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 767                 pmd_t *pmd, int flags)
 768 {
 769         unsigned long pfn = pmd_pfn(*pmd);
 770         struct mm_struct *mm = vma->vm_mm;
 771         struct dev_pagemap *pgmap;
 772         struct page *page;
 773
 774         assert_spin_locked(pmd_lockptr(mm, pmd));
 775
 776         if (flags & FOLL_WRITE && !pmd_write(*pmd))
 777                 return NULL;
 778
 779         if (pmd_present(*pmd) && pmd_devmap(*pmd))
 780                 /* pass */;
 781         else
 782                 return NULL;
 783
 784         if (flags & FOLL_TOUCH)
 785                 touch_pmd(vma, addr, pmd);
 786
 787         /*
 788          * device mapped pages can only be returned if the
 789          * caller will manage the page reference count.
 790          */
 791         if (!(flags & FOLL_GET))
 792                 return ERR_PTR(-EEXIST);
 793
 794         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
 795         pgmap = get_dev_pagemap(pfn, NULL);
 796         if (!pgmap)
 797                 return ERR_PTR(-EFAULT);
 798         page = pfn_to_page(pfn);
 799         get_page(page);
 800         put_dev_pagemap(pgmap);
 801
 802         return page;
 803 }
 804
 805 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 806                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 807                   struct vm_area_struct *vma)
 808 {
 809         spinlock_t *dst_ptl, *src_ptl;
 810         struct page *src_page;
 811         pmd_t pmd;
 812         pgtable_t pgtable = NULL;
 813         int ret = -ENOMEM;
 814
 815         /* Skip if can be re-fill on fault */
 816         if (!vma_is_anonymous(vma))
 817                 return 0;
 818
 819         pgtable = pte_alloc_one(dst_mm, addr);
 820         if (unlikely(!pgtable))
 821                 goto out;
 822
 823         dst_ptl = pmd_lock(dst_mm, dst_pmd);
 824         src_ptl = pmd_lockptr(src_mm, src_pmd);
 825         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 826
 827         ret = -EAGAIN;
 828         pmd = *src_pmd;
 829         if (unlikely(!pmd_trans_huge(pmd))) {
 830                 pte_free(dst_mm, pgtable);
 831                 goto out_unlock;
 832         }
 833         /*
 834          * When page table lock is held, the huge zero pmd should not be
 835          * under splitting since we don't split the page itself, only pmd to
 836          * a page table.
 837          */
 838         if (is_huge_zero_pmd(pmd)) {
 839                 struct page *zero_page;
 840                 /*
 841                  * get_huge_zero_page() will never allocate a new page here,
 842                  * since we already have a zero page to copy. It just takes a
 843                  * reference.
 844                  */
 845                 zero_page = mm_get_huge_zero_page(dst_mm);
 846                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
 847                                 zero_page);
 848                 ret = 0;
 849                 goto out_unlock;
 850         }
 851
 852         src_page = pmd_page(pmd);
 853         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
 854         get_page(src_page);
 855         page_dup_rmap(src_page, true);
 856         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 857         atomic_long_inc(&dst_mm->nr_ptes);
 858         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
 859
 860         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 861         pmd = pmd_mkold(pmd_wrprotect(pmd));
 862         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 863
 864         ret = 0;
 865 out_unlock:
 866         spin_unlock(src_ptl);
 867         spin_unlock(dst_ptl);
 868 out:
 869         return ret;
 870 }
 871
 872 void huge_pmd_set_accessed(struct fault_env *fe, pmd_t orig_pmd)
 873 {
 874         pmd_t entry;
 875         unsigned long haddr;
 876
 877         fe->ptl = pmd_lock(fe->vma->vm_mm, fe->pmd);
 878         if (unlikely(!pmd_same(*fe->pmd, orig_pmd)))
 879                 goto unlock;
 880
 881         entry = pmd_mkyoung(orig_pmd);
 882         haddr = fe->address & HPAGE_PMD_MASK;
 883         if (pmdp_set_access_flags(fe->vma, haddr, fe->pmd, entry,
 884                                 fe->flags & FAULT_FLAG_WRITE))
 885                 update_mmu_cache_pmd(fe->vma, fe->address, fe->pmd);
 886
 887 unlock:
 888         spin_unlock(fe->ptl);
 889 }
 890
 891 static int do_huge_pmd_wp_page_fallback(struct fault_env *fe, pmd_t orig_pmd,
 892                 struct page *page)
 893 {
 894         struct vm_area_struct *vma = fe->vma;
 895         unsigned long haddr = fe->address & HPAGE_PMD_MASK;
 896         struct mem_cgroup *memcg;
 897         pgtable_t pgtable;
 898         pmd_t _pmd;
 899         int ret = 0, i;
 900         struct page **pages;
 901         unsigned long mmun_start;       /* For mmu_notifiers */
 902         unsigned long mmun_end;         /* For mmu_notifiers */
 903
 904         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
 905                         GFP_KERNEL);
 906         if (unlikely(!pages)) {
 907                 ret |= VM_FAULT_OOM;
 908                 goto out;
 909         }
 910
 911         for (i = 0; i < HPAGE_PMD_NR; i++) {
 912                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
 913                                                __GFP_OTHER_NODE, vma,
 914                                                fe->address, page_to_nid(page));
 915                 if (unlikely(!pages[i] ||
 916                              mem_cgroup_try_charge(pages[i], vma->vm_mm,
 917                                      GFP_KERNEL, &memcg, false))) {
 918                         if (pages[i])
 919                                 put_page(pages[i]);
 920                         while (--i >= 0) {
 921                                 memcg = (void *)page_private(pages[i]);
 922                                 set_page_private(pages[i], 0);
 923                                 mem_cgroup_cancel_charge(pages[i], memcg,
 924                                                 false);
 925                                 put_page(pages[i]);
 926                         }
 927                         kfree(pages);
 928                         ret |= VM_FAULT_OOM;
 929                         goto out;
 930                 }
 931                 set_page_private(pages[i], (unsigned long)memcg);
 932         }
 933
 934         for (i = 0; i < HPAGE_PMD_NR; i++) {
 935                 copy_user_highpage(pages[i], page + i,
 936                                    haddr + PAGE_SIZE * i, vma);
 937                 __SetPageUptodate(pages[i]);
 938                 cond_resched();
 939         }
 940
 941         mmun_start = haddr;
 942         mmun_end   = haddr + HPAGE_PMD_SIZE;
 943         mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
 944
 945         fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
 946         if (unlikely(!pmd_same(*fe->pmd, orig_pmd)))
 947                 goto out_free_pages;
 948         VM_BUG_ON_PAGE(!PageHead(page), page);
 949
 950         pmdp_huge_clear_flush_notify(vma, haddr, fe->pmd);
 951         /* leave pmd empty until pte is filled */
 952
 953         pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, fe->pmd);
 954         pmd_populate(vma->vm_mm, &_pmd, pgtable);
 955
 956         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
 957                 pte_t entry;
 958                 entry = mk_pte(pages[i], vma->vm_page_prot);
 959                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
 960                 memcg = (void *)page_private(pages[i]);
 961                 set_page_private(pages[i], 0);
 962                 page_add_new_anon_rmap(pages[i], fe->vma, haddr, false);
 963                 mem_cgroup_commit_charge(pages[i], memcg, false, false);
 964                 lru_cache_add_active_or_unevictable(pages[i], vma);
 965                 fe->pte = pte_offset_map(&_pmd, haddr);
 966                 VM_BUG_ON(!pte_none(*fe->pte));
 967                 set_pte_at(vma->vm_mm, haddr, fe->pte, entry);
 968                 pte_unmap(fe->pte);
 969         }
 970         kfree(pages);
 971
 972         smp_wmb(); /* make pte visible before pmd */
 973         pmd_populate(vma->vm_mm, fe->pmd, pgtable);
 974         page_remove_rmap(page, true);
 975         spin_unlock(fe->ptl);
 976
 977         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
 978
 979         ret |= VM_FAULT_WRITE;
 980         put_page(page);
 981
 982 out:
 983         return ret;
 984
 985 out_free_pages:
 986         spin_unlock(fe->ptl);
 987         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
 988         for (i = 0; i < HPAGE_PMD_NR; i++) {
 989                 memcg = (void *)page_private(pages[i]);
 990                 set_page_private(pages[i], 0);
 991                 mem_cgroup_cancel_charge(pages[i], memcg, false);
 992                 put_page(pages[i]);
 993         }
 994         kfree(pages);
 995         goto out;
 996 }
 997
 998 int do_huge_pmd_wp_page(struct fault_env *fe, pmd_t orig_pmd)
 999 {
1000         struct vm_area_struct *vma = fe->vma;
1001         struct page *page = NULL, *new_page;
1002         struct mem_cgroup *memcg;
1003         unsigned long haddr = fe->address & HPAGE_PMD_MASK;
1004         unsigned long mmun_start;       /* For mmu_notifiers */
1005         unsigned long mmun_end;         /* For mmu_notifiers */
1006         gfp_t huge_gfp;                 /* for allocation and charge */
1007         int ret = 0;
1008
1009         fe->ptl = pmd_lockptr(vma->vm_mm, fe->pmd);
1010         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1011         if (is_huge_zero_pmd(orig_pmd))
1012                 goto alloc;
1013         spin_lock(fe->ptl);
1014         if (unlikely(!pmd_same(*fe->pmd, orig_pmd)))
1015                 goto out_unlock;
1016
1017         page = pmd_page(orig_pmd);
1018         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1019         /*
1020          * We can only reuse the page if nobody else maps the huge page or it's
1021          * part.
1022          */
1023         if (page_trans_huge_mapcount(page, NULL) == 1) {
1024                 pmd_t entry;
1025                 entry = pmd_mkyoung(orig_pmd);
1026                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1027                 if (pmdp_set_access_flags(vma, haddr, fe->pmd, entry,  1))
1028                         update_mmu_cache_pmd(vma, fe->address, fe->pmd);
1029                 ret |= VM_FAULT_WRITE;
1030                 goto out_unlock;
1031         }
1032         get_page(page);
1033         spin_unlock(fe->ptl);
1034 alloc:
1035         if (transparent_hugepage_enabled(vma) &&
1036             !transparent_hugepage_debug_cow()) {
1037                 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1038                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1039         } else
1040                 new_page = NULL;
1041
1042         if (likely(new_page)) {
1043                 prep_transhuge_page(new_page);
1044         } else {
1045                 if (!page) {
1046                         split_huge_pmd(vma, fe->pmd, fe->address);
1047                         ret |= VM_FAULT_FALLBACK;
1048                 } else {
1049                         ret = do_huge_pmd_wp_page_fallback(fe, orig_pmd, page);
1050                         if (ret & VM_FAULT_OOM) {
1051                                 split_huge_pmd(vma, fe->pmd, fe->address);
1052                                 ret |= VM_FAULT_FALLBACK;
1053                         }
1054                         put_page(page);
1055                 }
1056                 count_vm_event(THP_FAULT_FALLBACK);
1057                 goto out;
1058         }
1059
1060         if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm,
1061                                         huge_gfp, &memcg, true))) {
1062                 put_page(new_page);
1063                 split_huge_pmd(vma, fe->pmd, fe->address);
1064                 if (page)
1065                         put_page(page);
1066                 ret |= VM_FAULT_FALLBACK;
1067                 count_vm_event(THP_FAULT_FALLBACK);
1068                 goto out;
1069         }
1070
1071         count_vm_event(THP_FAULT_ALLOC);
1072
1073         if (!page)
1074                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1075         else
1076                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1077         __SetPageUptodate(new_page);
1078
1079         mmun_start = haddr;
1080         mmun_end   = haddr + HPAGE_PMD_SIZE;
1081         mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
1082
1083         spin_lock(fe->ptl);
1084         if (page)
1085                 put_page(page);
1086         if (unlikely(!pmd_same(*fe->pmd, orig_pmd))) {
1087                 spin_unlock(fe->ptl);
1088                 mem_cgroup_cancel_charge(new_page, memcg, true);
1089                 put_page(new_page);
1090                 goto out_mn;
1091         } else {
1092                 pmd_t entry;
1093                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1094                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1095                 pmdp_huge_clear_flush_notify(vma, haddr, fe->pmd);
1096                 page_add_new_anon_rmap(new_page, vma, haddr, true);
1097                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1098                 lru_cache_add_active_or_unevictable(new_page, vma);
1099                 set_pmd_at(vma->vm_mm, haddr, fe->pmd, entry);
1100                 update_mmu_cache_pmd(vma, fe->address, fe->pmd);
1101                 if (!page) {
1102                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1103                 } else {
1104                         VM_BUG_ON_PAGE(!PageHead(page), page);
1105                         page_remove_rmap(page, true);
1106                         put_page(page);
1107                 }
1108                 ret |= VM_FAULT_WRITE;
1109         }
1110         spin_unlock(fe->ptl);
1111 out_mn:
1112         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1113 out:
1114         return ret;
1115 out_unlock:
1116         spin_unlock(fe->ptl);
1117         return ret;
1118 }
1119
1120 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1121                                    unsigned long addr,
1122                                    pmd_t *pmd,
1123                                    unsigned int flags)
1124 {
1125         struct mm_struct *mm = vma->vm_mm;
1126         struct page *page = NULL;
1127
1128         assert_spin_locked(pmd_lockptr(mm, pmd));
1129
1130         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1131                 goto out;
1132
1133         /* Avoid dumping huge zero page */
1134         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1135                 return ERR_PTR(-EFAULT);
1136
1137         /* Full NUMA hinting faults to serialise migration in fault paths */
1138         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1139                 goto out;
1140
1141         page = pmd_page(*pmd);
1142         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1143         if (flags & FOLL_TOUCH)
1144                 touch_pmd(vma, addr, pmd);
1145         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1146                 /*
1147                  * We don't mlock() pte-mapped THPs. This way we can avoid
1148                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1149                  *
1150                  * For anon THP:
1151                  *
1152                  * In most cases the pmd is the only mapping of the page as we
1153                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1154                  * writable private mappings in populate_vma_page_range().
1155                  *
1156                  * The only scenario when we have the page shared here is if we
1157                  * mlocking read-only mapping shared over fork(). We skip
1158                  * mlocking such pages.
1159                  *
1160                  * For file THP:
1161                  *
1162                  * We can expect PageDoubleMap() to be stable under page lock:
1163                  * for file pages we set it in page_add_file_rmap(), which
1164                  * requires page to be locked.
1165                  */
1166
1167                 if (PageAnon(page) && compound_mapcount(page) != 1)
1168                         goto skip_mlock;
1169                 if (PageDoubleMap(page) || !page->mapping)
1170                         goto skip_mlock;
1171                 if (!trylock_page(page))
1172                         goto skip_mlock;
1173                 lru_add_drain();
1174                 if (page->mapping && !PageDoubleMap(page))
1175                         mlock_vma_page(page);
1176                 unlock_page(page);
1177         }
1178 skip_mlock:
1179         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1180         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1181         if (flags & FOLL_GET)
1182                 get_page(page);
1183
1184 out:
1185         return page;
1186 }
1187
1188 /* NUMA hinting page fault entry point for trans huge pmds */
1189 int do_huge_pmd_numa_page(struct fault_env *fe, pmd_t pmd)
1190 {
1191         struct vm_area_struct *vma = fe->vma;
1192         struct anon_vma *anon_vma = NULL;
1193         struct page *page;
1194         unsigned long haddr = fe->address & HPAGE_PMD_MASK;
1195         int page_nid = -1, this_nid = numa_node_id();
1196         int target_nid, last_cpupid = -1;
1197         bool page_locked;
1198         bool migrated = false;
1199         bool was_writable;
1200         int flags = 0;
1201
1202         fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
1203         if (unlikely(!pmd_same(pmd, *fe->pmd)))
1204                 goto out_unlock;
1205
1206         /*
1207          * If there are potential migrations, wait for completion and retry
1208          * without disrupting NUMA hinting information. Do not relock and
1209          * check_same as the page may no longer be mapped.
1210          */
1211         if (unlikely(pmd_trans_migrating(*fe->pmd))) {
1212                 page = pmd_page(*fe->pmd);
1213                 spin_unlock(fe->ptl);
1214                 wait_on_page_locked(page);
1215                 goto out;
1216         }
1217
1218         page = pmd_page(pmd);
1219         BUG_ON(is_huge_zero_page(page));
1220         page_nid = page_to_nid(page);
1221         last_cpupid = page_cpupid_last(page);
1222         count_vm_numa_event(NUMA_HINT_FAULTS);
1223         if (page_nid == this_nid) {
1224                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1225                 flags |= TNF_FAULT_LOCAL;
1226         }
1227
1228         /* See similar comment in do_numa_page for explanation */
1229         if (!pmd_write(pmd))
1230                 flags |= TNF_NO_GROUP;
1231
1232         /*
1233          * Acquire the page lock to serialise THP migrations but avoid dropping
1234          * page_table_lock if at all possible
1235          */
1236         page_locked = trylock_page(page);
1237         target_nid = mpol_misplaced(page, vma, haddr);
1238         if (target_nid == -1) {
1239                 /* If the page was locked, there are no parallel migrations */
1240                 if (page_locked)
1241                         goto clear_pmdnuma;
1242         }
1243
1244         /* Migration could have started since the pmd_trans_migrating check */
1245         if (!page_locked) {
1246                 spin_unlock(fe->ptl);
1247                 wait_on_page_locked(page);
1248                 page_nid = -1;
1249                 goto out;
1250         }
1251
1252         /*
1253          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1254          * to serialises splits
1255          */
1256         get_page(page);
1257         spin_unlock(fe->ptl);
1258         anon_vma = page_lock_anon_vma_read(page);
1259
1260         /* Confirm the PMD did not change while page_table_lock was released */
1261         spin_lock(fe->ptl);
1262         if (unlikely(!pmd_same(pmd, *fe->pmd))) {
1263                 unlock_page(page);
1264                 put_page(page);
1265                 page_nid = -1;
1266                 goto out_unlock;
1267         }
1268
1269         /* Bail if we fail to protect against THP splits for any reason */
1270         if (unlikely(!anon_vma)) {
1271                 put_page(page);
1272                 page_nid = -1;
1273                 goto clear_pmdnuma;
1274         }
1275
1276         /*
1277          * Migrate the THP to the requested node, returns with page unlocked
1278          * and access rights restored.
1279          */
1280         spin_unlock(fe->ptl);
1281         migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1282                                 fe->pmd, pmd, fe->address, page, target_nid);
1283         if (migrated) {
1284                 flags |= TNF_MIGRATED;
1285                 page_nid = target_nid;
1286         } else
1287                 flags |= TNF_MIGRATE_FAIL;
1288
1289         goto out;
1290 clear_pmdnuma:
1291         BUG_ON(!PageLocked(page));
1292         was_writable = pmd_write(pmd);
1293         pmd = pmd_modify(pmd, vma->vm_page_prot);
1294         pmd = pmd_mkyoung(pmd);
1295         if (was_writable)
1296                 pmd = pmd_mkwrite(pmd);
1297         set_pmd_at(vma->vm_mm, haddr, fe->pmd, pmd);
1298         update_mmu_cache_pmd(vma, fe->address, fe->pmd);
1299         unlock_page(page);
1300 out_unlock:
1301         spin_unlock(fe->ptl);
1302
1303 out:
1304         if (anon_vma)
1305                 page_unlock_anon_vma_read(anon_vma);
1306
1307         if (page_nid != -1)
1308                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, fe->flags);
1309
1310         return 0;
1311 }
1312
1313 /*
1314  * Return true if we do MADV_FREE successfully on entire pmd page.
1315  * Otherwise, return false.
1316  */
1317 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1318                 pmd_t *pmd, unsigned long addr, unsigned long next)
1319 {
1320         spinlock_t *ptl;
1321         pmd_t orig_pmd;
1322         struct page *page;
1323         struct mm_struct *mm = tlb->mm;
1324         bool ret = false;
1325
1326         ptl = pmd_trans_huge_lock(pmd, vma);
1327         if (!ptl)
1328                 goto out_unlocked;
1329
1330         orig_pmd = *pmd;
1331         if (is_huge_zero_pmd(orig_pmd))
1332                 goto out;
1333
1334         page = pmd_page(orig_pmd);
1335         /*
1336          * If other processes are mapping this page, we couldn't discard
1337          * the page unless they all do MADV_FREE so let's skip the page.
1338          */
1339         if (page_mapcount(page) != 1)
1340                 goto out;
1341
1342         if (!trylock_page(page))
1343                 goto out;
1344
1345         /*
1346          * If user want to discard part-pages of THP, split it so MADV_FREE
1347          * will deactivate only them.
1348          */
1349         if (next - addr != HPAGE_PMD_SIZE) {
1350                 get_page(page);
1351                 spin_unlock(ptl);
1352                 split_huge_page(page);
1353                 put_page(page);
1354                 unlock_page(page);
1355                 goto out_unlocked;
1356         }
1357
1358         if (PageDirty(page))
1359                 ClearPageDirty(page);
1360         unlock_page(page);
1361
1362         if (PageActive(page))
1363                 deactivate_page(page);
1364
1365         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1366                 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1367                         tlb->fullmm);
1368                 orig_pmd = pmd_mkold(orig_pmd);
1369                 orig_pmd = pmd_mkclean(orig_pmd);
1370
1371                 set_pmd_at(mm, addr, pmd, orig_pmd);
1372                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1373         }
1374         ret = true;
1375 out:
1376         spin_unlock(ptl);
1377 out_unlocked:
1378         return ret;
1379 }
1380
1381 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1382                  pmd_t *pmd, unsigned long addr)
1383 {
1384         pmd_t orig_pmd;
1385         spinlock_t *ptl;
1386
1387         ptl = __pmd_trans_huge_lock(pmd, vma);
1388         if (!ptl)
1389                 return 0;
1390         /*
1391          * For architectures like ppc64 we look at deposited pgtable
1392          * when calling pmdp_huge_get_and_clear. So do the
1393          * pgtable_trans_huge_withdraw after finishing pmdp related
1394          * operations.
1395          */
1396         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1397                         tlb->fullmm);
1398         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1399         if (vma_is_dax(vma)) {
1400                 spin_unlock(ptl);
1401                 if (is_huge_zero_pmd(orig_pmd))
1402                         tlb_remove_page(tlb, pmd_page(orig_pmd));
1403         } else if (is_huge_zero_pmd(orig_pmd)) {
1404                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1405                 atomic_long_dec(&tlb->mm->nr_ptes);
1406                 spin_unlock(ptl);
1407                 tlb_remove_page(tlb, pmd_page(orig_pmd));
1408         } else {
1409                 struct page *page = pmd_page(orig_pmd);
1410                 page_remove_rmap(page, true);
1411                 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1412                 VM_BUG_ON_PAGE(!PageHead(page), page);
1413                 if (PageAnon(page)) {
1414                         pgtable_t pgtable;
1415                         pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1416                         pte_free(tlb->mm, pgtable);
1417                         atomic_long_dec(&tlb->mm->nr_ptes);
1418                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1419                 } else {
1420                         add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1421                 }
1422                 spin_unlock(ptl);
1423                 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1424         }
1425         return 1;
1426 }
1427
1428 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1429                   unsigned long new_addr, unsigned long old_end,
1430                   pmd_t *old_pmd, pmd_t *new_pmd, bool *need_flush)
1431 {
1432         spinlock_t *old_ptl, *new_ptl;
1433         pmd_t pmd;
1434         struct mm_struct *mm = vma->vm_mm;
1435         bool force_flush = false;
1436
1437         if ((old_addr & ~HPAGE_PMD_MASK) ||
1438             (new_addr & ~HPAGE_PMD_MASK) ||
1439             old_end - old_addr < HPAGE_PMD_SIZE)
1440                 return false;
1441
1442         /*
1443          * The destination pmd shouldn't be established, free_pgtables()
1444          * should have release it.
1445          */
1446         if (WARN_ON(!pmd_none(*new_pmd))) {
1447                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1448                 return false;
1449         }
1450
1451         /*
1452          * We don't have to worry about the ordering of src and dst
1453          * ptlocks because exclusive mmap_sem prevents deadlock.
1454          */
1455         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1456         if (old_ptl) {
1457                 new_ptl = pmd_lockptr(mm, new_pmd);
1458                 if (new_ptl != old_ptl)
1459                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1460                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1461                 if (pmd_present(pmd) && pmd_dirty(pmd))
1462                         force_flush = true;
1463                 VM_BUG_ON(!pmd_none(*new_pmd));
1464
1465                 if (pmd_move_must_withdraw(new_ptl, old_ptl) &&
1466                                 vma_is_anonymous(vma)) {
1467                         pgtable_t pgtable;
1468                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1469                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1470                 }
1471                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1472                 if (new_ptl != old_ptl)
1473                         spin_unlock(new_ptl);
1474                 if (force_flush)
1475                         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1476                 else
1477                         *need_flush = true;
1478                 spin_unlock(old_ptl);
1479                 return true;
1480         }
1481         return false;
1482 }
1483
1484 /*
1485  * Returns
1486  *  - 0 if PMD could not be locked
1487  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1488  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1489  */
1490 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1491                 unsigned long addr, pgprot_t newprot, int prot_numa)
1492 {
1493         struct mm_struct *mm = vma->vm_mm;
1494         spinlock_t *ptl;
1495         int ret = 0;
1496
1497         ptl = __pmd_trans_huge_lock(pmd, vma);
1498         if (ptl) {
1499                 pmd_t entry;
1500                 bool preserve_write = prot_numa && pmd_write(*pmd);
1501                 ret = 1;
1502
1503                 /*
1504                  * Avoid trapping faults against the zero page. The read-only
1505                  * data is likely to be read-cached on the local CPU and
1506                  * local/remote hits to the zero page are not interesting.
1507                  */
1508                 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1509                         spin_unlock(ptl);
1510                         return ret;
1511                 }
1512
1513                 if (!prot_numa || !pmd_protnone(*pmd)) {
1514                         entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
1515                         entry = pmd_modify(entry, newprot);
1516                         if (preserve_write)
1517                                 entry = pmd_mkwrite(entry);
1518                         ret = HPAGE_PMD_NR;
1519                         set_pmd_at(mm, addr, pmd, entry);
1520                         BUG_ON(vma_is_anonymous(vma) && !preserve_write &&
1521                                         pmd_write(entry));
1522                 }
1523                 spin_unlock(ptl);
1524         }
1525
1526         return ret;
1527 }
1528
1529 /*
1530  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1531  *
1532  * Note that if it returns page table lock pointer, this routine returns without
1533  * unlocking page table lock. So callers must unlock it.
1534  */
1535 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1536 {
1537         spinlock_t *ptl;
1538         ptl = pmd_lock(vma->vm_mm, pmd);
1539         if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd)))
1540                 return ptl;
1541         spin_unlock(ptl);
1542         return NULL;
1543 }
1544
1545 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1546                 unsigned long haddr, pmd_t *pmd)
1547 {
1548         struct mm_struct *mm = vma->vm_mm;
1549         pgtable_t pgtable;
1550         pmd_t _pmd;
1551         int i;
1552
1553         /* leave pmd empty until pte is filled */
1554         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1555
1556         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1557         pmd_populate(mm, &_pmd, pgtable);
1558
1559         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1560                 pte_t *pte, entry;
1561                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
1562                 entry = pte_mkspecial(entry);
1563                 pte = pte_offset_map(&_pmd, haddr);
1564                 VM_BUG_ON(!pte_none(*pte));
1565                 set_pte_at(mm, haddr, pte, entry);
1566                 pte_unmap(pte);
1567         }
1568         smp_wmb(); /* make pte visible before pmd */
1569         pmd_populate(mm, pmd, pgtable);
1570 }
1571
1572 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
1573                 unsigned long haddr, bool freeze)
1574 {
1575         struct mm_struct *mm = vma->vm_mm;
1576         struct page *page;
1577         pgtable_t pgtable;
1578         pmd_t _pmd;
1579         bool young, write, dirty, soft_dirty;
1580         unsigned long addr;
1581         int i;
1582
1583         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
1584         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1585         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
1586         VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd));
1587
1588         count_vm_event(THP_SPLIT_PMD);
1589
1590         if (!vma_is_anonymous(vma)) {
1591                 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1592                 if (vma_is_dax(vma))
1593                         return;
1594                 page = pmd_page(_pmd);
1595                 if (!PageReferenced(page) && pmd_young(_pmd))
1596                         SetPageReferenced(page);
1597                 page_remove_rmap(page, true);
1598                 put_page(page);
1599                 add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1600                 return;
1601         } else if (is_huge_zero_pmd(*pmd)) {
1602                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
1603         }
1604
1605         page = pmd_page(*pmd);
1606         VM_BUG_ON_PAGE(!page_count(page), page);
1607         page_ref_add(page, HPAGE_PMD_NR - 1);
1608         write = pmd_write(*pmd);
1609         young = pmd_young(*pmd);
1610         dirty = pmd_dirty(*pmd);
1611         soft_dirty = pmd_soft_dirty(*pmd);
1612
1613         pmdp_huge_split_prepare(vma, haddr, pmd);
1614         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1615         pmd_populate(mm, &_pmd, pgtable);
1616
1617         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
1618                 pte_t entry, *pte;
1619                 /*
1620                  * Note that NUMA hinting access restrictions are not
1621                  * transferred to avoid any possibility of altering
1622                  * permissions across VMAs.
1623                  */
1624                 if (freeze) {
1625                         swp_entry_t swp_entry;
1626                         swp_entry = make_migration_entry(page + i, write);
1627                         entry = swp_entry_to_pte(swp_entry);
1628                         if (soft_dirty)
1629                                 entry = pte_swp_mksoft_dirty(entry);
1630                 } else {
1631                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
1632                         entry = maybe_mkwrite(entry, vma);
1633                         if (!write)
1634                                 entry = pte_wrprotect(entry);
1635                         if (!young)
1636                                 entry = pte_mkold(entry);
1637                         if (soft_dirty)
1638                                 entry = pte_mksoft_dirty(entry);
1639                 }
1640                 if (dirty)
1641                         SetPageDirty(page + i);
1642                 pte = pte_offset_map(&_pmd, addr);
1643                 BUG_ON(!pte_none(*pte));
1644                 set_pte_at(mm, addr, pte, entry);
1645                 atomic_inc(&page[i]._mapcount);
1646                 pte_unmap(pte);
1647         }
1648
1649         /*
1650          * Set PG_double_map before dropping compound_mapcount to avoid
1651          * false-negative page_mapped().
1652          */
1653         if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
1654                 for (i = 0; i < HPAGE_PMD_NR; i++)
1655                         atomic_inc(&page[i]._mapcount);
1656         }
1657
1658         if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
1659                 /* Last compound_mapcount is gone. */
1660                 __dec_node_page_state(page, NR_ANON_THPS);
1661                 if (TestClearPageDoubleMap(page)) {
1662                         /* No need in mapcount reference anymore */
1663                         for (i = 0; i < HPAGE_PMD_NR; i++)
1664                                 atomic_dec(&page[i]._mapcount);
1665                 }
1666         }
1667
1668         smp_wmb(); /* make pte visible before pmd */
1669         /*
1670          * Up to this point the pmd is present and huge and userland has the
1671          * whole access to the hugepage during the split (which happens in
1672          * place). If we overwrite the pmd with the not-huge version pointing
1673          * to the pte here (which of course we could if all CPUs were bug
1674          * free), userland could trigger a small page size TLB miss on the
1675          * small sized TLB while the hugepage TLB entry is still established in
1676          * the huge TLB. Some CPU doesn't like that.
1677          * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
1678          * 383 on page 93. Intel should be safe but is also warns that it's
1679          * only safe if the permission and cache attributes of the two entries
1680          * loaded in the two TLB is identical (which should be the case here).
1681          * But it is generally safer to never allow small and huge TLB entries
1682          * for the same virtual address to be loaded simultaneously. So instead
1683          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
1684          * current pmd notpresent (atomically because here the pmd_trans_huge
1685          * and pmd_trans_splitting must remain set at all times on the pmd
1686          * until the split is complete for this pmd), then we flush the SMP TLB
1687          * and finally we write the non-huge version of the pmd entry with
1688          * pmd_populate.
1689          */
1690         pmdp_invalidate(vma, haddr, pmd);
1691         pmd_populate(mm, pmd, pgtable);
1692
1693         if (freeze) {
1694                 for (i = 0; i < HPAGE_PMD_NR; i++) {
1695                         page_remove_rmap(page + i, false);
1696                         put_page(page + i);
1697                 }
1698         }
1699 }
1700
1701 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1702                 unsigned long address, bool freeze, struct page *page)
1703 {
1704         spinlock_t *ptl;
1705         struct mm_struct *mm = vma->vm_mm;
1706         unsigned long haddr = address & HPAGE_PMD_MASK;
1707
1708         mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
1709         ptl = pmd_lock(mm, pmd);
1710
1711         /*
1712          * If caller asks to setup a migration entries, we need a page to check
1713          * pmd against. Otherwise we can end up replacing wrong page.
1714          */
1715         VM_BUG_ON(freeze && !page);
1716         if (page && page != pmd_page(*pmd))
1717                 goto out;
1718
1719         if (pmd_trans_huge(*pmd)) {
1720                 page = pmd_page(*pmd);
1721                 if (PageMlocked(page))
1722                         clear_page_mlock(page);
1723         } else if (!pmd_devmap(*pmd))
1724                 goto out;
1725         __split_huge_pmd_locked(vma, pmd, haddr, freeze);
1726 out:
1727         spin_unlock(ptl);
1728         mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
1729 }
1730
1731 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
1732                 bool freeze, struct page *page)
1733 {
1734         pgd_t *pgd;
1735         pud_t *pud;
1736         pmd_t *pmd;
1737
1738         pgd = pgd_offset(vma->vm_mm, address);
1739         if (!pgd_present(*pgd))
1740                 return;
1741
1742         pud = pud_offset(pgd, address);
1743         if (!pud_present(*pud))
1744                 return;
1745
1746         pmd = pmd_offset(pud, address);
1747
1748         __split_huge_pmd(vma, pmd, address, freeze, page);
1749 }
1750
1751 void vma_adjust_trans_huge(struct vm_area_struct *vma,
1752                              unsigned long start,
1753                              unsigned long end,
1754                              long adjust_next)
1755 {
1756         /*
1757          * If the new start address isn't hpage aligned and it could
1758          * previously contain an hugepage: check if we need to split
1759          * an huge pmd.
1760          */
1761         if (start & ~HPAGE_PMD_MASK &&
1762             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
1763             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
1764                 split_huge_pmd_address(vma, start, false, NULL);
1765
1766         /*
1767          * If the new end address isn't hpage aligned and it could
1768          * previously contain an hugepage: check if we need to split
1769          * an huge pmd.
1770          */
1771         if (end & ~HPAGE_PMD_MASK &&
1772             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
1773             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
1774                 split_huge_pmd_address(vma, end, false, NULL);
1775
1776         /*
1777          * If we're also updating the vma->vm_next->vm_start, if the new
1778          * vm_next->vm_start isn't page aligned and it could previously
1779          * contain an hugepage: check if we need to split an huge pmd.
1780          */
1781         if (adjust_next > 0) {
1782                 struct vm_area_struct *next = vma->vm_next;
1783                 unsigned long nstart = next->vm_start;
1784                 nstart += adjust_next << PAGE_SHIFT;
1785                 if (nstart & ~HPAGE_PMD_MASK &&
1786                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
1787                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
1788                         split_huge_pmd_address(next, nstart, false, NULL);
1789         }
1790 }
1791
1792 static void freeze_page(struct page *page)
1793 {
1794         enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
1795                 TTU_RMAP_LOCKED;
1796         int i, ret;
1797
1798         VM_BUG_ON_PAGE(!PageHead(page), page);
1799
1800         if (PageAnon(page))
1801                 ttu_flags |= TTU_MIGRATION;
1802
1803         /* We only need TTU_SPLIT_HUGE_PMD once */
1804         ret = try_to_unmap(page, ttu_flags | TTU_SPLIT_HUGE_PMD);
1805         for (i = 1; !ret && i < HPAGE_PMD_NR; i++) {
1806                 /* Cut short if the page is unmapped */
1807                 if (page_count(page) == 1)
1808                         return;
1809
1810                 ret = try_to_unmap(page + i, ttu_flags);
1811         }
1812         VM_BUG_ON_PAGE(ret, page + i - 1);
1813 }
1814
1815 static void unfreeze_page(struct page *page)
1816 {
1817         int i;
1818
1819         for (i = 0; i < HPAGE_PMD_NR; i++)
1820                 remove_migration_ptes(page + i, page + i, true);
1821 }
1822
1823 static void __split_huge_page_tail(struct page *head, int tail,
1824                 struct lruvec *lruvec, struct list_head *list)
1825 {
1826         struct page *page_tail = head + tail;
1827
1828         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
1829         VM_BUG_ON_PAGE(page_ref_count(page_tail) != 0, page_tail);
1830
1831         /*
1832          * tail_page->_refcount is zero and not changing from under us. But
1833          * get_page_unless_zero() may be running from under us on the
1834          * tail_page. If we used atomic_set() below instead of atomic_inc() or
1835          * atomic_add(), we would then run atomic_set() concurrently with
1836          * get_page_unless_zero(), and atomic_set() is implemented in C not
1837          * using locked ops. spin_unlock on x86 sometime uses locked ops
1838          * because of PPro errata 66, 92, so unless somebody can guarantee
1839          * atomic_set() here would be safe on all archs (and not only on x86),
1840          * it's safer to use atomic_inc()/atomic_add().
1841          */
1842         if (PageAnon(head)) {
1843                 page_ref_inc(page_tail);
1844         } else {
1845                 /* Additional pin to radix tree */
1846                 page_ref_add(page_tail, 2);
1847         }
1848
1849         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1850         page_tail->flags |= (head->flags &
1851                         ((1L << PG_referenced) |
1852                          (1L << PG_swapbacked) |
1853                          (1L << PG_mlocked) |
1854                          (1L << PG_uptodate) |
1855                          (1L << PG_active) |
1856                          (1L << PG_locked) |
1857                          (1L << PG_unevictable) |
1858                          (1L << PG_dirty)));
1859
1860         /*
1861          * After clearing PageTail the gup refcount can be released.
1862          * Page flags also must be visible before we make the page non-compound.
1863          */
1864         smp_wmb();
1865
1866         clear_compound_head(page_tail);
1867
1868         if (page_is_young(head))
1869                 set_page_young(page_tail);
1870         if (page_is_idle(head))
1871                 set_page_idle(page_tail);
1872
1873         /* ->mapping in first tail page is compound_mapcount */
1874         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
1875                         page_tail);
1876         page_tail->mapping = head->mapping;
1877
1878         page_tail->index = head->index + tail;
1879         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
1880         lru_add_page_tail(head, page_tail, lruvec, list);
1881 }
1882
1883 static void __split_huge_page(struct page *page, struct list_head *list,
1884                 unsigned long flags)
1885 {
1886         struct page *head = compound_head(page);
1887         struct zone *zone = page_zone(head);
1888         struct lruvec *lruvec;
1889         pgoff_t end = -1;
1890         int i;
1891
1892         lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat);
1893
1894         /* complete memcg works before add pages to LRU */
1895         mem_cgroup_split_huge_fixup(head);
1896
1897         if (!PageAnon(page))
1898                 end = DIV_ROUND_UP(i_size_read(head->mapping->host), PAGE_SIZE);
1899
1900         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1901                 __split_huge_page_tail(head, i, lruvec, list);
1902                 /* Some pages can be beyond i_size: drop them from page cache */
1903                 if (head[i].index >= end) {
1904                         __ClearPageDirty(head + i);
1905                         __delete_from_page_cache(head + i, NULL);
1906                         if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
1907                                 shmem_uncharge(head->mapping->host, 1);
1908                         put_page(head + i);
1909                 }
1910         }
1911
1912         ClearPageCompound(head);
1913         /* See comment in __split_huge_page_tail() */
1914         if (PageAnon(head)) {
1915                 page_ref_inc(head);
1916         } else {
1917                 /* Additional pin to radix tree */
1918                 page_ref_add(head, 2);
1919                 spin_unlock(&head->mapping->tree_lock);
1920         }
1921
1922         spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
1923
1924         unfreeze_page(head);
1925
1926         for (i = 0; i < HPAGE_PMD_NR; i++) {
1927                 struct page *subpage = head + i;
1928                 if (subpage == page)
1929                         continue;
1930                 unlock_page(subpage);
1931
1932                 /*
1933                  * Subpages may be freed if there wasn't any mapping
1934                  * like if add_to_swap() is running on a lru page that
1935                  * had its mapping zapped. And freeing these pages
1936                  * requires taking the lru_lock so we do the put_page
1937                  * of the tail pages after the split is complete.
1938                  */
1939                 put_page(subpage);
1940         }
1941 }
1942
1943 int total_mapcount(struct page *page)
1944 {
1945         int i, compound, ret;
1946
1947         VM_BUG_ON_PAGE(PageTail(page), page);
1948
1949         if (likely(!PageCompound(page)))
1950                 return atomic_read(&page->_mapcount) + 1;
1951
1952         compound = compound_mapcount(page);
1953         if (PageHuge(page))
1954                 return compound;
1955         ret = compound;
1956         for (i = 0; i < HPAGE_PMD_NR; i++)
1957                 ret += atomic_read(&page[i]._mapcount) + 1;
1958         /* File pages has compound_mapcount included in _mapcount */
1959         if (!PageAnon(page))
1960                 return ret - compound * HPAGE_PMD_NR;
1961         if (PageDoubleMap(page))
1962                 ret -= HPAGE_PMD_NR;
1963         return ret;
1964 }
1965
1966 /*
1967  * This calculates accurately how many mappings a transparent hugepage
1968  * has (unlike page_mapcount() which isn't fully accurate). This full
1969  * accuracy is primarily needed to know if copy-on-write faults can
1970  * reuse the page and change the mapping to read-write instead of
1971  * copying them. At the same time this returns the total_mapcount too.
1972  *
1973  * The function returns the highest mapcount any one of the subpages
1974  * has. If the return value is one, even if different processes are
1975  * mapping different subpages of the transparent hugepage, they can
1976  * all reuse it, because each process is reusing a different subpage.
1977  *
1978  * The total_mapcount is instead counting all virtual mappings of the
1979  * subpages. If the total_mapcount is equal to "one", it tells the
1980  * caller all mappings belong to the same "mm" and in turn the
1981  * anon_vma of the transparent hugepage can become the vma->anon_vma
1982  * local one as no other process may be mapping any of the subpages.
1983  *
1984  * It would be more accurate to replace page_mapcount() with
1985  * page_trans_huge_mapcount(), however we only use
1986  * page_trans_huge_mapcount() in the copy-on-write faults where we
1987  * need full accuracy to avoid breaking page pinning, because
1988  * page_trans_huge_mapcount() is slower than page_mapcount().
1989  */
1990 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
1991 {
1992         int i, ret, _total_mapcount, mapcount;
1993
1994         /* hugetlbfs shouldn't call it */
1995         VM_BUG_ON_PAGE(PageHuge(page), page);
1996
1997         if (likely(!PageTransCompound(page))) {
1998                 mapcount = atomic_read(&page->_mapcount) + 1;
1999                 if (total_mapcount)
2000                         *total_mapcount = mapcount;
2001                 return mapcount;
2002         }
2003
2004         page = compound_head(page);
2005
2006         _total_mapcount = ret = 0;
2007         for (i = 0; i < HPAGE_PMD_NR; i++) {
2008                 mapcount = atomic_read(&page[i]._mapcount) + 1;
2009                 ret = max(ret, mapcount);
2010                 _total_mapcount += mapcount;
2011         }
2012         if (PageDoubleMap(page)) {
2013                 ret -= 1;
2014                 _total_mapcount -= HPAGE_PMD_NR;
2015         }
2016         mapcount = compound_mapcount(page);
2017         ret += mapcount;
2018         _total_mapcount += mapcount;
2019         if (total_mapcount)
2020                 *total_mapcount = _total_mapcount;
2021         return ret;
2022 }
2023
2024 /*
2025  * This function splits huge page into normal pages. @page can point to any
2026  * subpage of huge page to split. Split doesn't change the position of @page.
2027  *
2028  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2029  * The huge page must be locked.
2030  *
2031  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2032  *
2033  * Both head page and tail pages will inherit mapping, flags, and so on from
2034  * the hugepage.
2035  *
2036  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2037  * they are not mapped.
2038  *
2039  * Returns 0 if the hugepage is split successfully.
2040  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2041  * us.
2042  */
2043 int split_huge_page_to_list(struct page *page, struct list_head *list)
2044 {
2045         struct page *head = compound_head(page);
2046         struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2047         struct anon_vma *anon_vma = NULL;
2048         struct address_space *mapping = NULL;
2049         int count, mapcount, extra_pins, ret;
2050         bool mlocked;
2051         unsigned long flags;
2052
2053         VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
2054         VM_BUG_ON_PAGE(!PageLocked(page), page);
2055         VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2056         VM_BUG_ON_PAGE(!PageCompound(page), page);
2057
2058         if (PageAnon(head)) {
2059                 /*
2060                  * The caller does not necessarily hold an mmap_sem that would
2061                  * prevent the anon_vma disappearing so we first we take a
2062                  * reference to it and then lock the anon_vma for write. This
2063                  * is similar to page_lock_anon_vma_read except the write lock
2064                  * is taken to serialise against parallel split or collapse
2065                  * operations.
2066                  */
2067                 anon_vma = page_get_anon_vma(head);
2068                 if (!anon_vma) {
2069                         ret = -EBUSY;
2070                         goto out;
2071                 }
2072                 extra_pins = 0;
2073                 mapping = NULL;
2074                 anon_vma_lock_write(anon_vma);
2075         } else {
2076                 mapping = head->mapping;
2077
2078                 /* Truncated ? */
2079                 if (!mapping) {
2080                         ret = -EBUSY;
2081                         goto out;
2082                 }
2083
2084                 /* Addidional pins from radix tree */
2085                 extra_pins = HPAGE_PMD_NR;
2086                 anon_vma = NULL;
2087                 i_mmap_lock_read(mapping);
2088         }
2089
2090         /*
2091          * Racy check if we can split the page, before freeze_page() will
2092          * split PMDs
2093          */
2094         if (total_mapcount(head) != page_count(head) - extra_pins - 1) {
2095                 ret = -EBUSY;
2096                 goto out_unlock;
2097         }
2098
2099         mlocked = PageMlocked(page);
2100         freeze_page(head);
2101         VM_BUG_ON_PAGE(compound_mapcount(head), head);
2102
2103         /* Make sure the page is not on per-CPU pagevec as it takes pin */
2104         if (mlocked)
2105                 lru_add_drain();
2106
2107         /* prevent PageLRU to go away from under us, and freeze lru stats */
2108         spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags);
2109
2110         if (mapping) {
2111                 void **pslot;
2112
2113                 spin_lock(&mapping->tree_lock);
2114                 pslot = radix_tree_lookup_slot(&mapping->page_tree,
2115                                 page_index(head));
2116                 /*
2117                  * Check if the head page is present in radix tree.
2118                  * We assume all tail are present too, if head is there.
2119                  */
2120                 if (radix_tree_deref_slot_protected(pslot,
2121                                         &mapping->tree_lock) != head)
2122                         goto fail;
2123         }
2124
2125         /* Prevent deferred_split_scan() touching ->_refcount */
2126         spin_lock(&pgdata->split_queue_lock);
2127         count = page_count(head);
2128         mapcount = total_mapcount(head);
2129         if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2130                 if (!list_empty(page_deferred_list(head))) {
2131                         pgdata->split_queue_len--;
2132                         list_del(page_deferred_list(head));
2133                 }
2134                 if (mapping)
2135                         __dec_node_page_state(page, NR_SHMEM_THPS);
2136                 spin_unlock(&pgdata->split_queue_lock);
2137                 __split_huge_page(page, list, flags);
2138                 ret = 0;
2139         } else {
2140                 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2141                         pr_alert("total_mapcount: %u, page_count(): %u\n",
2142                                         mapcount, count);
2143                         if (PageTail(page))
2144                                 dump_page(head, NULL);
2145                         dump_page(page, "total_mapcount(head) > 0");
2146                         BUG();
2147                 }
2148                 spin_unlock(&pgdata->split_queue_lock);
2149 fail:           if (mapping)
2150                         spin_unlock(&mapping->tree_lock);
2151                 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
2152                 unfreeze_page(head);
2153                 ret = -EBUSY;
2154         }
2155
2156 out_unlock:
2157         if (anon_vma) {
2158                 anon_vma_unlock_write(anon_vma);
2159                 put_anon_vma(anon_vma);
2160         }
2161         if (mapping)
2162                 i_mmap_unlock_read(mapping);
2163 out:
2164         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2165         return ret;
2166 }
2167
2168 void free_transhuge_page(struct page *page)
2169 {
2170         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2171         unsigned long flags;
2172
2173         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2174         if (!list_empty(page_deferred_list(page))) {
2175                 pgdata->split_queue_len--;
2176                 list_del(page_deferred_list(page));
2177         }
2178         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2179         free_compound_page(page);
2180 }
2181
2182 void deferred_split_huge_page(struct page *page)
2183 {
2184         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2185         unsigned long flags;
2186
2187         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2188
2189         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2190         if (list_empty(page_deferred_list(page))) {
2191                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2192                 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2193                 pgdata->split_queue_len++;
2194         }
2195         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2196 }
2197
2198 static unsigned long deferred_split_count(struct shrinker *shrink,
2199                 struct shrink_control *sc)
2200 {
2201         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2202         return ACCESS_ONCE(pgdata->split_queue_len);
2203 }
2204
2205 static unsigned long deferred_split_scan(struct shrinker *shrink,
2206                 struct shrink_control *sc)
2207 {
2208         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2209         unsigned long flags;
2210         LIST_HEAD(list), *pos, *next;
2211         struct page *page;
2212         int split = 0;
2213
2214         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2215         /* Take pin on all head pages to avoid freeing them under us */
2216         list_for_each_safe(pos, next, &pgdata->split_queue) {
2217                 page = list_entry((void *)pos, struct page, mapping);
2218                 page = compound_head(page);
2219                 if (get_page_unless_zero(page)) {
2220                         list_move(page_deferred_list(page), &list);
2221                 } else {
2222                         /* We lost race with put_compound_page() */
2223                         list_del_init(page_deferred_list(page));
2224                         pgdata->split_queue_len--;
2225                 }
2226                 if (!--sc->nr_to_scan)
2227                         break;
2228         }
2229         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2230
2231         list_for_each_safe(pos, next, &list) {
2232                 page = list_entry((void *)pos, struct page, mapping);
2233                 lock_page(page);
2234                 /* split_huge_page() removes page from list on success */
2235                 if (!split_huge_page(page))
2236                         split++;
2237                 unlock_page(page);
2238                 put_page(page);
2239         }
2240
2241         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2242         list_splice_tail(&list, &pgdata->split_queue);
2243         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2244
2245         /*
2246          * Stop shrinker if we didn't split any page, but the queue is empty.
2247          * This can happen if pages were freed under us.
2248          */
2249         if (!split && list_empty(&pgdata->split_queue))
2250                 return SHRINK_STOP;
2251         return split;
2252 }
2253
2254 static struct shrinker deferred_split_shrinker = {
2255         .count_objects = deferred_split_count,
2256         .scan_objects = deferred_split_scan,
2257         .seeks = DEFAULT_SEEKS,
2258         .flags = SHRINKER_NUMA_AWARE,
2259 };
2260
2261 #ifdef CONFIG_DEBUG_FS
2262 static int split_huge_pages_set(void *data, u64 val)
2263 {
2264         struct zone *zone;
2265         struct page *page;
2266         unsigned long pfn, max_zone_pfn;
2267         unsigned long total = 0, split = 0;
2268
2269         if (val != 1)
2270                 return -EINVAL;
2271
2272         for_each_populated_zone(zone) {
2273                 max_zone_pfn = zone_end_pfn(zone);
2274                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2275                         if (!pfn_valid(pfn))
2276                                 continue;
2277
2278                         page = pfn_to_page(pfn);
2279                         if (!get_page_unless_zero(page))
2280                                 continue;
2281
2282                         if (zone != page_zone(page))
2283                                 goto next;
2284
2285                         if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2286                                 goto next;
2287
2288                         total++;
2289                         lock_page(page);
2290                         if (!split_huge_page(page))
2291                                 split++;
2292                         unlock_page(page);
2293 next:
2294                         put_page(page);
2295                 }
2296         }
2297
2298         pr_info("%lu of %lu THP split\n", split, total);
2299
2300         return 0;
2301 }
2302 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2303                 "%llu\n");
2304
2305 static int __init split_huge_pages_debugfs(void)
2306 {
2307         void *ret;
2308
2309         ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2310                         &split_huge_pages_fops);
2311         if (!ret)
2312                 pr_warn("Failed to create split_huge_pages in debugfs");
2313         return 0;
2314 }
2315 late_initcall(split_huge_pages_debugfs);
2316 #endif