kernel/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/kthread.h>
  20 #include <linux/khugepaged.h>
  21 #include <linux/freezer.h>
  22 #include <linux/mman.h>
  23 #include <linux/pagemap.h>
  24 #include <linux/migrate.h>
  25 #include <linux/hashtable.h>
  26
  27 #include <asm/tlb.h>
  28 #include <asm/pgalloc.h>
  29 #include "internal.h"
  30
  31 /*
  32  * By default transparent hugepage support is disabled in order that avoid
  33  * to risk increase the memory footprint of applications without a guaranteed
  34  * benefit. When transparent hugepage support is enabled, is for all mappings,
  35  * and khugepaged scans all mappings.
  36  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  37  * for all hugepage allocations.
  38  */
  39 unsigned long transparent_hugepage_flags __read_mostly =
  40 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  41         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  42 #endif
  43 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  44         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  45 #endif
  46         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
  47         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  48         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  49
  50 /* default scan 8*512 pte (or vmas) every 30 second */
  51 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
  52 static unsigned int khugepaged_pages_collapsed;
  53 static unsigned int khugepaged_full_scans;
  54 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  55 /* during fragmentation poll the hugepage allocator once every minute */
  56 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  57 static struct task_struct *khugepaged_thread __read_mostly;
  58 static DEFINE_MUTEX(khugepaged_mutex);
  59 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  60 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  61 /*
  62  * default collapse hugepages if there is at least one pte mapped like
  63  * it would have happened if the vma was large enough during page
  64  * fault.
  65  */
  66 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
  67
  68 static int khugepaged(void *none);
  69 static int khugepaged_slab_init(void);
  70 static void khugepaged_slab_exit(void);
  71
  72 #define MM_SLOTS_HASH_BITS 10
  73 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
  74
  75 static struct kmem_cache *mm_slot_cache __read_mostly;
  76
  77 /**
  78  * struct mm_slot - hash lookup from mm to mm_slot
  79  * @hash: hash collision list
  80  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  81  * @mm: the mm that this information is valid for
  82  */
  83 struct mm_slot {
  84         struct hlist_node hash;
  85         struct list_head mm_node;
  86         struct mm_struct *mm;
  87 };
  88
  89 /**
  90  * struct khugepaged_scan - cursor for scanning
  91  * @mm_head: the head of the mm list to scan
  92  * @mm_slot: the current mm_slot we are scanning
  93  * @address: the next address inside that to be scanned
  94  *
  95  * There is only the one khugepaged_scan instance of this cursor structure.
  96  */
  97 struct khugepaged_scan {
  98         struct list_head mm_head;
  99         struct mm_slot *mm_slot;
 100         unsigned long address;
 101 };
 102 static struct khugepaged_scan khugepaged_scan = {
 103         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 104 };
 105
 106
 107 static int set_recommended_min_free_kbytes(void)
 108 {
 109         struct zone *zone;
 110         int nr_zones = 0;
 111         unsigned long recommended_min;
 112
 113         for_each_populated_zone(zone)
 114                 nr_zones++;
 115
 116         /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
 117         recommended_min = pageblock_nr_pages * nr_zones * 2;
 118
 119         /*
 120          * Make sure that on average at least two pageblocks are almost free
 121          * of another type, one for a migratetype to fall back to and a
 122          * second to avoid subsequent fallbacks of other types There are 3
 123          * MIGRATE_TYPES we care about.
 124          */
 125         recommended_min += pageblock_nr_pages * nr_zones *
 126                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
 127
 128         /* don't ever allow to reserve more than 5% of the lowmem */
 129         recommended_min = min(recommended_min,
 130                               (unsigned long) nr_free_buffer_pages() / 20);
 131         recommended_min <<= (PAGE_SHIFT-10);
 132
 133         if (recommended_min > min_free_kbytes) {
 134                 if (user_min_free_kbytes >= 0)
 135                         pr_info("raising min_free_kbytes from %d to %lu "
 136                                 "to help transparent hugepage allocations\n",
 137                                 min_free_kbytes, recommended_min);
 138
 139                 min_free_kbytes = recommended_min;
 140         }
 141         setup_per_zone_wmarks();
 142         return 0;
 143 }
 144
 145 static int start_stop_khugepaged(void)
 146 {
 147         int err = 0;
 148         if (khugepaged_enabled()) {
 149                 if (!khugepaged_thread)
 150                         khugepaged_thread = kthread_run(khugepaged, NULL,
 151                                                         "khugepaged");
 152                 if (unlikely(IS_ERR(khugepaged_thread))) {
 153                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
 154                         err = PTR_ERR(khugepaged_thread);
 155                         khugepaged_thread = NULL;
 156                         goto fail;
 157                 }
 158
 159                 if (!list_empty(&khugepaged_scan.mm_head))
 160                         wake_up_interruptible(&khugepaged_wait);
 161
 162                 set_recommended_min_free_kbytes();
 163         } else if (khugepaged_thread) {
 164                 kthread_stop(khugepaged_thread);
 165                 khugepaged_thread = NULL;
 166         }
 167 fail:
 168         return err;
 169 }
 170
 171 static atomic_t huge_zero_refcount;
 172 struct page *huge_zero_page __read_mostly;
 173
 174 static inline bool is_huge_zero_pmd(pmd_t pmd)
 175 {
 176         return is_huge_zero_page(pmd_page(pmd));
 177 }
 178
 179 static struct page *get_huge_zero_page(void)
 180 {
 181         struct page *zero_page;
 182 retry:
 183         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
 184                 return READ_ONCE(huge_zero_page);
 185
 186         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
 187                         HPAGE_PMD_ORDER);
 188         if (!zero_page) {
 189                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
 190                 return NULL;
 191         }
 192         count_vm_event(THP_ZERO_PAGE_ALLOC);
 193         preempt_disable();
 194         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
 195                 preempt_enable();
 196                 __free_pages(zero_page, compound_order(zero_page));
 197                 goto retry;
 198         }
 199
 200         /* We take additional reference here. It will be put back by shrinker */
 201         atomic_set(&huge_zero_refcount, 2);
 202         preempt_enable();
 203         return READ_ONCE(huge_zero_page);
 204 }
 205
 206 static void put_huge_zero_page(void)
 207 {
 208         /*
 209          * Counter should never go to zero here. Only shrinker can put
 210          * last reference.
 211          */
 212         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 213 }
 214
 215 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 216                                         struct shrink_control *sc)
 217 {
 218         /* we can free zero page only if last reference remains */
 219         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 220 }
 221
 222 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 223                                        struct shrink_control *sc)
 224 {
 225         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 226                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 227                 BUG_ON(zero_page == NULL);
 228                 __free_pages(zero_page, compound_order(zero_page));
 229                 return HPAGE_PMD_NR;
 230         }
 231
 232         return 0;
 233 }
 234
 235 static struct shrinker huge_zero_page_shrinker = {
 236         .count_objects = shrink_huge_zero_page_count,
 237         .scan_objects = shrink_huge_zero_page_scan,
 238         .seeks = DEFAULT_SEEKS,
 239 };
 240
 241 #ifdef CONFIG_SYSFS
 242
 243 static ssize_t double_flag_show(struct kobject *kobj,
 244                                 struct kobj_attribute *attr, char *buf,
 245                                 enum transparent_hugepage_flag enabled,
 246                                 enum transparent_hugepage_flag req_madv)
 247 {
 248         if (test_bit(enabled, &transparent_hugepage_flags)) {
 249                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
 250                 return sprintf(buf, "[always] madvise never\n");
 251         } else if (test_bit(req_madv, &transparent_hugepage_flags))
 252                 return sprintf(buf, "always [madvise] never\n");
 253         else
 254                 return sprintf(buf, "always madvise [never]\n");
 255 }
 256 static ssize_t double_flag_store(struct kobject *kobj,
 257                                  struct kobj_attribute *attr,
 258                                  const char *buf, size_t count,
 259                                  enum transparent_hugepage_flag enabled,
 260                                  enum transparent_hugepage_flag req_madv)
 261 {
 262         if (!memcmp("always", buf,
 263                     min(sizeof("always")-1, count))) {
 264                 set_bit(enabled, &transparent_hugepage_flags);
 265                 clear_bit(req_madv, &transparent_hugepage_flags);
 266         } else if (!memcmp("madvise", buf,
 267                            min(sizeof("madvise")-1, count))) {
 268                 clear_bit(enabled, &transparent_hugepage_flags);
 269                 set_bit(req_madv, &transparent_hugepage_flags);
 270         } else if (!memcmp("never", buf,
 271                            min(sizeof("never")-1, count))) {
 272                 clear_bit(enabled, &transparent_hugepage_flags);
 273                 clear_bit(req_madv, &transparent_hugepage_flags);
 274         } else
 275                 return -EINVAL;
 276
 277         return count;
 278 }
 279
 280 static ssize_t enabled_show(struct kobject *kobj,
 281                             struct kobj_attribute *attr, char *buf)
 282 {
 283         return double_flag_show(kobj, attr, buf,
 284                                 TRANSPARENT_HUGEPAGE_FLAG,
 285                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 286 }
 287 static ssize_t enabled_store(struct kobject *kobj,
 288                              struct kobj_attribute *attr,
 289                              const char *buf, size_t count)
 290 {
 291         ssize_t ret;
 292
 293         ret = double_flag_store(kobj, attr, buf, count,
 294                                 TRANSPARENT_HUGEPAGE_FLAG,
 295                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 296
 297         if (ret > 0) {
 298                 int err;
 299
 300                 mutex_lock(&khugepaged_mutex);
 301                 err = start_stop_khugepaged();
 302                 mutex_unlock(&khugepaged_mutex);
 303
 304                 if (err)
 305                         ret = err;
 306         }
 307
 308         return ret;
 309 }
 310 static struct kobj_attribute enabled_attr =
 311         __ATTR(enabled, 0644, enabled_show, enabled_store);
 312
 313 static ssize_t single_flag_show(struct kobject *kobj,
 314                                 struct kobj_attribute *attr, char *buf,
 315                                 enum transparent_hugepage_flag flag)
 316 {
 317         return sprintf(buf, "%d\n",
 318                        !!test_bit(flag, &transparent_hugepage_flags));
 319 }
 320
 321 static ssize_t single_flag_store(struct kobject *kobj,
 322                                  struct kobj_attribute *attr,
 323                                  const char *buf, size_t count,
 324                                  enum transparent_hugepage_flag flag)
 325 {
 326         unsigned long value;
 327         int ret;
 328
 329         ret = kstrtoul(buf, 10, &value);
 330         if (ret < 0)
 331                 return ret;
 332         if (value > 1)
 333                 return -EINVAL;
 334
 335         if (value)
 336                 set_bit(flag, &transparent_hugepage_flags);
 337         else
 338                 clear_bit(flag, &transparent_hugepage_flags);
 339
 340         return count;
 341 }
 342
 343 /*
 344  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 345  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 346  * memory just to allocate one more hugepage.
 347  */
 348 static ssize_t defrag_show(struct kobject *kobj,
 349                            struct kobj_attribute *attr, char *buf)
 350 {
 351         return double_flag_show(kobj, attr, buf,
 352                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 353                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 354 }
 355 static ssize_t defrag_store(struct kobject *kobj,
 356                             struct kobj_attribute *attr,
 357                             const char *buf, size_t count)
 358 {
 359         return double_flag_store(kobj, attr, buf, count,
 360                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 361                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 362 }
 363 static struct kobj_attribute defrag_attr =
 364         __ATTR(defrag, 0644, defrag_show, defrag_store);
 365
 366 static ssize_t use_zero_page_show(struct kobject *kobj,
 367                 struct kobj_attribute *attr, char *buf)
 368 {
 369         return single_flag_show(kobj, attr, buf,
 370                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 371 }
 372 static ssize_t use_zero_page_store(struct kobject *kobj,
 373                 struct kobj_attribute *attr, const char *buf, size_t count)
 374 {
 375         return single_flag_store(kobj, attr, buf, count,
 376                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 377 }
 378 static struct kobj_attribute use_zero_page_attr =
 379         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 380 #ifdef CONFIG_DEBUG_VM
 381 static ssize_t debug_cow_show(struct kobject *kobj,
 382                                 struct kobj_attribute *attr, char *buf)
 383 {
 384         return single_flag_show(kobj, attr, buf,
 385                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 386 }
 387 static ssize_t debug_cow_store(struct kobject *kobj,
 388                                struct kobj_attribute *attr,
 389                                const char *buf, size_t count)
 390 {
 391         return single_flag_store(kobj, attr, buf, count,
 392                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 393 }
 394 static struct kobj_attribute debug_cow_attr =
 395         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 396 #endif /* CONFIG_DEBUG_VM */
 397
 398 static struct attribute *hugepage_attr[] = {
 399         &enabled_attr.attr,
 400         &defrag_attr.attr,
 401         &use_zero_page_attr.attr,
 402 #ifdef CONFIG_DEBUG_VM
 403         &debug_cow_attr.attr,
 404 #endif
 405         NULL,
 406 };
 407
 408 static struct attribute_group hugepage_attr_group = {
 409         .attrs = hugepage_attr,
 410 };
 411
 412 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 413                                          struct kobj_attribute *attr,
 414                                          char *buf)
 415 {
 416         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 417 }
 418
 419 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 420                                           struct kobj_attribute *attr,
 421                                           const char *buf, size_t count)
 422 {
 423         unsigned long msecs;
 424         int err;
 425
 426         err = kstrtoul(buf, 10, &msecs);
 427         if (err || msecs > UINT_MAX)
 428                 return -EINVAL;
 429
 430         khugepaged_scan_sleep_millisecs = msecs;
 431         wake_up_interruptible(&khugepaged_wait);
 432
 433         return count;
 434 }
 435 static struct kobj_attribute scan_sleep_millisecs_attr =
 436         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 437                scan_sleep_millisecs_store);
 438
 439 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 440                                           struct kobj_attribute *attr,
 441                                           char *buf)
 442 {
 443         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 444 }
 445
 446 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 447                                            struct kobj_attribute *attr,
 448                                            const char *buf, size_t count)
 449 {
 450         unsigned long msecs;
 451         int err;
 452
 453         err = kstrtoul(buf, 10, &msecs);
 454         if (err || msecs > UINT_MAX)
 455                 return -EINVAL;
 456
 457         khugepaged_alloc_sleep_millisecs = msecs;
 458         wake_up_interruptible(&khugepaged_wait);
 459
 460         return count;
 461 }
 462 static struct kobj_attribute alloc_sleep_millisecs_attr =
 463         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 464                alloc_sleep_millisecs_store);
 465
 466 static ssize_t pages_to_scan_show(struct kobject *kobj,
 467                                   struct kobj_attribute *attr,
 468                                   char *buf)
 469 {
 470         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 471 }
 472 static ssize_t pages_to_scan_store(struct kobject *kobj,
 473                                    struct kobj_attribute *attr,
 474                                    const char *buf, size_t count)
 475 {
 476         int err;
 477         unsigned long pages;
 478
 479         err = kstrtoul(buf, 10, &pages);
 480         if (err || !pages || pages > UINT_MAX)
 481                 return -EINVAL;
 482
 483         khugepaged_pages_to_scan = pages;
 484
 485         return count;
 486 }
 487 static struct kobj_attribute pages_to_scan_attr =
 488         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 489                pages_to_scan_store);
 490
 491 static ssize_t pages_collapsed_show(struct kobject *kobj,
 492                                     struct kobj_attribute *attr,
 493                                     char *buf)
 494 {
 495         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 496 }
 497 static struct kobj_attribute pages_collapsed_attr =
 498         __ATTR_RO(pages_collapsed);
 499
 500 static ssize_t full_scans_show(struct kobject *kobj,
 501                                struct kobj_attribute *attr,
 502                                char *buf)
 503 {
 504         return sprintf(buf, "%u\n", khugepaged_full_scans);
 505 }
 506 static struct kobj_attribute full_scans_attr =
 507         __ATTR_RO(full_scans);
 508
 509 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 510                                       struct kobj_attribute *attr, char *buf)
 511 {
 512         return single_flag_show(kobj, attr, buf,
 513                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 514 }
 515 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 516                                        struct kobj_attribute *attr,
 517                                        const char *buf, size_t count)
 518 {
 519         return single_flag_store(kobj, attr, buf, count,
 520                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 521 }
 522 static struct kobj_attribute khugepaged_defrag_attr =
 523         __ATTR(defrag, 0644, khugepaged_defrag_show,
 524                khugepaged_defrag_store);
 525
 526 /*
 527  * max_ptes_none controls if khugepaged should collapse hugepages over
 528  * any unmapped ptes in turn potentially increasing the memory
 529  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 530  * reduce the available free memory in the system as it
 531  * runs. Increasing max_ptes_none will instead potentially reduce the
 532  * free memory in the system during the khugepaged scan.
 533  */
 534 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 535                                              struct kobj_attribute *attr,
 536                                              char *buf)
 537 {
 538         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 539 }
 540 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 541                                               struct kobj_attribute *attr,
 542                                               const char *buf, size_t count)
 543 {
 544         int err;
 545         unsigned long max_ptes_none;
 546
 547         err = kstrtoul(buf, 10, &max_ptes_none);
 548         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 549                 return -EINVAL;
 550
 551         khugepaged_max_ptes_none = max_ptes_none;
 552
 553         return count;
 554 }
 555 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 556         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 557                khugepaged_max_ptes_none_store);
 558
 559 static struct attribute *khugepaged_attr[] = {
 560         &khugepaged_defrag_attr.attr,
 561         &khugepaged_max_ptes_none_attr.attr,
 562         &pages_to_scan_attr.attr,
 563         &pages_collapsed_attr.attr,
 564         &full_scans_attr.attr,
 565         &scan_sleep_millisecs_attr.attr,
 566         &alloc_sleep_millisecs_attr.attr,
 567         NULL,
 568 };
 569
 570 static struct attribute_group khugepaged_attr_group = {
 571         .attrs = khugepaged_attr,
 572         .name = "khugepaged",
 573 };
 574
 575 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 576 {
 577         int err;
 578
 579         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 580         if (unlikely(!*hugepage_kobj)) {
 581                 pr_err("failed to create transparent hugepage kobject\n");
 582                 return -ENOMEM;
 583         }
 584
 585         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 586         if (err) {
 587                 pr_err("failed to register transparent hugepage group\n");
 588                 goto delete_obj;
 589         }
 590
 591         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 592         if (err) {
 593                 pr_err("failed to register transparent hugepage group\n");
 594                 goto remove_hp_group;
 595         }
 596
 597         return 0;
 598
 599 remove_hp_group:
 600         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 601 delete_obj:
 602         kobject_put(*hugepage_kobj);
 603         return err;
 604 }
 605
 606 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 607 {
 608         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 609         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 610         kobject_put(hugepage_kobj);
 611 }
 612 #else
 613 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 614 {
 615         return 0;
 616 }
 617
 618 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 619 {
 620 }
 621 #endif /* CONFIG_SYSFS */
 622
 623 static int __init hugepage_init(void)
 624 {
 625         int err;
 626         struct kobject *hugepage_kobj;
 627
 628         if (!has_transparent_hugepage()) {
 629                 transparent_hugepage_flags = 0;
 630                 return -EINVAL;
 631         }
 632
 633         err = hugepage_init_sysfs(&hugepage_kobj);
 634         if (err)
 635                 goto err_sysfs;
 636
 637         err = khugepaged_slab_init();
 638         if (err)
 639                 goto err_slab;
 640
 641         err = register_shrinker(&huge_zero_page_shrinker);
 642         if (err)
 643                 goto err_hzp_shrinker;
 644
 645         /*
 646          * By default disable transparent hugepages on smaller systems,
 647          * where the extra memory used could hurt more than TLB overhead
 648          * is likely to save.  The admin can still enable it through /sys.
 649          */
 650         if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
 651                 transparent_hugepage_flags = 0;
 652                 return 0;
 653         }
 654
 655         err = start_stop_khugepaged();
 656         if (err)
 657                 goto err_khugepaged;
 658
 659         return 0;
 660 err_khugepaged:
 661         unregister_shrinker(&huge_zero_page_shrinker);
 662 err_hzp_shrinker:
 663         khugepaged_slab_exit();
 664 err_slab:
 665         hugepage_exit_sysfs(hugepage_kobj);
 666 err_sysfs:
 667         return err;
 668 }
 669 subsys_initcall(hugepage_init);
 670
 671 static int __init setup_transparent_hugepage(char *str)
 672 {
 673         int ret = 0;
 674         if (!str)
 675                 goto out;
 676         if (!strcmp(str, "always")) {
 677                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 678                         &transparent_hugepage_flags);
 679                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 680                           &transparent_hugepage_flags);
 681                 ret = 1;
 682         } else if (!strcmp(str, "madvise")) {
 683                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 684                           &transparent_hugepage_flags);
 685                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 686                         &transparent_hugepage_flags);
 687                 ret = 1;
 688         } else if (!strcmp(str, "never")) {
 689                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 690                           &transparent_hugepage_flags);
 691                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 692                           &transparent_hugepage_flags);
 693                 ret = 1;
 694         }
 695 out:
 696         if (!ret)
 697                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 698         return ret;
 699 }
 700 __setup("transparent_hugepage=", setup_transparent_hugepage);
 701
 702 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 703 {
 704         if (likely(vma->vm_flags & VM_WRITE))
 705                 pmd = pmd_mkwrite(pmd);
 706         return pmd;
 707 }
 708
 709 static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
 710 {
 711         pmd_t entry;
 712         entry = mk_pmd(page, prot);
 713         entry = pmd_mkhuge(entry);
 714         return entry;
 715 }
 716
 717 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
 718                                         struct vm_area_struct *vma,
 719                                         unsigned long haddr, pmd_t *pmd,
 720                                         struct page *page, gfp_t gfp)
 721 {
 722         struct mem_cgroup *memcg;
 723         pgtable_t pgtable;
 724         spinlock_t *ptl;
 725
 726         VM_BUG_ON_PAGE(!PageCompound(page), page);
 727
 728         if (mem_cgroup_try_charge(page, mm, gfp, &memcg))
 729                 return VM_FAULT_OOM;
 730
 731         pgtable = pte_alloc_one(mm, haddr);
 732         if (unlikely(!pgtable)) {
 733                 mem_cgroup_cancel_charge(page, memcg);
 734                 return VM_FAULT_OOM;
 735         }
 736
 737         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 738         /*
 739          * The memory barrier inside __SetPageUptodate makes sure that
 740          * clear_huge_page writes become visible before the set_pmd_at()
 741          * write.
 742          */
 743         __SetPageUptodate(page);
 744
 745         ptl = pmd_lock(mm, pmd);
 746         if (unlikely(!pmd_none(*pmd))) {
 747                 spin_unlock(ptl);
 748                 mem_cgroup_cancel_charge(page, memcg);
 749                 put_page(page);
 750                 pte_free(mm, pgtable);
 751         } else {
 752                 pmd_t entry;
 753                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 754                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 755                 page_add_new_anon_rmap(page, vma, haddr);
 756                 mem_cgroup_commit_charge(page, memcg, false);
 757                 lru_cache_add_active_or_unevictable(page, vma);
 758                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 759                 set_pmd_at(mm, haddr, pmd, entry);
 760                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 761                 atomic_long_inc(&mm->nr_ptes);
 762                 spin_unlock(ptl);
 763         }
 764
 765         return 0;
 766 }
 767
 768 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 769 {
 770         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
 771 }
 772
 773 /* Caller must hold page table lock. */
 774 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 775                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 776                 struct page *zero_page)
 777 {
 778         pmd_t entry;
 779         if (!pmd_none(*pmd))
 780                 return false;
 781         entry = mk_pmd(zero_page, vma->vm_page_prot);
 782         entry = pmd_mkhuge(entry);
 783         pgtable_trans_huge_deposit(mm, pmd, pgtable);
 784         set_pmd_at(mm, haddr, pmd, entry);
 785         atomic_long_inc(&mm->nr_ptes);
 786         return true;
 787 }
 788
 789 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 790                                unsigned long address, pmd_t *pmd,
 791                                unsigned int flags)
 792 {
 793         gfp_t gfp;
 794         struct page *page;
 795         unsigned long haddr = address & HPAGE_PMD_MASK;
 796
 797         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 798                 return VM_FAULT_FALLBACK;
 799         if (unlikely(anon_vma_prepare(vma)))
 800                 return VM_FAULT_OOM;
 801         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 802                 return VM_FAULT_OOM;
 803         if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm) &&
 804                         transparent_hugepage_use_zero_page()) {
 805                 spinlock_t *ptl;
 806                 pgtable_t pgtable;
 807                 struct page *zero_page;
 808                 bool set;
 809                 pgtable = pte_alloc_one(mm, haddr);
 810                 if (unlikely(!pgtable))
 811                         return VM_FAULT_OOM;
 812                 zero_page = get_huge_zero_page();
 813                 if (unlikely(!zero_page)) {
 814                         pte_free(mm, pgtable);
 815                         count_vm_event(THP_FAULT_FALLBACK);
 816                         return VM_FAULT_FALLBACK;
 817                 }
 818                 ptl = pmd_lock(mm, pmd);
 819                 set = set_huge_zero_page(pgtable, mm, vma, haddr, pmd,
 820                                 zero_page);
 821                 spin_unlock(ptl);
 822                 if (!set) {
 823                         pte_free(mm, pgtable);
 824                         put_huge_zero_page();
 825                 }
 826                 return 0;
 827         }
 828         gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
 829         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 830         if (unlikely(!page)) {
 831                 count_vm_event(THP_FAULT_FALLBACK);
 832                 return VM_FAULT_FALLBACK;
 833         }
 834         if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page, gfp))) {
 835                 put_page(page);
 836                 count_vm_event(THP_FAULT_FALLBACK);
 837                 return VM_FAULT_FALLBACK;
 838         }
 839
 840         count_vm_event(THP_FAULT_ALLOC);
 841         return 0;
 842 }
 843
 844 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 845                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 846                   struct vm_area_struct *vma)
 847 {
 848         spinlock_t *dst_ptl, *src_ptl;
 849         struct page *src_page;
 850         pmd_t pmd;
 851         pgtable_t pgtable;
 852         int ret;
 853
 854         ret = -ENOMEM;
 855         pgtable = pte_alloc_one(dst_mm, addr);
 856         if (unlikely(!pgtable))
 857                 goto out;
 858
 859         dst_ptl = pmd_lock(dst_mm, dst_pmd);
 860         src_ptl = pmd_lockptr(src_mm, src_pmd);
 861         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 862
 863         ret = -EAGAIN;
 864         pmd = *src_pmd;
 865         if (unlikely(!pmd_trans_huge(pmd))) {
 866                 pte_free(dst_mm, pgtable);
 867                 goto out_unlock;
 868         }
 869         /*
 870          * When page table lock is held, the huge zero pmd should not be
 871          * under splitting since we don't split the page itself, only pmd to
 872          * a page table.
 873          */
 874         if (is_huge_zero_pmd(pmd)) {
 875                 struct page *zero_page;
 876                 bool set;
 877                 /*
 878                  * get_huge_zero_page() will never allocate a new page here,
 879                  * since we already have a zero page to copy. It just takes a
 880                  * reference.
 881                  */
 882                 zero_page = get_huge_zero_page();
 883                 set = set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
 884                                 zero_page);
 885                 BUG_ON(!set); /* unexpected !pmd_none(dst_pmd) */
 886                 ret = 0;
 887                 goto out_unlock;
 888         }
 889
 890         if (unlikely(pmd_trans_splitting(pmd))) {
 891                 /* split huge page running from under us */
 892                 spin_unlock(src_ptl);
 893                 spin_unlock(dst_ptl);
 894                 pte_free(dst_mm, pgtable);
 895
 896                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 897                 goto out;
 898         }
 899         src_page = pmd_page(pmd);
 900         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
 901         get_page(src_page);
 902         page_dup_rmap(src_page);
 903         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 904
 905         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 906         pmd = pmd_mkold(pmd_wrprotect(pmd));
 907         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
 908         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 909         atomic_long_inc(&dst_mm->nr_ptes);
 910
 911         ret = 0;
 912 out_unlock:
 913         spin_unlock(src_ptl);
 914         spin_unlock(dst_ptl);
 915 out:
 916         return ret;
 917 }
 918
 919 void huge_pmd_set_accessed(struct mm_struct *mm,
 920                            struct vm_area_struct *vma,
 921                            unsigned long address,
 922                            pmd_t *pmd, pmd_t orig_pmd,
 923                            int dirty)
 924 {
 925         spinlock_t *ptl;
 926         pmd_t entry;
 927         unsigned long haddr;
 928
 929         ptl = pmd_lock(mm, pmd);
 930         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 931                 goto unlock;
 932
 933         entry = pmd_mkyoung(orig_pmd);
 934         haddr = address & HPAGE_PMD_MASK;
 935         if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
 936                 update_mmu_cache_pmd(vma, address, pmd);
 937
 938 unlock:
 939         spin_unlock(ptl);
 940 }
 941
 942 /*
 943  * Save CONFIG_DEBUG_PAGEALLOC from faulting falsely on tail pages
 944  * during copy_user_huge_page()'s copy_page_rep(): in the case when
 945  * the source page gets split and a tail freed before copy completes.
 946  * Called under pmd_lock of checked pmd, so safe from splitting itself.
 947  */
 948 static void get_user_huge_page(struct page *page)
 949 {
 950         if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
 951                 struct page *endpage = page + HPAGE_PMD_NR;
 952
 953                 atomic_add(HPAGE_PMD_NR, &page->_count);
 954                 while (++page < endpage)
 955                         get_huge_page_tail(page);
 956         } else {
 957                 get_page(page);
 958         }
 959 }
 960
 961 static void put_user_huge_page(struct page *page)
 962 {
 963         if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
 964                 struct page *endpage = page + HPAGE_PMD_NR;
 965
 966                 while (page < endpage)
 967                         put_page(page++);
 968         } else {
 969                 put_page(page);
 970         }
 971 }
 972
 973 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
 974                                         struct vm_area_struct *vma,
 975                                         unsigned long address,
 976                                         pmd_t *pmd, pmd_t orig_pmd,
 977                                         struct page *page,
 978                                         unsigned long haddr)
 979 {
 980         struct mem_cgroup *memcg;
 981         spinlock_t *ptl;
 982         pgtable_t pgtable;
 983         pmd_t _pmd;
 984         int ret = 0, i;
 985         struct page **pages;
 986         unsigned long mmun_start;       /* For mmu_notifiers */
 987         unsigned long mmun_end;         /* For mmu_notifiers */
 988
 989         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
 990                         GFP_KERNEL);
 991         if (unlikely(!pages)) {
 992                 ret |= VM_FAULT_OOM;
 993                 goto out;
 994         }
 995
 996         for (i = 0; i < HPAGE_PMD_NR; i++) {
 997                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
 998                                                __GFP_OTHER_NODE,
 999                                                vma, address, page_to_nid(page));
1000                 if (unlikely(!pages[i] ||
1001                              mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL,
1002                                                    &memcg))) {
1003                         if (pages[i])
1004                                 put_page(pages[i]);
1005                         while (--i >= 0) {
1006                                 memcg = (void *)page_private(pages[i]);
1007                                 set_page_private(pages[i], 0);
1008                                 mem_cgroup_cancel_charge(pages[i], memcg);
1009                                 put_page(pages[i]);
1010                         }
1011                         kfree(pages);
1012                         ret |= VM_FAULT_OOM;
1013                         goto out;
1014                 }
1015                 set_page_private(pages[i], (unsigned long)memcg);
1016         }
1017
1018         for (i = 0; i < HPAGE_PMD_NR; i++) {
1019                 copy_user_highpage(pages[i], page + i,
1020                                    haddr + PAGE_SIZE * i, vma);
1021                 __SetPageUptodate(pages[i]);
1022                 cond_resched();
1023         }
1024
1025         mmun_start = haddr;
1026         mmun_end   = haddr + HPAGE_PMD_SIZE;
1027         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1028
1029         ptl = pmd_lock(mm, pmd);
1030         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1031                 goto out_free_pages;
1032         VM_BUG_ON_PAGE(!PageHead(page), page);
1033
1034         pmdp_clear_flush_notify(vma, haddr, pmd);
1035         /* leave pmd empty until pte is filled */
1036
1037         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1038         pmd_populate(mm, &_pmd, pgtable);
1039
1040         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1041                 pte_t *pte, entry;
1042                 entry = mk_pte(pages[i], vma->vm_page_prot);
1043                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1044                 memcg = (void *)page_private(pages[i]);
1045                 set_page_private(pages[i], 0);
1046                 page_add_new_anon_rmap(pages[i], vma, haddr);
1047                 mem_cgroup_commit_charge(pages[i], memcg, false);
1048                 lru_cache_add_active_or_unevictable(pages[i], vma);
1049                 pte = pte_offset_map(&_pmd, haddr);
1050                 VM_BUG_ON(!pte_none(*pte));
1051                 set_pte_at(mm, haddr, pte, entry);
1052                 pte_unmap(pte);
1053         }
1054         kfree(pages);
1055
1056         smp_wmb(); /* make pte visible before pmd */
1057         pmd_populate(mm, pmd, pgtable);
1058         page_remove_rmap(page);
1059         spin_unlock(ptl);
1060
1061         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1062
1063         ret |= VM_FAULT_WRITE;
1064         put_page(page);
1065
1066 out:
1067         return ret;
1068
1069 out_free_pages:
1070         spin_unlock(ptl);
1071         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1072         for (i = 0; i < HPAGE_PMD_NR; i++) {
1073                 memcg = (void *)page_private(pages[i]);
1074                 set_page_private(pages[i], 0);
1075                 mem_cgroup_cancel_charge(pages[i], memcg);
1076                 put_page(pages[i]);
1077         }
1078         kfree(pages);
1079         goto out;
1080 }
1081
1082 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1083                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1084 {
1085         spinlock_t *ptl;
1086         int ret = 0;
1087         struct page *page = NULL, *new_page;
1088         struct mem_cgroup *memcg;
1089         unsigned long haddr;
1090         unsigned long mmun_start;       /* For mmu_notifiers */
1091         unsigned long mmun_end;         /* For mmu_notifiers */
1092         gfp_t huge_gfp;                 /* for allocation and charge */
1093
1094         ptl = pmd_lockptr(mm, pmd);
1095         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1096         haddr = address & HPAGE_PMD_MASK;
1097         if (is_huge_zero_pmd(orig_pmd))
1098                 goto alloc;
1099         spin_lock(ptl);
1100         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1101                 goto out_unlock;
1102
1103         page = pmd_page(orig_pmd);
1104         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1105         if (page_mapcount(page) == 1) {
1106                 pmd_t entry;
1107                 entry = pmd_mkyoung(orig_pmd);
1108                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1109                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1110                         update_mmu_cache_pmd(vma, address, pmd);
1111                 ret |= VM_FAULT_WRITE;
1112                 goto out_unlock;
1113         }
1114         get_user_huge_page(page);
1115         spin_unlock(ptl);
1116 alloc:
1117         if (transparent_hugepage_enabled(vma) &&
1118             !transparent_hugepage_debug_cow()) {
1119                 huge_gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
1120                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1121         } else
1122                 new_page = NULL;
1123
1124         if (unlikely(!new_page)) {
1125                 if (!page) {
1126                         split_huge_page_pmd(vma, address, pmd);
1127                         ret |= VM_FAULT_FALLBACK;
1128                 } else {
1129                         ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1130                                         pmd, orig_pmd, page, haddr);
1131                         if (ret & VM_FAULT_OOM) {
1132                                 split_huge_page(page);
1133                                 ret |= VM_FAULT_FALLBACK;
1134                         }
1135                         put_user_huge_page(page);
1136                 }
1137                 count_vm_event(THP_FAULT_FALLBACK);
1138                 goto out;
1139         }
1140
1141         if (unlikely(mem_cgroup_try_charge(new_page, mm, huge_gfp, &memcg))) {
1142                 put_page(new_page);
1143                 if (page) {
1144                         split_huge_page(page);
1145                         put_user_huge_page(page);
1146                 } else
1147                         split_huge_page_pmd(vma, address, pmd);
1148                 ret |= VM_FAULT_FALLBACK;
1149                 count_vm_event(THP_FAULT_FALLBACK);
1150                 goto out;
1151         }
1152
1153         count_vm_event(THP_FAULT_ALLOC);
1154
1155         if (!page)
1156                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1157         else
1158                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1159         __SetPageUptodate(new_page);
1160
1161         mmun_start = haddr;
1162         mmun_end   = haddr + HPAGE_PMD_SIZE;
1163         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1164
1165         spin_lock(ptl);
1166         if (page)
1167                 put_user_huge_page(page);
1168         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1169                 spin_unlock(ptl);
1170                 mem_cgroup_cancel_charge(new_page, memcg);
1171                 put_page(new_page);
1172                 goto out_mn;
1173         } else {
1174                 pmd_t entry;
1175                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1176                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1177                 pmdp_clear_flush_notify(vma, haddr, pmd);
1178                 page_add_new_anon_rmap(new_page, vma, haddr);
1179                 mem_cgroup_commit_charge(new_page, memcg, false);
1180                 lru_cache_add_active_or_unevictable(new_page, vma);
1181                 set_pmd_at(mm, haddr, pmd, entry);
1182                 update_mmu_cache_pmd(vma, address, pmd);
1183                 if (!page) {
1184                         add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1185                         put_huge_zero_page();
1186                 } else {
1187                         VM_BUG_ON_PAGE(!PageHead(page), page);
1188                         page_remove_rmap(page);
1189                         put_page(page);
1190                 }
1191                 ret |= VM_FAULT_WRITE;
1192         }
1193         spin_unlock(ptl);
1194 out_mn:
1195         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1196 out:
1197         return ret;
1198 out_unlock:
1199         spin_unlock(ptl);
1200         return ret;
1201 }
1202
1203 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1204                                    unsigned long addr,
1205                                    pmd_t *pmd,
1206                                    unsigned int flags)
1207 {
1208         struct mm_struct *mm = vma->vm_mm;
1209         struct page *page = NULL;
1210
1211         assert_spin_locked(pmd_lockptr(mm, pmd));
1212
1213         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1214                 goto out;
1215
1216         /* Avoid dumping huge zero page */
1217         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1218                 return ERR_PTR(-EFAULT);
1219
1220         /* Full NUMA hinting faults to serialise migration in fault paths */
1221         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1222                 goto out;
1223
1224         page = pmd_page(*pmd);
1225         VM_BUG_ON_PAGE(!PageHead(page), page);
1226         if (flags & FOLL_TOUCH) {
1227                 pmd_t _pmd;
1228                 /*
1229                  * We should set the dirty bit only for FOLL_WRITE but
1230                  * for now the dirty bit in the pmd is meaningless.
1231                  * And if the dirty bit will become meaningful and
1232                  * we'll only set it with FOLL_WRITE, an atomic
1233                  * set_bit will be required on the pmd to set the
1234                  * young bit, instead of the current set_pmd_at.
1235                  */
1236                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1237                 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1238                                           pmd, _pmd,  1))
1239                         update_mmu_cache_pmd(vma, addr, pmd);
1240         }
1241         if ((flags & FOLL_POPULATE) && (vma->vm_flags & VM_LOCKED)) {
1242                 if (page->mapping && trylock_page(page)) {
1243                         lru_add_drain();
1244                         if (page->mapping)
1245                                 mlock_vma_page(page);
1246                         unlock_page(page);
1247                 }
1248         }
1249         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1250         VM_BUG_ON_PAGE(!PageCompound(page), page);
1251         if (flags & FOLL_GET)
1252                 get_page_foll(page);
1253
1254 out:
1255         return page;
1256 }
1257
1258 /* NUMA hinting page fault entry point for trans huge pmds */
1259 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1260                                 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1261 {
1262         spinlock_t *ptl;
1263         struct anon_vma *anon_vma = NULL;
1264         struct page *page;
1265         unsigned long haddr = addr & HPAGE_PMD_MASK;
1266         int page_nid = -1, this_nid = numa_node_id();
1267         int target_nid, last_cpupid = -1;
1268         bool page_locked;
1269         bool migrated = false;
1270         bool was_writable;
1271         int flags = 0;
1272
1273         /* A PROT_NONE fault should not end up here */
1274         BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
1275
1276         ptl = pmd_lock(mm, pmdp);
1277         if (unlikely(!pmd_same(pmd, *pmdp)))
1278                 goto out_unlock;
1279
1280         /*
1281          * If there are potential migrations, wait for completion and retry
1282          * without disrupting NUMA hinting information. Do not relock and
1283          * check_same as the page may no longer be mapped.
1284          */
1285         if (unlikely(pmd_trans_migrating(*pmdp))) {
1286                 page = pmd_page(*pmdp);
1287                 spin_unlock(ptl);
1288                 wait_on_page_locked(page);
1289                 goto out;
1290         }
1291
1292         page = pmd_page(pmd);
1293         BUG_ON(is_huge_zero_page(page));
1294         page_nid = page_to_nid(page);
1295         last_cpupid = page_cpupid_last(page);
1296         count_vm_numa_event(NUMA_HINT_FAULTS);
1297         if (page_nid == this_nid) {
1298                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1299                 flags |= TNF_FAULT_LOCAL;
1300         }
1301
1302         /* See similar comment in do_numa_page for explanation */
1303         if (!(vma->vm_flags & VM_WRITE))
1304                 flags |= TNF_NO_GROUP;
1305
1306         /*
1307          * Acquire the page lock to serialise THP migrations but avoid dropping
1308          * page_table_lock if at all possible
1309          */
1310         page_locked = trylock_page(page);
1311         target_nid = mpol_misplaced(page, vma, haddr);
1312         if (target_nid == -1) {
1313                 /* If the page was locked, there are no parallel migrations */
1314                 if (page_locked)
1315                         goto clear_pmdnuma;
1316         }
1317
1318         /* Migration could have started since the pmd_trans_migrating check */
1319         if (!page_locked) {
1320                 spin_unlock(ptl);
1321                 wait_on_page_locked(page);
1322                 page_nid = -1;
1323                 goto out;
1324         }
1325
1326         /*
1327          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1328          * to serialises splits
1329          */
1330         get_page(page);
1331         spin_unlock(ptl);
1332         anon_vma = page_lock_anon_vma_read(page);
1333
1334         /* Confirm the PMD did not change while page_table_lock was released */
1335         spin_lock(ptl);
1336         if (unlikely(!pmd_same(pmd, *pmdp))) {
1337                 unlock_page(page);
1338                 put_page(page);
1339                 page_nid = -1;
1340                 goto out_unlock;
1341         }
1342
1343         /* Bail if we fail to protect against THP splits for any reason */
1344         if (unlikely(!anon_vma)) {
1345                 put_page(page);
1346                 page_nid = -1;
1347                 goto clear_pmdnuma;
1348         }
1349
1350         /*
1351          * Migrate the THP to the requested node, returns with page unlocked
1352          * and access rights restored.
1353          */
1354         spin_unlock(ptl);
1355         migrated = migrate_misplaced_transhuge_page(mm, vma,
1356                                 pmdp, pmd, addr, page, target_nid);
1357         if (migrated) {
1358                 flags |= TNF_MIGRATED;
1359                 page_nid = target_nid;
1360         } else
1361                 flags |= TNF_MIGRATE_FAIL;
1362
1363         goto out;
1364 clear_pmdnuma:
1365         BUG_ON(!PageLocked(page));
1366         was_writable = pmd_write(pmd);
1367         pmd = pmd_modify(pmd, vma->vm_page_prot);
1368         pmd = pmd_mkyoung(pmd);
1369         if (was_writable)
1370                 pmd = pmd_mkwrite(pmd);
1371         set_pmd_at(mm, haddr, pmdp, pmd);
1372         update_mmu_cache_pmd(vma, addr, pmdp);
1373         unlock_page(page);
1374 out_unlock:
1375         spin_unlock(ptl);
1376
1377 out:
1378         if (anon_vma)
1379                 page_unlock_anon_vma_read(anon_vma);
1380
1381         if (page_nid != -1)
1382                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
1383
1384         return 0;
1385 }
1386
1387 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1388                  pmd_t *pmd, unsigned long addr)
1389 {
1390         spinlock_t *ptl;
1391         int ret = 0;
1392
1393         if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1394                 struct page *page;
1395                 pgtable_t pgtable;
1396                 pmd_t orig_pmd;
1397                 /*
1398                  * For architectures like ppc64 we look at deposited pgtable
1399                  * when calling pmdp_get_and_clear. So do the
1400                  * pgtable_trans_huge_withdraw after finishing pmdp related
1401                  * operations.
1402                  */
1403                 orig_pmd = pmdp_get_and_clear_full(tlb->mm, addr, pmd,
1404                                                    tlb->fullmm);
1405                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1406                 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1407                 if (is_huge_zero_pmd(orig_pmd)) {
1408                         atomic_long_dec(&tlb->mm->nr_ptes);
1409                         spin_unlock(ptl);
1410                         put_huge_zero_page();
1411                 } else {
1412                         page = pmd_page(orig_pmd);
1413                         page_remove_rmap(page);
1414                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1415                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1416                         VM_BUG_ON_PAGE(!PageHead(page), page);
1417                         atomic_long_dec(&tlb->mm->nr_ptes);
1418                         spin_unlock(ptl);
1419                         tlb_remove_page(tlb, page);
1420                 }
1421                 pte_free(tlb->mm, pgtable);
1422                 ret = 1;
1423         }
1424         return ret;
1425 }
1426
1427 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1428                   unsigned long old_addr,
1429                   unsigned long new_addr, unsigned long old_end,
1430                   pmd_t *old_pmd, pmd_t *new_pmd)
1431 {
1432         spinlock_t *old_ptl, *new_ptl;
1433         int ret = 0;
1434         pmd_t pmd;
1435
1436         struct mm_struct *mm = vma->vm_mm;
1437
1438         if ((old_addr & ~HPAGE_PMD_MASK) ||
1439             (new_addr & ~HPAGE_PMD_MASK) ||
1440             old_end - old_addr < HPAGE_PMD_SIZE ||
1441             (new_vma->vm_flags & VM_NOHUGEPAGE))
1442                 goto out;
1443
1444         /*
1445          * The destination pmd shouldn't be established, free_pgtables()
1446          * should have release it.
1447          */
1448         if (WARN_ON(!pmd_none(*new_pmd))) {
1449                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1450                 goto out;
1451         }
1452
1453         /*
1454          * We don't have to worry about the ordering of src and dst
1455          * ptlocks because exclusive mmap_sem prevents deadlock.
1456          */
1457         ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl);
1458         if (ret == 1) {
1459                 new_ptl = pmd_lockptr(mm, new_pmd);
1460                 if (new_ptl != old_ptl)
1461                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1462                 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1463                 VM_BUG_ON(!pmd_none(*new_pmd));
1464
1465                 if (pmd_move_must_withdraw(new_ptl, old_ptl)) {
1466                         pgtable_t pgtable;
1467                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1468                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1469                 }
1470                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1471                 if (new_ptl != old_ptl)
1472                         spin_unlock(new_ptl);
1473                 spin_unlock(old_ptl);
1474         }
1475 out:
1476         return ret;
1477 }
1478
1479 /*
1480  * Returns
1481  *  - 0 if PMD could not be locked
1482  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1483  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1484  */
1485 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1486                 unsigned long addr, pgprot_t newprot, int prot_numa)
1487 {
1488         struct mm_struct *mm = vma->vm_mm;
1489         spinlock_t *ptl;
1490         int ret = 0;
1491
1492         if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1493                 pmd_t entry;
1494                 bool preserve_write = prot_numa && pmd_write(*pmd);
1495                 ret = 1;
1496
1497                 /*
1498                  * Avoid trapping faults against the zero page. The read-only
1499                  * data is likely to be read-cached on the local CPU and
1500                  * local/remote hits to the zero page are not interesting.
1501                  */
1502                 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1503                         spin_unlock(ptl);
1504                         return ret;
1505                 }
1506
1507                 if (!prot_numa || !pmd_protnone(*pmd)) {
1508                         entry = pmdp_get_and_clear_notify(mm, addr, pmd);
1509                         entry = pmd_modify(entry, newprot);
1510                         if (preserve_write)
1511                                 entry = pmd_mkwrite(entry);
1512                         ret = HPAGE_PMD_NR;
1513                         set_pmd_at(mm, addr, pmd, entry);
1514                         BUG_ON(!preserve_write && pmd_write(entry));
1515                 }
1516                 spin_unlock(ptl);
1517         }
1518
1519         return ret;
1520 }
1521
1522 /*
1523  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1524  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1525  *
1526  * Note that if it returns 1, this routine returns without unlocking page
1527  * table locks. So callers must unlock them.
1528  */
1529 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1530                 spinlock_t **ptl)
1531 {
1532         *ptl = pmd_lock(vma->vm_mm, pmd);
1533         if (likely(pmd_trans_huge(*pmd))) {
1534                 if (unlikely(pmd_trans_splitting(*pmd))) {
1535                         spin_unlock(*ptl);
1536                         wait_split_huge_page(vma->anon_vma, pmd);
1537                         return -1;
1538                 } else {
1539                         /* Thp mapped by 'pmd' is stable, so we can
1540                          * handle it as it is. */
1541                         return 1;
1542                 }
1543         }
1544         spin_unlock(*ptl);
1545         return 0;
1546 }
1547
1548 /*
1549  * This function returns whether a given @page is mapped onto the @address
1550  * in the virtual space of @mm.
1551  *
1552  * When it's true, this function returns *pmd with holding the page table lock
1553  * and passing it back to the caller via @ptl.
1554  * If it's false, returns NULL without holding the page table lock.
1555  */
1556 pmd_t *page_check_address_pmd(struct page *page,
1557                               struct mm_struct *mm,
1558                               unsigned long address,
1559                               enum page_check_address_pmd_flag flag,
1560                               spinlock_t **ptl)
1561 {
1562         pgd_t *pgd;
1563         pud_t *pud;
1564         pmd_t *pmd;
1565
1566         if (address & ~HPAGE_PMD_MASK)
1567                 return NULL;
1568
1569         pgd = pgd_offset(mm, address);
1570         if (!pgd_present(*pgd))
1571                 return NULL;
1572         pud = pud_offset(pgd, address);
1573         if (!pud_present(*pud))
1574                 return NULL;
1575         pmd = pmd_offset(pud, address);
1576
1577         *ptl = pmd_lock(mm, pmd);
1578         if (!pmd_present(*pmd))
1579                 goto unlock;
1580         if (pmd_page(*pmd) != page)
1581                 goto unlock;
1582         /*
1583          * split_vma() may create temporary aliased mappings. There is
1584          * no risk as long as all huge pmd are found and have their
1585          * splitting bit set before __split_huge_page_refcount
1586          * runs. Finding the same huge pmd more than once during the
1587          * same rmap walk is not a problem.
1588          */
1589         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1590             pmd_trans_splitting(*pmd))
1591                 goto unlock;
1592         if (pmd_trans_huge(*pmd)) {
1593                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1594                           !pmd_trans_splitting(*pmd));
1595                 return pmd;
1596         }
1597 unlock:
1598         spin_unlock(*ptl);
1599         return NULL;
1600 }
1601
1602 static int __split_huge_page_splitting(struct page *page,
1603                                        struct vm_area_struct *vma,
1604                                        unsigned long address)
1605 {
1606         struct mm_struct *mm = vma->vm_mm;
1607         spinlock_t *ptl;
1608         pmd_t *pmd;
1609         int ret = 0;
1610         /* For mmu_notifiers */
1611         const unsigned long mmun_start = address;
1612         const unsigned long mmun_end   = address + HPAGE_PMD_SIZE;
1613
1614         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1615         pmd = page_check_address_pmd(page, mm, address,
1616                         PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl);
1617         if (pmd) {
1618                 /*
1619                  * We can't temporarily set the pmd to null in order
1620                  * to split it, the pmd must remain marked huge at all
1621                  * times or the VM won't take the pmd_trans_huge paths
1622                  * and it won't wait on the anon_vma->root->rwsem to
1623                  * serialize against split_huge_page*.
1624                  */
1625                 pmdp_splitting_flush(vma, address, pmd);
1626
1627                 ret = 1;
1628                 spin_unlock(ptl);
1629         }
1630         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1631
1632         return ret;
1633 }
1634
1635 static void __split_huge_page_refcount(struct page *page,
1636                                        struct list_head *list)
1637 {
1638         int i;
1639         struct zone *zone = page_zone(page);
1640         struct lruvec *lruvec;
1641         int tail_count = 0;
1642
1643         /* prevent PageLRU to go away from under us, and freeze lru stats */
1644         spin_lock_irq(&zone->lru_lock);
1645         lruvec = mem_cgroup_page_lruvec(page, zone);
1646
1647         compound_lock(page);
1648         /* complete memcg works before add pages to LRU */
1649         mem_cgroup_split_huge_fixup(page);
1650
1651         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1652                 struct page *page_tail = page + i;
1653
1654                 /* tail_page->_mapcount cannot change */
1655                 BUG_ON(page_mapcount(page_tail) < 0);
1656                 tail_count += page_mapcount(page_tail);
1657                 /* check for overflow */
1658                 BUG_ON(tail_count < 0);
1659                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1660                 /*
1661                  * tail_page->_count is zero and not changing from
1662                  * under us. But get_page_unless_zero() may be running
1663                  * from under us on the tail_page. If we used
1664                  * atomic_set() below instead of atomic_add(), we
1665                  * would then run atomic_set() concurrently with
1666                  * get_page_unless_zero(), and atomic_set() is
1667                  * implemented in C not using locked ops. spin_unlock
1668                  * on x86 sometime uses locked ops because of PPro
1669                  * errata 66, 92, so unless somebody can guarantee
1670                  * atomic_set() here would be safe on all archs (and
1671                  * not only on x86), it's safer to use atomic_add().
1672                  */
1673                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1674                            &page_tail->_count);
1675
1676                 /* after clearing PageTail the gup refcount can be released */
1677                 smp_mb__after_atomic();
1678
1679                 /*
1680                  * retain hwpoison flag of the poisoned tail page:
1681                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1682                  *   by the memory-failure.
1683                  */
1684                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1685                 page_tail->flags |= (page->flags &
1686                                      ((1L << PG_referenced) |
1687                                       (1L << PG_swapbacked) |
1688                                       (1L << PG_mlocked) |
1689                                       (1L << PG_uptodate) |
1690                                       (1L << PG_active) |
1691                                       (1L << PG_unevictable)));
1692                 page_tail->flags |= (1L << PG_dirty);
1693
1694                 /* clear PageTail before overwriting first_page */
1695                 smp_wmb();
1696
1697                 /*
1698                  * __split_huge_page_splitting() already set the
1699                  * splitting bit in all pmd that could map this
1700                  * hugepage, that will ensure no CPU can alter the
1701                  * mapcount on the head page. The mapcount is only
1702                  * accounted in the head page and it has to be
1703                  * transferred to all tail pages in the below code. So
1704                  * for this code to be safe, the split the mapcount
1705                  * can't change. But that doesn't mean userland can't
1706                  * keep changing and reading the page contents while
1707                  * we transfer the mapcount, so the pmd splitting
1708                  * status is achieved setting a reserved bit in the
1709                  * pmd, not by clearing the present bit.
1710                 */
1711                 page_tail->_mapcount = page->_mapcount;
1712
1713                 BUG_ON(page_tail->mapping);
1714                 page_tail->mapping = page->mapping;
1715
1716                 page_tail->index = page->index + i;
1717                 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
1718
1719                 BUG_ON(!PageAnon(page_tail));
1720                 BUG_ON(!PageUptodate(page_tail));
1721                 BUG_ON(!PageDirty(page_tail));
1722                 BUG_ON(!PageSwapBacked(page_tail));
1723
1724                 lru_add_page_tail(page, page_tail, lruvec, list);
1725         }
1726         atomic_sub(tail_count, &page->_count);
1727         BUG_ON(atomic_read(&page->_count) <= 0);
1728
1729         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1730
1731         ClearPageCompound(page);
1732         compound_unlock(page);
1733         spin_unlock_irq(&zone->lru_lock);
1734
1735         for (i = 1; i < HPAGE_PMD_NR; i++) {
1736                 struct page *page_tail = page + i;
1737                 BUG_ON(page_count(page_tail) <= 0);
1738                 /*
1739                  * Tail pages may be freed if there wasn't any mapping
1740                  * like if add_to_swap() is running on a lru page that
1741                  * had its mapping zapped. And freeing these pages
1742                  * requires taking the lru_lock so we do the put_page
1743                  * of the tail pages after the split is complete.
1744                  */
1745                 put_page(page_tail);
1746         }
1747
1748         /*
1749          * Only the head page (now become a regular page) is required
1750          * to be pinned by the caller.
1751          */
1752         BUG_ON(page_count(page) <= 0);
1753 }
1754
1755 static int __split_huge_page_map(struct page *page,
1756                                  struct vm_area_struct *vma,
1757                                  unsigned long address)
1758 {
1759         struct mm_struct *mm = vma->vm_mm;
1760         spinlock_t *ptl;
1761         pmd_t *pmd, _pmd;
1762         int ret = 0, i;
1763         pgtable_t pgtable;
1764         unsigned long haddr;
1765
1766         pmd = page_check_address_pmd(page, mm, address,
1767                         PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl);
1768         if (pmd) {
1769                 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1770                 pmd_populate(mm, &_pmd, pgtable);
1771                 if (pmd_write(*pmd))
1772                         BUG_ON(page_mapcount(page) != 1);
1773
1774                 haddr = address;
1775                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1776                         pte_t *pte, entry;
1777                         BUG_ON(PageCompound(page+i));
1778                         /*
1779                          * Note that NUMA hinting access restrictions are not
1780                          * transferred to avoid any possibility of altering
1781                          * permissions across VMAs.
1782                          */
1783                         entry = mk_pte(page + i, vma->vm_page_prot);
1784                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1785                         if (!pmd_write(*pmd))
1786                                 entry = pte_wrprotect(entry);
1787                         if (!pmd_young(*pmd))
1788                                 entry = pte_mkold(entry);
1789                         pte = pte_offset_map(&_pmd, haddr);
1790                         BUG_ON(!pte_none(*pte));
1791                         set_pte_at(mm, haddr, pte, entry);
1792                         pte_unmap(pte);
1793                 }
1794
1795                 smp_wmb(); /* make pte visible before pmd */
1796                 /*
1797                  * Up to this point the pmd is present and huge and
1798                  * userland has the whole access to the hugepage
1799                  * during the split (which happens in place). If we
1800                  * overwrite the pmd with the not-huge version
1801                  * pointing to the pte here (which of course we could
1802                  * if all CPUs were bug free), userland could trigger
1803                  * a small page size TLB miss on the small sized TLB
1804                  * while the hugepage TLB entry is still established
1805                  * in the huge TLB. Some CPU doesn't like that. See
1806                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1807                  * Erratum 383 on page 93. Intel should be safe but is
1808                  * also warns that it's only safe if the permission
1809                  * and cache attributes of the two entries loaded in
1810                  * the two TLB is identical (which should be the case
1811                  * here). But it is generally safer to never allow
1812                  * small and huge TLB entries for the same virtual
1813                  * address to be loaded simultaneously. So instead of
1814                  * doing "pmd_populate(); flush_tlb_range();" we first
1815                  * mark the current pmd notpresent (atomically because
1816                  * here the pmd_trans_huge and pmd_trans_splitting
1817                  * must remain set at all times on the pmd until the
1818                  * split is complete for this pmd), then we flush the
1819                  * SMP TLB and finally we write the non-huge version
1820                  * of the pmd entry with pmd_populate.
1821                  */
1822                 pmdp_invalidate(vma, address, pmd);
1823                 pmd_populate(mm, pmd, pgtable);
1824                 ret = 1;
1825                 spin_unlock(ptl);
1826         }
1827
1828         return ret;
1829 }
1830
1831 /* must be called with anon_vma->root->rwsem held */
1832 static void __split_huge_page(struct page *page,
1833                               struct anon_vma *anon_vma,
1834                               struct list_head *list)
1835 {
1836         int mapcount, mapcount2;
1837         pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1838         struct anon_vma_chain *avc;
1839
1840         BUG_ON(!PageHead(page));
1841         BUG_ON(PageTail(page));
1842
1843         mapcount = 0;
1844         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1845                 struct vm_area_struct *vma = avc->vma;
1846                 unsigned long addr = vma_address(page, vma);
1847                 BUG_ON(is_vma_temporary_stack(vma));
1848                 mapcount += __split_huge_page_splitting(page, vma, addr);
1849         }
1850         /*
1851          * It is critical that new vmas are added to the tail of the
1852          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1853          * and establishes a child pmd before
1854          * __split_huge_page_splitting() freezes the parent pmd (so if
1855          * we fail to prevent copy_huge_pmd() from running until the
1856          * whole __split_huge_page() is complete), we will still see
1857          * the newly established pmd of the child later during the
1858          * walk, to be able to set it as pmd_trans_splitting too.
1859          */
1860         if (mapcount != page_mapcount(page)) {
1861                 pr_err("mapcount %d page_mapcount %d\n",
1862                         mapcount, page_mapcount(page));
1863                 BUG();
1864         }
1865
1866         __split_huge_page_refcount(page, list);
1867
1868         mapcount2 = 0;
1869         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1870                 struct vm_area_struct *vma = avc->vma;
1871                 unsigned long addr = vma_address(page, vma);
1872                 BUG_ON(is_vma_temporary_stack(vma));
1873                 mapcount2 += __split_huge_page_map(page, vma, addr);
1874         }
1875         if (mapcount != mapcount2) {
1876                 pr_err("mapcount %d mapcount2 %d page_mapcount %d\n",
1877                         mapcount, mapcount2, page_mapcount(page));
1878                 BUG();
1879         }
1880 }
1881
1882 /*
1883  * Split a hugepage into normal pages. This doesn't change the position of head
1884  * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1885  * @list. Both head page and tail pages will inherit mapping, flags, and so on
1886  * from the hugepage.
1887  * Return 0 if the hugepage is split successfully otherwise return 1.
1888  */
1889 int split_huge_page_to_list(struct page *page, struct list_head *list)
1890 {
1891         struct anon_vma *anon_vma;
1892         int ret = 1;
1893
1894         BUG_ON(is_huge_zero_page(page));
1895         BUG_ON(!PageAnon(page));
1896
1897         /*
1898          * The caller does not necessarily hold an mmap_sem that would prevent
1899          * the anon_vma disappearing so we first we take a reference to it
1900          * and then lock the anon_vma for write. This is similar to
1901          * page_lock_anon_vma_read except the write lock is taken to serialise
1902          * against parallel split or collapse operations.
1903          */
1904         anon_vma = page_get_anon_vma(page);
1905         if (!anon_vma)
1906                 goto out;
1907         anon_vma_lock_write(anon_vma);
1908
1909         ret = 0;
1910         if (!PageCompound(page))
1911                 goto out_unlock;
1912
1913         BUG_ON(!PageSwapBacked(page));
1914         __split_huge_page(page, anon_vma, list);
1915         count_vm_event(THP_SPLIT);
1916
1917         BUG_ON(PageCompound(page));
1918 out_unlock:
1919         anon_vma_unlock_write(anon_vma);
1920         put_anon_vma(anon_vma);
1921 out:
1922         return ret;
1923 }
1924
1925 #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
1926
1927 int hugepage_madvise(struct vm_area_struct *vma,
1928                      unsigned long *vm_flags, int advice)
1929 {
1930         switch (advice) {
1931         case MADV_HUGEPAGE:
1932 #ifdef CONFIG_S390
1933                 /*
1934                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1935                  * can't handle this properly after s390_enable_sie, so we simply
1936                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
1937                  */
1938                 if (mm_has_pgste(vma->vm_mm))
1939                         return 0;
1940 #endif
1941                 /*
1942                  * Be somewhat over-protective like KSM for now!
1943                  */
1944                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1945                         return -EINVAL;
1946                 *vm_flags &= ~VM_NOHUGEPAGE;
1947                 *vm_flags |= VM_HUGEPAGE;
1948                 /*
1949                  * If the vma become good for khugepaged to scan,
1950                  * register it here without waiting a page fault that
1951                  * may not happen any time soon.
1952                  */
1953                 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
1954                         return -ENOMEM;
1955                 break;
1956         case MADV_NOHUGEPAGE:
1957                 /*
1958                  * Be somewhat over-protective like KSM for now!
1959                  */
1960                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1961                         return -EINVAL;
1962                 *vm_flags &= ~VM_HUGEPAGE;
1963                 *vm_flags |= VM_NOHUGEPAGE;
1964                 /*
1965                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1966                  * this vma even if we leave the mm registered in khugepaged if
1967                  * it got registered before VM_NOHUGEPAGE was set.
1968                  */
1969                 break;
1970         }
1971
1972         return 0;
1973 }
1974
1975 static int __init khugepaged_slab_init(void)
1976 {
1977         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1978                                           sizeof(struct mm_slot),
1979                                           __alignof__(struct mm_slot), 0, NULL);
1980         if (!mm_slot_cache)
1981                 return -ENOMEM;
1982
1983         return 0;
1984 }
1985
1986 static void __init khugepaged_slab_exit(void)
1987 {
1988         kmem_cache_destroy(mm_slot_cache);
1989 }
1990
1991 static inline struct mm_slot *alloc_mm_slot(void)
1992 {
1993         if (!mm_slot_cache)     /* initialization failed */
1994                 return NULL;
1995         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1996 }
1997
1998 static inline void free_mm_slot(struct mm_slot *mm_slot)
1999 {
2000         kmem_cache_free(mm_slot_cache, mm_slot);
2001 }
2002
2003 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
2004 {
2005         struct mm_slot *mm_slot;
2006
2007         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
2008                 if (mm == mm_slot->mm)
2009                         return mm_slot;
2010
2011         return NULL;
2012 }
2013
2014 static void insert_to_mm_slots_hash(struct mm_struct *mm,
2015                                     struct mm_slot *mm_slot)
2016 {
2017         mm_slot->mm = mm;
2018         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
2019 }
2020
2021 static inline int khugepaged_test_exit(struct mm_struct *mm)
2022 {
2023         return atomic_read(&mm->mm_users) == 0;
2024 }
2025
2026 int __khugepaged_enter(struct mm_struct *mm)
2027 {
2028         struct mm_slot *mm_slot;
2029         int wakeup;
2030
2031         mm_slot = alloc_mm_slot();
2032         if (!mm_slot)
2033                 return -ENOMEM;
2034
2035         /* __khugepaged_exit() must not run from under us */
2036         VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
2037         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2038                 free_mm_slot(mm_slot);
2039                 return 0;
2040         }
2041
2042         spin_lock(&khugepaged_mm_lock);
2043         insert_to_mm_slots_hash(mm, mm_slot);
2044         /*
2045          * Insert just behind the scanning cursor, to let the area settle
2046          * down a little.
2047          */
2048         wakeup = list_empty(&khugepaged_scan.mm_head);
2049         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2050         spin_unlock(&khugepaged_mm_lock);
2051
2052         atomic_inc(&mm->mm_count);
2053         if (wakeup)
2054                 wake_up_interruptible(&khugepaged_wait);
2055
2056         return 0;
2057 }
2058
2059 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
2060                                unsigned long vm_flags)
2061 {
2062         unsigned long hstart, hend;
2063         if (!vma->anon_vma)
2064                 /*
2065                  * Not yet faulted in so we will register later in the
2066                  * page fault if needed.
2067                  */
2068                 return 0;
2069         if (vma->vm_ops)
2070                 /* khugepaged not yet working on file or special mappings */
2071                 return 0;
2072         VM_BUG_ON_VMA(vm_flags & VM_NO_THP, vma);
2073         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2074         hend = vma->vm_end & HPAGE_PMD_MASK;
2075         if (hstart < hend)
2076                 return khugepaged_enter(vma, vm_flags);
2077         return 0;
2078 }
2079
2080 void __khugepaged_exit(struct mm_struct *mm)
2081 {
2082         struct mm_slot *mm_slot;
2083         int free = 0;
2084
2085         spin_lock(&khugepaged_mm_lock);
2086         mm_slot = get_mm_slot(mm);
2087         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
2088                 hash_del(&mm_slot->hash);
2089                 list_del(&mm_slot->mm_node);
2090                 free = 1;
2091         }
2092         spin_unlock(&khugepaged_mm_lock);
2093
2094         if (free) {
2095                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2096                 free_mm_slot(mm_slot);
2097                 mmdrop(mm);
2098         } else if (mm_slot) {
2099                 /*
2100                  * This is required to serialize against
2101                  * khugepaged_test_exit() (which is guaranteed to run
2102                  * under mmap sem read mode). Stop here (after we
2103                  * return all pagetables will be destroyed) until
2104                  * khugepaged has finished working on the pagetables
2105                  * under the mmap_sem.
2106                  */
2107                 down_write(&mm->mmap_sem);
2108                 up_write(&mm->mmap_sem);
2109         }
2110 }
2111
2112 static void release_pte_page(struct page *page)
2113 {
2114         /* 0 stands for page_is_file_cache(page) == false */
2115         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2116         unlock_page(page);
2117         putback_lru_page(page);
2118 }
2119
2120 static void release_pte_pages(pte_t *pte, pte_t *_pte)
2121 {
2122         while (--_pte >= pte) {
2123                 pte_t pteval = *_pte;
2124                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
2125                         release_pte_page(pte_page(pteval));
2126         }
2127 }
2128
2129 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2130                                         unsigned long address,
2131                                         pte_t *pte)
2132 {
2133         struct page *page;
2134         pte_t *_pte;
2135         int none_or_zero = 0;
2136         bool referenced = false, writable = false;
2137         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2138              _pte++, address += PAGE_SIZE) {
2139                 pte_t pteval = *_pte;
2140                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2141                         if (++none_or_zero <= khugepaged_max_ptes_none)
2142                                 continue;
2143                         else
2144                                 goto out;
2145                 }
2146                 if (!pte_present(pteval))
2147                         goto out;
2148                 page = vm_normal_page(vma, address, pteval);
2149                 if (unlikely(!page))
2150                         goto out;
2151
2152                 VM_BUG_ON_PAGE(PageCompound(page), page);
2153                 VM_BUG_ON_PAGE(!PageAnon(page), page);
2154                 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2155
2156                 /*
2157                  * We can do it before isolate_lru_page because the
2158                  * page can't be freed from under us. NOTE: PG_lock
2159                  * is needed to serialize against split_huge_page
2160                  * when invoked from the VM.
2161                  */
2162                 if (!trylock_page(page))
2163                         goto out;
2164
2165                 /*
2166                  * cannot use mapcount: can't collapse if there's a gup pin.
2167                  * The page must only be referenced by the scanned process
2168                  * and page swap cache.
2169                  */
2170                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2171                         unlock_page(page);
2172                         goto out;
2173                 }
2174                 if (pte_write(pteval)) {
2175                         writable = true;
2176                 } else {
2177                         if (PageSwapCache(page) && !reuse_swap_page(page)) {
2178                                 unlock_page(page);
2179                                 goto out;
2180                         }
2181                         /*
2182                          * Page is not in the swap cache. It can be collapsed
2183                          * into a THP.
2184                          */
2185                 }
2186
2187                 /*
2188                  * Isolate the page to avoid collapsing an hugepage
2189                  * currently in use by the VM.
2190                  */
2191                 if (isolate_lru_page(page)) {
2192                         unlock_page(page);
2193                         goto out;
2194                 }
2195                 /* 0 stands for page_is_file_cache(page) == false */
2196                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2197                 VM_BUG_ON_PAGE(!PageLocked(page), page);
2198                 VM_BUG_ON_PAGE(PageLRU(page), page);
2199
2200                 /* If there is no mapped pte young don't collapse the page */
2201                 if (pte_young(pteval) || PageReferenced(page) ||
2202                     mmu_notifier_test_young(vma->vm_mm, address))
2203                         referenced = true;
2204         }
2205         if (likely(referenced && writable))
2206                 return 1;
2207 out:
2208         release_pte_pages(pte, _pte);
2209         return 0;
2210 }
2211
2212 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2213                                       struct vm_area_struct *vma,
2214                                       unsigned long address,
2215                                       spinlock_t *ptl)
2216 {
2217         pte_t *_pte;
2218         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2219                 pte_t pteval = *_pte;
2220                 struct page *src_page;
2221
2222                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2223                         clear_user_highpage(page, address);
2224                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2225                         if (is_zero_pfn(pte_pfn(pteval))) {
2226                                 /*
2227                                  * ptl mostly unnecessary.
2228                                  */
2229                                 spin_lock(ptl);
2230                                 /*
2231                                  * paravirt calls inside pte_clear here are
2232                                  * superfluous.
2233                                  */
2234                                 pte_clear(vma->vm_mm, address, _pte);
2235                                 spin_unlock(ptl);
2236                         }
2237                 } else {
2238                         src_page = pte_page(pteval);
2239                         copy_user_highpage(page, src_page, address, vma);
2240                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
2241                         release_pte_page(src_page);
2242                         /*
2243                          * ptl mostly unnecessary, but preempt has to
2244                          * be disabled to update the per-cpu stats
2245                          * inside page_remove_rmap().
2246                          */
2247                         spin_lock(ptl);
2248                         /*
2249                          * paravirt calls inside pte_clear here are
2250                          * superfluous.
2251                          */
2252                         pte_clear(vma->vm_mm, address, _pte);
2253                         page_remove_rmap(src_page);
2254                         spin_unlock(ptl);
2255                         free_page_and_swap_cache(src_page);
2256                 }
2257
2258                 address += PAGE_SIZE;
2259                 page++;
2260         }
2261 }
2262
2263 static void khugepaged_alloc_sleep(void)
2264 {
2265         wait_event_freezable_timeout(khugepaged_wait, false,
2266                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2267 }
2268
2269 static int khugepaged_node_load[MAX_NUMNODES];
2270
2271 static bool khugepaged_scan_abort(int nid)
2272 {
2273         int i;
2274
2275         /*
2276          * If zone_reclaim_mode is disabled, then no extra effort is made to
2277          * allocate memory locally.
2278          */
2279         if (!zone_reclaim_mode)
2280                 return false;
2281
2282         /* If there is a count for this node already, it must be acceptable */
2283         if (khugepaged_node_load[nid])
2284                 return false;
2285
2286         for (i = 0; i < MAX_NUMNODES; i++) {
2287                 if (!khugepaged_node_load[i])
2288                         continue;
2289                 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2290                         return true;
2291         }
2292         return false;
2293 }
2294
2295 #ifdef CONFIG_NUMA
2296 static int khugepaged_find_target_node(void)
2297 {
2298         static int last_khugepaged_target_node = NUMA_NO_NODE;
2299         int nid, target_node = 0, max_value = 0;
2300
2301         /* find first node with max normal pages hit */
2302         for (nid = 0; nid < MAX_NUMNODES; nid++)
2303                 if (khugepaged_node_load[nid] > max_value) {
2304                         max_value = khugepaged_node_load[nid];
2305                         target_node = nid;
2306                 }
2307
2308         /* do some balance if several nodes have the same hit record */
2309         if (target_node <= last_khugepaged_target_node)
2310                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2311                                 nid++)
2312                         if (max_value == khugepaged_node_load[nid]) {
2313                                 target_node = nid;
2314                                 break;
2315                         }
2316
2317         last_khugepaged_target_node = target_node;
2318         return target_node;
2319 }
2320
2321 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2322 {
2323         if (IS_ERR(*hpage)) {
2324                 if (!*wait)
2325                         return false;
2326
2327                 *wait = false;
2328                 *hpage = NULL;
2329                 khugepaged_alloc_sleep();
2330         } else if (*hpage) {
2331                 put_page(*hpage);
2332                 *hpage = NULL;
2333         }
2334
2335         return true;
2336 }
2337
2338 static struct page *
2339 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2340                        struct vm_area_struct *vma, unsigned long address,
2341                        int node)
2342 {
2343         VM_BUG_ON_PAGE(*hpage, *hpage);
2344
2345         /*
2346          * Before allocating the hugepage, release the mmap_sem read lock.
2347          * The allocation can take potentially a long time if it involves
2348          * sync compaction, and we do not need to hold the mmap_sem during
2349          * that. We will recheck the vma after taking it again in write mode.
2350          */
2351         up_read(&mm->mmap_sem);
2352
2353         *hpage = alloc_pages_exact_node(node, gfp, HPAGE_PMD_ORDER);
2354         if (unlikely(!*hpage)) {
2355                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2356                 *hpage = ERR_PTR(-ENOMEM);
2357                 return NULL;
2358         }
2359
2360         count_vm_event(THP_COLLAPSE_ALLOC);
2361         return *hpage;
2362 }
2363 #else
2364 static int khugepaged_find_target_node(void)
2365 {
2366         return 0;
2367 }
2368
2369 static inline struct page *alloc_hugepage(int defrag)
2370 {
2371         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2372                            HPAGE_PMD_ORDER);
2373 }
2374
2375 static struct page *khugepaged_alloc_hugepage(bool *wait)
2376 {
2377         struct page *hpage;
2378
2379         do {
2380                 hpage = alloc_hugepage(khugepaged_defrag());
2381                 if (!hpage) {
2382                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2383                         if (!*wait)
2384                                 return NULL;
2385
2386                         *wait = false;
2387                         khugepaged_alloc_sleep();
2388                 } else
2389                         count_vm_event(THP_COLLAPSE_ALLOC);
2390         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2391
2392         return hpage;
2393 }
2394
2395 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2396 {
2397         if (!*hpage)
2398                 *hpage = khugepaged_alloc_hugepage(wait);
2399
2400         if (unlikely(!*hpage))
2401                 return false;
2402
2403         return true;
2404 }
2405
2406 static struct page *
2407 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2408                        struct vm_area_struct *vma, unsigned long address,
2409                        int node)
2410 {
2411         up_read(&mm->mmap_sem);
2412         VM_BUG_ON(!*hpage);
2413
2414         return  *hpage;
2415 }
2416 #endif
2417
2418 static bool hugepage_vma_check(struct vm_area_struct *vma)
2419 {
2420         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2421             (vma->vm_flags & VM_NOHUGEPAGE))
2422                 return false;
2423
2424         if (!vma->anon_vma || vma->vm_ops)
2425                 return false;
2426         if (is_vma_temporary_stack(vma))
2427                 return false;
2428         VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma);
2429         return true;
2430 }
2431
2432 static void collapse_huge_page(struct mm_struct *mm,
2433                                    unsigned long address,
2434                                    struct page **hpage,
2435                                    struct vm_area_struct *vma,
2436                                    int node)
2437 {
2438         pmd_t *pmd, _pmd;
2439         pte_t *pte;
2440         pgtable_t pgtable;
2441         struct page *new_page;
2442         spinlock_t *pmd_ptl, *pte_ptl;
2443         int isolated;
2444         unsigned long hstart, hend;
2445         struct mem_cgroup *memcg;
2446         unsigned long mmun_start;       /* For mmu_notifiers */
2447         unsigned long mmun_end;         /* For mmu_notifiers */
2448         gfp_t gfp;
2449
2450         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2451
2452         /* Only allocate from the target node */
2453         gfp = alloc_hugepage_gfpmask(khugepaged_defrag(), __GFP_OTHER_NODE) |
2454                 __GFP_THISNODE;
2455
2456         /* release the mmap_sem read lock. */
2457         new_page = khugepaged_alloc_page(hpage, gfp, mm, vma, address, node);
2458         if (!new_page)
2459                 return;
2460
2461         if (unlikely(mem_cgroup_try_charge(new_page, mm,
2462                                            gfp, &memcg)))
2463                 return;
2464
2465         /*
2466          * Prevent all access to pagetables with the exception of
2467          * gup_fast later hanlded by the ptep_clear_flush and the VM
2468          * handled by the anon_vma lock + PG_lock.
2469          */
2470         down_write(&mm->mmap_sem);
2471         if (unlikely(khugepaged_test_exit(mm)))
2472                 goto out;
2473
2474         vma = find_vma(mm, address);
2475         if (!vma)
2476                 goto out;
2477         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2478         hend = vma->vm_end & HPAGE_PMD_MASK;
2479         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2480                 goto out;
2481         if (!hugepage_vma_check(vma))
2482                 goto out;
2483         pmd = mm_find_pmd(mm, address);
2484         if (!pmd)
2485                 goto out;
2486
2487         anon_vma_lock_write(vma->anon_vma);
2488
2489         pte = pte_offset_map(pmd, address);
2490         pte_ptl = pte_lockptr(mm, pmd);
2491
2492         mmun_start = address;
2493         mmun_end   = address + HPAGE_PMD_SIZE;
2494         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2495         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2496         /*
2497          * After this gup_fast can't run anymore. This also removes
2498          * any huge TLB entry from the CPU so we won't allow
2499          * huge and small TLB entries for the same virtual address
2500          * to avoid the risk of CPU bugs in that area.
2501          */
2502         _pmd = pmdp_clear_flush(vma, address, pmd);
2503         spin_unlock(pmd_ptl);
2504         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2505
2506         spin_lock(pte_ptl);
2507         isolated = __collapse_huge_page_isolate(vma, address, pte);
2508         spin_unlock(pte_ptl);
2509
2510         if (unlikely(!isolated)) {
2511                 pte_unmap(pte);
2512                 spin_lock(pmd_ptl);
2513                 BUG_ON(!pmd_none(*pmd));
2514                 /*
2515                  * We can only use set_pmd_at when establishing
2516                  * hugepmds and never for establishing regular pmds that
2517                  * points to regular pagetables. Use pmd_populate for that
2518                  */
2519                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2520                 spin_unlock(pmd_ptl);
2521                 anon_vma_unlock_write(vma->anon_vma);
2522                 goto out;
2523         }
2524
2525         /*
2526          * All pages are isolated and locked so anon_vma rmap
2527          * can't run anymore.
2528          */
2529         anon_vma_unlock_write(vma->anon_vma);
2530
2531         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2532         pte_unmap(pte);
2533         __SetPageUptodate(new_page);
2534         pgtable = pmd_pgtable(_pmd);
2535
2536         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2537         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2538
2539         /*
2540          * spin_lock() below is not the equivalent of smp_wmb(), so
2541          * this is needed to avoid the copy_huge_page writes to become
2542          * visible after the set_pmd_at() write.
2543          */
2544         smp_wmb();
2545
2546         spin_lock(pmd_ptl);
2547         BUG_ON(!pmd_none(*pmd));
2548         page_add_new_anon_rmap(new_page, vma, address);
2549         mem_cgroup_commit_charge(new_page, memcg, false);
2550         lru_cache_add_active_or_unevictable(new_page, vma);
2551         pgtable_trans_huge_deposit(mm, pmd, pgtable);
2552         set_pmd_at(mm, address, pmd, _pmd);
2553         update_mmu_cache_pmd(vma, address, pmd);
2554         spin_unlock(pmd_ptl);
2555
2556         *hpage = NULL;
2557
2558         khugepaged_pages_collapsed++;
2559 out_up_write:
2560         up_write(&mm->mmap_sem);
2561         return;
2562
2563 out:
2564         mem_cgroup_cancel_charge(new_page, memcg);
2565         goto out_up_write;
2566 }
2567
2568 static int khugepaged_scan_pmd(struct mm_struct *mm,
2569                                struct vm_area_struct *vma,
2570                                unsigned long address,
2571                                struct page **hpage)
2572 {
2573         pmd_t *pmd;
2574         pte_t *pte, *_pte;
2575         int ret = 0, none_or_zero = 0;
2576         struct page *page;
2577         unsigned long _address;
2578         spinlock_t *ptl;
2579         int node = NUMA_NO_NODE;
2580         bool writable = false, referenced = false;
2581
2582         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2583
2584         pmd = mm_find_pmd(mm, address);
2585         if (!pmd)
2586                 goto out;
2587
2588         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2589         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2590         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2591              _pte++, _address += PAGE_SIZE) {
2592                 pte_t pteval = *_pte;
2593                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2594                         if (++none_or_zero <= khugepaged_max_ptes_none)
2595                                 continue;
2596                         else
2597                                 goto out_unmap;
2598                 }
2599                 if (!pte_present(pteval))
2600                         goto out_unmap;
2601                 if (pte_write(pteval))
2602                         writable = true;
2603
2604                 page = vm_normal_page(vma, _address, pteval);
2605                 if (unlikely(!page))
2606                         goto out_unmap;
2607                 /*
2608                  * Record which node the original page is from and save this
2609                  * information to khugepaged_node_load[].
2610                  * Khupaged will allocate hugepage from the node has the max
2611                  * hit record.
2612                  */
2613                 node = page_to_nid(page);
2614                 if (khugepaged_scan_abort(node))
2615                         goto out_unmap;
2616                 khugepaged_node_load[node]++;
2617                 VM_BUG_ON_PAGE(PageCompound(page), page);
2618                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2619                         goto out_unmap;
2620                 /*
2621                  * cannot use mapcount: can't collapse if there's a gup pin.
2622                  * The page must only be referenced by the scanned process
2623                  * and page swap cache.
2624                  */
2625                 if (page_count(page) != 1 + !!PageSwapCache(page))
2626                         goto out_unmap;
2627                 if (pte_young(pteval) || PageReferenced(page) ||
2628                     mmu_notifier_test_young(vma->vm_mm, address))
2629                         referenced = true;
2630         }
2631         if (referenced && writable)
2632                 ret = 1;
2633 out_unmap:
2634         pte_unmap_unlock(pte, ptl);
2635         if (ret) {
2636                 node = khugepaged_find_target_node();
2637                 /* collapse_huge_page will return with the mmap_sem released */
2638                 collapse_huge_page(mm, address, hpage, vma, node);
2639         }
2640 out:
2641         return ret;
2642 }
2643
2644 static void collect_mm_slot(struct mm_slot *mm_slot)
2645 {
2646         struct mm_struct *mm = mm_slot->mm;
2647
2648         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2649
2650         if (khugepaged_test_exit(mm)) {
2651                 /* free mm_slot */
2652                 hash_del(&mm_slot->hash);
2653                 list_del(&mm_slot->mm_node);
2654
2655                 /*
2656                  * Not strictly needed because the mm exited already.
2657                  *
2658                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2659                  */
2660
2661                 /* khugepaged_mm_lock actually not necessary for the below */
2662                 free_mm_slot(mm_slot);
2663                 mmdrop(mm);
2664         }
2665 }
2666
2667 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2668                                             struct page **hpage)
2669         __releases(&khugepaged_mm_lock)
2670         __acquires(&khugepaged_mm_lock)
2671 {
2672         struct mm_slot *mm_slot;
2673         struct mm_struct *mm;
2674         struct vm_area_struct *vma;
2675         int progress = 0;
2676
2677         VM_BUG_ON(!pages);
2678         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2679
2680         if (khugepaged_scan.mm_slot)
2681                 mm_slot = khugepaged_scan.mm_slot;
2682         else {
2683                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2684                                      struct mm_slot, mm_node);
2685                 khugepaged_scan.address = 0;
2686                 khugepaged_scan.mm_slot = mm_slot;
2687         }
2688         spin_unlock(&khugepaged_mm_lock);
2689
2690         mm = mm_slot->mm;
2691         down_read(&mm->mmap_sem);
2692         if (unlikely(khugepaged_test_exit(mm)))
2693                 vma = NULL;
2694         else
2695                 vma = find_vma(mm, khugepaged_scan.address);
2696
2697         progress++;
2698         for (; vma; vma = vma->vm_next) {
2699                 unsigned long hstart, hend;
2700
2701                 cond_resched();
2702                 if (unlikely(khugepaged_test_exit(mm))) {
2703                         progress++;
2704                         break;
2705                 }
2706                 if (!hugepage_vma_check(vma)) {
2707 skip:
2708                         progress++;
2709                         continue;
2710                 }
2711                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2712                 hend = vma->vm_end & HPAGE_PMD_MASK;
2713                 if (hstart >= hend)
2714                         goto skip;
2715                 if (khugepaged_scan.address > hend)
2716                         goto skip;
2717                 if (khugepaged_scan.address < hstart)
2718                         khugepaged_scan.address = hstart;
2719                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2720
2721                 while (khugepaged_scan.address < hend) {
2722                         int ret;
2723                         cond_resched();
2724                         if (unlikely(khugepaged_test_exit(mm)))
2725                                 goto breakouterloop;
2726
2727                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2728                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2729                                   hend);
2730                         ret = khugepaged_scan_pmd(mm, vma,
2731                                                   khugepaged_scan.address,
2732                                                   hpage);
2733                         /* move to next address */
2734                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2735                         progress += HPAGE_PMD_NR;
2736                         if (ret)
2737                                 /* we released mmap_sem so break loop */
2738                                 goto breakouterloop_mmap_sem;
2739                         if (progress >= pages)
2740                                 goto breakouterloop;
2741                 }
2742         }
2743 breakouterloop:
2744         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2745 breakouterloop_mmap_sem:
2746
2747         spin_lock(&khugepaged_mm_lock);
2748         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2749         /*
2750          * Release the current mm_slot if this mm is about to die, or
2751          * if we scanned all vmas of this mm.
2752          */
2753         if (khugepaged_test_exit(mm) || !vma) {
2754                 /*
2755                  * Make sure that if mm_users is reaching zero while
2756                  * khugepaged runs here, khugepaged_exit will find
2757                  * mm_slot not pointing to the exiting mm.
2758                  */
2759                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2760                         khugepaged_scan.mm_slot = list_entry(
2761                                 mm_slot->mm_node.next,
2762                                 struct mm_slot, mm_node);
2763                         khugepaged_scan.address = 0;
2764                 } else {
2765                         khugepaged_scan.mm_slot = NULL;
2766                         khugepaged_full_scans++;
2767                 }
2768
2769                 collect_mm_slot(mm_slot);
2770         }
2771
2772         return progress;
2773 }
2774
2775 static int khugepaged_has_work(void)
2776 {
2777         return !list_empty(&khugepaged_scan.mm_head) &&
2778                 khugepaged_enabled();
2779 }
2780
2781 static int khugepaged_wait_event(void)
2782 {
2783         return !list_empty(&khugepaged_scan.mm_head) ||
2784                 kthread_should_stop();
2785 }
2786
2787 static void khugepaged_do_scan(void)
2788 {
2789         struct page *hpage = NULL;
2790         unsigned int progress = 0, pass_through_head = 0;
2791         unsigned int pages = khugepaged_pages_to_scan;
2792         bool wait = true;
2793
2794         barrier(); /* write khugepaged_pages_to_scan to local stack */
2795
2796         while (progress < pages) {
2797                 if (!khugepaged_prealloc_page(&hpage, &wait))
2798                         break;
2799
2800                 cond_resched();
2801
2802                 if (unlikely(kthread_should_stop() || freezing(current)))
2803                         break;
2804
2805                 spin_lock(&khugepaged_mm_lock);
2806                 if (!khugepaged_scan.mm_slot)
2807                         pass_through_head++;
2808                 if (khugepaged_has_work() &&
2809                     pass_through_head < 2)
2810                         progress += khugepaged_scan_mm_slot(pages - progress,
2811                                                             &hpage);
2812                 else
2813                         progress = pages;
2814                 spin_unlock(&khugepaged_mm_lock);
2815         }
2816
2817         if (!IS_ERR_OR_NULL(hpage))
2818                 put_page(hpage);
2819 }
2820
2821 static void khugepaged_wait_work(void)
2822 {
2823         try_to_freeze();
2824
2825         if (khugepaged_has_work()) {
2826                 if (!khugepaged_scan_sleep_millisecs)
2827                         return;
2828
2829                 wait_event_freezable_timeout(khugepaged_wait,
2830                                              kthread_should_stop(),
2831                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2832                 return;
2833         }
2834
2835         if (khugepaged_enabled())
2836                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2837 }
2838
2839 static int khugepaged(void *none)
2840 {
2841         struct mm_slot *mm_slot;
2842
2843         set_freezable();
2844         set_user_nice(current, MAX_NICE);
2845
2846         while (!kthread_should_stop()) {
2847                 khugepaged_do_scan();
2848                 khugepaged_wait_work();
2849         }
2850
2851         spin_lock(&khugepaged_mm_lock);
2852         mm_slot = khugepaged_scan.mm_slot;
2853         khugepaged_scan.mm_slot = NULL;
2854         if (mm_slot)
2855                 collect_mm_slot(mm_slot);
2856         spin_unlock(&khugepaged_mm_lock);
2857         return 0;
2858 }
2859
2860 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2861                 unsigned long haddr, pmd_t *pmd)
2862 {
2863         struct mm_struct *mm = vma->vm_mm;
2864         pgtable_t pgtable;
2865         pmd_t _pmd;
2866         int i;
2867
2868         pmdp_clear_flush_notify(vma, haddr, pmd);
2869         /* leave pmd empty until pte is filled */
2870
2871         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2872         pmd_populate(mm, &_pmd, pgtable);
2873
2874         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2875                 pte_t *pte, entry;
2876                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2877                 entry = pte_mkspecial(entry);
2878                 pte = pte_offset_map(&_pmd, haddr);
2879                 VM_BUG_ON(!pte_none(*pte));
2880                 set_pte_at(mm, haddr, pte, entry);
2881                 pte_unmap(pte);
2882         }
2883         smp_wmb(); /* make pte visible before pmd */
2884         pmd_populate(mm, pmd, pgtable);
2885         put_huge_zero_page();
2886 }
2887
2888 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2889                 pmd_t *pmd)
2890 {
2891         spinlock_t *ptl;
2892         struct page *page;
2893         struct mm_struct *mm = vma->vm_mm;
2894         unsigned long haddr = address & HPAGE_PMD_MASK;
2895         unsigned long mmun_start;       /* For mmu_notifiers */
2896         unsigned long mmun_end;         /* For mmu_notifiers */
2897
2898         BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
2899
2900         mmun_start = haddr;
2901         mmun_end   = haddr + HPAGE_PMD_SIZE;
2902 again:
2903         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2904         ptl = pmd_lock(mm, pmd);
2905         if (unlikely(!pmd_trans_huge(*pmd))) {
2906                 spin_unlock(ptl);
2907                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2908                 return;
2909         }
2910         if (is_huge_zero_pmd(*pmd)) {
2911                 __split_huge_zero_page_pmd(vma, haddr, pmd);
2912                 spin_unlock(ptl);
2913                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2914                 return;
2915         }
2916         page = pmd_page(*pmd);
2917         VM_BUG_ON_PAGE(!page_count(page), page);
2918         get_page(page);
2919         spin_unlock(ptl);
2920         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2921
2922         split_huge_page(page);
2923
2924         put_page(page);
2925
2926         /*
2927          * We don't always have down_write of mmap_sem here: a racing
2928          * do_huge_pmd_wp_page() might have copied-on-write to another
2929          * huge page before our split_huge_page() got the anon_vma lock.
2930          */
2931         if (unlikely(pmd_trans_huge(*pmd)))
2932                 goto again;
2933 }
2934
2935 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2936                 pmd_t *pmd)
2937 {
2938         struct vm_area_struct *vma;
2939
2940         vma = find_vma(mm, address);
2941         BUG_ON(vma == NULL);
2942         split_huge_page_pmd(vma, address, pmd);
2943 }
2944
2945 static void split_huge_page_address(struct mm_struct *mm,
2946                                     unsigned long address)
2947 {
2948         pgd_t *pgd;
2949         pud_t *pud;
2950         pmd_t *pmd;
2951
2952         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2953
2954         pgd = pgd_offset(mm, address);
2955         if (!pgd_present(*pgd))
2956                 return;
2957
2958         pud = pud_offset(pgd, address);
2959         if (!pud_present(*pud))
2960                 return;
2961
2962         pmd = pmd_offset(pud, address);
2963         if (!pmd_present(*pmd))
2964                 return;
2965         /*
2966          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2967          * materialize from under us.
2968          */
2969         split_huge_page_pmd_mm(mm, address, pmd);
2970 }
2971
2972 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2973                              unsigned long start,
2974                              unsigned long end,
2975                              long adjust_next)
2976 {
2977         /*
2978          * If the new start address isn't hpage aligned and it could
2979          * previously contain an hugepage: check if we need to split
2980          * an huge pmd.
2981          */
2982         if (start & ~HPAGE_PMD_MASK &&
2983             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2984             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2985                 split_huge_page_address(vma->vm_mm, start);
2986
2987         /*
2988          * If the new end address isn't hpage aligned and it could
2989          * previously contain an hugepage: check if we need to split
2990          * an huge pmd.
2991          */
2992         if (end & ~HPAGE_PMD_MASK &&
2993             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2994             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2995                 split_huge_page_address(vma->vm_mm, end);
2996
2997         /*
2998          * If we're also updating the vma->vm_next->vm_start, if the new
2999          * vm_next->vm_start isn't page aligned and it could previously
3000          * contain an hugepage: check if we need to split an huge pmd.
3001          */
3002         if (adjust_next > 0) {
3003                 struct vm_area_struct *next = vma->vm_next;
3004                 unsigned long nstart = next->vm_start;
3005                 nstart += adjust_next << PAGE_SHIFT;
3006                 if (nstart & ~HPAGE_PMD_MASK &&
3007                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
3008                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
3009                         split_huge_page_address(next->vm_mm, nstart);
3010         }
3011 }