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_huge_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_huge_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_huge_get_and_clear. So do the
1400                  * pgtable_trans_huge_withdraw after finishing pmdp related
1401                  * operations.
1402                  */
1403                 orig_pmd = pmdp_huge_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_huge_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_huge_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                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1680                 page_tail->flags |= (page->flags &
1681                                      ((1L << PG_referenced) |
1682                                       (1L << PG_swapbacked) |
1683                                       (1L << PG_mlocked) |
1684                                       (1L << PG_uptodate) |
1685                                       (1L << PG_active) |
1686                                       (1L << PG_unevictable)));
1687                 page_tail->flags |= (1L << PG_dirty);
1688
1689                 /* clear PageTail before overwriting first_page */
1690                 smp_wmb();
1691
1692                 /*
1693                  * __split_huge_page_splitting() already set the
1694                  * splitting bit in all pmd that could map this
1695                  * hugepage, that will ensure no CPU can alter the
1696                  * mapcount on the head page. The mapcount is only
1697                  * accounted in the head page and it has to be
1698                  * transferred to all tail pages in the below code. So
1699                  * for this code to be safe, the split the mapcount
1700                  * can't change. But that doesn't mean userland can't
1701                  * keep changing and reading the page contents while
1702                  * we transfer the mapcount, so the pmd splitting
1703                  * status is achieved setting a reserved bit in the
1704                  * pmd, not by clearing the present bit.
1705                 */
1706                 page_tail->_mapcount = page->_mapcount;
1707
1708                 BUG_ON(page_tail->mapping);
1709                 page_tail->mapping = page->mapping;
1710
1711                 page_tail->index = page->index + i;
1712                 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
1713
1714                 BUG_ON(!PageAnon(page_tail));
1715                 BUG_ON(!PageUptodate(page_tail));
1716                 BUG_ON(!PageDirty(page_tail));
1717                 BUG_ON(!PageSwapBacked(page_tail));
1718
1719                 lru_add_page_tail(page, page_tail, lruvec, list);
1720         }
1721         atomic_sub(tail_count, &page->_count);
1722         BUG_ON(atomic_read(&page->_count) <= 0);
1723
1724         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1725
1726         ClearPageCompound(page);
1727         compound_unlock(page);
1728         spin_unlock_irq(&zone->lru_lock);
1729
1730         for (i = 1; i < HPAGE_PMD_NR; i++) {
1731                 struct page *page_tail = page + i;
1732                 BUG_ON(page_count(page_tail) <= 0);
1733                 /*
1734                  * Tail pages may be freed if there wasn't any mapping
1735                  * like if add_to_swap() is running on a lru page that
1736                  * had its mapping zapped. And freeing these pages
1737                  * requires taking the lru_lock so we do the put_page
1738                  * of the tail pages after the split is complete.
1739                  */
1740                 put_page(page_tail);
1741         }
1742
1743         /*
1744          * Only the head page (now become a regular page) is required
1745          * to be pinned by the caller.
1746          */
1747         BUG_ON(page_count(page) <= 0);
1748 }
1749
1750 static int __split_huge_page_map(struct page *page,
1751                                  struct vm_area_struct *vma,
1752                                  unsigned long address)
1753 {
1754         struct mm_struct *mm = vma->vm_mm;
1755         spinlock_t *ptl;
1756         pmd_t *pmd, _pmd;
1757         int ret = 0, i;
1758         pgtable_t pgtable;
1759         unsigned long haddr;
1760
1761         pmd = page_check_address_pmd(page, mm, address,
1762                         PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl);
1763         if (pmd) {
1764                 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1765                 pmd_populate(mm, &_pmd, pgtable);
1766                 if (pmd_write(*pmd))
1767                         BUG_ON(page_mapcount(page) != 1);
1768
1769                 haddr = address;
1770                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1771                         pte_t *pte, entry;
1772                         BUG_ON(PageCompound(page+i));
1773                         /*
1774                          * Note that NUMA hinting access restrictions are not
1775                          * transferred to avoid any possibility of altering
1776                          * permissions across VMAs.
1777                          */
1778                         entry = mk_pte(page + i, vma->vm_page_prot);
1779                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1780                         if (!pmd_write(*pmd))
1781                                 entry = pte_wrprotect(entry);
1782                         if (!pmd_young(*pmd))
1783                                 entry = pte_mkold(entry);
1784                         pte = pte_offset_map(&_pmd, haddr);
1785                         BUG_ON(!pte_none(*pte));
1786                         set_pte_at(mm, haddr, pte, entry);
1787                         pte_unmap(pte);
1788                 }
1789
1790                 smp_wmb(); /* make pte visible before pmd */
1791                 /*
1792                  * Up to this point the pmd is present and huge and
1793                  * userland has the whole access to the hugepage
1794                  * during the split (which happens in place). If we
1795                  * overwrite the pmd with the not-huge version
1796                  * pointing to the pte here (which of course we could
1797                  * if all CPUs were bug free), userland could trigger
1798                  * a small page size TLB miss on the small sized TLB
1799                  * while the hugepage TLB entry is still established
1800                  * in the huge TLB. Some CPU doesn't like that. See
1801                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1802                  * Erratum 383 on page 93. Intel should be safe but is
1803                  * also warns that it's only safe if the permission
1804                  * and cache attributes of the two entries loaded in
1805                  * the two TLB is identical (which should be the case
1806                  * here). But it is generally safer to never allow
1807                  * small and huge TLB entries for the same virtual
1808                  * address to be loaded simultaneously. So instead of
1809                  * doing "pmd_populate(); flush_tlb_range();" we first
1810                  * mark the current pmd notpresent (atomically because
1811                  * here the pmd_trans_huge and pmd_trans_splitting
1812                  * must remain set at all times on the pmd until the
1813                  * split is complete for this pmd), then we flush the
1814                  * SMP TLB and finally we write the non-huge version
1815                  * of the pmd entry with pmd_populate.
1816                  */
1817                 pmdp_invalidate(vma, address, pmd);
1818                 pmd_populate(mm, pmd, pgtable);
1819                 ret = 1;
1820                 spin_unlock(ptl);
1821         }
1822
1823         return ret;
1824 }
1825
1826 /* must be called with anon_vma->root->rwsem held */
1827 static void __split_huge_page(struct page *page,
1828                               struct anon_vma *anon_vma,
1829                               struct list_head *list)
1830 {
1831         int mapcount, mapcount2;
1832         pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1833         struct anon_vma_chain *avc;
1834
1835         BUG_ON(!PageHead(page));
1836         BUG_ON(PageTail(page));
1837
1838         mapcount = 0;
1839         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1840                 struct vm_area_struct *vma = avc->vma;
1841                 unsigned long addr = vma_address(page, vma);
1842                 BUG_ON(is_vma_temporary_stack(vma));
1843                 mapcount += __split_huge_page_splitting(page, vma, addr);
1844         }
1845         /*
1846          * It is critical that new vmas are added to the tail of the
1847          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1848          * and establishes a child pmd before
1849          * __split_huge_page_splitting() freezes the parent pmd (so if
1850          * we fail to prevent copy_huge_pmd() from running until the
1851          * whole __split_huge_page() is complete), we will still see
1852          * the newly established pmd of the child later during the
1853          * walk, to be able to set it as pmd_trans_splitting too.
1854          */
1855         if (mapcount != page_mapcount(page)) {
1856                 pr_err("mapcount %d page_mapcount %d\n",
1857                         mapcount, page_mapcount(page));
1858                 BUG();
1859         }
1860
1861         __split_huge_page_refcount(page, list);
1862
1863         mapcount2 = 0;
1864         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1865                 struct vm_area_struct *vma = avc->vma;
1866                 unsigned long addr = vma_address(page, vma);
1867                 BUG_ON(is_vma_temporary_stack(vma));
1868                 mapcount2 += __split_huge_page_map(page, vma, addr);
1869         }
1870         if (mapcount != mapcount2) {
1871                 pr_err("mapcount %d mapcount2 %d page_mapcount %d\n",
1872                         mapcount, mapcount2, page_mapcount(page));
1873                 BUG();
1874         }
1875 }
1876
1877 /*
1878  * Split a hugepage into normal pages. This doesn't change the position of head
1879  * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1880  * @list. Both head page and tail pages will inherit mapping, flags, and so on
1881  * from the hugepage.
1882  * Return 0 if the hugepage is split successfully otherwise return 1.
1883  */
1884 int split_huge_page_to_list(struct page *page, struct list_head *list)
1885 {
1886         struct anon_vma *anon_vma;
1887         int ret = 1;
1888
1889         BUG_ON(is_huge_zero_page(page));
1890         BUG_ON(!PageAnon(page));
1891
1892         /*
1893          * The caller does not necessarily hold an mmap_sem that would prevent
1894          * the anon_vma disappearing so we first we take a reference to it
1895          * and then lock the anon_vma for write. This is similar to
1896          * page_lock_anon_vma_read except the write lock is taken to serialise
1897          * against parallel split or collapse operations.
1898          */
1899         anon_vma = page_get_anon_vma(page);
1900         if (!anon_vma)
1901                 goto out;
1902         anon_vma_lock_write(anon_vma);
1903
1904         ret = 0;
1905         if (!PageCompound(page))
1906                 goto out_unlock;
1907
1908         BUG_ON(!PageSwapBacked(page));
1909         __split_huge_page(page, anon_vma, list);
1910         count_vm_event(THP_SPLIT);
1911
1912         BUG_ON(PageCompound(page));
1913 out_unlock:
1914         anon_vma_unlock_write(anon_vma);
1915         put_anon_vma(anon_vma);
1916 out:
1917         return ret;
1918 }
1919
1920 #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
1921
1922 int hugepage_madvise(struct vm_area_struct *vma,
1923                      unsigned long *vm_flags, int advice)
1924 {
1925         switch (advice) {
1926         case MADV_HUGEPAGE:
1927 #ifdef CONFIG_S390
1928                 /*
1929                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1930                  * can't handle this properly after s390_enable_sie, so we simply
1931                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
1932                  */
1933                 if (mm_has_pgste(vma->vm_mm))
1934                         return 0;
1935 #endif
1936                 /*
1937                  * Be somewhat over-protective like KSM for now!
1938                  */
1939                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1940                         return -EINVAL;
1941                 *vm_flags &= ~VM_NOHUGEPAGE;
1942                 *vm_flags |= VM_HUGEPAGE;
1943                 /*
1944                  * If the vma become good for khugepaged to scan,
1945                  * register it here without waiting a page fault that
1946                  * may not happen any time soon.
1947                  */
1948                 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
1949                         return -ENOMEM;
1950                 break;
1951         case MADV_NOHUGEPAGE:
1952                 /*
1953                  * Be somewhat over-protective like KSM for now!
1954                  */
1955                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1956                         return -EINVAL;
1957                 *vm_flags &= ~VM_HUGEPAGE;
1958                 *vm_flags |= VM_NOHUGEPAGE;
1959                 /*
1960                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1961                  * this vma even if we leave the mm registered in khugepaged if
1962                  * it got registered before VM_NOHUGEPAGE was set.
1963                  */
1964                 break;
1965         }
1966
1967         return 0;
1968 }
1969
1970 static int __init khugepaged_slab_init(void)
1971 {
1972         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1973                                           sizeof(struct mm_slot),
1974                                           __alignof__(struct mm_slot), 0, NULL);
1975         if (!mm_slot_cache)
1976                 return -ENOMEM;
1977
1978         return 0;
1979 }
1980
1981 static void __init khugepaged_slab_exit(void)
1982 {
1983         kmem_cache_destroy(mm_slot_cache);
1984 }
1985
1986 static inline struct mm_slot *alloc_mm_slot(void)
1987 {
1988         if (!mm_slot_cache)     /* initialization failed */
1989                 return NULL;
1990         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1991 }
1992
1993 static inline void free_mm_slot(struct mm_slot *mm_slot)
1994 {
1995         kmem_cache_free(mm_slot_cache, mm_slot);
1996 }
1997
1998 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1999 {
2000         struct mm_slot *mm_slot;
2001
2002         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
2003                 if (mm == mm_slot->mm)
2004                         return mm_slot;
2005
2006         return NULL;
2007 }
2008
2009 static void insert_to_mm_slots_hash(struct mm_struct *mm,
2010                                     struct mm_slot *mm_slot)
2011 {
2012         mm_slot->mm = mm;
2013         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
2014 }
2015
2016 static inline int khugepaged_test_exit(struct mm_struct *mm)
2017 {
2018         return atomic_read(&mm->mm_users) == 0;
2019 }
2020
2021 int __khugepaged_enter(struct mm_struct *mm)
2022 {
2023         struct mm_slot *mm_slot;
2024         int wakeup;
2025
2026         mm_slot = alloc_mm_slot();
2027         if (!mm_slot)
2028                 return -ENOMEM;
2029
2030         /* __khugepaged_exit() must not run from under us */
2031         VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
2032         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2033                 free_mm_slot(mm_slot);
2034                 return 0;
2035         }
2036
2037         spin_lock(&khugepaged_mm_lock);
2038         insert_to_mm_slots_hash(mm, mm_slot);
2039         /*
2040          * Insert just behind the scanning cursor, to let the area settle
2041          * down a little.
2042          */
2043         wakeup = list_empty(&khugepaged_scan.mm_head);
2044         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2045         spin_unlock(&khugepaged_mm_lock);
2046
2047         atomic_inc(&mm->mm_count);
2048         if (wakeup)
2049                 wake_up_interruptible(&khugepaged_wait);
2050
2051         return 0;
2052 }
2053
2054 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
2055                                unsigned long vm_flags)
2056 {
2057         unsigned long hstart, hend;
2058         if (!vma->anon_vma)
2059                 /*
2060                  * Not yet faulted in so we will register later in the
2061                  * page fault if needed.
2062                  */
2063                 return 0;
2064         if (vma->vm_ops)
2065                 /* khugepaged not yet working on file or special mappings */
2066                 return 0;
2067         VM_BUG_ON_VMA(vm_flags & VM_NO_THP, vma);
2068         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2069         hend = vma->vm_end & HPAGE_PMD_MASK;
2070         if (hstart < hend)
2071                 return khugepaged_enter(vma, vm_flags);
2072         return 0;
2073 }
2074
2075 void __khugepaged_exit(struct mm_struct *mm)
2076 {
2077         struct mm_slot *mm_slot;
2078         int free = 0;
2079
2080         spin_lock(&khugepaged_mm_lock);
2081         mm_slot = get_mm_slot(mm);
2082         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
2083                 hash_del(&mm_slot->hash);
2084                 list_del(&mm_slot->mm_node);
2085                 free = 1;
2086         }
2087         spin_unlock(&khugepaged_mm_lock);
2088
2089         if (free) {
2090                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2091                 free_mm_slot(mm_slot);
2092                 mmdrop(mm);
2093         } else if (mm_slot) {
2094                 /*
2095                  * This is required to serialize against
2096                  * khugepaged_test_exit() (which is guaranteed to run
2097                  * under mmap sem read mode). Stop here (after we
2098                  * return all pagetables will be destroyed) until
2099                  * khugepaged has finished working on the pagetables
2100                  * under the mmap_sem.
2101                  */
2102                 down_write(&mm->mmap_sem);
2103                 up_write(&mm->mmap_sem);
2104         }
2105 }
2106
2107 static void release_pte_page(struct page *page)
2108 {
2109         /* 0 stands for page_is_file_cache(page) == false */
2110         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2111         unlock_page(page);
2112         putback_lru_page(page);
2113 }
2114
2115 static void release_pte_pages(pte_t *pte, pte_t *_pte)
2116 {
2117         while (--_pte >= pte) {
2118                 pte_t pteval = *_pte;
2119                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
2120                         release_pte_page(pte_page(pteval));
2121         }
2122 }
2123
2124 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2125                                         unsigned long address,
2126                                         pte_t *pte)
2127 {
2128         struct page *page;
2129         pte_t *_pte;
2130         int none_or_zero = 0;
2131         bool referenced = false, writable = false;
2132         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2133              _pte++, address += PAGE_SIZE) {
2134                 pte_t pteval = *_pte;
2135                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2136                         if (++none_or_zero <= khugepaged_max_ptes_none)
2137                                 continue;
2138                         else
2139                                 goto out;
2140                 }
2141                 if (!pte_present(pteval))
2142                         goto out;
2143                 page = vm_normal_page(vma, address, pteval);
2144                 if (unlikely(!page))
2145                         goto out;
2146
2147                 VM_BUG_ON_PAGE(PageCompound(page), page);
2148                 VM_BUG_ON_PAGE(!PageAnon(page), page);
2149                 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2150
2151                 /*
2152                  * We can do it before isolate_lru_page because the
2153                  * page can't be freed from under us. NOTE: PG_lock
2154                  * is needed to serialize against split_huge_page
2155                  * when invoked from the VM.
2156                  */
2157                 if (!trylock_page(page))
2158                         goto out;
2159
2160                 /*
2161                  * cannot use mapcount: can't collapse if there's a gup pin.
2162                  * The page must only be referenced by the scanned process
2163                  * and page swap cache.
2164                  */
2165                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2166                         unlock_page(page);
2167                         goto out;
2168                 }
2169                 if (pte_write(pteval)) {
2170                         writable = true;
2171                 } else {
2172                         if (PageSwapCache(page) && !reuse_swap_page(page)) {
2173                                 unlock_page(page);
2174                                 goto out;
2175                         }
2176                         /*
2177                          * Page is not in the swap cache. It can be collapsed
2178                          * into a THP.
2179                          */
2180                 }
2181
2182                 /*
2183                  * Isolate the page to avoid collapsing an hugepage
2184                  * currently in use by the VM.
2185                  */
2186                 if (isolate_lru_page(page)) {
2187                         unlock_page(page);
2188                         goto out;
2189                 }
2190                 /* 0 stands for page_is_file_cache(page) == false */
2191                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2192                 VM_BUG_ON_PAGE(!PageLocked(page), page);
2193                 VM_BUG_ON_PAGE(PageLRU(page), page);
2194
2195                 /* If there is no mapped pte young don't collapse the page */
2196                 if (pte_young(pteval) || PageReferenced(page) ||
2197                     mmu_notifier_test_young(vma->vm_mm, address))
2198                         referenced = true;
2199         }
2200         if (likely(referenced && writable))
2201                 return 1;
2202 out:
2203         release_pte_pages(pte, _pte);
2204         return 0;
2205 }
2206
2207 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2208                                       struct vm_area_struct *vma,
2209                                       unsigned long address,
2210                                       spinlock_t *ptl)
2211 {
2212         pte_t *_pte;
2213         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2214                 pte_t pteval = *_pte;
2215                 struct page *src_page;
2216
2217                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2218                         clear_user_highpage(page, address);
2219                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2220                         if (is_zero_pfn(pte_pfn(pteval))) {
2221                                 /*
2222                                  * ptl mostly unnecessary.
2223                                  */
2224                                 spin_lock(ptl);
2225                                 /*
2226                                  * paravirt calls inside pte_clear here are
2227                                  * superfluous.
2228                                  */
2229                                 pte_clear(vma->vm_mm, address, _pte);
2230                                 spin_unlock(ptl);
2231                         }
2232                 } else {
2233                         src_page = pte_page(pteval);
2234                         copy_user_highpage(page, src_page, address, vma);
2235                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
2236                         release_pte_page(src_page);
2237                         /*
2238                          * ptl mostly unnecessary, but preempt has to
2239                          * be disabled to update the per-cpu stats
2240                          * inside page_remove_rmap().
2241                          */
2242                         spin_lock(ptl);
2243                         /*
2244                          * paravirt calls inside pte_clear here are
2245                          * superfluous.
2246                          */
2247                         pte_clear(vma->vm_mm, address, _pte);
2248                         page_remove_rmap(src_page);
2249                         spin_unlock(ptl);
2250                         free_page_and_swap_cache(src_page);
2251                 }
2252
2253                 address += PAGE_SIZE;
2254                 page++;
2255         }
2256 }
2257
2258 static void khugepaged_alloc_sleep(void)
2259 {
2260         wait_event_freezable_timeout(khugepaged_wait, false,
2261                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2262 }
2263
2264 static int khugepaged_node_load[MAX_NUMNODES];
2265
2266 static bool khugepaged_scan_abort(int nid)
2267 {
2268         int i;
2269
2270         /*
2271          * If zone_reclaim_mode is disabled, then no extra effort is made to
2272          * allocate memory locally.
2273          */
2274         if (!zone_reclaim_mode)
2275                 return false;
2276
2277         /* If there is a count for this node already, it must be acceptable */
2278         if (khugepaged_node_load[nid])
2279                 return false;
2280
2281         for (i = 0; i < MAX_NUMNODES; i++) {
2282                 if (!khugepaged_node_load[i])
2283                         continue;
2284                 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2285                         return true;
2286         }
2287         return false;
2288 }
2289
2290 #ifdef CONFIG_NUMA
2291 static int khugepaged_find_target_node(void)
2292 {
2293         static int last_khugepaged_target_node = NUMA_NO_NODE;
2294         int nid, target_node = 0, max_value = 0;
2295
2296         /* find first node with max normal pages hit */
2297         for (nid = 0; nid < MAX_NUMNODES; nid++)
2298                 if (khugepaged_node_load[nid] > max_value) {
2299                         max_value = khugepaged_node_load[nid];
2300                         target_node = nid;
2301                 }
2302
2303         /* do some balance if several nodes have the same hit record */
2304         if (target_node <= last_khugepaged_target_node)
2305                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2306                                 nid++)
2307                         if (max_value == khugepaged_node_load[nid]) {
2308                                 target_node = nid;
2309                                 break;
2310                         }
2311
2312         last_khugepaged_target_node = target_node;
2313         return target_node;
2314 }
2315
2316 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2317 {
2318         if (IS_ERR(*hpage)) {
2319                 if (!*wait)
2320                         return false;
2321
2322                 *wait = false;
2323                 *hpage = NULL;
2324                 khugepaged_alloc_sleep();
2325         } else if (*hpage) {
2326                 put_page(*hpage);
2327                 *hpage = NULL;
2328         }
2329
2330         return true;
2331 }
2332
2333 static struct page *
2334 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2335                        struct vm_area_struct *vma, unsigned long address,
2336                        int node)
2337 {
2338         VM_BUG_ON_PAGE(*hpage, *hpage);
2339
2340         /*
2341          * Before allocating the hugepage, release the mmap_sem read lock.
2342          * The allocation can take potentially a long time if it involves
2343          * sync compaction, and we do not need to hold the mmap_sem during
2344          * that. We will recheck the vma after taking it again in write mode.
2345          */
2346         up_read(&mm->mmap_sem);
2347
2348         *hpage = alloc_pages_exact_node(node, gfp, HPAGE_PMD_ORDER);
2349         if (unlikely(!*hpage)) {
2350                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2351                 *hpage = ERR_PTR(-ENOMEM);
2352                 return NULL;
2353         }
2354
2355         count_vm_event(THP_COLLAPSE_ALLOC);
2356         return *hpage;
2357 }
2358 #else
2359 static int khugepaged_find_target_node(void)
2360 {
2361         return 0;
2362 }
2363
2364 static inline struct page *alloc_hugepage(int defrag)
2365 {
2366         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2367                            HPAGE_PMD_ORDER);
2368 }
2369
2370 static struct page *khugepaged_alloc_hugepage(bool *wait)
2371 {
2372         struct page *hpage;
2373
2374         do {
2375                 hpage = alloc_hugepage(khugepaged_defrag());
2376                 if (!hpage) {
2377                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2378                         if (!*wait)
2379                                 return NULL;
2380
2381                         *wait = false;
2382                         khugepaged_alloc_sleep();
2383                 } else
2384                         count_vm_event(THP_COLLAPSE_ALLOC);
2385         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2386
2387         return hpage;
2388 }
2389
2390 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2391 {
2392         if (!*hpage)
2393                 *hpage = khugepaged_alloc_hugepage(wait);
2394
2395         if (unlikely(!*hpage))
2396                 return false;
2397
2398         return true;
2399 }
2400
2401 static struct page *
2402 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2403                        struct vm_area_struct *vma, unsigned long address,
2404                        int node)
2405 {
2406         up_read(&mm->mmap_sem);
2407         VM_BUG_ON(!*hpage);
2408
2409         return  *hpage;
2410 }
2411 #endif
2412
2413 static bool hugepage_vma_check(struct vm_area_struct *vma)
2414 {
2415         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2416             (vma->vm_flags & VM_NOHUGEPAGE))
2417                 return false;
2418
2419         if (!vma->anon_vma || vma->vm_ops)
2420                 return false;
2421         if (is_vma_temporary_stack(vma))
2422                 return false;
2423         VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma);
2424         return true;
2425 }
2426
2427 static void collapse_huge_page(struct mm_struct *mm,
2428                                    unsigned long address,
2429                                    struct page **hpage,
2430                                    struct vm_area_struct *vma,
2431                                    int node)
2432 {
2433         pmd_t *pmd, _pmd;
2434         pte_t *pte;
2435         pgtable_t pgtable;
2436         struct page *new_page;
2437         spinlock_t *pmd_ptl, *pte_ptl;
2438         int isolated;
2439         unsigned long hstart, hend;
2440         struct mem_cgroup *memcg;
2441         unsigned long mmun_start;       /* For mmu_notifiers */
2442         unsigned long mmun_end;         /* For mmu_notifiers */
2443         gfp_t gfp;
2444
2445         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2446
2447         /* Only allocate from the target node */
2448         gfp = alloc_hugepage_gfpmask(khugepaged_defrag(), __GFP_OTHER_NODE) |
2449                 __GFP_THISNODE;
2450
2451         /* release the mmap_sem read lock. */
2452         new_page = khugepaged_alloc_page(hpage, gfp, mm, vma, address, node);
2453         if (!new_page)
2454                 return;
2455
2456         if (unlikely(mem_cgroup_try_charge(new_page, mm,
2457                                            gfp, &memcg)))
2458                 return;
2459
2460         /*
2461          * Prevent all access to pagetables with the exception of
2462          * gup_fast later hanlded by the ptep_clear_flush and the VM
2463          * handled by the anon_vma lock + PG_lock.
2464          */
2465         down_write(&mm->mmap_sem);
2466         if (unlikely(khugepaged_test_exit(mm)))
2467                 goto out;
2468
2469         vma = find_vma(mm, address);
2470         if (!vma)
2471                 goto out;
2472         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2473         hend = vma->vm_end & HPAGE_PMD_MASK;
2474         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2475                 goto out;
2476         if (!hugepage_vma_check(vma))
2477                 goto out;
2478         pmd = mm_find_pmd(mm, address);
2479         if (!pmd)
2480                 goto out;
2481
2482         anon_vma_lock_write(vma->anon_vma);
2483
2484         pte = pte_offset_map(pmd, address);
2485         pte_ptl = pte_lockptr(mm, pmd);
2486
2487         mmun_start = address;
2488         mmun_end   = address + HPAGE_PMD_SIZE;
2489         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2490         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2491         /*
2492          * After this gup_fast can't run anymore. This also removes
2493          * any huge TLB entry from the CPU so we won't allow
2494          * huge and small TLB entries for the same virtual address
2495          * to avoid the risk of CPU bugs in that area.
2496          */
2497         _pmd = pmdp_collapse_flush(vma, address, pmd);
2498         spin_unlock(pmd_ptl);
2499         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2500
2501         spin_lock(pte_ptl);
2502         isolated = __collapse_huge_page_isolate(vma, address, pte);
2503         spin_unlock(pte_ptl);
2504
2505         if (unlikely(!isolated)) {
2506                 pte_unmap(pte);
2507                 spin_lock(pmd_ptl);
2508                 BUG_ON(!pmd_none(*pmd));
2509                 /*
2510                  * We can only use set_pmd_at when establishing
2511                  * hugepmds and never for establishing regular pmds that
2512                  * points to regular pagetables. Use pmd_populate for that
2513                  */
2514                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2515                 spin_unlock(pmd_ptl);
2516                 anon_vma_unlock_write(vma->anon_vma);
2517                 goto out;
2518         }
2519
2520         /*
2521          * All pages are isolated and locked so anon_vma rmap
2522          * can't run anymore.
2523          */
2524         anon_vma_unlock_write(vma->anon_vma);
2525
2526         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2527         pte_unmap(pte);
2528         __SetPageUptodate(new_page);
2529         pgtable = pmd_pgtable(_pmd);
2530
2531         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2532         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2533
2534         /*
2535          * spin_lock() below is not the equivalent of smp_wmb(), so
2536          * this is needed to avoid the copy_huge_page writes to become
2537          * visible after the set_pmd_at() write.
2538          */
2539         smp_wmb();
2540
2541         spin_lock(pmd_ptl);
2542         BUG_ON(!pmd_none(*pmd));
2543         page_add_new_anon_rmap(new_page, vma, address);
2544         mem_cgroup_commit_charge(new_page, memcg, false);
2545         lru_cache_add_active_or_unevictable(new_page, vma);
2546         pgtable_trans_huge_deposit(mm, pmd, pgtable);
2547         set_pmd_at(mm, address, pmd, _pmd);
2548         update_mmu_cache_pmd(vma, address, pmd);
2549         spin_unlock(pmd_ptl);
2550
2551         *hpage = NULL;
2552
2553         khugepaged_pages_collapsed++;
2554 out_up_write:
2555         up_write(&mm->mmap_sem);
2556         return;
2557
2558 out:
2559         mem_cgroup_cancel_charge(new_page, memcg);
2560         goto out_up_write;
2561 }
2562
2563 static int khugepaged_scan_pmd(struct mm_struct *mm,
2564                                struct vm_area_struct *vma,
2565                                unsigned long address,
2566                                struct page **hpage)
2567 {
2568         pmd_t *pmd;
2569         pte_t *pte, *_pte;
2570         int ret = 0, none_or_zero = 0;
2571         struct page *page;
2572         unsigned long _address;
2573         spinlock_t *ptl;
2574         int node = NUMA_NO_NODE;
2575         bool writable = false, referenced = false;
2576
2577         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2578
2579         pmd = mm_find_pmd(mm, address);
2580         if (!pmd)
2581                 goto out;
2582
2583         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2584         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2585         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2586              _pte++, _address += PAGE_SIZE) {
2587                 pte_t pteval = *_pte;
2588                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2589                         if (++none_or_zero <= khugepaged_max_ptes_none)
2590                                 continue;
2591                         else
2592                                 goto out_unmap;
2593                 }
2594                 if (!pte_present(pteval))
2595                         goto out_unmap;
2596                 if (pte_write(pteval))
2597                         writable = true;
2598
2599                 page = vm_normal_page(vma, _address, pteval);
2600                 if (unlikely(!page))
2601                         goto out_unmap;
2602                 /*
2603                  * Record which node the original page is from and save this
2604                  * information to khugepaged_node_load[].
2605                  * Khupaged will allocate hugepage from the node has the max
2606                  * hit record.
2607                  */
2608                 node = page_to_nid(page);
2609                 if (khugepaged_scan_abort(node))
2610                         goto out_unmap;
2611                 khugepaged_node_load[node]++;
2612                 VM_BUG_ON_PAGE(PageCompound(page), page);
2613                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2614                         goto out_unmap;
2615                 /*
2616                  * cannot use mapcount: can't collapse if there's a gup pin.
2617                  * The page must only be referenced by the scanned process
2618                  * and page swap cache.
2619                  */
2620                 if (page_count(page) != 1 + !!PageSwapCache(page))
2621                         goto out_unmap;
2622                 if (pte_young(pteval) || PageReferenced(page) ||
2623                     mmu_notifier_test_young(vma->vm_mm, address))
2624                         referenced = true;
2625         }
2626         if (referenced && writable)
2627                 ret = 1;
2628 out_unmap:
2629         pte_unmap_unlock(pte, ptl);
2630         if (ret) {
2631                 node = khugepaged_find_target_node();
2632                 /* collapse_huge_page will return with the mmap_sem released */
2633                 collapse_huge_page(mm, address, hpage, vma, node);
2634         }
2635 out:
2636         return ret;
2637 }
2638
2639 static void collect_mm_slot(struct mm_slot *mm_slot)
2640 {
2641         struct mm_struct *mm = mm_slot->mm;
2642
2643         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2644
2645         if (khugepaged_test_exit(mm)) {
2646                 /* free mm_slot */
2647                 hash_del(&mm_slot->hash);
2648                 list_del(&mm_slot->mm_node);
2649
2650                 /*
2651                  * Not strictly needed because the mm exited already.
2652                  *
2653                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2654                  */
2655
2656                 /* khugepaged_mm_lock actually not necessary for the below */
2657                 free_mm_slot(mm_slot);
2658                 mmdrop(mm);
2659         }
2660 }
2661
2662 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2663                                             struct page **hpage)
2664         __releases(&khugepaged_mm_lock)
2665         __acquires(&khugepaged_mm_lock)
2666 {
2667         struct mm_slot *mm_slot;
2668         struct mm_struct *mm;
2669         struct vm_area_struct *vma;
2670         int progress = 0;
2671
2672         VM_BUG_ON(!pages);
2673         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2674
2675         if (khugepaged_scan.mm_slot)
2676                 mm_slot = khugepaged_scan.mm_slot;
2677         else {
2678                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2679                                      struct mm_slot, mm_node);
2680                 khugepaged_scan.address = 0;
2681                 khugepaged_scan.mm_slot = mm_slot;
2682         }
2683         spin_unlock(&khugepaged_mm_lock);
2684
2685         mm = mm_slot->mm;
2686         down_read(&mm->mmap_sem);
2687         if (unlikely(khugepaged_test_exit(mm)))
2688                 vma = NULL;
2689         else
2690                 vma = find_vma(mm, khugepaged_scan.address);
2691
2692         progress++;
2693         for (; vma; vma = vma->vm_next) {
2694                 unsigned long hstart, hend;
2695
2696                 cond_resched();
2697                 if (unlikely(khugepaged_test_exit(mm))) {
2698                         progress++;
2699                         break;
2700                 }
2701                 if (!hugepage_vma_check(vma)) {
2702 skip:
2703                         progress++;
2704                         continue;
2705                 }
2706                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2707                 hend = vma->vm_end & HPAGE_PMD_MASK;
2708                 if (hstart >= hend)
2709                         goto skip;
2710                 if (khugepaged_scan.address > hend)
2711                         goto skip;
2712                 if (khugepaged_scan.address < hstart)
2713                         khugepaged_scan.address = hstart;
2714                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2715
2716                 while (khugepaged_scan.address < hend) {
2717                         int ret;
2718                         cond_resched();
2719                         if (unlikely(khugepaged_test_exit(mm)))
2720                                 goto breakouterloop;
2721
2722                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2723                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2724                                   hend);
2725                         ret = khugepaged_scan_pmd(mm, vma,
2726                                                   khugepaged_scan.address,
2727                                                   hpage);
2728                         /* move to next address */
2729                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2730                         progress += HPAGE_PMD_NR;
2731                         if (ret)
2732                                 /* we released mmap_sem so break loop */
2733                                 goto breakouterloop_mmap_sem;
2734                         if (progress >= pages)
2735                                 goto breakouterloop;
2736                 }
2737         }
2738 breakouterloop:
2739         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2740 breakouterloop_mmap_sem:
2741
2742         spin_lock(&khugepaged_mm_lock);
2743         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2744         /*
2745          * Release the current mm_slot if this mm is about to die, or
2746          * if we scanned all vmas of this mm.
2747          */
2748         if (khugepaged_test_exit(mm) || !vma) {
2749                 /*
2750                  * Make sure that if mm_users is reaching zero while
2751                  * khugepaged runs here, khugepaged_exit will find
2752                  * mm_slot not pointing to the exiting mm.
2753                  */
2754                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2755                         khugepaged_scan.mm_slot = list_entry(
2756                                 mm_slot->mm_node.next,
2757                                 struct mm_slot, mm_node);
2758                         khugepaged_scan.address = 0;
2759                 } else {
2760                         khugepaged_scan.mm_slot = NULL;
2761                         khugepaged_full_scans++;
2762                 }
2763
2764                 collect_mm_slot(mm_slot);
2765         }
2766
2767         return progress;
2768 }
2769
2770 static int khugepaged_has_work(void)
2771 {
2772         return !list_empty(&khugepaged_scan.mm_head) &&
2773                 khugepaged_enabled();
2774 }
2775
2776 static int khugepaged_wait_event(void)
2777 {
2778         return !list_empty(&khugepaged_scan.mm_head) ||
2779                 kthread_should_stop();
2780 }
2781
2782 static void khugepaged_do_scan(void)
2783 {
2784         struct page *hpage = NULL;
2785         unsigned int progress = 0, pass_through_head = 0;
2786         unsigned int pages = khugepaged_pages_to_scan;
2787         bool wait = true;
2788
2789         barrier(); /* write khugepaged_pages_to_scan to local stack */
2790
2791         while (progress < pages) {
2792                 if (!khugepaged_prealloc_page(&hpage, &wait))
2793                         break;
2794
2795                 cond_resched();
2796
2797                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2798                         break;
2799
2800                 spin_lock(&khugepaged_mm_lock);
2801                 if (!khugepaged_scan.mm_slot)
2802                         pass_through_head++;
2803                 if (khugepaged_has_work() &&
2804                     pass_through_head < 2)
2805                         progress += khugepaged_scan_mm_slot(pages - progress,
2806                                                             &hpage);
2807                 else
2808                         progress = pages;
2809                 spin_unlock(&khugepaged_mm_lock);
2810         }
2811
2812         if (!IS_ERR_OR_NULL(hpage))
2813                 put_page(hpage);
2814 }
2815
2816 static void khugepaged_wait_work(void)
2817 {
2818         if (khugepaged_has_work()) {
2819                 if (!khugepaged_scan_sleep_millisecs)
2820                         return;
2821
2822                 wait_event_freezable_timeout(khugepaged_wait,
2823                                              kthread_should_stop(),
2824                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2825                 return;
2826         }
2827
2828         if (khugepaged_enabled())
2829                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2830 }
2831
2832 static int khugepaged(void *none)
2833 {
2834         struct mm_slot *mm_slot;
2835
2836         set_freezable();
2837         set_user_nice(current, MAX_NICE);
2838
2839         while (!kthread_should_stop()) {
2840                 khugepaged_do_scan();
2841                 khugepaged_wait_work();
2842         }
2843
2844         spin_lock(&khugepaged_mm_lock);
2845         mm_slot = khugepaged_scan.mm_slot;
2846         khugepaged_scan.mm_slot = NULL;
2847         if (mm_slot)
2848                 collect_mm_slot(mm_slot);
2849         spin_unlock(&khugepaged_mm_lock);
2850         return 0;
2851 }
2852
2853 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2854                 unsigned long haddr, pmd_t *pmd)
2855 {
2856         struct mm_struct *mm = vma->vm_mm;
2857         pgtable_t pgtable;
2858         pmd_t _pmd;
2859         int i;
2860
2861         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2862         /* leave pmd empty until pte is filled */
2863
2864         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2865         pmd_populate(mm, &_pmd, pgtable);
2866
2867         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2868                 pte_t *pte, entry;
2869                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2870                 entry = pte_mkspecial(entry);
2871                 pte = pte_offset_map(&_pmd, haddr);
2872                 VM_BUG_ON(!pte_none(*pte));
2873                 set_pte_at(mm, haddr, pte, entry);
2874                 pte_unmap(pte);
2875         }
2876         smp_wmb(); /* make pte visible before pmd */
2877         pmd_populate(mm, pmd, pgtable);
2878         put_huge_zero_page();
2879 }
2880
2881 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2882                 pmd_t *pmd)
2883 {
2884         spinlock_t *ptl;
2885         struct page *page;
2886         struct mm_struct *mm = vma->vm_mm;
2887         unsigned long haddr = address & HPAGE_PMD_MASK;
2888         unsigned long mmun_start;       /* For mmu_notifiers */
2889         unsigned long mmun_end;         /* For mmu_notifiers */
2890
2891         BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
2892
2893         mmun_start = haddr;
2894         mmun_end   = haddr + HPAGE_PMD_SIZE;
2895 again:
2896         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2897         ptl = pmd_lock(mm, pmd);
2898         if (unlikely(!pmd_trans_huge(*pmd))) {
2899                 spin_unlock(ptl);
2900                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2901                 return;
2902         }
2903         if (is_huge_zero_pmd(*pmd)) {
2904                 __split_huge_zero_page_pmd(vma, haddr, pmd);
2905                 spin_unlock(ptl);
2906                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2907                 return;
2908         }
2909         page = pmd_page(*pmd);
2910         VM_BUG_ON_PAGE(!page_count(page), page);
2911         get_page(page);
2912         spin_unlock(ptl);
2913         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2914
2915         split_huge_page(page);
2916
2917         put_page(page);
2918
2919         /*
2920          * We don't always have down_write of mmap_sem here: a racing
2921          * do_huge_pmd_wp_page() might have copied-on-write to another
2922          * huge page before our split_huge_page() got the anon_vma lock.
2923          */
2924         if (unlikely(pmd_trans_huge(*pmd)))
2925                 goto again;
2926 }
2927
2928 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2929                 pmd_t *pmd)
2930 {
2931         struct vm_area_struct *vma;
2932
2933         vma = find_vma(mm, address);
2934         BUG_ON(vma == NULL);
2935         split_huge_page_pmd(vma, address, pmd);
2936 }
2937
2938 static void split_huge_page_address(struct mm_struct *mm,
2939                                     unsigned long address)
2940 {
2941         pgd_t *pgd;
2942         pud_t *pud;
2943         pmd_t *pmd;
2944
2945         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2946
2947         pgd = pgd_offset(mm, address);
2948         if (!pgd_present(*pgd))
2949                 return;
2950
2951         pud = pud_offset(pgd, address);
2952         if (!pud_present(*pud))
2953                 return;
2954
2955         pmd = pmd_offset(pud, address);
2956         if (!pmd_present(*pmd))
2957                 return;
2958         /*
2959          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2960          * materialize from under us.
2961          */
2962         split_huge_page_pmd_mm(mm, address, pmd);
2963 }
2964
2965 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2966                              unsigned long start,
2967                              unsigned long end,
2968                              long adjust_next)
2969 {
2970         /*
2971          * If the new start address isn't hpage aligned and it could
2972          * previously contain an hugepage: check if we need to split
2973          * an huge pmd.
2974          */
2975         if (start & ~HPAGE_PMD_MASK &&
2976             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2977             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2978                 split_huge_page_address(vma->vm_mm, start);
2979
2980         /*
2981          * If the new end address isn't hpage aligned and it could
2982          * previously contain an hugepage: check if we need to split
2983          * an huge pmd.
2984          */
2985         if (end & ~HPAGE_PMD_MASK &&
2986             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2987             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2988                 split_huge_page_address(vma->vm_mm, end);
2989
2990         /*
2991          * If we're also updating the vma->vm_next->vm_start, if the new
2992          * vm_next->vm_start isn't page aligned and it could previously
2993          * contain an hugepage: check if we need to split an huge pmd.
2994          */
2995         if (adjust_next > 0) {
2996                 struct vm_area_struct *next = vma->vm_next;
2997                 unsigned long nstart = next->vm_start;
2998                 nstart += adjust_next << PAGE_SHIFT;
2999                 if (nstart & ~HPAGE_PMD_MASK &&
3000                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
3001                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
3002                         split_huge_page_address(next->vm_mm, nstart);
3003         }
3004 }