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