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         src_page = pmd_page(pmd);
 990         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
 991         get_page(src_page);
 992         page_dup_rmap(src_page, true);
 993         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 994
 995         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 996         pmd = pmd_mkold(pmd_wrprotect(pmd));
 997         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
 998         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 999         atomic_long_inc(&dst_mm->nr_ptes);
1000
1001         ret = 0;
1002 out_unlock:
1003         spin_unlock(src_ptl);
1004         spin_unlock(dst_ptl);
1005 out:
1006         return ret;
1007 }
1008
1009 void huge_pmd_set_accessed(struct mm_struct *mm,
1010                            struct vm_area_struct *vma,
1011                            unsigned long address,
1012                            pmd_t *pmd, pmd_t orig_pmd,
1013                            int dirty)
1014 {
1015         spinlock_t *ptl;
1016         pmd_t entry;
1017         unsigned long haddr;
1018
1019         ptl = pmd_lock(mm, pmd);
1020         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1021                 goto unlock;
1022
1023         entry = pmd_mkyoung(orig_pmd);
1024         haddr = address & HPAGE_PMD_MASK;
1025         if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
1026                 update_mmu_cache_pmd(vma, address, pmd);
1027
1028 unlock:
1029         spin_unlock(ptl);
1030 }
1031
1032 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1033                                         struct vm_area_struct *vma,
1034                                         unsigned long address,
1035                                         pmd_t *pmd, pmd_t orig_pmd,
1036                                         struct page *page,
1037                                         unsigned long haddr)
1038 {
1039         struct mem_cgroup *memcg;
1040         spinlock_t *ptl;
1041         pgtable_t pgtable;
1042         pmd_t _pmd;
1043         int ret = 0, i;
1044         struct page **pages;
1045         unsigned long mmun_start;       /* For mmu_notifiers */
1046         unsigned long mmun_end;         /* For mmu_notifiers */
1047
1048         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1049                         GFP_KERNEL);
1050         if (unlikely(!pages)) {
1051                 ret |= VM_FAULT_OOM;
1052                 goto out;
1053         }
1054
1055         for (i = 0; i < HPAGE_PMD_NR; i++) {
1056                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1057                                                __GFP_OTHER_NODE,
1058                                                vma, address, page_to_nid(page));
1059                 if (unlikely(!pages[i] ||
1060                              mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL,
1061                                                    &memcg, false))) {
1062                         if (pages[i])
1063                                 put_page(pages[i]);
1064                         while (--i >= 0) {
1065                                 memcg = (void *)page_private(pages[i]);
1066                                 set_page_private(pages[i], 0);
1067                                 mem_cgroup_cancel_charge(pages[i], memcg,
1068                                                 false);
1069                                 put_page(pages[i]);
1070                         }
1071                         kfree(pages);
1072                         ret |= VM_FAULT_OOM;
1073                         goto out;
1074                 }
1075                 set_page_private(pages[i], (unsigned long)memcg);
1076         }
1077
1078         for (i = 0; i < HPAGE_PMD_NR; i++) {
1079                 copy_user_highpage(pages[i], page + i,
1080                                    haddr + PAGE_SIZE * i, vma);
1081                 __SetPageUptodate(pages[i]);
1082                 cond_resched();
1083         }
1084
1085         mmun_start = haddr;
1086         mmun_end   = haddr + HPAGE_PMD_SIZE;
1087         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1088
1089         ptl = pmd_lock(mm, pmd);
1090         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1091                 goto out_free_pages;
1092         VM_BUG_ON_PAGE(!PageHead(page), page);
1093
1094         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1095         /* leave pmd empty until pte is filled */
1096
1097         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1098         pmd_populate(mm, &_pmd, pgtable);
1099
1100         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1101                 pte_t *pte, entry;
1102                 entry = mk_pte(pages[i], vma->vm_page_prot);
1103                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1104                 memcg = (void *)page_private(pages[i]);
1105                 set_page_private(pages[i], 0);
1106                 page_add_new_anon_rmap(pages[i], vma, haddr, false);
1107                 mem_cgroup_commit_charge(pages[i], memcg, false, false);
1108                 lru_cache_add_active_or_unevictable(pages[i], vma);
1109                 pte = pte_offset_map(&_pmd, haddr);
1110                 VM_BUG_ON(!pte_none(*pte));
1111                 set_pte_at(mm, haddr, pte, entry);
1112                 pte_unmap(pte);
1113         }
1114         kfree(pages);
1115
1116         smp_wmb(); /* make pte visible before pmd */
1117         pmd_populate(mm, pmd, pgtable);
1118         page_remove_rmap(page, true);
1119         spin_unlock(ptl);
1120
1121         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1122
1123         ret |= VM_FAULT_WRITE;
1124         put_page(page);
1125
1126 out:
1127         return ret;
1128
1129 out_free_pages:
1130         spin_unlock(ptl);
1131         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1132         for (i = 0; i < HPAGE_PMD_NR; i++) {
1133                 memcg = (void *)page_private(pages[i]);
1134                 set_page_private(pages[i], 0);
1135                 mem_cgroup_cancel_charge(pages[i], memcg, false);
1136                 put_page(pages[i]);
1137         }
1138         kfree(pages);
1139         goto out;
1140 }
1141
1142 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1143                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1144 {
1145         spinlock_t *ptl;
1146         int ret = 0;
1147         struct page *page = NULL, *new_page;
1148         struct mem_cgroup *memcg;
1149         unsigned long haddr;
1150         unsigned long mmun_start;       /* For mmu_notifiers */
1151         unsigned long mmun_end;         /* For mmu_notifiers */
1152         gfp_t huge_gfp;                 /* for allocation and charge */
1153
1154         ptl = pmd_lockptr(mm, pmd);
1155         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1156         haddr = address & HPAGE_PMD_MASK;
1157         if (is_huge_zero_pmd(orig_pmd))
1158                 goto alloc;
1159         spin_lock(ptl);
1160         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1161                 goto out_unlock;
1162
1163         page = pmd_page(orig_pmd);
1164         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1165         /*
1166          * We can only reuse the page if nobody else maps the huge page or it's
1167          * part. We can do it by checking page_mapcount() on each sub-page, but
1168          * it's expensive.
1169          * The cheaper way is to check page_count() to be equal 1: every
1170          * mapcount takes page reference reference, so this way we can
1171          * guarantee, that the PMD is the only mapping.
1172          * This can give false negative if somebody pinned the page, but that's
1173          * fine.
1174          */
1175         if (page_mapcount(page) == 1 && page_count(page) == 1) {
1176                 pmd_t entry;
1177                 entry = pmd_mkyoung(orig_pmd);
1178                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1179                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1180                         update_mmu_cache_pmd(vma, address, pmd);
1181                 ret |= VM_FAULT_WRITE;
1182                 goto out_unlock;
1183         }
1184         get_page(page);
1185         spin_unlock(ptl);
1186 alloc:
1187         if (transparent_hugepage_enabled(vma) &&
1188             !transparent_hugepage_debug_cow()) {
1189                 huge_gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
1190                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1191         } else
1192                 new_page = NULL;
1193
1194         if (unlikely(!new_page)) {
1195                 if (!page) {
1196                         split_huge_pmd(vma, pmd, address);
1197                         ret |= VM_FAULT_FALLBACK;
1198                 } else {
1199                         ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1200                                         pmd, orig_pmd, page, haddr);
1201                         if (ret & VM_FAULT_OOM) {
1202                                 split_huge_pmd(vma, pmd, address);
1203                                 ret |= VM_FAULT_FALLBACK;
1204                         }
1205                         put_page(page);
1206                 }
1207                 count_vm_event(THP_FAULT_FALLBACK);
1208                 goto out;
1209         }
1210
1211         if (unlikely(mem_cgroup_try_charge(new_page, mm, huge_gfp, &memcg,
1212                                            true))) {
1213                 put_page(new_page);
1214                 if (page) {
1215                         split_huge_pmd(vma, pmd, address);
1216                         put_page(page);
1217                 } else
1218                         split_huge_pmd(vma, pmd, address);
1219                 ret |= VM_FAULT_FALLBACK;
1220                 count_vm_event(THP_FAULT_FALLBACK);
1221                 goto out;
1222         }
1223
1224         count_vm_event(THP_FAULT_ALLOC);
1225
1226         if (!page)
1227                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1228         else
1229                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1230         __SetPageUptodate(new_page);
1231
1232         mmun_start = haddr;
1233         mmun_end   = haddr + HPAGE_PMD_SIZE;
1234         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1235
1236         spin_lock(ptl);
1237         if (page)
1238                 put_page(page);
1239         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1240                 spin_unlock(ptl);
1241                 mem_cgroup_cancel_charge(new_page, memcg, true);
1242                 put_page(new_page);
1243                 goto out_mn;
1244         } else {
1245                 pmd_t entry;
1246                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1247                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1248                 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1249                 page_add_new_anon_rmap(new_page, vma, haddr, true);
1250                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1251                 lru_cache_add_active_or_unevictable(new_page, vma);
1252                 set_pmd_at(mm, haddr, pmd, entry);
1253                 update_mmu_cache_pmd(vma, address, pmd);
1254                 if (!page) {
1255                         add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1256                         put_huge_zero_page();
1257                 } else {
1258                         VM_BUG_ON_PAGE(!PageHead(page), page);
1259                         page_remove_rmap(page, true);
1260                         put_page(page);
1261                 }
1262                 ret |= VM_FAULT_WRITE;
1263         }
1264         spin_unlock(ptl);
1265 out_mn:
1266         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1267 out:
1268         return ret;
1269 out_unlock:
1270         spin_unlock(ptl);
1271         return ret;
1272 }
1273
1274 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1275                                    unsigned long addr,
1276                                    pmd_t *pmd,
1277                                    unsigned int flags)
1278 {
1279         struct mm_struct *mm = vma->vm_mm;
1280         struct page *page = NULL;
1281
1282         assert_spin_locked(pmd_lockptr(mm, pmd));
1283
1284         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1285                 goto out;
1286
1287         /* Avoid dumping huge zero page */
1288         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1289                 return ERR_PTR(-EFAULT);
1290
1291         /* Full NUMA hinting faults to serialise migration in fault paths */
1292         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1293                 goto out;
1294
1295         page = pmd_page(*pmd);
1296         VM_BUG_ON_PAGE(!PageHead(page), page);
1297         if (flags & FOLL_TOUCH) {
1298                 pmd_t _pmd;
1299                 /*
1300                  * We should set the dirty bit only for FOLL_WRITE but
1301                  * for now the dirty bit in the pmd is meaningless.
1302                  * And if the dirty bit will become meaningful and
1303                  * we'll only set it with FOLL_WRITE, an atomic
1304                  * set_bit will be required on the pmd to set the
1305                  * young bit, instead of the current set_pmd_at.
1306                  */
1307                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1308                 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1309                                           pmd, _pmd,  1))
1310                         update_mmu_cache_pmd(vma, addr, pmd);
1311         }
1312         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1313                 if (page->mapping && trylock_page(page)) {
1314                         lru_add_drain();
1315                         if (page->mapping)
1316                                 mlock_vma_page(page);
1317                         unlock_page(page);
1318                 }
1319         }
1320         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1321         VM_BUG_ON_PAGE(!PageCompound(page), page);
1322         if (flags & FOLL_GET)
1323                 get_page(page);
1324
1325 out:
1326         return page;
1327 }
1328
1329 /* NUMA hinting page fault entry point for trans huge pmds */
1330 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1331                                 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1332 {
1333         spinlock_t *ptl;
1334         struct anon_vma *anon_vma = NULL;
1335         struct page *page;
1336         unsigned long haddr = addr & HPAGE_PMD_MASK;
1337         int page_nid = -1, this_nid = numa_node_id();
1338         int target_nid, last_cpupid = -1;
1339         bool page_locked;
1340         bool migrated = false;
1341         bool was_writable;
1342         int flags = 0;
1343
1344         /* A PROT_NONE fault should not end up here */
1345         BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
1346
1347         ptl = pmd_lock(mm, pmdp);
1348         if (unlikely(!pmd_same(pmd, *pmdp)))
1349                 goto out_unlock;
1350
1351         /*
1352          * If there are potential migrations, wait for completion and retry
1353          * without disrupting NUMA hinting information. Do not relock and
1354          * check_same as the page may no longer be mapped.
1355          */
1356         if (unlikely(pmd_trans_migrating(*pmdp))) {
1357                 page = pmd_page(*pmdp);
1358                 spin_unlock(ptl);
1359                 wait_on_page_locked(page);
1360                 goto out;
1361         }
1362
1363         page = pmd_page(pmd);
1364         BUG_ON(is_huge_zero_page(page));
1365         page_nid = page_to_nid(page);
1366         last_cpupid = page_cpupid_last(page);
1367         count_vm_numa_event(NUMA_HINT_FAULTS);
1368         if (page_nid == this_nid) {
1369                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1370                 flags |= TNF_FAULT_LOCAL;
1371         }
1372
1373         /* See similar comment in do_numa_page for explanation */
1374         if (!(vma->vm_flags & VM_WRITE))
1375                 flags |= TNF_NO_GROUP;
1376
1377         /*
1378          * Acquire the page lock to serialise THP migrations but avoid dropping
1379          * page_table_lock if at all possible
1380          */
1381         page_locked = trylock_page(page);
1382         target_nid = mpol_misplaced(page, vma, haddr);
1383         if (target_nid == -1) {
1384                 /* If the page was locked, there are no parallel migrations */
1385                 if (page_locked)
1386                         goto clear_pmdnuma;
1387         }
1388
1389         /* Migration could have started since the pmd_trans_migrating check */
1390         if (!page_locked) {
1391                 spin_unlock(ptl);
1392                 wait_on_page_locked(page);
1393                 page_nid = -1;
1394                 goto out;
1395         }
1396
1397         /*
1398          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1399          * to serialises splits
1400          */
1401         get_page(page);
1402         spin_unlock(ptl);
1403         anon_vma = page_lock_anon_vma_read(page);
1404
1405         /* Confirm the PMD did not change while page_table_lock was released */
1406         spin_lock(ptl);
1407         if (unlikely(!pmd_same(pmd, *pmdp))) {
1408                 unlock_page(page);
1409                 put_page(page);
1410                 page_nid = -1;
1411                 goto out_unlock;
1412         }
1413
1414         /* Bail if we fail to protect against THP splits for any reason */
1415         if (unlikely(!anon_vma)) {
1416                 put_page(page);
1417                 page_nid = -1;
1418                 goto clear_pmdnuma;
1419         }
1420
1421         /*
1422          * Migrate the THP to the requested node, returns with page unlocked
1423          * and access rights restored.
1424          */
1425         spin_unlock(ptl);
1426         migrated = migrate_misplaced_transhuge_page(mm, vma,
1427                                 pmdp, pmd, addr, page, target_nid);
1428         if (migrated) {
1429                 flags |= TNF_MIGRATED;
1430                 page_nid = target_nid;
1431         } else
1432                 flags |= TNF_MIGRATE_FAIL;
1433
1434         goto out;
1435 clear_pmdnuma:
1436         BUG_ON(!PageLocked(page));
1437         was_writable = pmd_write(pmd);
1438         pmd = pmd_modify(pmd, vma->vm_page_prot);
1439         pmd = pmd_mkyoung(pmd);
1440         if (was_writable)
1441                 pmd = pmd_mkwrite(pmd);
1442         set_pmd_at(mm, haddr, pmdp, pmd);
1443         update_mmu_cache_pmd(vma, addr, pmdp);
1444         unlock_page(page);
1445 out_unlock:
1446         spin_unlock(ptl);
1447
1448 out:
1449         if (anon_vma)
1450                 page_unlock_anon_vma_read(anon_vma);
1451
1452         if (page_nid != -1)
1453                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
1454
1455         return 0;
1456 }
1457
1458 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1459                  pmd_t *pmd, unsigned long addr)
1460 {
1461         pmd_t orig_pmd;
1462         spinlock_t *ptl;
1463
1464         if (!__pmd_trans_huge_lock(pmd, vma, &ptl))
1465                 return 0;
1466         /*
1467          * For architectures like ppc64 we look at deposited pgtable
1468          * when calling pmdp_huge_get_and_clear. So do the
1469          * pgtable_trans_huge_withdraw after finishing pmdp related
1470          * operations.
1471          */
1472         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1473                         tlb->fullmm);
1474         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1475         if (vma_is_dax(vma)) {
1476                 spin_unlock(ptl);
1477                 if (is_huge_zero_pmd(orig_pmd))
1478                         put_huge_zero_page();
1479         } else if (is_huge_zero_pmd(orig_pmd)) {
1480                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1481                 atomic_long_dec(&tlb->mm->nr_ptes);
1482                 spin_unlock(ptl);
1483                 put_huge_zero_page();
1484         } else {
1485                 struct page *page = pmd_page(orig_pmd);
1486                 page_remove_rmap(page, true);
1487                 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1488                 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1489                 VM_BUG_ON_PAGE(!PageHead(page), page);
1490                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1491                 atomic_long_dec(&tlb->mm->nr_ptes);
1492                 spin_unlock(ptl);
1493                 tlb_remove_page(tlb, page);
1494         }
1495         return 1;
1496 }
1497
1498 bool move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1499                   unsigned long old_addr,
1500                   unsigned long new_addr, unsigned long old_end,
1501                   pmd_t *old_pmd, pmd_t *new_pmd)
1502 {
1503         spinlock_t *old_ptl, *new_ptl;
1504         pmd_t pmd;
1505
1506         struct mm_struct *mm = vma->vm_mm;
1507
1508         if ((old_addr & ~HPAGE_PMD_MASK) ||
1509             (new_addr & ~HPAGE_PMD_MASK) ||
1510             old_end - old_addr < HPAGE_PMD_SIZE ||
1511             (new_vma->vm_flags & VM_NOHUGEPAGE))
1512                 return false;
1513
1514         /*
1515          * The destination pmd shouldn't be established, free_pgtables()
1516          * should have release it.
1517          */
1518         if (WARN_ON(!pmd_none(*new_pmd))) {
1519                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1520                 return false;
1521         }
1522
1523         /*
1524          * We don't have to worry about the ordering of src and dst
1525          * ptlocks because exclusive mmap_sem prevents deadlock.
1526          */
1527         if (__pmd_trans_huge_lock(old_pmd, vma, &old_ptl)) {
1528                 new_ptl = pmd_lockptr(mm, new_pmd);
1529                 if (new_ptl != old_ptl)
1530                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1531                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1532                 VM_BUG_ON(!pmd_none(*new_pmd));
1533
1534                 if (pmd_move_must_withdraw(new_ptl, old_ptl)) {
1535                         pgtable_t pgtable;
1536                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1537                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1538                 }
1539                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1540                 if (new_ptl != old_ptl)
1541                         spin_unlock(new_ptl);
1542                 spin_unlock(old_ptl);
1543                 return true;
1544         }
1545         return false;
1546 }
1547
1548 /*
1549  * Returns
1550  *  - 0 if PMD could not be locked
1551  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1552  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1553  */
1554 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1555                 unsigned long addr, pgprot_t newprot, int prot_numa)
1556 {
1557         struct mm_struct *mm = vma->vm_mm;
1558         spinlock_t *ptl;
1559         int ret = 0;
1560
1561         if (__pmd_trans_huge_lock(pmd, vma, &ptl)) {
1562                 pmd_t entry;
1563                 bool preserve_write = prot_numa && pmd_write(*pmd);
1564                 ret = 1;
1565
1566                 /*
1567                  * Avoid trapping faults against the zero page. The read-only
1568                  * data is likely to be read-cached on the local CPU and
1569                  * local/remote hits to the zero page are not interesting.
1570                  */
1571                 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1572                         spin_unlock(ptl);
1573                         return ret;
1574                 }
1575
1576                 if (!prot_numa || !pmd_protnone(*pmd)) {
1577                         entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
1578                         entry = pmd_modify(entry, newprot);
1579                         if (preserve_write)
1580                                 entry = pmd_mkwrite(entry);
1581                         ret = HPAGE_PMD_NR;
1582                         set_pmd_at(mm, addr, pmd, entry);
1583                         BUG_ON(!preserve_write && pmd_write(entry));
1584                 }
1585                 spin_unlock(ptl);
1586         }
1587
1588         return ret;
1589 }
1590
1591 /*
1592  * Returns true if a given pmd maps a thp, false otherwise.
1593  *
1594  * Note that if it returns true, this routine returns without unlocking page
1595  * table lock. So callers must unlock it.
1596  */
1597 bool __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1598                 spinlock_t **ptl)
1599 {
1600         *ptl = pmd_lock(vma->vm_mm, pmd);
1601         if (likely(pmd_trans_huge(*pmd)))
1602                 return true;
1603         spin_unlock(*ptl);
1604         return false;
1605 }
1606
1607 /*
1608  * This function returns whether a given @page is mapped onto the @address
1609  * in the virtual space of @mm.
1610  *
1611  * When it's true, this function returns *pmd with holding the page table lock
1612  * and passing it back to the caller via @ptl.
1613  * If it's false, returns NULL without holding the page table lock.
1614  */
1615 pmd_t *page_check_address_pmd(struct page *page,
1616                               struct mm_struct *mm,
1617                               unsigned long address,
1618                               spinlock_t **ptl)
1619 {
1620         pgd_t *pgd;
1621         pud_t *pud;
1622         pmd_t *pmd;
1623
1624         if (address & ~HPAGE_PMD_MASK)
1625                 return NULL;
1626
1627         pgd = pgd_offset(mm, address);
1628         if (!pgd_present(*pgd))
1629                 return NULL;
1630         pud = pud_offset(pgd, address);
1631         if (!pud_present(*pud))
1632                 return NULL;
1633         pmd = pmd_offset(pud, address);
1634
1635         *ptl = pmd_lock(mm, pmd);
1636         if (!pmd_present(*pmd))
1637                 goto unlock;
1638         if (pmd_page(*pmd) != page)
1639                 goto unlock;
1640         if (pmd_trans_huge(*pmd))
1641                 return pmd;
1642 unlock:
1643         spin_unlock(*ptl);
1644         return NULL;
1645 }
1646
1647 #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
1648
1649 int hugepage_madvise(struct vm_area_struct *vma,
1650                      unsigned long *vm_flags, int advice)
1651 {
1652         switch (advice) {
1653         case MADV_HUGEPAGE:
1654 #ifdef CONFIG_S390
1655                 /*
1656                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1657                  * can't handle this properly after s390_enable_sie, so we simply
1658                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
1659                  */
1660                 if (mm_has_pgste(vma->vm_mm))
1661                         return 0;
1662 #endif
1663                 /*
1664                  * Be somewhat over-protective like KSM for now!
1665                  */
1666                 if (*vm_flags & VM_NO_THP)
1667                         return -EINVAL;
1668                 *vm_flags &= ~VM_NOHUGEPAGE;
1669                 *vm_flags |= VM_HUGEPAGE;
1670                 /*
1671                  * If the vma become good for khugepaged to scan,
1672                  * register it here without waiting a page fault that
1673                  * may not happen any time soon.
1674                  */
1675                 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
1676                         return -ENOMEM;
1677                 break;
1678         case MADV_NOHUGEPAGE:
1679                 /*
1680                  * Be somewhat over-protective like KSM for now!
1681                  */
1682                 if (*vm_flags & VM_NO_THP)
1683                         return -EINVAL;
1684                 *vm_flags &= ~VM_HUGEPAGE;
1685                 *vm_flags |= VM_NOHUGEPAGE;
1686                 /*
1687                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1688                  * this vma even if we leave the mm registered in khugepaged if
1689                  * it got registered before VM_NOHUGEPAGE was set.
1690                  */
1691                 break;
1692         }
1693
1694         return 0;
1695 }
1696
1697 static int __init khugepaged_slab_init(void)
1698 {
1699         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1700                                           sizeof(struct mm_slot),
1701                                           __alignof__(struct mm_slot), 0, NULL);
1702         if (!mm_slot_cache)
1703                 return -ENOMEM;
1704
1705         return 0;
1706 }
1707
1708 static void __init khugepaged_slab_exit(void)
1709 {
1710         kmem_cache_destroy(mm_slot_cache);
1711 }
1712
1713 static inline struct mm_slot *alloc_mm_slot(void)
1714 {
1715         if (!mm_slot_cache)     /* initialization failed */
1716                 return NULL;
1717         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1718 }
1719
1720 static inline void free_mm_slot(struct mm_slot *mm_slot)
1721 {
1722         kmem_cache_free(mm_slot_cache, mm_slot);
1723 }
1724
1725 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1726 {
1727         struct mm_slot *mm_slot;
1728
1729         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
1730                 if (mm == mm_slot->mm)
1731                         return mm_slot;
1732
1733         return NULL;
1734 }
1735
1736 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1737                                     struct mm_slot *mm_slot)
1738 {
1739         mm_slot->mm = mm;
1740         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
1741 }
1742
1743 static inline int khugepaged_test_exit(struct mm_struct *mm)
1744 {
1745         return atomic_read(&mm->mm_users) == 0;
1746 }
1747
1748 int __khugepaged_enter(struct mm_struct *mm)
1749 {
1750         struct mm_slot *mm_slot;
1751         int wakeup;
1752
1753         mm_slot = alloc_mm_slot();
1754         if (!mm_slot)
1755                 return -ENOMEM;
1756
1757         /* __khugepaged_exit() must not run from under us */
1758         VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
1759         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1760                 free_mm_slot(mm_slot);
1761                 return 0;
1762         }
1763
1764         spin_lock(&khugepaged_mm_lock);
1765         insert_to_mm_slots_hash(mm, mm_slot);
1766         /*
1767          * Insert just behind the scanning cursor, to let the area settle
1768          * down a little.
1769          */
1770         wakeup = list_empty(&khugepaged_scan.mm_head);
1771         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1772         spin_unlock(&khugepaged_mm_lock);
1773
1774         atomic_inc(&mm->mm_count);
1775         if (wakeup)
1776                 wake_up_interruptible(&khugepaged_wait);
1777
1778         return 0;
1779 }
1780
1781 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
1782                                unsigned long vm_flags)
1783 {
1784         unsigned long hstart, hend;
1785         if (!vma->anon_vma)
1786                 /*
1787                  * Not yet faulted in so we will register later in the
1788                  * page fault if needed.
1789                  */
1790                 return 0;
1791         if (vma->vm_ops)
1792                 /* khugepaged not yet working on file or special mappings */
1793                 return 0;
1794         VM_BUG_ON_VMA(vm_flags & VM_NO_THP, vma);
1795         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1796         hend = vma->vm_end & HPAGE_PMD_MASK;
1797         if (hstart < hend)
1798                 return khugepaged_enter(vma, vm_flags);
1799         return 0;
1800 }
1801
1802 void __khugepaged_exit(struct mm_struct *mm)
1803 {
1804         struct mm_slot *mm_slot;
1805         int free = 0;
1806
1807         spin_lock(&khugepaged_mm_lock);
1808         mm_slot = get_mm_slot(mm);
1809         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1810                 hash_del(&mm_slot->hash);
1811                 list_del(&mm_slot->mm_node);
1812                 free = 1;
1813         }
1814         spin_unlock(&khugepaged_mm_lock);
1815
1816         if (free) {
1817                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1818                 free_mm_slot(mm_slot);
1819                 mmdrop(mm);
1820         } else if (mm_slot) {
1821                 /*
1822                  * This is required to serialize against
1823                  * khugepaged_test_exit() (which is guaranteed to run
1824                  * under mmap sem read mode). Stop here (after we
1825                  * return all pagetables will be destroyed) until
1826                  * khugepaged has finished working on the pagetables
1827                  * under the mmap_sem.
1828                  */
1829                 down_write(&mm->mmap_sem);
1830                 up_write(&mm->mmap_sem);
1831         }
1832 }
1833
1834 static void release_pte_page(struct page *page)
1835 {
1836         /* 0 stands for page_is_file_cache(page) == false */
1837         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1838         unlock_page(page);
1839         putback_lru_page(page);
1840 }
1841
1842 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1843 {
1844         while (--_pte >= pte) {
1845                 pte_t pteval = *_pte;
1846                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
1847                         release_pte_page(pte_page(pteval));
1848         }
1849 }
1850
1851 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1852                                         unsigned long address,
1853                                         pte_t *pte)
1854 {
1855         struct page *page = NULL;
1856         pte_t *_pte;
1857         int none_or_zero = 0, result = 0;
1858         bool referenced = false, writable = false;
1859
1860         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1861              _pte++, address += PAGE_SIZE) {
1862                 pte_t pteval = *_pte;
1863                 if (pte_none(pteval) || (pte_present(pteval) &&
1864                                 is_zero_pfn(pte_pfn(pteval)))) {
1865                         if (!userfaultfd_armed(vma) &&
1866                             ++none_or_zero <= khugepaged_max_ptes_none) {
1867                                 continue;
1868                         } else {
1869                                 result = SCAN_EXCEED_NONE_PTE;
1870                                 goto out;
1871                         }
1872                 }
1873                 if (!pte_present(pteval)) {
1874                         result = SCAN_PTE_NON_PRESENT;
1875                         goto out;
1876                 }
1877                 page = vm_normal_page(vma, address, pteval);
1878                 if (unlikely(!page)) {
1879                         result = SCAN_PAGE_NULL;
1880                         goto out;
1881                 }
1882
1883                 VM_BUG_ON_PAGE(PageCompound(page), page);
1884                 VM_BUG_ON_PAGE(!PageAnon(page), page);
1885                 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
1886
1887                 /*
1888                  * We can do it before isolate_lru_page because the
1889                  * page can't be freed from under us. NOTE: PG_lock
1890                  * is needed to serialize against split_huge_page
1891                  * when invoked from the VM.
1892                  */
1893                 if (!trylock_page(page)) {
1894                         result = SCAN_PAGE_LOCK;
1895                         goto out;
1896                 }
1897
1898                 /*
1899                  * cannot use mapcount: can't collapse if there's a gup pin.
1900                  * The page must only be referenced by the scanned process
1901                  * and page swap cache.
1902                  */
1903                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
1904                         unlock_page(page);
1905                         result = SCAN_PAGE_COUNT;
1906                         goto out;
1907                 }
1908                 if (pte_write(pteval)) {
1909                         writable = true;
1910                 } else {
1911                         if (PageSwapCache(page) && !reuse_swap_page(page)) {
1912                                 unlock_page(page);
1913                                 result = SCAN_SWAP_CACHE_PAGE;
1914                                 goto out;
1915                         }
1916                         /*
1917                          * Page is not in the swap cache. It can be collapsed
1918                          * into a THP.
1919                          */
1920                 }
1921
1922                 /*
1923                  * Isolate the page to avoid collapsing an hugepage
1924                  * currently in use by the VM.
1925                  */
1926                 if (isolate_lru_page(page)) {
1927                         unlock_page(page);
1928                         result = SCAN_DEL_PAGE_LRU;
1929                         goto out;
1930                 }
1931                 /* 0 stands for page_is_file_cache(page) == false */
1932                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
1933                 VM_BUG_ON_PAGE(!PageLocked(page), page);
1934                 VM_BUG_ON_PAGE(PageLRU(page), page);
1935
1936                 /* If there is no mapped pte young don't collapse the page */
1937                 if (pte_young(pteval) ||
1938                     page_is_young(page) || PageReferenced(page) ||
1939                     mmu_notifier_test_young(vma->vm_mm, address))
1940                         referenced = true;
1941         }
1942         if (likely(writable)) {
1943                 if (likely(referenced)) {
1944                         result = SCAN_SUCCEED;
1945                         trace_mm_collapse_huge_page_isolate(page_to_pfn(page), none_or_zero,
1946                                                             referenced, writable, result);
1947                         return 1;
1948                 }
1949         } else {
1950                 result = SCAN_PAGE_RO;
1951         }
1952
1953 out:
1954         release_pte_pages(pte, _pte);
1955         trace_mm_collapse_huge_page_isolate(page_to_pfn(page), none_or_zero,
1956                                             referenced, writable, result);
1957         return 0;
1958 }
1959
1960 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
1961                                       struct vm_area_struct *vma,
1962                                       unsigned long address,
1963                                       spinlock_t *ptl)
1964 {
1965         pte_t *_pte;
1966         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
1967                 pte_t pteval = *_pte;
1968                 struct page *src_page;
1969
1970                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1971                         clear_user_highpage(page, address);
1972                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
1973                         if (is_zero_pfn(pte_pfn(pteval))) {
1974                                 /*
1975                                  * ptl mostly unnecessary.
1976                                  */
1977                                 spin_lock(ptl);
1978                                 /*
1979                                  * paravirt calls inside pte_clear here are
1980                                  * superfluous.
1981                                  */
1982                                 pte_clear(vma->vm_mm, address, _pte);
1983                                 spin_unlock(ptl);
1984                         }
1985                 } else {
1986                         src_page = pte_page(pteval);
1987                         copy_user_highpage(page, src_page, address, vma);
1988                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
1989                         release_pte_page(src_page);
1990                         /*
1991                          * ptl mostly unnecessary, but preempt has to
1992                          * be disabled to update the per-cpu stats
1993                          * inside page_remove_rmap().
1994                          */
1995                         spin_lock(ptl);
1996                         /*
1997                          * paravirt calls inside pte_clear here are
1998                          * superfluous.
1999                          */
2000                         pte_clear(vma->vm_mm, address, _pte);
2001                         page_remove_rmap(src_page, false);
2002                         spin_unlock(ptl);
2003                         free_page_and_swap_cache(src_page);
2004                 }
2005
2006                 address += PAGE_SIZE;
2007                 page++;
2008         }
2009 }
2010
2011 static void khugepaged_alloc_sleep(void)
2012 {
2013         DEFINE_WAIT(wait);
2014
2015         add_wait_queue(&khugepaged_wait, &wait);
2016         freezable_schedule_timeout_interruptible(
2017                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2018         remove_wait_queue(&khugepaged_wait, &wait);
2019 }
2020
2021 static int khugepaged_node_load[MAX_NUMNODES];
2022
2023 static bool khugepaged_scan_abort(int nid)
2024 {
2025         int i;
2026
2027         /*
2028          * If zone_reclaim_mode is disabled, then no extra effort is made to
2029          * allocate memory locally.
2030          */
2031         if (!zone_reclaim_mode)
2032                 return false;
2033
2034         /* If there is a count for this node already, it must be acceptable */
2035         if (khugepaged_node_load[nid])
2036                 return false;
2037
2038         for (i = 0; i < MAX_NUMNODES; i++) {
2039                 if (!khugepaged_node_load[i])
2040                         continue;
2041                 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2042                         return true;
2043         }
2044         return false;
2045 }
2046
2047 #ifdef CONFIG_NUMA
2048 static int khugepaged_find_target_node(void)
2049 {
2050         static int last_khugepaged_target_node = NUMA_NO_NODE;
2051         int nid, target_node = 0, max_value = 0;
2052
2053         /* find first node with max normal pages hit */
2054         for (nid = 0; nid < MAX_NUMNODES; nid++)
2055                 if (khugepaged_node_load[nid] > max_value) {
2056                         max_value = khugepaged_node_load[nid];
2057                         target_node = nid;
2058                 }
2059
2060         /* do some balance if several nodes have the same hit record */
2061         if (target_node <= last_khugepaged_target_node)
2062                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2063                                 nid++)
2064                         if (max_value == khugepaged_node_load[nid]) {
2065                                 target_node = nid;
2066                                 break;
2067                         }
2068
2069         last_khugepaged_target_node = target_node;
2070         return target_node;
2071 }
2072
2073 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2074 {
2075         if (IS_ERR(*hpage)) {
2076                 if (!*wait)
2077                         return false;
2078
2079                 *wait = false;
2080                 *hpage = NULL;
2081                 khugepaged_alloc_sleep();
2082         } else if (*hpage) {
2083                 put_page(*hpage);
2084                 *hpage = NULL;
2085         }
2086
2087         return true;
2088 }
2089
2090 static struct page *
2091 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2092                        unsigned long address, int node)
2093 {
2094         VM_BUG_ON_PAGE(*hpage, *hpage);
2095
2096         /*
2097          * Before allocating the hugepage, release the mmap_sem read lock.
2098          * The allocation can take potentially a long time if it involves
2099          * sync compaction, and we do not need to hold the mmap_sem during
2100          * that. We will recheck the vma after taking it again in write mode.
2101          */
2102         up_read(&mm->mmap_sem);
2103
2104         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
2105         if (unlikely(!*hpage)) {
2106                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2107                 *hpage = ERR_PTR(-ENOMEM);
2108                 return NULL;
2109         }
2110
2111         count_vm_event(THP_COLLAPSE_ALLOC);
2112         return *hpage;
2113 }
2114 #else
2115 static int khugepaged_find_target_node(void)
2116 {
2117         return 0;
2118 }
2119
2120 static inline struct page *alloc_hugepage(int defrag)
2121 {
2122         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2123                            HPAGE_PMD_ORDER);
2124 }
2125
2126 static struct page *khugepaged_alloc_hugepage(bool *wait)
2127 {
2128         struct page *hpage;
2129
2130         do {
2131                 hpage = alloc_hugepage(khugepaged_defrag());
2132                 if (!hpage) {
2133                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2134                         if (!*wait)
2135                                 return NULL;
2136
2137                         *wait = false;
2138                         khugepaged_alloc_sleep();
2139                 } else
2140                         count_vm_event(THP_COLLAPSE_ALLOC);
2141         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2142
2143         return hpage;
2144 }
2145
2146 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2147 {
2148         if (!*hpage)
2149                 *hpage = khugepaged_alloc_hugepage(wait);
2150
2151         if (unlikely(!*hpage))
2152                 return false;
2153
2154         return true;
2155 }
2156
2157 static struct page *
2158 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2159                        unsigned long address, int node)
2160 {
2161         up_read(&mm->mmap_sem);
2162         VM_BUG_ON(!*hpage);
2163
2164         return  *hpage;
2165 }
2166 #endif
2167
2168 static bool hugepage_vma_check(struct vm_area_struct *vma)
2169 {
2170         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2171             (vma->vm_flags & VM_NOHUGEPAGE))
2172                 return false;
2173         if (vma->vm_flags & VM_LOCKED)
2174                 return false;
2175         if (!vma->anon_vma || vma->vm_ops)
2176                 return false;
2177         if (is_vma_temporary_stack(vma))
2178                 return false;
2179         VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma);
2180         return true;
2181 }
2182
2183 static void collapse_huge_page(struct mm_struct *mm,
2184                                    unsigned long address,
2185                                    struct page **hpage,
2186                                    struct vm_area_struct *vma,
2187                                    int node)
2188 {
2189         pmd_t *pmd, _pmd;
2190         pte_t *pte;
2191         pgtable_t pgtable;
2192         struct page *new_page;
2193         spinlock_t *pmd_ptl, *pte_ptl;
2194         int isolated, result = 0;
2195         unsigned long hstart, hend;
2196         struct mem_cgroup *memcg;
2197         unsigned long mmun_start;       /* For mmu_notifiers */
2198         unsigned long mmun_end;         /* For mmu_notifiers */
2199         gfp_t gfp;
2200
2201         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2202
2203         /* Only allocate from the target node */
2204         gfp = alloc_hugepage_gfpmask(khugepaged_defrag(), __GFP_OTHER_NODE) |
2205                 __GFP_THISNODE;
2206
2207         /* release the mmap_sem read lock. */
2208         new_page = khugepaged_alloc_page(hpage, gfp, mm, address, node);
2209         if (!new_page) {
2210                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
2211                 goto out_nolock;
2212         }
2213
2214         if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
2215                 result = SCAN_CGROUP_CHARGE_FAIL;
2216                 goto out_nolock;
2217         }
2218
2219         /*
2220          * Prevent all access to pagetables with the exception of
2221          * gup_fast later hanlded by the ptep_clear_flush and the VM
2222          * handled by the anon_vma lock + PG_lock.
2223          */
2224         down_write(&mm->mmap_sem);
2225         if (unlikely(khugepaged_test_exit(mm))) {
2226                 result = SCAN_ANY_PROCESS;
2227                 goto out;
2228         }
2229
2230         vma = find_vma(mm, address);
2231         if (!vma) {
2232                 result = SCAN_VMA_NULL;
2233                 goto out;
2234         }
2235         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2236         hend = vma->vm_end & HPAGE_PMD_MASK;
2237         if (address < hstart || address + HPAGE_PMD_SIZE > hend) {
2238                 result = SCAN_ADDRESS_RANGE;
2239                 goto out;
2240         }
2241         if (!hugepage_vma_check(vma)) {
2242                 result = SCAN_VMA_CHECK;
2243                 goto out;
2244         }
2245         pmd = mm_find_pmd(mm, address);
2246         if (!pmd) {
2247                 result = SCAN_PMD_NULL;
2248                 goto out;
2249         }
2250
2251         anon_vma_lock_write(vma->anon_vma);
2252
2253         pte = pte_offset_map(pmd, address);
2254         pte_ptl = pte_lockptr(mm, pmd);
2255
2256         mmun_start = address;
2257         mmun_end   = address + HPAGE_PMD_SIZE;
2258         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2259         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2260         /*
2261          * After this gup_fast can't run anymore. This also removes
2262          * any huge TLB entry from the CPU so we won't allow
2263          * huge and small TLB entries for the same virtual address
2264          * to avoid the risk of CPU bugs in that area.
2265          */
2266         _pmd = pmdp_collapse_flush(vma, address, pmd);
2267         spin_unlock(pmd_ptl);
2268         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2269
2270         spin_lock(pte_ptl);
2271         isolated = __collapse_huge_page_isolate(vma, address, pte);
2272         spin_unlock(pte_ptl);
2273
2274         if (unlikely(!isolated)) {
2275                 pte_unmap(pte);
2276                 spin_lock(pmd_ptl);
2277                 BUG_ON(!pmd_none(*pmd));
2278                 /*
2279                  * We can only use set_pmd_at when establishing
2280                  * hugepmds and never for establishing regular pmds that
2281                  * points to regular pagetables. Use pmd_populate for that
2282                  */
2283                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2284                 spin_unlock(pmd_ptl);
2285                 anon_vma_unlock_write(vma->anon_vma);
2286                 result = SCAN_FAIL;
2287                 goto out;
2288         }
2289
2290         /*
2291          * All pages are isolated and locked so anon_vma rmap
2292          * can't run anymore.
2293          */
2294         anon_vma_unlock_write(vma->anon_vma);
2295
2296         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2297         pte_unmap(pte);
2298         __SetPageUptodate(new_page);
2299         pgtable = pmd_pgtable(_pmd);
2300
2301         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2302         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2303
2304         /*
2305          * spin_lock() below is not the equivalent of smp_wmb(), so
2306          * this is needed to avoid the copy_huge_page writes to become
2307          * visible after the set_pmd_at() write.
2308          */
2309         smp_wmb();
2310
2311         spin_lock(pmd_ptl);
2312         BUG_ON(!pmd_none(*pmd));
2313         page_add_new_anon_rmap(new_page, vma, address, true);
2314         mem_cgroup_commit_charge(new_page, memcg, false, true);
2315         lru_cache_add_active_or_unevictable(new_page, vma);
2316         pgtable_trans_huge_deposit(mm, pmd, pgtable);
2317         set_pmd_at(mm, address, pmd, _pmd);
2318         update_mmu_cache_pmd(vma, address, pmd);
2319         spin_unlock(pmd_ptl);
2320
2321         *hpage = NULL;
2322
2323         khugepaged_pages_collapsed++;
2324         result = SCAN_SUCCEED;
2325 out_up_write:
2326         up_write(&mm->mmap_sem);
2327         trace_mm_collapse_huge_page(mm, isolated, result);
2328         return;
2329
2330 out_nolock:
2331         trace_mm_collapse_huge_page(mm, isolated, result);
2332         return;
2333 out:
2334         mem_cgroup_cancel_charge(new_page, memcg, true);
2335         goto out_up_write;
2336 }
2337
2338 static int khugepaged_scan_pmd(struct mm_struct *mm,
2339                                struct vm_area_struct *vma,
2340                                unsigned long address,
2341                                struct page **hpage)
2342 {
2343         pmd_t *pmd;
2344         pte_t *pte, *_pte;
2345         int ret = 0, none_or_zero = 0, result = 0;
2346         struct page *page = NULL;
2347         unsigned long _address;
2348         spinlock_t *ptl;
2349         int node = NUMA_NO_NODE;
2350         bool writable = false, referenced = false;
2351
2352         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2353
2354         pmd = mm_find_pmd(mm, address);
2355         if (!pmd) {
2356                 result = SCAN_PMD_NULL;
2357                 goto out;
2358         }
2359
2360         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2361         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2362         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2363              _pte++, _address += PAGE_SIZE) {
2364                 pte_t pteval = *_pte;
2365                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2366                         if (!userfaultfd_armed(vma) &&
2367                             ++none_or_zero <= khugepaged_max_ptes_none) {
2368                                 continue;
2369                         } else {
2370                                 result = SCAN_EXCEED_NONE_PTE;
2371                                 goto out_unmap;
2372                         }
2373                 }
2374                 if (!pte_present(pteval)) {
2375                         result = SCAN_PTE_NON_PRESENT;
2376                         goto out_unmap;
2377                 }
2378                 if (pte_write(pteval))
2379                         writable = true;
2380
2381                 page = vm_normal_page(vma, _address, pteval);
2382                 if (unlikely(!page)) {
2383                         result = SCAN_PAGE_NULL;
2384                         goto out_unmap;
2385                 }
2386
2387                 /* TODO: teach khugepaged to collapse THP mapped with pte */
2388                 if (PageCompound(page)) {
2389                         result = SCAN_PAGE_COMPOUND;
2390                         goto out_unmap;
2391                 }
2392
2393                 /*
2394                  * Record which node the original page is from and save this
2395                  * information to khugepaged_node_load[].
2396                  * Khupaged will allocate hugepage from the node has the max
2397                  * hit record.
2398                  */
2399                 node = page_to_nid(page);
2400                 if (khugepaged_scan_abort(node)) {
2401                         result = SCAN_SCAN_ABORT;
2402                         goto out_unmap;
2403                 }
2404                 khugepaged_node_load[node]++;
2405                 if (!PageLRU(page)) {
2406                         result = SCAN_SCAN_ABORT;
2407                         goto out_unmap;
2408                 }
2409                 if (PageLocked(page)) {
2410                         result = SCAN_PAGE_LOCK;
2411                         goto out_unmap;
2412                 }
2413                 if (!PageAnon(page)) {
2414                         result = SCAN_PAGE_ANON;
2415                         goto out_unmap;
2416                 }
2417
2418                 /*
2419                  * cannot use mapcount: can't collapse if there's a gup pin.
2420                  * The page must only be referenced by the scanned process
2421                  * and page swap cache.
2422                  */
2423                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2424                         result = SCAN_PAGE_COUNT;
2425                         goto out_unmap;
2426                 }
2427                 if (pte_young(pteval) ||
2428                     page_is_young(page) || PageReferenced(page) ||
2429                     mmu_notifier_test_young(vma->vm_mm, address))
2430                         referenced = true;
2431         }
2432         if (writable) {
2433                 if (referenced) {
2434                         result = SCAN_SUCCEED;
2435                         ret = 1;
2436                 } else {
2437                         result = SCAN_NO_REFERENCED_PAGE;
2438                 }
2439         } else {
2440                 result = SCAN_PAGE_RO;
2441         }
2442 out_unmap:
2443         pte_unmap_unlock(pte, ptl);
2444         if (ret) {
2445                 node = khugepaged_find_target_node();
2446                 /* collapse_huge_page will return with the mmap_sem released */
2447                 collapse_huge_page(mm, address, hpage, vma, node);
2448         }
2449 out:
2450         trace_mm_khugepaged_scan_pmd(mm, page_to_pfn(page), writable, referenced,
2451                                      none_or_zero, result);
2452         return ret;
2453 }
2454
2455 static void collect_mm_slot(struct mm_slot *mm_slot)
2456 {
2457         struct mm_struct *mm = mm_slot->mm;
2458
2459         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2460
2461         if (khugepaged_test_exit(mm)) {
2462                 /* free mm_slot */
2463                 hash_del(&mm_slot->hash);
2464                 list_del(&mm_slot->mm_node);
2465
2466                 /*
2467                  * Not strictly needed because the mm exited already.
2468                  *
2469                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2470                  */
2471
2472                 /* khugepaged_mm_lock actually not necessary for the below */
2473                 free_mm_slot(mm_slot);
2474                 mmdrop(mm);
2475         }
2476 }
2477
2478 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2479                                             struct page **hpage)
2480         __releases(&khugepaged_mm_lock)
2481         __acquires(&khugepaged_mm_lock)
2482 {
2483         struct mm_slot *mm_slot;
2484         struct mm_struct *mm;
2485         struct vm_area_struct *vma;
2486         int progress = 0;
2487
2488         VM_BUG_ON(!pages);
2489         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2490
2491         if (khugepaged_scan.mm_slot)
2492                 mm_slot = khugepaged_scan.mm_slot;
2493         else {
2494                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2495                                      struct mm_slot, mm_node);
2496                 khugepaged_scan.address = 0;
2497                 khugepaged_scan.mm_slot = mm_slot;
2498         }
2499         spin_unlock(&khugepaged_mm_lock);
2500
2501         mm = mm_slot->mm;
2502         down_read(&mm->mmap_sem);
2503         if (unlikely(khugepaged_test_exit(mm)))
2504                 vma = NULL;
2505         else
2506                 vma = find_vma(mm, khugepaged_scan.address);
2507
2508         progress++;
2509         for (; vma; vma = vma->vm_next) {
2510                 unsigned long hstart, hend;
2511
2512                 cond_resched();
2513                 if (unlikely(khugepaged_test_exit(mm))) {
2514                         progress++;
2515                         break;
2516                 }
2517                 if (!hugepage_vma_check(vma)) {
2518 skip:
2519                         progress++;
2520                         continue;
2521                 }
2522                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2523                 hend = vma->vm_end & HPAGE_PMD_MASK;
2524                 if (hstart >= hend)
2525                         goto skip;
2526                 if (khugepaged_scan.address > hend)
2527                         goto skip;
2528                 if (khugepaged_scan.address < hstart)
2529                         khugepaged_scan.address = hstart;
2530                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2531
2532                 while (khugepaged_scan.address < hend) {
2533                         int ret;
2534                         cond_resched();
2535                         if (unlikely(khugepaged_test_exit(mm)))
2536                                 goto breakouterloop;
2537
2538                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2539                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2540                                   hend);
2541                         ret = khugepaged_scan_pmd(mm, vma,
2542                                                   khugepaged_scan.address,
2543                                                   hpage);
2544                         /* move to next address */
2545                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2546                         progress += HPAGE_PMD_NR;
2547                         if (ret)
2548                                 /* we released mmap_sem so break loop */
2549                                 goto breakouterloop_mmap_sem;
2550                         if (progress >= pages)
2551                                 goto breakouterloop;
2552                 }
2553         }
2554 breakouterloop:
2555         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2556 breakouterloop_mmap_sem:
2557
2558         spin_lock(&khugepaged_mm_lock);
2559         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2560         /*
2561          * Release the current mm_slot if this mm is about to die, or
2562          * if we scanned all vmas of this mm.
2563          */
2564         if (khugepaged_test_exit(mm) || !vma) {
2565                 /*
2566                  * Make sure that if mm_users is reaching zero while
2567                  * khugepaged runs here, khugepaged_exit will find
2568                  * mm_slot not pointing to the exiting mm.
2569                  */
2570                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2571                         khugepaged_scan.mm_slot = list_entry(
2572                                 mm_slot->mm_node.next,
2573                                 struct mm_slot, mm_node);
2574                         khugepaged_scan.address = 0;
2575                 } else {
2576                         khugepaged_scan.mm_slot = NULL;
2577                         khugepaged_full_scans++;
2578                 }
2579
2580                 collect_mm_slot(mm_slot);
2581         }
2582
2583         return progress;
2584 }
2585
2586 static int khugepaged_has_work(void)
2587 {
2588         return !list_empty(&khugepaged_scan.mm_head) &&
2589                 khugepaged_enabled();
2590 }
2591
2592 static int khugepaged_wait_event(void)
2593 {
2594         return !list_empty(&khugepaged_scan.mm_head) ||
2595                 kthread_should_stop();
2596 }
2597
2598 static void khugepaged_do_scan(void)
2599 {
2600         struct page *hpage = NULL;
2601         unsigned int progress = 0, pass_through_head = 0;
2602         unsigned int pages = khugepaged_pages_to_scan;
2603         bool wait = true;
2604
2605         barrier(); /* write khugepaged_pages_to_scan to local stack */
2606
2607         while (progress < pages) {
2608                 if (!khugepaged_prealloc_page(&hpage, &wait))
2609                         break;
2610
2611                 cond_resched();
2612
2613                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2614                         break;
2615
2616                 spin_lock(&khugepaged_mm_lock);
2617                 if (!khugepaged_scan.mm_slot)
2618                         pass_through_head++;
2619                 if (khugepaged_has_work() &&
2620                     pass_through_head < 2)
2621                         progress += khugepaged_scan_mm_slot(pages - progress,
2622                                                             &hpage);
2623                 else
2624                         progress = pages;
2625                 spin_unlock(&khugepaged_mm_lock);
2626         }
2627
2628         if (!IS_ERR_OR_NULL(hpage))
2629                 put_page(hpage);
2630 }
2631
2632 static void khugepaged_wait_work(void)
2633 {
2634         if (khugepaged_has_work()) {
2635                 if (!khugepaged_scan_sleep_millisecs)
2636                         return;
2637
2638                 wait_event_freezable_timeout(khugepaged_wait,
2639                                              kthread_should_stop(),
2640                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2641                 return;
2642         }
2643
2644         if (khugepaged_enabled())
2645                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2646 }
2647
2648 static int khugepaged(void *none)
2649 {
2650         struct mm_slot *mm_slot;
2651
2652         set_freezable();
2653         set_user_nice(current, MAX_NICE);
2654
2655         while (!kthread_should_stop()) {
2656                 khugepaged_do_scan();
2657                 khugepaged_wait_work();
2658         }
2659
2660         spin_lock(&khugepaged_mm_lock);
2661         mm_slot = khugepaged_scan.mm_slot;
2662         khugepaged_scan.mm_slot = NULL;
2663         if (mm_slot)
2664                 collect_mm_slot(mm_slot);
2665         spin_unlock(&khugepaged_mm_lock);
2666         return 0;
2667 }
2668
2669 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2670                 unsigned long haddr, pmd_t *pmd)
2671 {
2672         struct mm_struct *mm = vma->vm_mm;
2673         pgtable_t pgtable;
2674         pmd_t _pmd;
2675         int i;
2676
2677         /* leave pmd empty until pte is filled */
2678         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2679
2680         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2681         pmd_populate(mm, &_pmd, pgtable);
2682
2683         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2684                 pte_t *pte, entry;
2685                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2686                 entry = pte_mkspecial(entry);
2687                 pte = pte_offset_map(&_pmd, haddr);
2688                 VM_BUG_ON(!pte_none(*pte));
2689                 set_pte_at(mm, haddr, pte, entry);
2690                 pte_unmap(pte);
2691         }
2692         smp_wmb(); /* make pte visible before pmd */
2693         pmd_populate(mm, pmd, pgtable);
2694         put_huge_zero_page();
2695 }
2696
2697 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2698                 unsigned long haddr)
2699 {
2700         struct mm_struct *mm = vma->vm_mm;
2701         struct page *page;
2702         pgtable_t pgtable;
2703         pmd_t _pmd;
2704         bool young, write;
2705         int i;
2706
2707         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2708         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2709         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2710         VM_BUG_ON(!pmd_trans_huge(*pmd));
2711
2712         count_vm_event(THP_SPLIT_PMD);
2713
2714         if (vma_is_dax(vma)) {
2715                 pmd_t _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2716                 if (is_huge_zero_pmd(_pmd))
2717                         put_huge_zero_page();
2718                 return;
2719         } else if (is_huge_zero_pmd(*pmd)) {
2720                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2721         }
2722
2723         page = pmd_page(*pmd);
2724         VM_BUG_ON_PAGE(!page_count(page), page);
2725         atomic_add(HPAGE_PMD_NR - 1, &page->_count);
2726         write = pmd_write(*pmd);
2727         young = pmd_young(*pmd);
2728
2729         /* leave pmd empty until pte is filled */
2730         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2731
2732         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2733         pmd_populate(mm, &_pmd, pgtable);
2734
2735         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2736                 pte_t entry, *pte;
2737                 /*
2738                  * Note that NUMA hinting access restrictions are not
2739                  * transferred to avoid any possibility of altering
2740                  * permissions across VMAs.
2741                  */
2742                 entry = mk_pte(page + i, vma->vm_page_prot);
2743                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2744                 if (!write)
2745                         entry = pte_wrprotect(entry);
2746                 if (!young)
2747                         entry = pte_mkold(entry);
2748                 pte = pte_offset_map(&_pmd, haddr);
2749                 BUG_ON(!pte_none(*pte));
2750                 set_pte_at(mm, haddr, pte, entry);
2751                 atomic_inc(&page[i]._mapcount);
2752                 pte_unmap(pte);
2753         }
2754
2755         /*
2756          * Set PG_double_map before dropping compound_mapcount to avoid
2757          * false-negative page_mapped().
2758          */
2759         if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2760                 for (i = 0; i < HPAGE_PMD_NR; i++)
2761                         atomic_inc(&page[i]._mapcount);
2762         }
2763
2764         if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2765                 /* Last compound_mapcount is gone. */
2766                 __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
2767                 if (TestClearPageDoubleMap(page)) {
2768                         /* No need in mapcount reference anymore */
2769                         for (i = 0; i < HPAGE_PMD_NR; i++)
2770                                 atomic_dec(&page[i]._mapcount);
2771                 }
2772         }
2773
2774         smp_wmb(); /* make pte visible before pmd */
2775         pmd_populate(mm, pmd, pgtable);
2776 }
2777
2778 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2779                 unsigned long address)
2780 {
2781         spinlock_t *ptl;
2782         struct mm_struct *mm = vma->vm_mm;
2783         unsigned long haddr = address & HPAGE_PMD_MASK;
2784
2785         mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
2786         ptl = pmd_lock(mm, pmd);
2787         if (likely(pmd_trans_huge(*pmd)))
2788                 __split_huge_pmd_locked(vma, pmd, haddr);
2789         spin_unlock(ptl);
2790         mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
2791 }
2792
2793 static void split_huge_pmd_address(struct vm_area_struct *vma,
2794                                     unsigned long address)
2795 {
2796         pgd_t *pgd;
2797         pud_t *pud;
2798         pmd_t *pmd;
2799
2800         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2801
2802         pgd = pgd_offset(vma->vm_mm, address);
2803         if (!pgd_present(*pgd))
2804                 return;
2805
2806         pud = pud_offset(pgd, address);
2807         if (!pud_present(*pud))
2808                 return;
2809
2810         pmd = pmd_offset(pud, address);
2811         if (!pmd_present(*pmd) || !pmd_trans_huge(*pmd))
2812                 return;
2813         /*
2814          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2815          * materialize from under us.
2816          */
2817         split_huge_pmd(vma, pmd, address);
2818 }
2819
2820 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2821                              unsigned long start,
2822                              unsigned long end,
2823                              long adjust_next)
2824 {
2825         /*
2826          * If the new start address isn't hpage aligned and it could
2827          * previously contain an hugepage: check if we need to split
2828          * an huge pmd.
2829          */
2830         if (start & ~HPAGE_PMD_MASK &&
2831             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2832             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2833                 split_huge_pmd_address(vma, start);
2834
2835         /*
2836          * If the new end address isn't hpage aligned and it could
2837          * previously contain an hugepage: check if we need to split
2838          * an huge pmd.
2839          */
2840         if (end & ~HPAGE_PMD_MASK &&
2841             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2842             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2843                 split_huge_pmd_address(vma, end);
2844
2845         /*
2846          * If we're also updating the vma->vm_next->vm_start, if the new
2847          * vm_next->vm_start isn't page aligned and it could previously
2848          * contain an hugepage: check if we need to split an huge pmd.
2849          */
2850         if (adjust_next > 0) {
2851                 struct vm_area_struct *next = vma->vm_next;
2852                 unsigned long nstart = next->vm_start;
2853                 nstart += adjust_next << PAGE_SHIFT;
2854                 if (nstart & ~HPAGE_PMD_MASK &&
2855                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2856                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2857                         split_huge_pmd_address(next, nstart);
2858         }
2859 }