mm/huge_memory.c

   1 /*
   2  *  Copyright (C) 2009  Red Hat, Inc.
   3  *
   4  *  This work is licensed under the terms of the GNU GPL, version 2. See
   5  *  the COPYING file in the top-level directory.
   6  */
   7
   8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   9
  10 #include <linux/mm.h>
  11 #include <linux/sched.h>
  12 #include <linux/highmem.h>
  13 #include <linux/hugetlb.h>
  14 #include <linux/mmu_notifier.h>
  15 #include <linux/rmap.h>
  16 #include <linux/swap.h>
  17 #include <linux/shrinker.h>
  18 #include <linux/mm_inline.h>
  19 #include <linux/swapops.h>
  20 #include <linux/dax.h>
  21 #include <linux/kthread.h>
  22 #include <linux/khugepaged.h>
  23 #include <linux/freezer.h>
  24 #include <linux/pfn_t.h>
  25 #include <linux/mman.h>
  26 #include <linux/memremap.h>
  27 #include <linux/pagemap.h>
  28 #include <linux/debugfs.h>
  29 #include <linux/migrate.h>
  30 #include <linux/hashtable.h>
  31 #include <linux/userfaultfd_k.h>
  32 #include <linux/page_idle.h>
  33
  34 #include <asm/tlb.h>
  35 #include <asm/pgalloc.h>
  36 #include "internal.h"
  37
  38 enum scan_result {
  39         SCAN_FAIL,
  40         SCAN_SUCCEED,
  41         SCAN_PMD_NULL,
  42         SCAN_EXCEED_NONE_PTE,
  43         SCAN_PTE_NON_PRESENT,
  44         SCAN_PAGE_RO,
  45         SCAN_NO_REFERENCED_PAGE,
  46         SCAN_PAGE_NULL,
  47         SCAN_SCAN_ABORT,
  48         SCAN_PAGE_COUNT,
  49         SCAN_PAGE_LRU,
  50         SCAN_PAGE_LOCK,
  51         SCAN_PAGE_ANON,
  52         SCAN_PAGE_COMPOUND,
  53         SCAN_ANY_PROCESS,
  54         SCAN_VMA_NULL,
  55         SCAN_VMA_CHECK,
  56         SCAN_ADDRESS_RANGE,
  57         SCAN_SWAP_CACHE_PAGE,
  58         SCAN_DEL_PAGE_LRU,
  59         SCAN_ALLOC_HUGE_PAGE_FAIL,
  60         SCAN_CGROUP_CHARGE_FAIL
  61 };
  62
  63 #define CREATE_TRACE_POINTS
  64 #include <trace/events/huge_memory.h>
  65
  66 /*
  67  * By default transparent hugepage support is disabled in order that avoid
  68  * to risk increase the memory footprint of applications without a guaranteed
  69  * benefit. When transparent hugepage support is enabled, is for all mappings,
  70  * and khugepaged scans all mappings.
  71  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  72  * for all hugepage allocations.
  73  */
  74 unsigned long transparent_hugepage_flags __read_mostly =
  75 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  76         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  77 #endif
  78 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  79         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  80 #endif
  81         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
  82         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  83         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  84
  85 /* default scan 8*512 pte (or vmas) every 30 second */
  86 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
  87 static unsigned int khugepaged_pages_collapsed;
  88 static unsigned int khugepaged_full_scans;
  89 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  90 /* during fragmentation poll the hugepage allocator once every minute */
  91 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  92 static struct task_struct *khugepaged_thread __read_mostly;
  93 static DEFINE_MUTEX(khugepaged_mutex);
  94 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  95 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  96 /*
  97  * default collapse hugepages if there is at least one pte mapped like
  98  * it would have happened if the vma was large enough during page
  99  * fault.
 100  */
 101 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
 102
 103 static int khugepaged(void *none);
 104 static int khugepaged_slab_init(void);
 105 static void khugepaged_slab_exit(void);
 106
 107 #define MM_SLOTS_HASH_BITS 10
 108 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
 109
 110 static struct kmem_cache *mm_slot_cache __read_mostly;
 111
 112 /**
 113  * struct mm_slot - hash lookup from mm to mm_slot
 114  * @hash: hash collision list
 115  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
 116  * @mm: the mm that this information is valid for
 117  */
 118 struct mm_slot {
 119         struct hlist_node hash;
 120         struct list_head mm_node;
 121         struct mm_struct *mm;
 122 };
 123
 124 /**
 125  * struct khugepaged_scan - cursor for scanning
 126  * @mm_head: the head of the mm list to scan
 127  * @mm_slot: the current mm_slot we are scanning
 128  * @address: the next address inside that to be scanned
 129  *
 130  * There is only the one khugepaged_scan instance of this cursor structure.
 131  */
 132 struct khugepaged_scan {
 133         struct list_head mm_head;
 134         struct mm_slot *mm_slot;
 135         unsigned long address;
 136 };
 137 static struct khugepaged_scan khugepaged_scan = {
 138         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 139 };
 140
 141 static DEFINE_SPINLOCK(split_queue_lock);
 142 static LIST_HEAD(split_queue);
 143 static unsigned long split_queue_len;
 144 static struct shrinker deferred_split_shrinker;
 145
 146 static void set_recommended_min_free_kbytes(void)
 147 {
 148         struct zone *zone;
 149         int nr_zones = 0;
 150         unsigned long recommended_min;
 151
 152         for_each_populated_zone(zone)
 153                 nr_zones++;
 154
 155         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
 156         recommended_min = pageblock_nr_pages * nr_zones * 2;
 157
 158         /*
 159          * Make sure that on average at least two pageblocks are almost free
 160          * of another type, one for a migratetype to fall back to and a
 161          * second to avoid subsequent fallbacks of other types There are 3
 162          * MIGRATE_TYPES we care about.
 163          */
 164         recommended_min += pageblock_nr_pages * nr_zones *
 165                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
 166
 167         /* don't ever allow to reserve more than 5% of the lowmem */
 168         recommended_min = min(recommended_min,
 169                               (unsigned long) nr_free_buffer_pages() / 20);
 170         recommended_min <<= (PAGE_SHIFT-10);
 171
 172         if (recommended_min > min_free_kbytes) {
 173                 if (user_min_free_kbytes >= 0)
 174                         pr_info("raising min_free_kbytes from %d to %lu "
 175                                 "to help transparent hugepage allocations\n",
 176                                 min_free_kbytes, recommended_min);
 177
 178                 min_free_kbytes = recommended_min;
 179         }
 180         setup_per_zone_wmarks();
 181 }
 182
 183 static int start_stop_khugepaged(void)
 184 {
 185         int err = 0;
 186         if (khugepaged_enabled()) {
 187                 if (!khugepaged_thread)
 188                         khugepaged_thread = kthread_run(khugepaged, NULL,
 189                                                         "khugepaged");
 190                 if (IS_ERR(khugepaged_thread)) {
 191                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
 192                         err = PTR_ERR(khugepaged_thread);
 193                         khugepaged_thread = NULL;
 194                         goto fail;
 195                 }
 196
 197                 if (!list_empty(&khugepaged_scan.mm_head))
 198                         wake_up_interruptible(&khugepaged_wait);
 199
 200                 set_recommended_min_free_kbytes();
 201         } else if (khugepaged_thread) {
 202                 kthread_stop(khugepaged_thread);
 203                 khugepaged_thread = NULL;
 204         }
 205 fail:
 206         return err;
 207 }
 208
 209 static atomic_t huge_zero_refcount;
 210 struct page *huge_zero_page __read_mostly;
 211
 212 struct page *get_huge_zero_page(void)
 213 {
 214         struct page *zero_page;
 215 retry:
 216         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
 217                 return READ_ONCE(huge_zero_page);
 218
 219         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
 220                         HPAGE_PMD_ORDER);
 221         if (!zero_page) {
 222                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
 223                 return NULL;
 224         }
 225         count_vm_event(THP_ZERO_PAGE_ALLOC);
 226         preempt_disable();
 227         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
 228                 preempt_enable();
 229                 __free_pages(zero_page, compound_order(zero_page));
 230                 goto retry;
 231         }
 232
 233         /* We take additional reference here. It will be put back by shrinker */
 234         atomic_set(&huge_zero_refcount, 2);
 235         preempt_enable();
 236         return READ_ONCE(huge_zero_page);
 237 }
 238
 239 static void put_huge_zero_page(void)
 240 {
 241         /*
 242          * Counter should never go to zero here. Only shrinker can put
 243          * last reference.
 244          */
 245         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 246 }
 247
 248 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 249                                         struct shrink_control *sc)
 250 {
 251         /* we can free zero page only if last reference remains */
 252         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 253 }
 254
 255 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 256                                        struct shrink_control *sc)
 257 {
 258         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 259                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 260                 BUG_ON(zero_page == NULL);
 261                 __free_pages(zero_page, compound_order(zero_page));
 262                 return HPAGE_PMD_NR;
 263         }
 264
 265         return 0;
 266 }
 267
 268 static struct shrinker huge_zero_page_shrinker = {
 269         .count_objects = shrink_huge_zero_page_count,
 270         .scan_objects = shrink_huge_zero_page_scan,
 271         .seeks = DEFAULT_SEEKS,
 272 };
 273
 274 #ifdef CONFIG_SYSFS
 275
 276 static ssize_t double_flag_show(struct kobject *kobj,
 277                                 struct kobj_attribute *attr, char *buf,
 278                                 enum transparent_hugepage_flag enabled,
 279                                 enum transparent_hugepage_flag req_madv)
 280 {
 281         if (test_bit(enabled, &transparent_hugepage_flags)) {
 282                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
 283                 return sprintf(buf, "[always] madvise never\n");
 284         } else if (test_bit(req_madv, &transparent_hugepage_flags))
 285                 return sprintf(buf, "always [madvise] never\n");
 286         else
 287                 return sprintf(buf, "always madvise [never]\n");
 288 }
 289 static ssize_t double_flag_store(struct kobject *kobj,
 290                                  struct kobj_attribute *attr,
 291                                  const char *buf, size_t count,
 292                                  enum transparent_hugepage_flag enabled,
 293                                  enum transparent_hugepage_flag req_madv)
 294 {
 295         if (!memcmp("always", buf,
 296                     min(sizeof("always")-1, count))) {
 297                 set_bit(enabled, &transparent_hugepage_flags);
 298                 clear_bit(req_madv, &transparent_hugepage_flags);
 299         } else if (!memcmp("madvise", buf,
 300                            min(sizeof("madvise")-1, count))) {
 301                 clear_bit(enabled, &transparent_hugepage_flags);
 302                 set_bit(req_madv, &transparent_hugepage_flags);
 303         } else if (!memcmp("never", buf,
 304                            min(sizeof("never")-1, count))) {
 305                 clear_bit(enabled, &transparent_hugepage_flags);
 306                 clear_bit(req_madv, &transparent_hugepage_flags);
 307         } else
 308                 return -EINVAL;
 309
 310         return count;
 311 }
 312
 313 static ssize_t enabled_show(struct kobject *kobj,
 314                             struct kobj_attribute *attr, char *buf)
 315 {
 316         return double_flag_show(kobj, attr, buf,
 317                                 TRANSPARENT_HUGEPAGE_FLAG,
 318                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 319 }
 320 static ssize_t enabled_store(struct kobject *kobj,
 321                              struct kobj_attribute *attr,
 322                              const char *buf, size_t count)
 323 {
 324         ssize_t ret;
 325
 326         ret = double_flag_store(kobj, attr, buf, count,
 327                                 TRANSPARENT_HUGEPAGE_FLAG,
 328                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 329
 330         if (ret > 0) {
 331                 int err;
 332
 333                 mutex_lock(&khugepaged_mutex);
 334                 err = start_stop_khugepaged();
 335                 mutex_unlock(&khugepaged_mutex);
 336
 337                 if (err)
 338                         ret = err;
 339         }
 340
 341         return ret;
 342 }
 343 static struct kobj_attribute enabled_attr =
 344         __ATTR(enabled, 0644, enabled_show, enabled_store);
 345
 346 static ssize_t single_flag_show(struct kobject *kobj,
 347                                 struct kobj_attribute *attr, char *buf,
 348                                 enum transparent_hugepage_flag flag)
 349 {
 350         return sprintf(buf, "%d\n",
 351                        !!test_bit(flag, &transparent_hugepage_flags));
 352 }
 353
 354 static ssize_t single_flag_store(struct kobject *kobj,
 355                                  struct kobj_attribute *attr,
 356                                  const char *buf, size_t count,
 357                                  enum transparent_hugepage_flag flag)
 358 {
 359         unsigned long value;
 360         int ret;
 361
 362         ret = kstrtoul(buf, 10, &value);
 363         if (ret < 0)
 364                 return ret;
 365         if (value > 1)
 366                 return -EINVAL;
 367
 368         if (value)
 369                 set_bit(flag, &transparent_hugepage_flags);
 370         else
 371                 clear_bit(flag, &transparent_hugepage_flags);
 372
 373         return count;
 374 }
 375
 376 /*
 377  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 378  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 379  * memory just to allocate one more hugepage.
 380  */
 381 static ssize_t defrag_show(struct kobject *kobj,
 382                            struct kobj_attribute *attr, char *buf)
 383 {
 384         return double_flag_show(kobj, attr, buf,
 385                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 386                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 387 }
 388 static ssize_t defrag_store(struct kobject *kobj,
 389                             struct kobj_attribute *attr,
 390                             const char *buf, size_t count)
 391 {
 392         return double_flag_store(kobj, attr, buf, count,
 393                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 394                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 395 }
 396 static struct kobj_attribute defrag_attr =
 397         __ATTR(defrag, 0644, defrag_show, defrag_store);
 398
 399 static ssize_t use_zero_page_show(struct kobject *kobj,
 400                 struct kobj_attribute *attr, char *buf)
 401 {
 402         return single_flag_show(kobj, attr, buf,
 403                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 404 }
 405 static ssize_t use_zero_page_store(struct kobject *kobj,
 406                 struct kobj_attribute *attr, const char *buf, size_t count)
 407 {
 408         return single_flag_store(kobj, attr, buf, count,
 409                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 410 }
 411 static struct kobj_attribute use_zero_page_attr =
 412         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 413 #ifdef CONFIG_DEBUG_VM
 414 static ssize_t debug_cow_show(struct kobject *kobj,
 415                                 struct kobj_attribute *attr, char *buf)
 416 {
 417         return single_flag_show(kobj, attr, buf,
 418                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 419 }
 420 static ssize_t debug_cow_store(struct kobject *kobj,
 421                                struct kobj_attribute *attr,
 422                                const char *buf, size_t count)
 423 {
 424         return single_flag_store(kobj, attr, buf, count,
 425                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 426 }
 427 static struct kobj_attribute debug_cow_attr =
 428         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 429 #endif /* CONFIG_DEBUG_VM */
 430
 431 static struct attribute *hugepage_attr[] = {
 432         &enabled_attr.attr,
 433         &defrag_attr.attr,
 434         &use_zero_page_attr.attr,
 435 #ifdef CONFIG_DEBUG_VM
 436         &debug_cow_attr.attr,
 437 #endif
 438         NULL,
 439 };
 440
 441 static struct attribute_group hugepage_attr_group = {
 442         .attrs = hugepage_attr,
 443 };
 444
 445 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 446                                          struct kobj_attribute *attr,
 447                                          char *buf)
 448 {
 449         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 450 }
 451
 452 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 453                                           struct kobj_attribute *attr,
 454                                           const char *buf, size_t count)
 455 {
 456         unsigned long msecs;
 457         int err;
 458
 459         err = kstrtoul(buf, 10, &msecs);
 460         if (err || msecs > UINT_MAX)
 461                 return -EINVAL;
 462
 463         khugepaged_scan_sleep_millisecs = msecs;
 464         wake_up_interruptible(&khugepaged_wait);
 465
 466         return count;
 467 }
 468 static struct kobj_attribute scan_sleep_millisecs_attr =
 469         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 470                scan_sleep_millisecs_store);
 471
 472 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 473                                           struct kobj_attribute *attr,
 474                                           char *buf)
 475 {
 476         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 477 }
 478
 479 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 480                                            struct kobj_attribute *attr,
 481                                            const char *buf, size_t count)
 482 {
 483         unsigned long msecs;
 484         int err;
 485
 486         err = kstrtoul(buf, 10, &msecs);
 487         if (err || msecs > UINT_MAX)
 488                 return -EINVAL;
 489
 490         khugepaged_alloc_sleep_millisecs = msecs;
 491         wake_up_interruptible(&khugepaged_wait);
 492
 493         return count;
 494 }
 495 static struct kobj_attribute alloc_sleep_millisecs_attr =
 496         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 497                alloc_sleep_millisecs_store);
 498
 499 static ssize_t pages_to_scan_show(struct kobject *kobj,
 500                                   struct kobj_attribute *attr,
 501                                   char *buf)
 502 {
 503         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 504 }
 505 static ssize_t pages_to_scan_store(struct kobject *kobj,
 506                                    struct kobj_attribute *attr,
 507                                    const char *buf, size_t count)
 508 {
 509         int err;
 510         unsigned long pages;
 511
 512         err = kstrtoul(buf, 10, &pages);
 513         if (err || !pages || pages > UINT_MAX)
 514                 return -EINVAL;
 515
 516         khugepaged_pages_to_scan = pages;
 517
 518         return count;
 519 }
 520 static struct kobj_attribute pages_to_scan_attr =
 521         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 522                pages_to_scan_store);
 523
 524 static ssize_t pages_collapsed_show(struct kobject *kobj,
 525                                     struct kobj_attribute *attr,
 526                                     char *buf)
 527 {
 528         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 529 }
 530 static struct kobj_attribute pages_collapsed_attr =
 531         __ATTR_RO(pages_collapsed);
 532
 533 static ssize_t full_scans_show(struct kobject *kobj,
 534                                struct kobj_attribute *attr,
 535                                char *buf)
 536 {
 537         return sprintf(buf, "%u\n", khugepaged_full_scans);
 538 }
 539 static struct kobj_attribute full_scans_attr =
 540         __ATTR_RO(full_scans);
 541
 542 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 543                                       struct kobj_attribute *attr, char *buf)
 544 {
 545         return single_flag_show(kobj, attr, buf,
 546                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 547 }
 548 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 549                                        struct kobj_attribute *attr,
 550                                        const char *buf, size_t count)
 551 {
 552         return single_flag_store(kobj, attr, buf, count,
 553                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 554 }
 555 static struct kobj_attribute khugepaged_defrag_attr =
 556         __ATTR(defrag, 0644, khugepaged_defrag_show,
 557                khugepaged_defrag_store);
 558
 559 /*
 560  * max_ptes_none controls if khugepaged should collapse hugepages over
 561  * any unmapped ptes in turn potentially increasing the memory
 562  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 563  * reduce the available free memory in the system as it
 564  * runs. Increasing max_ptes_none will instead potentially reduce the
 565  * free memory in the system during the khugepaged scan.
 566  */
 567 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 568                                              struct kobj_attribute *attr,
 569                                              char *buf)
 570 {
 571         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 572 }
 573 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 574                                               struct kobj_attribute *attr,
 575                                               const char *buf, size_t count)
 576 {
 577         int err;
 578         unsigned long max_ptes_none;
 579
 580         err = kstrtoul(buf, 10, &max_ptes_none);
 581         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 582                 return -EINVAL;
 583
 584         khugepaged_max_ptes_none = max_ptes_none;
 585
 586         return count;
 587 }
 588 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 589         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 590                khugepaged_max_ptes_none_store);
 591
 592 static struct attribute *khugepaged_attr[] = {
 593         &khugepaged_defrag_attr.attr,
 594         &khugepaged_max_ptes_none_attr.attr,
 595         &pages_to_scan_attr.attr,
 596         &pages_collapsed_attr.attr,
 597         &full_scans_attr.attr,
 598         &scan_sleep_millisecs_attr.attr,
 599         &alloc_sleep_millisecs_attr.attr,
 600         NULL,
 601 };
 602
 603 static struct attribute_group khugepaged_attr_group = {
 604         .attrs = khugepaged_attr,
 605         .name = "khugepaged",
 606 };
 607
 608 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 609 {
 610         int err;
 611
 612         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 613         if (unlikely(!*hugepage_kobj)) {
 614                 pr_err("failed to create transparent hugepage kobject\n");
 615                 return -ENOMEM;
 616         }
 617
 618         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 619         if (err) {
 620                 pr_err("failed to register transparent hugepage group\n");
 621                 goto delete_obj;
 622         }
 623
 624         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 625         if (err) {
 626                 pr_err("failed to register transparent hugepage group\n");
 627                 goto remove_hp_group;
 628         }
 629
 630         return 0;
 631
 632 remove_hp_group:
 633         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 634 delete_obj:
 635         kobject_put(*hugepage_kobj);
 636         return err;
 637 }
 638
 639 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 640 {
 641         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 642         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 643         kobject_put(hugepage_kobj);
 644 }
 645 #else
 646 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 647 {
 648         return 0;
 649 }
 650
 651 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 652 {
 653 }
 654 #endif /* CONFIG_SYSFS */
 655
 656 static int __init hugepage_init(void)
 657 {
 658         int err;
 659         struct kobject *hugepage_kobj;
 660
 661         if (!has_transparent_hugepage()) {
 662                 transparent_hugepage_flags = 0;
 663                 return -EINVAL;
 664         }
 665
 666         err = hugepage_init_sysfs(&hugepage_kobj);
 667         if (err)
 668                 goto err_sysfs;
 669
 670         err = khugepaged_slab_init();
 671         if (err)
 672                 goto err_slab;
 673
 674         err = register_shrinker(&huge_zero_page_shrinker);
 675         if (err)
 676                 goto err_hzp_shrinker;
 677         err = register_shrinker(&deferred_split_shrinker);
 678         if (err)
 679                 goto err_split_shrinker;
 680
 681         /*
 682          * By default disable transparent hugepages on smaller systems,
 683          * where the extra memory used could hurt more than TLB overhead
 684          * is likely to save.  The admin can still enable it through /sys.
 685          */
 686         if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
 687                 transparent_hugepage_flags = 0;
 688                 return 0;
 689         }
 690
 691         err = start_stop_khugepaged();
 692         if (err)
 693                 goto err_khugepaged;
 694
 695         return 0;
 696 err_khugepaged:
 697         unregister_shrinker(&deferred_split_shrinker);
 698 err_split_shrinker:
 699         unregister_shrinker(&huge_zero_page_shrinker);
 700 err_hzp_shrinker:
 701         khugepaged_slab_exit();
 702 err_slab:
 703         hugepage_exit_sysfs(hugepage_kobj);
 704 err_sysfs:
 705         return err;
 706 }
 707 subsys_initcall(hugepage_init);
 708
 709 static int __init setup_transparent_hugepage(char *str)
 710 {
 711         int ret = 0;
 712         if (!str)
 713                 goto out;
 714         if (!strcmp(str, "always")) {
 715                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 716                         &transparent_hugepage_flags);
 717                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 718                           &transparent_hugepage_flags);
 719                 ret = 1;
 720         } else if (!strcmp(str, "madvise")) {
 721                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 722                           &transparent_hugepage_flags);
 723                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 724                         &transparent_hugepage_flags);
 725                 ret = 1;
 726         } else if (!strcmp(str, "never")) {
 727                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 728                           &transparent_hugepage_flags);
 729                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 730                           &transparent_hugepage_flags);
 731                 ret = 1;
 732         }
 733 out:
 734         if (!ret)
 735                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 736         return ret;
 737 }
 738 __setup("transparent_hugepage=", setup_transparent_hugepage);
 739
 740 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 741 {
 742         if (likely(vma->vm_flags & VM_WRITE))
 743                 pmd = pmd_mkwrite(pmd);
 744         return pmd;
 745 }
 746
 747 static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
 748 {
 749         pmd_t entry;
 750         entry = mk_pmd(page, prot);
 751         entry = pmd_mkhuge(entry);
 752         return entry;
 753 }
 754
 755 static inline struct list_head *page_deferred_list(struct page *page)
 756 {
 757         /*
 758          * ->lru in the tail pages is occupied by compound_head.
 759          * Let's use ->mapping + ->index in the second tail page as list_head.
 760          */
 761         return (struct list_head *)&page[2].mapping;
 762 }
 763
 764 void prep_transhuge_page(struct page *page)
 765 {
 766         /*
 767          * we use page->mapping and page->indexlru in second tail page
 768          * as list_head: assuming THP order >= 2
 769          */
 770         BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 771
 772         INIT_LIST_HEAD(page_deferred_list(page));
 773         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 774 }
 775
 776 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
 777                                         struct vm_area_struct *vma,
 778                                         unsigned long address, pmd_t *pmd,
 779                                         struct page *page, gfp_t gfp,
 780                                         unsigned int flags)
 781 {
 782         struct mem_cgroup *memcg;
 783         pgtable_t pgtable;
 784         spinlock_t *ptl;
 785         unsigned long haddr = address & HPAGE_PMD_MASK;
 786
 787         VM_BUG_ON_PAGE(!PageCompound(page), page);
 788
 789         if (mem_cgroup_try_charge(page, mm, gfp, &memcg, true)) {
 790                 put_page(page);
 791                 count_vm_event(THP_FAULT_FALLBACK);
 792                 return VM_FAULT_FALLBACK;
 793         }
 794
 795         pgtable = pte_alloc_one(mm, haddr);
 796         if (unlikely(!pgtable)) {
 797                 mem_cgroup_cancel_charge(page, memcg, true);
 798                 put_page(page);
 799                 return VM_FAULT_OOM;
 800         }
 801
 802         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 803         /*
 804          * The memory barrier inside __SetPageUptodate makes sure that
 805          * clear_huge_page writes become visible before the set_pmd_at()
 806          * write.
 807          */
 808         __SetPageUptodate(page);
 809
 810         ptl = pmd_lock(mm, pmd);
 811         if (unlikely(!pmd_none(*pmd))) {
 812                 spin_unlock(ptl);
 813                 mem_cgroup_cancel_charge(page, memcg, true);
 814                 put_page(page);
 815                 pte_free(mm, pgtable);
 816         } else {
 817                 pmd_t entry;
 818
 819                 /* Deliver the page fault to userland */
 820                 if (userfaultfd_missing(vma)) {
 821                         int ret;
 822
 823                         spin_unlock(ptl);
 824                         mem_cgroup_cancel_charge(page, memcg, true);
 825                         put_page(page);
 826                         pte_free(mm, pgtable);
 827                         ret = handle_userfault(vma, address, flags,
 828                                                VM_UFFD_MISSING);
 829                         VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 830                         return ret;
 831                 }
 832
 833                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 834                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 835                 page_add_new_anon_rmap(page, vma, haddr, true);
 836                 mem_cgroup_commit_charge(page, memcg, false, true);
 837                 lru_cache_add_active_or_unevictable(page, vma);
 838                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 839                 set_pmd_at(mm, haddr, pmd, entry);
 840                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 841                 atomic_long_inc(&mm->nr_ptes);
 842                 spin_unlock(ptl);
 843                 count_vm_event(THP_FAULT_ALLOC);
 844         }
 845
 846         return 0;
 847 }
 848
 849 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 850 {
 851         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_RECLAIM)) | extra_gfp;
 852 }
 853
 854 /* Caller must hold page table lock. */
 855 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 856                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 857                 struct page *zero_page)
 858 {
 859         pmd_t entry;
 860         if (!pmd_none(*pmd))
 861                 return false;
 862         entry = mk_pmd(zero_page, vma->vm_page_prot);
 863         entry = pmd_mkhuge(entry);
 864         pgtable_trans_huge_deposit(mm, pmd, pgtable);
 865         set_pmd_at(mm, haddr, pmd, entry);
 866         atomic_long_inc(&mm->nr_ptes);
 867         return true;
 868 }
 869
 870 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 871                                unsigned long address, pmd_t *pmd,
 872                                unsigned int flags)
 873 {
 874         gfp_t gfp;
 875         struct page *page;
 876         unsigned long haddr = address & HPAGE_PMD_MASK;
 877
 878         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 879                 return VM_FAULT_FALLBACK;
 880         if (unlikely(anon_vma_prepare(vma)))
 881                 return VM_FAULT_OOM;
 882         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 883                 return VM_FAULT_OOM;
 884         if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm) &&
 885                         transparent_hugepage_use_zero_page()) {
 886                 spinlock_t *ptl;
 887                 pgtable_t pgtable;
 888                 struct page *zero_page;
 889                 bool set;
 890                 int ret;
 891                 pgtable = pte_alloc_one(mm, haddr);
 892                 if (unlikely(!pgtable))
 893                         return VM_FAULT_OOM;
 894                 zero_page = get_huge_zero_page();
 895                 if (unlikely(!zero_page)) {
 896                         pte_free(mm, pgtable);
 897                         count_vm_event(THP_FAULT_FALLBACK);
 898                         return VM_FAULT_FALLBACK;
 899                 }
 900                 ptl = pmd_lock(mm, pmd);
 901                 ret = 0;
 902                 set = false;
 903                 if (pmd_none(*pmd)) {
 904                         if (userfaultfd_missing(vma)) {
 905                                 spin_unlock(ptl);
 906                                 ret = handle_userfault(vma, address, flags,
 907                                                        VM_UFFD_MISSING);
 908                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 909                         } else {
 910                                 set_huge_zero_page(pgtable, mm, vma,
 911                                                    haddr, pmd,
 912                                                    zero_page);
 913                                 spin_unlock(ptl);
 914                                 set = true;
 915                         }
 916                 } else
 917                         spin_unlock(ptl);
 918                 if (!set) {
 919                         pte_free(mm, pgtable);
 920                         put_huge_zero_page();
 921                 }
 922                 return ret;
 923         }
 924         gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
 925         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 926         if (unlikely(!page)) {
 927                 count_vm_event(THP_FAULT_FALLBACK);
 928                 return VM_FAULT_FALLBACK;
 929         }
 930         prep_transhuge_page(page);
 931         return __do_huge_pmd_anonymous_page(mm, vma, address, pmd, page, gfp,
 932                                             flags);
 933 }
 934
 935 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 936                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write)
 937 {
 938         struct mm_struct *mm = vma->vm_mm;
 939         pmd_t entry;
 940         spinlock_t *ptl;
 941
 942         ptl = pmd_lock(mm, pmd);
 943         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 944         if (pfn_t_devmap(pfn))
 945                 entry = pmd_mkdevmap(entry);
 946         if (write) {
 947                 entry = pmd_mkyoung(pmd_mkdirty(entry));
 948                 entry = maybe_pmd_mkwrite(entry, vma);
 949         }
 950         set_pmd_at(mm, addr, pmd, entry);
 951         update_mmu_cache_pmd(vma, addr, pmd);
 952         spin_unlock(ptl);
 953 }
 954
 955 int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 956                         pmd_t *pmd, pfn_t pfn, bool write)
 957 {
 958         pgprot_t pgprot = vma->vm_page_prot;
 959         /*
 960          * If we had pmd_special, we could avoid all these restrictions,
 961          * but we need to be consistent with PTEs and architectures that
 962          * can't support a 'special' bit.
 963          */
 964         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
 965         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 966                                                 (VM_PFNMAP|VM_MIXEDMAP));
 967         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 968         BUG_ON(!pfn_t_devmap(pfn));
 969
 970         if (addr < vma->vm_start || addr >= vma->vm_end)
 971                 return VM_FAULT_SIGBUS;
 972         if (track_pfn_insert(vma, &pgprot, pfn))
 973                 return VM_FAULT_SIGBUS;
 974         insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
 975         return VM_FAULT_NOPAGE;
 976 }
 977
 978 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
 979                 pmd_t *pmd)
 980 {
 981         pmd_t _pmd;
 982
 983         /*
 984          * We should set the dirty bit only for FOLL_WRITE but for now
 985          * the dirty bit in the pmd is meaningless.  And if the dirty
 986          * bit will become meaningful and we'll only set it with
 987          * FOLL_WRITE, an atomic set_bit will be required on the pmd to
 988          * set the young bit, instead of the current set_pmd_at.
 989          */
 990         _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
 991         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
 992                                 pmd, _pmd,  1))
 993                 update_mmu_cache_pmd(vma, addr, pmd);
 994 }
 995
 996 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 997                 pmd_t *pmd, int flags)
 998 {
 999         unsigned long pfn = pmd_pfn(*pmd);
1000         struct mm_struct *mm = vma->vm_mm;
1001         struct dev_pagemap *pgmap;
1002         struct page *page;
1003
1004         assert_spin_locked(pmd_lockptr(mm, pmd));
1005
1006         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1007                 return NULL;
1008
1009         if (pmd_present(*pmd) && pmd_devmap(*pmd))
1010                 /* pass */;
1011         else
1012                 return NULL;
1013
1014         if (flags & FOLL_TOUCH)
1015                 touch_pmd(vma, addr, pmd);
1016
1017         /*
1018          * device mapped pages can only be returned if the
1019          * caller will manage the page reference count.
1020          */
1021         if (!(flags & FOLL_GET))
1022                 return ERR_PTR(-EEXIST);
1023
1024         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1025         pgmap = get_dev_pagemap(pfn, NULL);
1026         if (!pgmap)
1027                 return ERR_PTR(-EFAULT);
1028         page = pfn_to_page(pfn);
1029         get_page(page);
1030         put_dev_pagemap(pgmap);
1031
1032         return page;
1033 }
1034
1035 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1036                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1037                   struct vm_area_struct *vma)
1038 {
1039         spinlock_t *dst_ptl, *src_ptl;
1040         struct page *src_page;
1041         pmd_t pmd;
1042         pgtable_t pgtable;
1043         int ret;
1044
1045         ret = -ENOMEM;
1046         pgtable = pte_alloc_one(dst_mm, addr);
1047         if (unlikely(!pgtable))
1048                 goto out;
1049
1050         dst_ptl = pmd_lock(dst_mm, dst_pmd);
1051         src_ptl = pmd_lockptr(src_mm, src_pmd);
1052         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1053
1054         ret = -EAGAIN;
1055         pmd = *src_pmd;
1056         if (unlikely(!pmd_trans_huge(pmd) && !pmd_devmap(pmd))) {
1057                 pte_free(dst_mm, pgtable);
1058                 goto out_unlock;
1059         }
1060         /*
1061          * When page table lock is held, the huge zero pmd should not be
1062          * under splitting since we don't split the page itself, only pmd to
1063          * a page table.
1064          */
1065         if (is_huge_zero_pmd(pmd)) {
1066                 struct page *zero_page;
1067                 /*
1068                  * get_huge_zero_page() will never allocate a new page here,
1069                  * since we already have a zero page to copy. It just takes a
1070                  * reference.
1071                  */
1072                 zero_page = get_huge_zero_page();
1073                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1074                                 zero_page);
1075                 ret = 0;
1076                 goto out_unlock;
1077         }
1078
1079         if (pmd_trans_huge(pmd)) {
1080                 /* thp accounting separate from pmd_devmap accounting */
1081                 src_page = pmd_page(pmd);
1082                 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1083                 get_page(src_page);
1084                 page_dup_rmap(src_page, true);
1085                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1086                 atomic_long_inc(&dst_mm->nr_ptes);
1087                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1088         }
1089
1090         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1091         pmd = pmd_mkold(pmd_wrprotect(pmd));
1092         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1093
1094         ret = 0;
1095 out_unlock:
1096         spin_unlock(src_ptl);
1097         spin_unlock(dst_ptl);
1098 out:
1099         return ret;
1100 }
1101
1102 void huge_pmd_set_accessed(struct mm_struct *mm,
1103                            struct vm_area_struct *vma,
1104                            unsigned long address,
1105                            pmd_t *pmd, pmd_t orig_pmd,
1106                            int dirty)
1107 {
1108         spinlock_t *ptl;
1109         pmd_t entry;
1110         unsigned long haddr;
1111
1112         ptl = pmd_lock(mm, pmd);
1113         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1114                 goto unlock;
1115
1116         entry = pmd_mkyoung(orig_pmd);
1117         haddr = address & HPAGE_PMD_MASK;
1118         if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
1119                 update_mmu_cache_pmd(vma, address, pmd);
1120
1121 unlock:
1122         spin_unlock(ptl);
1123 }
1124
1125 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1126                                         struct vm_area_struct *vma,
1127                                         unsigned long address,
1128                                         pmd_t *pmd, pmd_t orig_pmd,
1129                                         struct page *page,
1130                                         unsigned long haddr)
1131 {
1132         struct mem_cgroup *memcg;
1133         spinlock_t *ptl;
1134         pgtable_t pgtable;
1135         pmd_t _pmd;
1136         int ret = 0, i;
1137         struct page **pages;
1138         unsigned long mmun_start;       /* For mmu_notifiers */
1139         unsigned long mmun_end;         /* For mmu_notifiers */
1140
1141         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1142                         GFP_KERNEL);
1143         if (unlikely(!pages)) {
1144                 ret |= VM_FAULT_OOM;
1145                 goto out;
1146         }
1147
1148         for (i = 0; i < HPAGE_PMD_NR; i++) {
1149                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1150                                                __GFP_OTHER_NODE,
1151                                                vma, address, page_to_nid(page));
1152                 if (unlikely(!pages[i] ||
1153                              mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL,
1154                                                    &memcg, false))) {
1155                         if (pages[i])
1156                                 put_page(pages[i]);
1157                         while (--i >= 0) {
1158                                 memcg = (void *)page_private(pages[i]);
1159                                 set_page_private(pages[i], 0);
1160                                 mem_cgroup_cancel_charge(pages[i], memcg,
1161                                                 false);
1162                                 put_page(pages[i]);
1163                         }
1164                         kfree(pages);
1165                         ret |= VM_FAULT_OOM;
1166                         goto out;
1167                 }
1168                 set_page_private(pages[i], (unsigned long)memcg);
1169         }
1170
1171         for (i = 0; i < HPAGE_PMD_NR; i++) {
1172                 copy_user_highpage(pages[i], page + i,
1173                                    haddr + PAGE_SIZE * i, vma);
1174                 __SetPageUptodate(pages[i]);
1175                 cond_resched();
1176         }
1177
1178         mmun_start = haddr;
1179         mmun_end   = haddr + HPAGE_PMD_SIZE;
1180         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1181
1182         ptl = pmd_lock(mm, pmd);
1183         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1184                 goto out_free_pages;
1185         VM_BUG_ON_PAGE(!PageHead(page), page);
1186
1187         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1188         /* leave pmd empty until pte is filled */
1189
1190         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1191         pmd_populate(mm, &_pmd, pgtable);
1192
1193         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1194                 pte_t *pte, entry;
1195                 entry = mk_pte(pages[i], vma->vm_page_prot);
1196                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1197                 memcg = (void *)page_private(pages[i]);
1198                 set_page_private(pages[i], 0);
1199                 page_add_new_anon_rmap(pages[i], vma, haddr, false);
1200                 mem_cgroup_commit_charge(pages[i], memcg, false, false);
1201                 lru_cache_add_active_or_unevictable(pages[i], vma);
1202                 pte = pte_offset_map(&_pmd, haddr);
1203                 VM_BUG_ON(!pte_none(*pte));
1204                 set_pte_at(mm, haddr, pte, entry);
1205                 pte_unmap(pte);
1206         }
1207         kfree(pages);
1208
1209         smp_wmb(); /* make pte visible before pmd */
1210         pmd_populate(mm, pmd, pgtable);
1211         page_remove_rmap(page, true);
1212         spin_unlock(ptl);
1213
1214         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1215
1216         ret |= VM_FAULT_WRITE;
1217         put_page(page);
1218
1219 out:
1220         return ret;
1221
1222 out_free_pages:
1223         spin_unlock(ptl);
1224         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1225         for (i = 0; i < HPAGE_PMD_NR; i++) {
1226                 memcg = (void *)page_private(pages[i]);
1227                 set_page_private(pages[i], 0);
1228                 mem_cgroup_cancel_charge(pages[i], memcg, false);
1229                 put_page(pages[i]);
1230         }
1231         kfree(pages);
1232         goto out;
1233 }
1234
1235 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1236                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1237 {
1238         spinlock_t *ptl;
1239         int ret = 0;
1240         struct page *page = NULL, *new_page;
1241         struct mem_cgroup *memcg;
1242         unsigned long haddr;
1243         unsigned long mmun_start;       /* For mmu_notifiers */
1244         unsigned long mmun_end;         /* For mmu_notifiers */
1245         gfp_t huge_gfp;                 /* for allocation and charge */
1246
1247         ptl = pmd_lockptr(mm, pmd);
1248         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1249         haddr = address & HPAGE_PMD_MASK;
1250         if (is_huge_zero_pmd(orig_pmd))
1251                 goto alloc;
1252         spin_lock(ptl);
1253         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1254                 goto out_unlock;
1255
1256         page = pmd_page(orig_pmd);
1257         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1258         /*
1259          * We can only reuse the page if nobody else maps the huge page or it's
1260          * part. We can do it by checking page_mapcount() on each sub-page, but
1261          * it's expensive.
1262          * The cheaper way is to check page_count() to be equal 1: every
1263          * mapcount takes page reference reference, so this way we can
1264          * guarantee, that the PMD is the only mapping.
1265          * This can give false negative if somebody pinned the page, but that's
1266          * fine.
1267          */
1268         if (page_mapcount(page) == 1 && page_count(page) == 1) {
1269                 pmd_t entry;
1270                 entry = pmd_mkyoung(orig_pmd);
1271                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1272                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1273                         update_mmu_cache_pmd(vma, address, pmd);
1274                 ret |= VM_FAULT_WRITE;
1275                 goto out_unlock;
1276         }
1277         get_page(page);
1278         spin_unlock(ptl);
1279 alloc:
1280         if (transparent_hugepage_enabled(vma) &&
1281             !transparent_hugepage_debug_cow()) {
1282                 huge_gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
1283                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1284         } else
1285                 new_page = NULL;
1286
1287         if (likely(new_page)) {
1288                 prep_transhuge_page(new_page);
1289         } else {
1290                 if (!page) {
1291                         split_huge_pmd(vma, pmd, address);
1292                         ret |= VM_FAULT_FALLBACK;
1293                 } else {
1294                         ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1295                                         pmd, orig_pmd, page, haddr);
1296                         if (ret & VM_FAULT_OOM) {
1297                                 split_huge_pmd(vma, pmd, address);
1298                                 ret |= VM_FAULT_FALLBACK;
1299                         }
1300                         put_page(page);
1301                 }
1302                 count_vm_event(THP_FAULT_FALLBACK);
1303                 goto out;
1304         }
1305
1306         if (unlikely(mem_cgroup_try_charge(new_page, mm, huge_gfp, &memcg,
1307                                            true))) {
1308                 put_page(new_page);
1309                 if (page) {
1310                         split_huge_pmd(vma, pmd, address);
1311                         put_page(page);
1312                 } else
1313                         split_huge_pmd(vma, pmd, address);
1314                 ret |= VM_FAULT_FALLBACK;
1315                 count_vm_event(THP_FAULT_FALLBACK);
1316                 goto out;
1317         }
1318
1319         count_vm_event(THP_FAULT_ALLOC);
1320
1321         if (!page)
1322                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1323         else
1324                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1325         __SetPageUptodate(new_page);
1326
1327         mmun_start = haddr;
1328         mmun_end   = haddr + HPAGE_PMD_SIZE;
1329         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1330
1331         spin_lock(ptl);
1332         if (page)
1333                 put_page(page);
1334         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1335                 spin_unlock(ptl);
1336                 mem_cgroup_cancel_charge(new_page, memcg, true);
1337                 put_page(new_page);
1338                 goto out_mn;
1339         } else {
1340                 pmd_t entry;
1341                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1342                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1343                 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1344                 page_add_new_anon_rmap(new_page, vma, haddr, true);
1345                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1346                 lru_cache_add_active_or_unevictable(new_page, vma);
1347                 set_pmd_at(mm, haddr, pmd, entry);
1348                 update_mmu_cache_pmd(vma, address, pmd);
1349                 if (!page) {
1350                         add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1351                         put_huge_zero_page();
1352                 } else {
1353                         VM_BUG_ON_PAGE(!PageHead(page), page);
1354                         page_remove_rmap(page, true);
1355                         put_page(page);
1356                 }
1357                 ret |= VM_FAULT_WRITE;
1358         }
1359         spin_unlock(ptl);
1360 out_mn:
1361         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1362 out:
1363         return ret;
1364 out_unlock:
1365         spin_unlock(ptl);
1366         return ret;
1367 }
1368
1369 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1370                                    unsigned long addr,
1371                                    pmd_t *pmd,
1372                                    unsigned int flags)
1373 {
1374         struct mm_struct *mm = vma->vm_mm;
1375         struct page *page = NULL;
1376
1377         assert_spin_locked(pmd_lockptr(mm, pmd));
1378
1379         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1380                 goto out;
1381
1382         /* Avoid dumping huge zero page */
1383         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1384                 return ERR_PTR(-EFAULT);
1385
1386         /* Full NUMA hinting faults to serialise migration in fault paths */
1387         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1388                 goto out;
1389
1390         page = pmd_page(*pmd);
1391         VM_BUG_ON_PAGE(!PageHead(page), page);
1392         if (flags & FOLL_TOUCH)
1393                 touch_pmd(vma, addr, pmd);
1394         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1395                 /*
1396                  * We don't mlock() pte-mapped THPs. This way we can avoid
1397                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1398                  *
1399                  * In most cases the pmd is the only mapping of the page as we
1400                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1401                  * writable private mappings in populate_vma_page_range().
1402                  *
1403                  * The only scenario when we have the page shared here is if we
1404                  * mlocking read-only mapping shared over fork(). We skip
1405                  * mlocking such pages.
1406                  */
1407                 if (compound_mapcount(page) == 1 && !PageDoubleMap(page) &&
1408                                 page->mapping && trylock_page(page)) {
1409                         lru_add_drain();
1410                         if (page->mapping)
1411                                 mlock_vma_page(page);
1412                         unlock_page(page);
1413                 }
1414         }
1415         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1416         VM_BUG_ON_PAGE(!PageCompound(page), page);
1417         if (flags & FOLL_GET)
1418                 get_page(page);
1419
1420 out:
1421         return page;
1422 }
1423
1424 /* NUMA hinting page fault entry point for trans huge pmds */
1425 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1426                                 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1427 {
1428         spinlock_t *ptl;
1429         struct anon_vma *anon_vma = NULL;
1430         struct page *page;
1431         unsigned long haddr = addr & HPAGE_PMD_MASK;
1432         int page_nid = -1, this_nid = numa_node_id();
1433         int target_nid, last_cpupid = -1;
1434         bool page_locked;
1435         bool migrated = false;
1436         bool was_writable;
1437         int flags = 0;
1438
1439         /* A PROT_NONE fault should not end up here */
1440         BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
1441
1442         ptl = pmd_lock(mm, pmdp);
1443         if (unlikely(!pmd_same(pmd, *pmdp)))
1444                 goto out_unlock;
1445
1446         /*
1447          * If there are potential migrations, wait for completion and retry
1448          * without disrupting NUMA hinting information. Do not relock and
1449          * check_same as the page may no longer be mapped.
1450          */
1451         if (unlikely(pmd_trans_migrating(*pmdp))) {
1452                 page = pmd_page(*pmdp);
1453                 spin_unlock(ptl);
1454                 wait_on_page_locked(page);
1455                 goto out;
1456         }
1457
1458         page = pmd_page(pmd);
1459         BUG_ON(is_huge_zero_page(page));
1460         page_nid = page_to_nid(page);
1461         last_cpupid = page_cpupid_last(page);
1462         count_vm_numa_event(NUMA_HINT_FAULTS);
1463         if (page_nid == this_nid) {
1464                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1465                 flags |= TNF_FAULT_LOCAL;
1466         }
1467
1468         /* See similar comment in do_numa_page for explanation */
1469         if (!(vma->vm_flags & VM_WRITE))
1470                 flags |= TNF_NO_GROUP;
1471
1472         /*
1473          * Acquire the page lock to serialise THP migrations but avoid dropping
1474          * page_table_lock if at all possible
1475          */
1476         page_locked = trylock_page(page);
1477         target_nid = mpol_misplaced(page, vma, haddr);
1478         if (target_nid == -1) {
1479                 /* If the page was locked, there are no parallel migrations */
1480                 if (page_locked)
1481                         goto clear_pmdnuma;
1482         }
1483
1484         /* Migration could have started since the pmd_trans_migrating check */
1485         if (!page_locked) {
1486                 spin_unlock(ptl);
1487                 wait_on_page_locked(page);
1488                 page_nid = -1;
1489                 goto out;
1490         }
1491
1492         /*
1493          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1494          * to serialises splits
1495          */
1496         get_page(page);
1497         spin_unlock(ptl);
1498         anon_vma = page_lock_anon_vma_read(page);
1499
1500         /* Confirm the PMD did not change while page_table_lock was released */
1501         spin_lock(ptl);
1502         if (unlikely(!pmd_same(pmd, *pmdp))) {
1503                 unlock_page(page);
1504                 put_page(page);
1505                 page_nid = -1;
1506                 goto out_unlock;
1507         }
1508
1509         /* Bail if we fail to protect against THP splits for any reason */
1510         if (unlikely(!anon_vma)) {
1511                 put_page(page);
1512                 page_nid = -1;
1513                 goto clear_pmdnuma;
1514         }
1515
1516         /*
1517          * Migrate the THP to the requested node, returns with page unlocked
1518          * and access rights restored.
1519          */
1520         spin_unlock(ptl);
1521         migrated = migrate_misplaced_transhuge_page(mm, vma,
1522                                 pmdp, pmd, addr, page, target_nid);
1523         if (migrated) {
1524                 flags |= TNF_MIGRATED;
1525                 page_nid = target_nid;
1526         } else
1527                 flags |= TNF_MIGRATE_FAIL;
1528
1529         goto out;
1530 clear_pmdnuma:
1531         BUG_ON(!PageLocked(page));
1532         was_writable = pmd_write(pmd);
1533         pmd = pmd_modify(pmd, vma->vm_page_prot);
1534         pmd = pmd_mkyoung(pmd);
1535         if (was_writable)
1536                 pmd = pmd_mkwrite(pmd);
1537         set_pmd_at(mm, haddr, pmdp, pmd);
1538         update_mmu_cache_pmd(vma, addr, pmdp);
1539         unlock_page(page);
1540 out_unlock:
1541         spin_unlock(ptl);
1542
1543 out:
1544         if (anon_vma)
1545                 page_unlock_anon_vma_read(anon_vma);
1546
1547         if (page_nid != -1)
1548                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
1549
1550         return 0;
1551 }
1552
1553 int madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1554                 pmd_t *pmd, unsigned long addr, unsigned long next)
1555
1556 {
1557         spinlock_t *ptl;
1558         pmd_t orig_pmd;
1559         struct page *page;
1560         struct mm_struct *mm = tlb->mm;
1561         int ret = 0;
1562
1563         if (!pmd_trans_huge_lock(pmd, vma, &ptl))
1564                 goto out_unlocked;
1565
1566         orig_pmd = *pmd;
1567         if (is_huge_zero_pmd(orig_pmd)) {
1568                 ret = 1;
1569                 goto out;
1570         }
1571
1572         page = pmd_page(orig_pmd);
1573         /*
1574          * If other processes are mapping this page, we couldn't discard
1575          * the page unless they all do MADV_FREE so let's skip the page.
1576          */
1577         if (page_mapcount(page) != 1)
1578                 goto out;
1579
1580         if (!trylock_page(page))
1581                 goto out;
1582
1583         /*
1584          * If user want to discard part-pages of THP, split it so MADV_FREE
1585          * will deactivate only them.
1586          */
1587         if (next - addr != HPAGE_PMD_SIZE) {
1588                 get_page(page);
1589                 spin_unlock(ptl);
1590                 if (split_huge_page(page)) {
1591                         put_page(page);
1592                         unlock_page(page);
1593                         goto out_unlocked;
1594                 }
1595                 put_page(page);
1596                 unlock_page(page);
1597                 ret = 1;
1598                 goto out_unlocked;
1599         }
1600
1601         if (PageDirty(page))
1602                 ClearPageDirty(page);
1603         unlock_page(page);
1604
1605         if (PageActive(page))
1606                 deactivate_page(page);
1607
1608         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1609                 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1610                         tlb->fullmm);
1611                 orig_pmd = pmd_mkold(orig_pmd);
1612                 orig_pmd = pmd_mkclean(orig_pmd);
1613
1614                 set_pmd_at(mm, addr, pmd, orig_pmd);
1615                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1616         }
1617         ret = 1;
1618 out:
1619         spin_unlock(ptl);
1620 out_unlocked:
1621         return ret;
1622 }
1623
1624 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1625                  pmd_t *pmd, unsigned long addr)
1626 {
1627         pmd_t orig_pmd;
1628         spinlock_t *ptl;
1629
1630         if (!__pmd_trans_huge_lock(pmd, vma, &ptl))
1631                 return 0;
1632         /*
1633          * For architectures like ppc64 we look at deposited pgtable
1634          * when calling pmdp_huge_get_and_clear. So do the
1635          * pgtable_trans_huge_withdraw after finishing pmdp related
1636          * operations.
1637          */
1638         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1639                         tlb->fullmm);
1640         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1641         if (vma_is_dax(vma)) {
1642                 spin_unlock(ptl);
1643                 if (is_huge_zero_pmd(orig_pmd))
1644                         put_huge_zero_page();
1645         } else if (is_huge_zero_pmd(orig_pmd)) {
1646                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1647                 atomic_long_dec(&tlb->mm->nr_ptes);
1648                 spin_unlock(ptl);
1649                 put_huge_zero_page();
1650         } else {
1651                 struct page *page = pmd_page(orig_pmd);
1652                 page_remove_rmap(page, true);
1653                 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1654                 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1655                 VM_BUG_ON_PAGE(!PageHead(page), page);
1656                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1657                 atomic_long_dec(&tlb->mm->nr_ptes);
1658                 spin_unlock(ptl);
1659                 tlb_remove_page(tlb, page);
1660         }
1661         return 1;
1662 }
1663
1664 bool move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1665                   unsigned long old_addr,
1666                   unsigned long new_addr, unsigned long old_end,
1667                   pmd_t *old_pmd, pmd_t *new_pmd)
1668 {
1669         spinlock_t *old_ptl, *new_ptl;
1670         pmd_t pmd;
1671
1672         struct mm_struct *mm = vma->vm_mm;
1673
1674         if ((old_addr & ~HPAGE_PMD_MASK) ||
1675             (new_addr & ~HPAGE_PMD_MASK) ||
1676             old_end - old_addr < HPAGE_PMD_SIZE ||
1677             (new_vma->vm_flags & VM_NOHUGEPAGE))
1678                 return false;
1679
1680         /*
1681          * The destination pmd shouldn't be established, free_pgtables()
1682          * should have release it.
1683          */
1684         if (WARN_ON(!pmd_none(*new_pmd))) {
1685                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1686                 return false;
1687         }
1688
1689         /*
1690          * We don't have to worry about the ordering of src and dst
1691          * ptlocks because exclusive mmap_sem prevents deadlock.
1692          */
1693         if (__pmd_trans_huge_lock(old_pmd, vma, &old_ptl)) {
1694                 new_ptl = pmd_lockptr(mm, new_pmd);
1695                 if (new_ptl != old_ptl)
1696                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1697                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1698                 VM_BUG_ON(!pmd_none(*new_pmd));
1699
1700                 if (pmd_move_must_withdraw(new_ptl, old_ptl)) {
1701                         pgtable_t pgtable;
1702                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1703                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1704                 }
1705                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1706                 if (new_ptl != old_ptl)
1707                         spin_unlock(new_ptl);
1708                 spin_unlock(old_ptl);
1709                 return true;
1710         }
1711         return false;
1712 }
1713
1714 /*
1715  * Returns
1716  *  - 0 if PMD could not be locked
1717  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1718  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1719  */
1720 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1721                 unsigned long addr, pgprot_t newprot, int prot_numa)
1722 {
1723         struct mm_struct *mm = vma->vm_mm;
1724         spinlock_t *ptl;
1725         int ret = 0;
1726
1727         if (__pmd_trans_huge_lock(pmd, vma, &ptl)) {
1728                 pmd_t entry;
1729                 bool preserve_write = prot_numa && pmd_write(*pmd);
1730                 ret = 1;
1731
1732                 /*
1733                  * Avoid trapping faults against the zero page. The read-only
1734                  * data is likely to be read-cached on the local CPU and
1735                  * local/remote hits to the zero page are not interesting.
1736                  */
1737                 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1738                         spin_unlock(ptl);
1739                         return ret;
1740                 }
1741
1742                 if (!prot_numa || !pmd_protnone(*pmd)) {
1743                         entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
1744                         entry = pmd_modify(entry, newprot);
1745                         if (preserve_write)
1746                                 entry = pmd_mkwrite(entry);
1747                         ret = HPAGE_PMD_NR;
1748                         set_pmd_at(mm, addr, pmd, entry);
1749                         BUG_ON(!preserve_write && pmd_write(entry));
1750                 }
1751                 spin_unlock(ptl);
1752         }
1753
1754         return ret;
1755 }
1756
1757 /*
1758  * Returns true if a given pmd maps a thp, false otherwise.
1759  *
1760  * Note that if it returns true, this routine returns without unlocking page
1761  * table lock. So callers must unlock it.
1762  */
1763 bool __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1764                 spinlock_t **ptl)
1765 {
1766         *ptl = pmd_lock(vma->vm_mm, pmd);
1767         if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd)))
1768                 return true;
1769         spin_unlock(*ptl);
1770         return false;
1771 }
1772
1773 #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
1774
1775 int hugepage_madvise(struct vm_area_struct *vma,
1776                      unsigned long *vm_flags, int advice)
1777 {
1778         switch (advice) {
1779         case MADV_HUGEPAGE:
1780 #ifdef CONFIG_S390
1781                 /*
1782                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1783                  * can't handle this properly after s390_enable_sie, so we simply
1784                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
1785                  */
1786                 if (mm_has_pgste(vma->vm_mm))
1787                         return 0;
1788 #endif
1789                 /*
1790                  * Be somewhat over-protective like KSM for now!
1791                  */
1792                 if (*vm_flags & VM_NO_THP)
1793                         return -EINVAL;
1794                 *vm_flags &= ~VM_NOHUGEPAGE;
1795                 *vm_flags |= VM_HUGEPAGE;
1796                 /*
1797                  * If the vma become good for khugepaged to scan,
1798                  * register it here without waiting a page fault that
1799                  * may not happen any time soon.
1800                  */
1801                 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
1802                         return -ENOMEM;
1803                 break;
1804         case MADV_NOHUGEPAGE:
1805                 /*
1806                  * Be somewhat over-protective like KSM for now!
1807                  */
1808                 if (*vm_flags & VM_NO_THP)
1809                         return -EINVAL;
1810                 *vm_flags &= ~VM_HUGEPAGE;
1811                 *vm_flags |= VM_NOHUGEPAGE;
1812                 /*
1813                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1814                  * this vma even if we leave the mm registered in khugepaged if
1815                  * it got registered before VM_NOHUGEPAGE was set.
1816                  */
1817                 break;
1818         }
1819
1820         return 0;
1821 }
1822
1823 static int __init khugepaged_slab_init(void)
1824 {
1825         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1826                                           sizeof(struct mm_slot),
1827                                           __alignof__(struct mm_slot), 0, NULL);
1828         if (!mm_slot_cache)
1829                 return -ENOMEM;
1830
1831         return 0;
1832 }
1833
1834 static void __init khugepaged_slab_exit(void)
1835 {
1836         kmem_cache_destroy(mm_slot_cache);
1837 }
1838
1839 static inline struct mm_slot *alloc_mm_slot(void)
1840 {
1841         if (!mm_slot_cache)     /* initialization failed */
1842                 return NULL;
1843         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1844 }
1845
1846 static inline void free_mm_slot(struct mm_slot *mm_slot)
1847 {
1848         kmem_cache_free(mm_slot_cache, mm_slot);
1849 }
1850
1851 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1852 {
1853         struct mm_slot *mm_slot;
1854
1855         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
1856                 if (mm == mm_slot->mm)
1857                         return mm_slot;
1858
1859         return NULL;
1860 }
1861
1862 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1863                                     struct mm_slot *mm_slot)
1864 {
1865         mm_slot->mm = mm;
1866         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
1867 }
1868
1869 static inline int khugepaged_test_exit(struct mm_struct *mm)
1870 {
1871         return atomic_read(&mm->mm_users) == 0;
1872 }
1873
1874 int __khugepaged_enter(struct mm_struct *mm)
1875 {
1876         struct mm_slot *mm_slot;
1877         int wakeup;
1878
1879         mm_slot = alloc_mm_slot();
1880         if (!mm_slot)
1881                 return -ENOMEM;
1882
1883         /* __khugepaged_exit() must not run from under us */
1884         VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
1885         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1886                 free_mm_slot(mm_slot);
1887                 return 0;
1888         }
1889
1890         spin_lock(&khugepaged_mm_lock);
1891         insert_to_mm_slots_hash(mm, mm_slot);
1892         /*
1893          * Insert just behind the scanning cursor, to let the area settle
1894          * down a little.
1895          */
1896         wakeup = list_empty(&khugepaged_scan.mm_head);
1897         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1898         spin_unlock(&khugepaged_mm_lock);
1899
1900         atomic_inc(&mm->mm_count);
1901         if (wakeup)
1902                 wake_up_interruptible(&khugepaged_wait);
1903
1904         return 0;
1905 }
1906
1907 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
1908                                unsigned long vm_flags)
1909 {
1910         unsigned long hstart, hend;
1911         if (!vma->anon_vma)
1912                 /*
1913                  * Not yet faulted in so we will register later in the
1914                  * page fault if needed.
1915                  */
1916                 return 0;
1917         if (vma->vm_ops)
1918                 /* khugepaged not yet working on file or special mappings */
1919                 return 0;
1920         VM_BUG_ON_VMA(vm_flags & VM_NO_THP, vma);
1921         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1922         hend = vma->vm_end & HPAGE_PMD_MASK;
1923         if (hstart < hend)
1924                 return khugepaged_enter(vma, vm_flags);
1925         return 0;
1926 }
1927
1928 void __khugepaged_exit(struct mm_struct *mm)
1929 {
1930         struct mm_slot *mm_slot;
1931         int free = 0;
1932
1933         spin_lock(&khugepaged_mm_lock);
1934         mm_slot = get_mm_slot(mm);
1935         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1936                 hash_del(&mm_slot->hash);
1937                 list_del(&mm_slot->mm_node);
1938                 free = 1;
1939         }
1940         spin_unlock(&khugepaged_mm_lock);
1941
1942         if (free) {
1943                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1944                 free_mm_slot(mm_slot);
1945                 mmdrop(mm);
1946         } else if (mm_slot) {
1947                 /*
1948                  * This is required to serialize against
1949                  * khugepaged_test_exit() (which is guaranteed to run
1950                  * under mmap sem read mode). Stop here (after we
1951                  * return all pagetables will be destroyed) until
1952                  * khugepaged has finished working on the pagetables
1953                  * under the mmap_sem.
1954                  */
1955                 down_write(&mm->mmap_sem);
1956                 up_write(&mm->mmap_sem);
1957         }
1958 }
1959
1960 static void release_pte_page(struct page *page)
1961 {
1962         /* 0 stands for page_is_file_cache(page) == false */
1963         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1964         unlock_page(page);
1965         putback_lru_page(page);
1966 }
1967
1968 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1969 {
1970         while (--_pte >= pte) {
1971                 pte_t pteval = *_pte;
1972                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
1973                         release_pte_page(pte_page(pteval));
1974         }
1975 }
1976
1977 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1978                                         unsigned long address,
1979                                         pte_t *pte)
1980 {
1981         struct page *page = NULL;
1982         pte_t *_pte;
1983         int none_or_zero = 0, result = 0;
1984         bool referenced = false, writable = false;
1985
1986         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1987              _pte++, address += PAGE_SIZE) {
1988                 pte_t pteval = *_pte;
1989                 if (pte_none(pteval) || (pte_present(pteval) &&
1990                                 is_zero_pfn(pte_pfn(pteval)))) {
1991                         if (!userfaultfd_armed(vma) &&
1992                             ++none_or_zero <= khugepaged_max_ptes_none) {
1993                                 continue;
1994                         } else {
1995                                 result = SCAN_EXCEED_NONE_PTE;
1996                                 goto out;
1997                         }
1998                 }
1999                 if (!pte_present(pteval)) {
2000                         result = SCAN_PTE_NON_PRESENT;
2001                         goto out;
2002                 }
2003                 page = vm_normal_page(vma, address, pteval);
2004                 if (unlikely(!page)) {
2005                         result = SCAN_PAGE_NULL;
2006                         goto out;
2007                 }
2008
2009                 VM_BUG_ON_PAGE(PageCompound(page), page);
2010                 VM_BUG_ON_PAGE(!PageAnon(page), page);
2011                 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2012
2013                 /*
2014                  * We can do it before isolate_lru_page because the
2015                  * page can't be freed from under us. NOTE: PG_lock
2016                  * is needed to serialize against split_huge_page
2017                  * when invoked from the VM.
2018                  */
2019                 if (!trylock_page(page)) {
2020                         result = SCAN_PAGE_LOCK;
2021                         goto out;
2022                 }
2023
2024                 /*
2025                  * cannot use mapcount: can't collapse if there's a gup pin.
2026                  * The page must only be referenced by the scanned process
2027                  * and page swap cache.
2028                  */
2029                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2030                         unlock_page(page);
2031                         result = SCAN_PAGE_COUNT;
2032                         goto out;
2033                 }
2034                 if (pte_write(pteval)) {
2035                         writable = true;
2036                 } else {
2037                         if (PageSwapCache(page) && !reuse_swap_page(page)) {
2038                                 unlock_page(page);
2039                                 result = SCAN_SWAP_CACHE_PAGE;
2040                                 goto out;
2041                         }
2042                         /*
2043                          * Page is not in the swap cache. It can be collapsed
2044                          * into a THP.
2045                          */
2046                 }
2047
2048                 /*
2049                  * Isolate the page to avoid collapsing an hugepage
2050                  * currently in use by the VM.
2051                  */
2052                 if (isolate_lru_page(page)) {
2053                         unlock_page(page);
2054                         result = SCAN_DEL_PAGE_LRU;
2055                         goto out;
2056                 }
2057                 /* 0 stands for page_is_file_cache(page) == false */
2058                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2059                 VM_BUG_ON_PAGE(!PageLocked(page), page);
2060                 VM_BUG_ON_PAGE(PageLRU(page), page);
2061
2062                 /* If there is no mapped pte young don't collapse the page */
2063                 if (pte_young(pteval) ||
2064                     page_is_young(page) || PageReferenced(page) ||
2065                     mmu_notifier_test_young(vma->vm_mm, address))
2066                         referenced = true;
2067         }
2068         if (likely(writable)) {
2069                 if (likely(referenced)) {
2070                         result = SCAN_SUCCEED;
2071                         trace_mm_collapse_huge_page_isolate(page_to_pfn(page), none_or_zero,
2072                                                             referenced, writable, result);
2073                         return 1;
2074                 }
2075         } else {
2076                 result = SCAN_PAGE_RO;
2077         }
2078
2079 out:
2080         release_pte_pages(pte, _pte);
2081         trace_mm_collapse_huge_page_isolate(page_to_pfn(page), none_or_zero,
2082                                             referenced, writable, result);
2083         return 0;
2084 }
2085
2086 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2087                                       struct vm_area_struct *vma,
2088                                       unsigned long address,
2089                                       spinlock_t *ptl)
2090 {
2091         pte_t *_pte;
2092         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2093                 pte_t pteval = *_pte;
2094                 struct page *src_page;
2095
2096                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2097                         clear_user_highpage(page, address);
2098                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2099                         if (is_zero_pfn(pte_pfn(pteval))) {
2100                                 /*
2101                                  * ptl mostly unnecessary.
2102                                  */
2103                                 spin_lock(ptl);
2104                                 /*
2105                                  * paravirt calls inside pte_clear here are
2106                                  * superfluous.
2107                                  */
2108                                 pte_clear(vma->vm_mm, address, _pte);
2109                                 spin_unlock(ptl);
2110                         }
2111                 } else {
2112                         src_page = pte_page(pteval);
2113                         copy_user_highpage(page, src_page, address, vma);
2114                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
2115                         release_pte_page(src_page);
2116                         /*
2117                          * ptl mostly unnecessary, but preempt has to
2118                          * be disabled to update the per-cpu stats
2119                          * inside page_remove_rmap().
2120                          */
2121                         spin_lock(ptl);
2122                         /*
2123                          * paravirt calls inside pte_clear here are
2124                          * superfluous.
2125                          */
2126                         pte_clear(vma->vm_mm, address, _pte);
2127                         page_remove_rmap(src_page, false);
2128                         spin_unlock(ptl);
2129                         free_page_and_swap_cache(src_page);
2130                 }
2131
2132                 address += PAGE_SIZE;
2133                 page++;
2134         }
2135 }
2136
2137 static void khugepaged_alloc_sleep(void)
2138 {
2139         DEFINE_WAIT(wait);
2140
2141         add_wait_queue(&khugepaged_wait, &wait);
2142         freezable_schedule_timeout_interruptible(
2143                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2144         remove_wait_queue(&khugepaged_wait, &wait);
2145 }
2146
2147 static int khugepaged_node_load[MAX_NUMNODES];
2148
2149 static bool khugepaged_scan_abort(int nid)
2150 {
2151         int i;
2152
2153         /*
2154          * If zone_reclaim_mode is disabled, then no extra effort is made to
2155          * allocate memory locally.
2156          */
2157         if (!zone_reclaim_mode)
2158                 return false;
2159
2160         /* If there is a count for this node already, it must be acceptable */
2161         if (khugepaged_node_load[nid])
2162                 return false;
2163
2164         for (i = 0; i < MAX_NUMNODES; i++) {
2165                 if (!khugepaged_node_load[i])
2166                         continue;
2167                 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2168                         return true;
2169         }
2170         return false;
2171 }
2172
2173 #ifdef CONFIG_NUMA
2174 static int khugepaged_find_target_node(void)
2175 {
2176         static int last_khugepaged_target_node = NUMA_NO_NODE;
2177         int nid, target_node = 0, max_value = 0;
2178
2179         /* find first node with max normal pages hit */
2180         for (nid = 0; nid < MAX_NUMNODES; nid++)
2181                 if (khugepaged_node_load[nid] > max_value) {
2182                         max_value = khugepaged_node_load[nid];
2183                         target_node = nid;
2184                 }
2185
2186         /* do some balance if several nodes have the same hit record */
2187         if (target_node <= last_khugepaged_target_node)
2188                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2189                                 nid++)
2190                         if (max_value == khugepaged_node_load[nid]) {
2191                                 target_node = nid;
2192                                 break;
2193                         }
2194
2195         last_khugepaged_target_node = target_node;
2196         return target_node;
2197 }
2198
2199 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2200 {
2201         if (IS_ERR(*hpage)) {
2202                 if (!*wait)
2203                         return false;
2204
2205                 *wait = false;
2206                 *hpage = NULL;
2207                 khugepaged_alloc_sleep();
2208         } else if (*hpage) {
2209                 put_page(*hpage);
2210                 *hpage = NULL;
2211         }
2212
2213         return true;
2214 }
2215
2216 static struct page *
2217 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2218                        unsigned long address, int node)
2219 {
2220         VM_BUG_ON_PAGE(*hpage, *hpage);
2221
2222         /*
2223          * Before allocating the hugepage, release the mmap_sem read lock.
2224          * The allocation can take potentially a long time if it involves
2225          * sync compaction, and we do not need to hold the mmap_sem during
2226          * that. We will recheck the vma after taking it again in write mode.
2227          */
2228         up_read(&mm->mmap_sem);
2229
2230         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
2231         if (unlikely(!*hpage)) {
2232                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2233                 *hpage = ERR_PTR(-ENOMEM);
2234                 return NULL;
2235         }
2236
2237         prep_transhuge_page(*hpage);
2238         count_vm_event(THP_COLLAPSE_ALLOC);
2239         return *hpage;
2240 }
2241 #else
2242 static int khugepaged_find_target_node(void)
2243 {
2244         return 0;
2245 }
2246
2247 static inline struct page *alloc_hugepage(int defrag)
2248 {
2249         struct page *page;
2250
2251         page = alloc_pages(alloc_hugepage_gfpmask(defrag, 0), HPAGE_PMD_ORDER);
2252         if (page)
2253                 prep_transhuge_page(page);
2254         return page;
2255 }
2256
2257 static struct page *khugepaged_alloc_hugepage(bool *wait)
2258 {
2259         struct page *hpage;
2260
2261         do {
2262                 hpage = alloc_hugepage(khugepaged_defrag());
2263                 if (!hpage) {
2264                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2265                         if (!*wait)
2266                                 return NULL;
2267
2268                         *wait = false;
2269                         khugepaged_alloc_sleep();
2270                 } else
2271                         count_vm_event(THP_COLLAPSE_ALLOC);
2272         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2273
2274         return hpage;
2275 }
2276
2277 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2278 {
2279         if (!*hpage)
2280                 *hpage = khugepaged_alloc_hugepage(wait);
2281
2282         if (unlikely(!*hpage))
2283                 return false;
2284
2285         return true;
2286 }
2287
2288 static struct page *
2289 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2290                        unsigned long address, int node)
2291 {
2292         up_read(&mm->mmap_sem);
2293         VM_BUG_ON(!*hpage);
2294
2295         return  *hpage;
2296 }
2297 #endif
2298
2299 static bool hugepage_vma_check(struct vm_area_struct *vma)
2300 {
2301         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2302             (vma->vm_flags & VM_NOHUGEPAGE))
2303                 return false;
2304         if (!vma->anon_vma || vma->vm_ops)
2305                 return false;
2306         if (is_vma_temporary_stack(vma))
2307                 return false;
2308         VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma);
2309         return true;
2310 }
2311
2312 static void collapse_huge_page(struct mm_struct *mm,
2313                                    unsigned long address,
2314                                    struct page **hpage,
2315                                    struct vm_area_struct *vma,
2316                                    int node)
2317 {
2318         pmd_t *pmd, _pmd;
2319         pte_t *pte;
2320         pgtable_t pgtable;
2321         struct page *new_page;
2322         spinlock_t *pmd_ptl, *pte_ptl;
2323         int isolated = 0, result = 0;
2324         unsigned long hstart, hend;
2325         struct mem_cgroup *memcg;
2326         unsigned long mmun_start;       /* For mmu_notifiers */
2327         unsigned long mmun_end;         /* For mmu_notifiers */
2328         gfp_t gfp;
2329
2330         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2331
2332         /* Only allocate from the target node */
2333         gfp = alloc_hugepage_gfpmask(khugepaged_defrag(), __GFP_OTHER_NODE) |
2334                 __GFP_THISNODE;
2335
2336         /* release the mmap_sem read lock. */
2337         new_page = khugepaged_alloc_page(hpage, gfp, mm, address, node);
2338         if (!new_page) {
2339                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
2340                 goto out_nolock;
2341         }
2342
2343         if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
2344                 result = SCAN_CGROUP_CHARGE_FAIL;
2345                 goto out_nolock;
2346         }
2347
2348         /*
2349          * Prevent all access to pagetables with the exception of
2350          * gup_fast later hanlded by the ptep_clear_flush and the VM
2351          * handled by the anon_vma lock + PG_lock.
2352          */
2353         down_write(&mm->mmap_sem);
2354         if (unlikely(khugepaged_test_exit(mm))) {
2355                 result = SCAN_ANY_PROCESS;
2356                 goto out;
2357         }
2358
2359         vma = find_vma(mm, address);
2360         if (!vma) {
2361                 result = SCAN_VMA_NULL;
2362                 goto out;
2363         }
2364         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2365         hend = vma->vm_end & HPAGE_PMD_MASK;
2366         if (address < hstart || address + HPAGE_PMD_SIZE > hend) {
2367                 result = SCAN_ADDRESS_RANGE;
2368                 goto out;
2369         }
2370         if (!hugepage_vma_check(vma)) {
2371                 result = SCAN_VMA_CHECK;
2372                 goto out;
2373         }
2374         pmd = mm_find_pmd(mm, address);
2375         if (!pmd) {
2376                 result = SCAN_PMD_NULL;
2377                 goto out;
2378         }
2379
2380         anon_vma_lock_write(vma->anon_vma);
2381
2382         pte = pte_offset_map(pmd, address);
2383         pte_ptl = pte_lockptr(mm, pmd);
2384
2385         mmun_start = address;
2386         mmun_end   = address + HPAGE_PMD_SIZE;
2387         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2388         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2389         /*
2390          * After this gup_fast can't run anymore. This also removes
2391          * any huge TLB entry from the CPU so we won't allow
2392          * huge and small TLB entries for the same virtual address
2393          * to avoid the risk of CPU bugs in that area.
2394          */
2395         _pmd = pmdp_collapse_flush(vma, address, pmd);
2396         spin_unlock(pmd_ptl);
2397         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2398
2399         spin_lock(pte_ptl);
2400         isolated = __collapse_huge_page_isolate(vma, address, pte);
2401         spin_unlock(pte_ptl);
2402
2403         if (unlikely(!isolated)) {
2404                 pte_unmap(pte);
2405                 spin_lock(pmd_ptl);
2406                 BUG_ON(!pmd_none(*pmd));
2407                 /*
2408                  * We can only use set_pmd_at when establishing
2409                  * hugepmds and never for establishing regular pmds that
2410                  * points to regular pagetables. Use pmd_populate for that
2411                  */
2412                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2413                 spin_unlock(pmd_ptl);
2414                 anon_vma_unlock_write(vma->anon_vma);
2415                 result = SCAN_FAIL;
2416                 goto out;
2417         }
2418
2419         /*
2420          * All pages are isolated and locked so anon_vma rmap
2421          * can't run anymore.
2422          */
2423         anon_vma_unlock_write(vma->anon_vma);
2424
2425         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2426         pte_unmap(pte);
2427         __SetPageUptodate(new_page);
2428         pgtable = pmd_pgtable(_pmd);
2429
2430         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2431         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2432
2433         /*
2434          * spin_lock() below is not the equivalent of smp_wmb(), so
2435          * this is needed to avoid the copy_huge_page writes to become
2436          * visible after the set_pmd_at() write.
2437          */
2438         smp_wmb();
2439
2440         spin_lock(pmd_ptl);
2441         BUG_ON(!pmd_none(*pmd));
2442         page_add_new_anon_rmap(new_page, vma, address, true);
2443         mem_cgroup_commit_charge(new_page, memcg, false, true);
2444         lru_cache_add_active_or_unevictable(new_page, vma);
2445         pgtable_trans_huge_deposit(mm, pmd, pgtable);
2446         set_pmd_at(mm, address, pmd, _pmd);
2447         update_mmu_cache_pmd(vma, address, pmd);
2448         spin_unlock(pmd_ptl);
2449
2450         *hpage = NULL;
2451
2452         khugepaged_pages_collapsed++;
2453         result = SCAN_SUCCEED;
2454 out_up_write:
2455         up_write(&mm->mmap_sem);
2456         trace_mm_collapse_huge_page(mm, isolated, result);
2457         return;
2458
2459 out_nolock:
2460         trace_mm_collapse_huge_page(mm, isolated, result);
2461         return;
2462 out:
2463         mem_cgroup_cancel_charge(new_page, memcg, true);
2464         goto out_up_write;
2465 }
2466
2467 static int khugepaged_scan_pmd(struct mm_struct *mm,
2468                                struct vm_area_struct *vma,
2469                                unsigned long address,
2470                                struct page **hpage)
2471 {
2472         pmd_t *pmd;
2473         pte_t *pte, *_pte;
2474         int ret = 0, none_or_zero = 0, result = 0;
2475         struct page *page = NULL;
2476         unsigned long _address;
2477         spinlock_t *ptl;
2478         int node = NUMA_NO_NODE;
2479         bool writable = false, referenced = false;
2480
2481         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2482
2483         pmd = mm_find_pmd(mm, address);
2484         if (!pmd) {
2485                 result = SCAN_PMD_NULL;
2486                 goto out;
2487         }
2488
2489         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2490         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2491         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2492              _pte++, _address += PAGE_SIZE) {
2493                 pte_t pteval = *_pte;
2494                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2495                         if (!userfaultfd_armed(vma) &&
2496                             ++none_or_zero <= khugepaged_max_ptes_none) {
2497                                 continue;
2498                         } else {
2499                                 result = SCAN_EXCEED_NONE_PTE;
2500                                 goto out_unmap;
2501                         }
2502                 }
2503                 if (!pte_present(pteval)) {
2504                         result = SCAN_PTE_NON_PRESENT;
2505                         goto out_unmap;
2506                 }
2507                 if (pte_write(pteval))
2508                         writable = true;
2509
2510                 page = vm_normal_page(vma, _address, pteval);
2511                 if (unlikely(!page)) {
2512                         result = SCAN_PAGE_NULL;
2513                         goto out_unmap;
2514                 }
2515
2516                 /* TODO: teach khugepaged to collapse THP mapped with pte */
2517                 if (PageCompound(page)) {
2518                         result = SCAN_PAGE_COMPOUND;
2519                         goto out_unmap;
2520                 }
2521
2522                 /*
2523                  * Record which node the original page is from and save this
2524                  * information to khugepaged_node_load[].
2525                  * Khupaged will allocate hugepage from the node has the max
2526                  * hit record.
2527                  */
2528                 node = page_to_nid(page);
2529                 if (khugepaged_scan_abort(node)) {
2530                         result = SCAN_SCAN_ABORT;
2531                         goto out_unmap;
2532                 }
2533                 khugepaged_node_load[node]++;
2534                 if (!PageLRU(page)) {
2535                         result = SCAN_SCAN_ABORT;
2536                         goto out_unmap;
2537                 }
2538                 if (PageLocked(page)) {
2539                         result = SCAN_PAGE_LOCK;
2540                         goto out_unmap;
2541                 }
2542                 if (!PageAnon(page)) {
2543                         result = SCAN_PAGE_ANON;
2544                         goto out_unmap;
2545                 }
2546
2547                 /*
2548                  * cannot use mapcount: can't collapse if there's a gup pin.
2549                  * The page must only be referenced by the scanned process
2550                  * and page swap cache.
2551                  */
2552                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2553                         result = SCAN_PAGE_COUNT;
2554                         goto out_unmap;
2555                 }
2556                 if (pte_young(pteval) ||
2557                     page_is_young(page) || PageReferenced(page) ||
2558                     mmu_notifier_test_young(vma->vm_mm, address))
2559                         referenced = true;
2560         }
2561         if (writable) {
2562                 if (referenced) {
2563                         result = SCAN_SUCCEED;
2564                         ret = 1;
2565                 } else {
2566                         result = SCAN_NO_REFERENCED_PAGE;
2567                 }
2568         } else {
2569                 result = SCAN_PAGE_RO;
2570         }
2571 out_unmap:
2572         pte_unmap_unlock(pte, ptl);
2573         if (ret) {
2574                 node = khugepaged_find_target_node();
2575                 /* collapse_huge_page will return with the mmap_sem released */
2576                 collapse_huge_page(mm, address, hpage, vma, node);
2577         }
2578 out:
2579         trace_mm_khugepaged_scan_pmd(mm, page_to_pfn(page), writable, referenced,
2580                                      none_or_zero, result);
2581         return ret;
2582 }
2583
2584 static void collect_mm_slot(struct mm_slot *mm_slot)
2585 {
2586         struct mm_struct *mm = mm_slot->mm;
2587
2588         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2589
2590         if (khugepaged_test_exit(mm)) {
2591                 /* free mm_slot */
2592                 hash_del(&mm_slot->hash);
2593                 list_del(&mm_slot->mm_node);
2594
2595                 /*
2596                  * Not strictly needed because the mm exited already.
2597                  *
2598                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2599                  */
2600
2601                 /* khugepaged_mm_lock actually not necessary for the below */
2602                 free_mm_slot(mm_slot);
2603                 mmdrop(mm);
2604         }
2605 }
2606
2607 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2608                                             struct page **hpage)
2609         __releases(&khugepaged_mm_lock)
2610         __acquires(&khugepaged_mm_lock)
2611 {
2612         struct mm_slot *mm_slot;
2613         struct mm_struct *mm;
2614         struct vm_area_struct *vma;
2615         int progress = 0;
2616
2617         VM_BUG_ON(!pages);
2618         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2619
2620         if (khugepaged_scan.mm_slot)
2621                 mm_slot = khugepaged_scan.mm_slot;
2622         else {
2623                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2624                                      struct mm_slot, mm_node);
2625                 khugepaged_scan.address = 0;
2626                 khugepaged_scan.mm_slot = mm_slot;
2627         }
2628         spin_unlock(&khugepaged_mm_lock);
2629
2630         mm = mm_slot->mm;
2631         down_read(&mm->mmap_sem);
2632         if (unlikely(khugepaged_test_exit(mm)))
2633                 vma = NULL;
2634         else
2635                 vma = find_vma(mm, khugepaged_scan.address);
2636
2637         progress++;
2638         for (; vma; vma = vma->vm_next) {
2639                 unsigned long hstart, hend;
2640
2641                 cond_resched();
2642                 if (unlikely(khugepaged_test_exit(mm))) {
2643                         progress++;
2644                         break;
2645                 }
2646                 if (!hugepage_vma_check(vma)) {
2647 skip:
2648                         progress++;
2649                         continue;
2650                 }
2651                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2652                 hend = vma->vm_end & HPAGE_PMD_MASK;
2653                 if (hstart >= hend)
2654                         goto skip;
2655                 if (khugepaged_scan.address > hend)
2656                         goto skip;
2657                 if (khugepaged_scan.address < hstart)
2658                         khugepaged_scan.address = hstart;
2659                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2660
2661                 while (khugepaged_scan.address < hend) {
2662                         int ret;
2663                         cond_resched();
2664                         if (unlikely(khugepaged_test_exit(mm)))
2665                                 goto breakouterloop;
2666
2667                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2668                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2669                                   hend);
2670                         ret = khugepaged_scan_pmd(mm, vma,
2671                                                   khugepaged_scan.address,
2672                                                   hpage);
2673                         /* move to next address */
2674                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2675                         progress += HPAGE_PMD_NR;
2676                         if (ret)
2677                                 /* we released mmap_sem so break loop */
2678                                 goto breakouterloop_mmap_sem;
2679                         if (progress >= pages)
2680                                 goto breakouterloop;
2681                 }
2682         }
2683 breakouterloop:
2684         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2685 breakouterloop_mmap_sem:
2686
2687         spin_lock(&khugepaged_mm_lock);
2688         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2689         /*
2690          * Release the current mm_slot if this mm is about to die, or
2691          * if we scanned all vmas of this mm.
2692          */
2693         if (khugepaged_test_exit(mm) || !vma) {
2694                 /*
2695                  * Make sure that if mm_users is reaching zero while
2696                  * khugepaged runs here, khugepaged_exit will find
2697                  * mm_slot not pointing to the exiting mm.
2698                  */
2699                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2700                         khugepaged_scan.mm_slot = list_entry(
2701                                 mm_slot->mm_node.next,
2702                                 struct mm_slot, mm_node);
2703                         khugepaged_scan.address = 0;
2704                 } else {
2705                         khugepaged_scan.mm_slot = NULL;
2706                         khugepaged_full_scans++;
2707                 }
2708
2709                 collect_mm_slot(mm_slot);
2710         }
2711
2712         return progress;
2713 }
2714
2715 static int khugepaged_has_work(void)
2716 {
2717         return !list_empty(&khugepaged_scan.mm_head) &&
2718                 khugepaged_enabled();
2719 }
2720
2721 static int khugepaged_wait_event(void)
2722 {
2723         return !list_empty(&khugepaged_scan.mm_head) ||
2724                 kthread_should_stop();
2725 }
2726
2727 static void khugepaged_do_scan(void)
2728 {
2729         struct page *hpage = NULL;
2730         unsigned int progress = 0, pass_through_head = 0;
2731         unsigned int pages = khugepaged_pages_to_scan;
2732         bool wait = true;
2733
2734         barrier(); /* write khugepaged_pages_to_scan to local stack */
2735
2736         while (progress < pages) {
2737                 if (!khugepaged_prealloc_page(&hpage, &wait))
2738                         break;
2739
2740                 cond_resched();
2741
2742                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2743                         break;
2744
2745                 spin_lock(&khugepaged_mm_lock);
2746                 if (!khugepaged_scan.mm_slot)
2747                         pass_through_head++;
2748                 if (khugepaged_has_work() &&
2749                     pass_through_head < 2)
2750                         progress += khugepaged_scan_mm_slot(pages - progress,
2751                                                             &hpage);
2752                 else
2753                         progress = pages;
2754                 spin_unlock(&khugepaged_mm_lock);
2755         }
2756
2757         if (!IS_ERR_OR_NULL(hpage))
2758                 put_page(hpage);
2759 }
2760
2761 static void khugepaged_wait_work(void)
2762 {
2763         if (khugepaged_has_work()) {
2764                 if (!khugepaged_scan_sleep_millisecs)
2765                         return;
2766
2767                 wait_event_freezable_timeout(khugepaged_wait,
2768                                              kthread_should_stop(),
2769                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2770                 return;
2771         }
2772
2773         if (khugepaged_enabled())
2774                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2775 }
2776
2777 static int khugepaged(void *none)
2778 {
2779         struct mm_slot *mm_slot;
2780
2781         set_freezable();
2782         set_user_nice(current, MAX_NICE);
2783
2784         while (!kthread_should_stop()) {
2785                 khugepaged_do_scan();
2786                 khugepaged_wait_work();
2787         }
2788
2789         spin_lock(&khugepaged_mm_lock);
2790         mm_slot = khugepaged_scan.mm_slot;
2791         khugepaged_scan.mm_slot = NULL;
2792         if (mm_slot)
2793                 collect_mm_slot(mm_slot);
2794         spin_unlock(&khugepaged_mm_lock);
2795         return 0;
2796 }
2797
2798 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2799                 unsigned long haddr, pmd_t *pmd)
2800 {
2801         struct mm_struct *mm = vma->vm_mm;
2802         pgtable_t pgtable;
2803         pmd_t _pmd;
2804         int i;
2805
2806         /* leave pmd empty until pte is filled */
2807         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2808
2809         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2810         pmd_populate(mm, &_pmd, pgtable);
2811
2812         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2813                 pte_t *pte, entry;
2814                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2815                 entry = pte_mkspecial(entry);
2816                 pte = pte_offset_map(&_pmd, haddr);
2817                 VM_BUG_ON(!pte_none(*pte));
2818                 set_pte_at(mm, haddr, pte, entry);
2819                 pte_unmap(pte);
2820         }
2821         smp_wmb(); /* make pte visible before pmd */
2822         pmd_populate(mm, pmd, pgtable);
2823         put_huge_zero_page();
2824 }
2825
2826 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2827                 unsigned long haddr, bool freeze)
2828 {
2829         struct mm_struct *mm = vma->vm_mm;
2830         struct page *page;
2831         pgtable_t pgtable;
2832         pmd_t _pmd;
2833         bool young, write, dirty;
2834         int i;
2835
2836         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2837         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2838         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2839         VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd));
2840
2841         count_vm_event(THP_SPLIT_PMD);
2842
2843         if (vma_is_dax(vma)) {
2844                 pmd_t _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2845                 if (is_huge_zero_pmd(_pmd))
2846                         put_huge_zero_page();
2847                 return;
2848         } else if (is_huge_zero_pmd(*pmd)) {
2849                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2850         }
2851
2852         page = pmd_page(*pmd);
2853         VM_BUG_ON_PAGE(!page_count(page), page);
2854         atomic_add(HPAGE_PMD_NR - 1, &page->_count);
2855         write = pmd_write(*pmd);
2856         young = pmd_young(*pmd);
2857         dirty = pmd_dirty(*pmd);
2858
2859         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2860         pmd_populate(mm, &_pmd, pgtable);
2861
2862         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2863                 pte_t entry, *pte;
2864                 /*
2865                  * Note that NUMA hinting access restrictions are not
2866                  * transferred to avoid any possibility of altering
2867                  * permissions across VMAs.
2868                  */
2869                 if (freeze) {
2870                         swp_entry_t swp_entry;
2871                         swp_entry = make_migration_entry(page + i, write);
2872                         entry = swp_entry_to_pte(swp_entry);
2873                 } else {
2874                         entry = mk_pte(page + i, vma->vm_page_prot);
2875                         entry = maybe_mkwrite(entry, vma);
2876                         if (!write)
2877                                 entry = pte_wrprotect(entry);
2878                         if (!young)
2879                                 entry = pte_mkold(entry);
2880                 }
2881                 if (dirty)
2882                         SetPageDirty(page + i);
2883                 pte = pte_offset_map(&_pmd, haddr);
2884                 BUG_ON(!pte_none(*pte));
2885                 set_pte_at(mm, haddr, pte, entry);
2886                 atomic_inc(&page[i]._mapcount);
2887                 pte_unmap(pte);
2888         }
2889
2890         /*
2891          * Set PG_double_map before dropping compound_mapcount to avoid
2892          * false-negative page_mapped().
2893          */
2894         if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2895                 for (i = 0; i < HPAGE_PMD_NR; i++)
2896                         atomic_inc(&page[i]._mapcount);
2897         }
2898
2899         if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2900                 /* Last compound_mapcount is gone. */
2901                 __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
2902                 if (TestClearPageDoubleMap(page)) {
2903                         /* No need in mapcount reference anymore */
2904                         for (i = 0; i < HPAGE_PMD_NR; i++)
2905                                 atomic_dec(&page[i]._mapcount);
2906                 }
2907         }
2908
2909         smp_wmb(); /* make pte visible before pmd */
2910         /*
2911          * Up to this point the pmd is present and huge and userland has the
2912          * whole access to the hugepage during the split (which happens in
2913          * place). If we overwrite the pmd with the not-huge version pointing
2914          * to the pte here (which of course we could if all CPUs were bug
2915          * free), userland could trigger a small page size TLB miss on the
2916          * small sized TLB while the hugepage TLB entry is still established in
2917          * the huge TLB. Some CPU doesn't like that.
2918          * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2919          * 383 on page 93. Intel should be safe but is also warns that it's
2920          * only safe if the permission and cache attributes of the two entries
2921          * loaded in the two TLB is identical (which should be the case here).
2922          * But it is generally safer to never allow small and huge TLB entries
2923          * for the same virtual address to be loaded simultaneously. So instead
2924          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2925          * current pmd notpresent (atomically because here the pmd_trans_huge
2926          * and pmd_trans_splitting must remain set at all times on the pmd
2927          * until the split is complete for this pmd), then we flush the SMP TLB
2928          * and finally we write the non-huge version of the pmd entry with
2929          * pmd_populate.
2930          */
2931         pmdp_invalidate(vma, haddr, pmd);
2932         pmd_populate(mm, pmd, pgtable);
2933
2934         if (freeze) {
2935                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2936                         page_remove_rmap(page + i, false);
2937                         put_page(page + i);
2938                 }
2939         }
2940 }
2941
2942 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2943                 unsigned long address)
2944 {
2945         spinlock_t *ptl;
2946         struct mm_struct *mm = vma->vm_mm;
2947         struct page *page = NULL;
2948         unsigned long haddr = address & HPAGE_PMD_MASK;
2949
2950         mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
2951         ptl = pmd_lock(mm, pmd);
2952         if (pmd_trans_huge(*pmd)) {
2953                 page = pmd_page(*pmd);
2954                 if (PageMlocked(page))
2955                         get_page(page);
2956                 else
2957                         page = NULL;
2958         } else if (!pmd_devmap(*pmd))
2959                 goto out;
2960         __split_huge_pmd_locked(vma, pmd, haddr, false);
2961 out:
2962         spin_unlock(ptl);
2963         mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
2964         if (page) {
2965                 lock_page(page);
2966                 munlock_vma_page(page);
2967                 unlock_page(page);
2968                 put_page(page);
2969         }
2970 }
2971
2972 static void split_huge_pmd_address(struct vm_area_struct *vma,
2973                                     unsigned long address)
2974 {
2975         pgd_t *pgd;
2976         pud_t *pud;
2977         pmd_t *pmd;
2978
2979         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2980
2981         pgd = pgd_offset(vma->vm_mm, address);
2982         if (!pgd_present(*pgd))
2983                 return;
2984
2985         pud = pud_offset(pgd, address);
2986         if (!pud_present(*pud))
2987                 return;
2988
2989         pmd = pmd_offset(pud, address);
2990         if (!pmd_present(*pmd) || (!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd)))
2991                 return;
2992         /*
2993          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2994          * materialize from under us.
2995          */
2996         split_huge_pmd(vma, pmd, address);
2997 }
2998
2999 void vma_adjust_trans_huge(struct vm_area_struct *vma,
3000                              unsigned long start,
3001                              unsigned long end,
3002                              long adjust_next)
3003 {
3004         /*
3005          * If the new start address isn't hpage aligned and it could
3006          * previously contain an hugepage: check if we need to split
3007          * an huge pmd.
3008          */
3009         if (start & ~HPAGE_PMD_MASK &&
3010             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
3011             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3012                 split_huge_pmd_address(vma, start);
3013
3014         /*
3015          * If the new end address isn't hpage aligned and it could
3016          * previously contain an hugepage: check if we need to split
3017          * an huge pmd.
3018          */
3019         if (end & ~HPAGE_PMD_MASK &&
3020             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
3021             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3022                 split_huge_pmd_address(vma, end);
3023
3024         /*
3025          * If we're also updating the vma->vm_next->vm_start, if the new
3026          * vm_next->vm_start isn't page aligned and it could previously
3027          * contain an hugepage: check if we need to split an huge pmd.
3028          */
3029         if (adjust_next > 0) {
3030                 struct vm_area_struct *next = vma->vm_next;
3031                 unsigned long nstart = next->vm_start;
3032                 nstart += adjust_next << PAGE_SHIFT;
3033                 if (nstart & ~HPAGE_PMD_MASK &&
3034                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
3035                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
3036                         split_huge_pmd_address(next, nstart);
3037         }
3038 }
3039
3040 static void freeze_page_vma(struct vm_area_struct *vma, struct page *page,
3041                 unsigned long address)
3042 {
3043         unsigned long haddr = address & HPAGE_PMD_MASK;
3044         spinlock_t *ptl;
3045         pgd_t *pgd;
3046         pud_t *pud;
3047         pmd_t *pmd;
3048         pte_t *pte;
3049         int i, nr = HPAGE_PMD_NR;
3050
3051         /* Skip pages which doesn't belong to the VMA */
3052         if (address < vma->vm_start) {
3053                 int off = (vma->vm_start - address) >> PAGE_SHIFT;
3054                 page += off;
3055                 nr -= off;
3056                 address = vma->vm_start;
3057         }
3058
3059         pgd = pgd_offset(vma->vm_mm, address);
3060         if (!pgd_present(*pgd))
3061                 return;
3062         pud = pud_offset(pgd, address);
3063         if (!pud_present(*pud))
3064                 return;
3065         pmd = pmd_offset(pud, address);
3066         ptl = pmd_lock(vma->vm_mm, pmd);
3067         if (!pmd_present(*pmd)) {
3068                 spin_unlock(ptl);
3069                 return;
3070         }
3071         if (pmd_trans_huge(*pmd)) {
3072                 if (page == pmd_page(*pmd))
3073                         __split_huge_pmd_locked(vma, pmd, haddr, true);
3074                 spin_unlock(ptl);
3075                 return;
3076         }
3077         spin_unlock(ptl);
3078
3079         pte = pte_offset_map_lock(vma->vm_mm, pmd, address, &ptl);
3080         for (i = 0; i < nr; i++, address += PAGE_SIZE, page++, pte++) {
3081                 pte_t entry, swp_pte;
3082                 swp_entry_t swp_entry;
3083
3084                 /*
3085                  * We've just crossed page table boundary: need to map next one.
3086                  * It can happen if THP was mremaped to non PMD-aligned address.
3087                  */
3088                 if (unlikely(address == haddr + HPAGE_PMD_SIZE)) {
3089                         pte_unmap_unlock(pte - 1, ptl);
3090                         pmd = mm_find_pmd(vma->vm_mm, address);
3091                         if (!pmd)
3092                                 return;
3093                         pte = pte_offset_map_lock(vma->vm_mm, pmd,
3094                                         address, &ptl);
3095                 }
3096
3097                 if (!pte_present(*pte))
3098                         continue;
3099                 if (page_to_pfn(page) != pte_pfn(*pte))
3100                         continue;
3101                 flush_cache_page(vma, address, page_to_pfn(page));
3102                 entry = ptep_clear_flush(vma, address, pte);
3103                 if (pte_dirty(entry))
3104                         SetPageDirty(page);
3105                 swp_entry = make_migration_entry(page, pte_write(entry));
3106                 swp_pte = swp_entry_to_pte(swp_entry);
3107                 if (pte_soft_dirty(entry))
3108                         swp_pte = pte_swp_mksoft_dirty(swp_pte);
3109                 set_pte_at(vma->vm_mm, address, pte, swp_pte);
3110                 page_remove_rmap(page, false);
3111                 put_page(page);
3112         }
3113         pte_unmap_unlock(pte - 1, ptl);
3114 }
3115
3116 static void freeze_page(struct anon_vma *anon_vma, struct page *page)
3117 {
3118         struct anon_vma_chain *avc;
3119         pgoff_t pgoff = page_to_pgoff(page);
3120
3121         VM_BUG_ON_PAGE(!PageHead(page), page);
3122
3123         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff,
3124                         pgoff + HPAGE_PMD_NR - 1) {
3125                 unsigned long address = __vma_address(page, avc->vma);
3126
3127                 mmu_notifier_invalidate_range_start(avc->vma->vm_mm,
3128                                 address, address + HPAGE_PMD_SIZE);
3129                 freeze_page_vma(avc->vma, page, address);
3130                 mmu_notifier_invalidate_range_end(avc->vma->vm_mm,
3131                                 address, address + HPAGE_PMD_SIZE);
3132         }
3133 }
3134
3135 static void unfreeze_page_vma(struct vm_area_struct *vma, struct page *page,
3136                 unsigned long address)
3137 {
3138         spinlock_t *ptl;
3139         pmd_t *pmd;
3140         pte_t *pte, entry;
3141         swp_entry_t swp_entry;
3142         unsigned long haddr = address & HPAGE_PMD_MASK;
3143         int i, nr = HPAGE_PMD_NR;
3144
3145         /* Skip pages which doesn't belong to the VMA */
3146         if (address < vma->vm_start) {
3147                 int off = (vma->vm_start - address) >> PAGE_SHIFT;
3148                 page += off;
3149                 nr -= off;
3150                 address = vma->vm_start;
3151         }
3152
3153         pmd = mm_find_pmd(vma->vm_mm, address);
3154         if (!pmd)
3155                 return;
3156
3157         pte = pte_offset_map_lock(vma->vm_mm, pmd, address, &ptl);
3158         for (i = 0; i < nr; i++, address += PAGE_SIZE, page++, pte++) {
3159                 /*
3160                  * We've just crossed page table boundary: need to map next one.
3161                  * It can happen if THP was mremaped to non-PMD aligned address.
3162                  */
3163                 if (unlikely(address == haddr + HPAGE_PMD_SIZE)) {
3164                         pte_unmap_unlock(pte - 1, ptl);
3165                         pmd = mm_find_pmd(vma->vm_mm, address);
3166                         if (!pmd)
3167                                 return;
3168                         pte = pte_offset_map_lock(vma->vm_mm, pmd,
3169                                         address, &ptl);
3170                 }
3171
3172                 if (!is_swap_pte(*pte))
3173                         continue;
3174
3175                 swp_entry = pte_to_swp_entry(*pte);
3176                 if (!is_migration_entry(swp_entry))
3177                         continue;
3178                 if (migration_entry_to_page(swp_entry) != page)
3179                         continue;
3180
3181                 get_page(page);
3182                 page_add_anon_rmap(page, vma, address, false);
3183
3184                 entry = pte_mkold(mk_pte(page, vma->vm_page_prot));
3185                 if (PageDirty(page))
3186                         entry = pte_mkdirty(entry);
3187                 if (is_write_migration_entry(swp_entry))
3188                         entry = maybe_mkwrite(entry, vma);
3189
3190                 flush_dcache_page(page);
3191                 set_pte_at(vma->vm_mm, address, pte, entry);
3192
3193                 /* No need to invalidate - it was non-present before */
3194                 update_mmu_cache(vma, address, pte);
3195         }
3196         pte_unmap_unlock(pte - 1, ptl);
3197 }
3198
3199 static void unfreeze_page(struct anon_vma *anon_vma, struct page *page)
3200 {
3201         struct anon_vma_chain *avc;
3202         pgoff_t pgoff = page_to_pgoff(page);
3203
3204         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
3205                         pgoff, pgoff + HPAGE_PMD_NR - 1) {
3206                 unsigned long address = __vma_address(page, avc->vma);
3207
3208                 mmu_notifier_invalidate_range_start(avc->vma->vm_mm,
3209                                 address, address + HPAGE_PMD_SIZE);
3210                 unfreeze_page_vma(avc->vma, page, address);
3211                 mmu_notifier_invalidate_range_end(avc->vma->vm_mm,
3212                                 address, address + HPAGE_PMD_SIZE);
3213         }
3214 }
3215
3216 static int __split_huge_page_tail(struct page *head, int tail,
3217                 struct lruvec *lruvec, struct list_head *list)
3218 {
3219         int mapcount;
3220         struct page *page_tail = head + tail;
3221
3222         mapcount = atomic_read(&page_tail->_mapcount) + 1;
3223         VM_BUG_ON_PAGE(atomic_read(&page_tail->_count) != 0, page_tail);
3224
3225         /*
3226          * tail_page->_count is zero and not changing from under us. But
3227          * get_page_unless_zero() may be running from under us on the
3228          * tail_page. If we used atomic_set() below instead of atomic_add(), we
3229          * would then run atomic_set() concurrently with
3230          * get_page_unless_zero(), and atomic_set() is implemented in C not
3231          * using locked ops. spin_unlock on x86 sometime uses locked ops
3232          * because of PPro errata 66, 92, so unless somebody can guarantee
3233          * atomic_set() here would be safe on all archs (and not only on x86),
3234          * it's safer to use atomic_add().
3235          */
3236         atomic_add(mapcount + 1, &page_tail->_count);
3237
3238
3239         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
3240         page_tail->flags |= (head->flags &
3241                         ((1L << PG_referenced) |
3242                          (1L << PG_swapbacked) |
3243                          (1L << PG_mlocked) |
3244                          (1L << PG_uptodate) |
3245                          (1L << PG_active) |
3246                          (1L << PG_locked) |
3247                          (1L << PG_unevictable) |
3248                          (1L << PG_dirty)));
3249
3250         /*
3251          * After clearing PageTail the gup refcount can be released.
3252          * Page flags also must be visible before we make the page non-compound.
3253          */
3254         smp_wmb();
3255
3256         clear_compound_head(page_tail);
3257
3258         if (page_is_young(head))
3259                 set_page_young(page_tail);
3260         if (page_is_idle(head))
3261                 set_page_idle(page_tail);
3262
3263         /* ->mapping in first tail page is compound_mapcount */
3264         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
3265                         page_tail);
3266         page_tail->mapping = head->mapping;
3267
3268         page_tail->index = head->index + tail;
3269         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
3270         lru_add_page_tail(head, page_tail, lruvec, list);
3271
3272         return mapcount;
3273 }
3274
3275 static void __split_huge_page(struct page *page, struct list_head *list)
3276 {
3277         struct page *head = compound_head(page);
3278         struct zone *zone = page_zone(head);
3279         struct lruvec *lruvec;
3280         int i, tail_mapcount;
3281
3282         /* prevent PageLRU to go away from under us, and freeze lru stats */
3283         spin_lock_irq(&zone->lru_lock);
3284         lruvec = mem_cgroup_page_lruvec(head, zone);
3285
3286         /* complete memcg works before add pages to LRU */
3287         mem_cgroup_split_huge_fixup(head);
3288
3289         tail_mapcount = 0;
3290         for (i = HPAGE_PMD_NR - 1; i >= 1; i--)
3291                 tail_mapcount += __split_huge_page_tail(head, i, lruvec, list);
3292         atomic_sub(tail_mapcount, &head->_count);
3293
3294         ClearPageCompound(head);
3295         spin_unlock_irq(&zone->lru_lock);
3296
3297         unfreeze_page(page_anon_vma(head), head);
3298
3299         for (i = 0; i < HPAGE_PMD_NR; i++) {
3300                 struct page *subpage = head + i;
3301                 if (subpage == page)
3302                         continue;
3303                 unlock_page(subpage);
3304
3305                 /*
3306                  * Subpages may be freed if there wasn't any mapping
3307                  * like if add_to_swap() is running on a lru page that
3308                  * had its mapping zapped. And freeing these pages
3309                  * requires taking the lru_lock so we do the put_page
3310                  * of the tail pages after the split is complete.
3311                  */
3312                 put_page(subpage);
3313         }
3314 }
3315
3316 int total_mapcount(struct page *page)
3317 {
3318         int i, ret;
3319
3320         VM_BUG_ON_PAGE(PageTail(page), page);
3321
3322         if (likely(!PageCompound(page)))
3323                 return atomic_read(&page->_mapcount) + 1;
3324
3325         ret = compound_mapcount(page);
3326         if (PageHuge(page))
3327                 return ret;
3328         for (i = 0; i < HPAGE_PMD_NR; i++)
3329                 ret += atomic_read(&page[i]._mapcount) + 1;
3330         if (PageDoubleMap(page))
3331                 ret -= HPAGE_PMD_NR;
3332         return ret;
3333 }
3334
3335 /*
3336  * This function splits huge page into normal pages. @page can point to any
3337  * subpage of huge page to split. Split doesn't change the position of @page.
3338  *
3339  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
3340  * The huge page must be locked.
3341  *
3342  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
3343  *
3344  * Both head page and tail pages will inherit mapping, flags, and so on from
3345  * the hugepage.
3346  *
3347  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
3348  * they are not mapped.
3349  *
3350  * Returns 0 if the hugepage is split successfully.
3351  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
3352  * us.
3353  */
3354 int split_huge_page_to_list(struct page *page, struct list_head *list)
3355 {
3356         struct page *head = compound_head(page);
3357         struct anon_vma *anon_vma;
3358         int count, mapcount, ret;
3359         bool mlocked;
3360
3361         VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
3362         VM_BUG_ON_PAGE(!PageAnon(page), page);
3363         VM_BUG_ON_PAGE(!PageLocked(page), page);
3364         VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
3365         VM_BUG_ON_PAGE(!PageCompound(page), page);
3366
3367         /*
3368          * The caller does not necessarily hold an mmap_sem that would prevent
3369          * the anon_vma disappearing so we first we take a reference to it
3370          * and then lock the anon_vma for write. This is similar to
3371          * page_lock_anon_vma_read except the write lock is taken to serialise
3372          * against parallel split or collapse operations.
3373          */
3374         anon_vma = page_get_anon_vma(head);
3375         if (!anon_vma) {
3376                 ret = -EBUSY;
3377                 goto out;
3378         }
3379         anon_vma_lock_write(anon_vma);
3380
3381         /*
3382          * Racy check if we can split the page, before freeze_page() will
3383          * split PMDs
3384          */
3385         if (total_mapcount(head) != page_count(head) - 1) {
3386                 ret = -EBUSY;
3387                 goto out_unlock;
3388         }
3389
3390         mlocked = PageMlocked(page);
3391         freeze_page(anon_vma, head);
3392         VM_BUG_ON_PAGE(compound_mapcount(head), head);
3393
3394         /* Make sure the page is not on per-CPU pagevec as it takes pin */
3395         if (mlocked)
3396                 lru_add_drain();
3397
3398         /* Prevent deferred_split_scan() touching ->_count */
3399         spin_lock(&split_queue_lock);
3400         count = page_count(head);
3401         mapcount = total_mapcount(head);
3402         if (!mapcount && count == 1) {
3403                 if (!list_empty(page_deferred_list(head))) {
3404                         split_queue_len--;
3405                         list_del(page_deferred_list(head));
3406                 }
3407                 spin_unlock(&split_queue_lock);
3408                 __split_huge_page(page, list);
3409                 ret = 0;
3410         } else if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
3411                 spin_unlock(&split_queue_lock);
3412                 pr_alert("total_mapcount: %u, page_count(): %u\n",
3413                                 mapcount, count);
3414                 if (PageTail(page))
3415                         dump_page(head, NULL);
3416                 dump_page(page, "total_mapcount(head) > 0");
3417                 BUG();
3418         } else {
3419                 spin_unlock(&split_queue_lock);
3420                 unfreeze_page(anon_vma, head);
3421                 ret = -EBUSY;
3422         }
3423
3424 out_unlock:
3425         anon_vma_unlock_write(anon_vma);
3426         put_anon_vma(anon_vma);
3427 out:
3428         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
3429         return ret;
3430 }
3431
3432 void free_transhuge_page(struct page *page)
3433 {
3434         unsigned long flags;
3435
3436         spin_lock_irqsave(&split_queue_lock, flags);
3437         if (!list_empty(page_deferred_list(page))) {
3438                 split_queue_len--;
3439                 list_del(page_deferred_list(page));
3440         }
3441         spin_unlock_irqrestore(&split_queue_lock, flags);
3442         free_compound_page(page);
3443 }
3444
3445 void deferred_split_huge_page(struct page *page)
3446 {
3447         unsigned long flags;
3448
3449         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3450
3451         spin_lock_irqsave(&split_queue_lock, flags);
3452         if (list_empty(page_deferred_list(page))) {
3453                 list_add_tail(page_deferred_list(page), &split_queue);
3454                 split_queue_len++;
3455         }
3456         spin_unlock_irqrestore(&split_queue_lock, flags);
3457 }
3458
3459 static unsigned long deferred_split_count(struct shrinker *shrink,
3460                 struct shrink_control *sc)
3461 {
3462         /*
3463          * Split a page from split_queue will free up at least one page,
3464          * at most HPAGE_PMD_NR - 1. We don't track exact number.
3465          * Let's use HPAGE_PMD_NR / 2 as ballpark.
3466          */
3467         return ACCESS_ONCE(split_queue_len) * HPAGE_PMD_NR / 2;
3468 }
3469
3470 static unsigned long deferred_split_scan(struct shrinker *shrink,
3471                 struct shrink_control *sc)
3472 {
3473         unsigned long flags;
3474         LIST_HEAD(list), *pos, *next;
3475         struct page *page;
3476         int split = 0;
3477
3478         spin_lock_irqsave(&split_queue_lock, flags);
3479         list_splice_init(&split_queue, &list);
3480
3481         /* Take pin on all head pages to avoid freeing them under us */
3482         list_for_each_safe(pos, next, &list) {
3483                 page = list_entry((void *)pos, struct page, mapping);
3484                 page = compound_head(page);
3485                 /* race with put_compound_page() */
3486                 if (!get_page_unless_zero(page)) {
3487                         list_del_init(page_deferred_list(page));
3488                         split_queue_len--;
3489                 }
3490         }
3491         spin_unlock_irqrestore(&split_queue_lock, flags);
3492
3493         list_for_each_safe(pos, next, &list) {
3494                 page = list_entry((void *)pos, struct page, mapping);
3495                 lock_page(page);
3496                 /* split_huge_page() removes page from list on success */
3497                 if (!split_huge_page(page))
3498                         split++;
3499                 unlock_page(page);
3500                 put_page(page);
3501         }
3502
3503         spin_lock_irqsave(&split_queue_lock, flags);
3504         list_splice_tail(&list, &split_queue);
3505         spin_unlock_irqrestore(&split_queue_lock, flags);
3506
3507         return split * HPAGE_PMD_NR / 2;
3508 }
3509
3510 static struct shrinker deferred_split_shrinker = {
3511         .count_objects = deferred_split_count,
3512         .scan_objects = deferred_split_scan,
3513         .seeks = DEFAULT_SEEKS,
3514 };
3515
3516 #ifdef CONFIG_DEBUG_FS
3517 static int split_huge_pages_set(void *data, u64 val)
3518 {
3519         struct zone *zone;
3520         struct page *page;
3521         unsigned long pfn, max_zone_pfn;
3522         unsigned long total = 0, split = 0;
3523
3524         if (val != 1)
3525                 return -EINVAL;
3526
3527         for_each_populated_zone(zone) {
3528                 max_zone_pfn = zone_end_pfn(zone);
3529                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
3530                         if (!pfn_valid(pfn))
3531                                 continue;
3532
3533                         page = pfn_to_page(pfn);
3534                         if (!get_page_unless_zero(page))
3535                                 continue;
3536
3537                         if (zone != page_zone(page))
3538                                 goto next;
3539
3540                         if (!PageHead(page) || !PageAnon(page) ||
3541                                         PageHuge(page))
3542                                 goto next;
3543
3544                         total++;
3545                         lock_page(page);
3546                         if (!split_huge_page(page))
3547                                 split++;
3548                         unlock_page(page);
3549 next:
3550                         put_page(page);
3551                 }
3552         }
3553
3554         pr_info("%lu of %lu THP split", split, total);
3555
3556         return 0;
3557 }
3558 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
3559                 "%llu\n");
3560
3561 static int __init split_huge_pages_debugfs(void)
3562 {
3563         void *ret;
3564
3565         ret = debugfs_create_file("split_huge_pages", 0644, NULL, NULL,
3566                         &split_huge_pages_fops);
3567         if (!ret)
3568                 pr_warn("Failed to create split_huge_pages in debugfs");
3569         return 0;
3570 }
3571 late_initcall(split_huge_pages_debugfs);
3572 #endif