mm/huge_memory.c

   1 /*
   2  *  Copyright (C) 2009  Red Hat, Inc.
   3  *
   4  *  This work is licensed under the terms of the GNU GPL, version 2. See
   5  *  the COPYING file in the top-level directory.
   6  */
   7
   8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   9
  10 #include <linux/mm.h>
  11 #include <linux/sched.h>
  12 #include <linux/highmem.h>
  13 #include <linux/hugetlb.h>
  14 #include <linux/mmu_notifier.h>
  15 #include <linux/rmap.h>
  16 #include <linux/swap.h>
  17 #include <linux/shrinker.h>
  18 #include <linux/mm_inline.h>
  19 #include <linux/kthread.h>
  20 #include <linux/khugepaged.h>
  21 #include <linux/freezer.h>
  22 #include <linux/mman.h>
  23 #include <linux/pagemap.h>
  24 #include <linux/migrate.h>
  25 #include <linux/hashtable.h>
  26 #include <linux/userfaultfd_k.h>
  27
  28 #include <asm/tlb.h>
  29 #include <asm/pgalloc.h>
  30 #include "internal.h"
  31
  32 /*
  33  * By default transparent hugepage support is disabled in order that avoid
  34  * to risk increase the memory footprint of applications without a guaranteed
  35  * benefit. When transparent hugepage support is enabled, is for all mappings,
  36  * and khugepaged scans all mappings.
  37  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  38  * for all hugepage allocations.
  39  */
  40 unsigned long transparent_hugepage_flags __read_mostly =
  41 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  42         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  43 #endif
  44 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  45         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  46 #endif
  47         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
  48         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  49         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  50
  51 /* default scan 8*512 pte (or vmas) every 30 second */
  52 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
  53 static unsigned int khugepaged_pages_collapsed;
  54 static unsigned int khugepaged_full_scans;
  55 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  56 /* during fragmentation poll the hugepage allocator once every minute */
  57 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  58 static struct task_struct *khugepaged_thread __read_mostly;
  59 static DEFINE_MUTEX(khugepaged_mutex);
  60 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  61 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  62 /*
  63  * default collapse hugepages if there is at least one pte mapped like
  64  * it would have happened if the vma was large enough during page
  65  * fault.
  66  */
  67 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
  68
  69 static int khugepaged(void *none);
  70 static int khugepaged_slab_init(void);
  71 static void khugepaged_slab_exit(void);
  72
  73 #define MM_SLOTS_HASH_BITS 10
  74 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
  75
  76 static struct kmem_cache *mm_slot_cache __read_mostly;
  77
  78 /**
  79  * struct mm_slot - hash lookup from mm to mm_slot
  80  * @hash: hash collision list
  81  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  82  * @mm: the mm that this information is valid for
  83  */
  84 struct mm_slot {
  85         struct hlist_node hash;
  86         struct list_head mm_node;
  87         struct mm_struct *mm;
  88 };
  89
  90 /**
  91  * struct khugepaged_scan - cursor for scanning
  92  * @mm_head: the head of the mm list to scan
  93  * @mm_slot: the current mm_slot we are scanning
  94  * @address: the next address inside that to be scanned
  95  *
  96  * There is only the one khugepaged_scan instance of this cursor structure.
  97  */
  98 struct khugepaged_scan {
  99         struct list_head mm_head;
 100         struct mm_slot *mm_slot;
 101         unsigned long address;
 102 };
 103 static struct khugepaged_scan khugepaged_scan = {
 104         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 105 };
 106
 107
 108 static int set_recommended_min_free_kbytes(void)
 109 {
 110         struct zone *zone;
 111         int nr_zones = 0;
 112         unsigned long recommended_min;
 113
 114         for_each_populated_zone(zone)
 115                 nr_zones++;
 116
 117         /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
 118         recommended_min = pageblock_nr_pages * nr_zones * 2;
 119
 120         /*
 121          * Make sure that on average at least two pageblocks are almost free
 122          * of another type, one for a migratetype to fall back to and a
 123          * second to avoid subsequent fallbacks of other types There are 3
 124          * MIGRATE_TYPES we care about.
 125          */
 126         recommended_min += pageblock_nr_pages * nr_zones *
 127                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
 128
 129         /* don't ever allow to reserve more than 5% of the lowmem */
 130         recommended_min = min(recommended_min,
 131                               (unsigned long) nr_free_buffer_pages() / 20);
 132         recommended_min <<= (PAGE_SHIFT-10);
 133
 134         if (recommended_min > min_free_kbytes) {
 135                 if (user_min_free_kbytes >= 0)
 136                         pr_info("raising min_free_kbytes from %d to %lu "
 137                                 "to help transparent hugepage allocations\n",
 138                                 min_free_kbytes, recommended_min);
 139
 140                 min_free_kbytes = recommended_min;
 141         }
 142         setup_per_zone_wmarks();
 143         return 0;
 144 }
 145
 146 static int start_stop_khugepaged(void)
 147 {
 148         int err = 0;
 149         if (khugepaged_enabled()) {
 150                 if (!khugepaged_thread)
 151                         khugepaged_thread = kthread_run(khugepaged, NULL,
 152                                                         "khugepaged");
 153                 if (unlikely(IS_ERR(khugepaged_thread))) {
 154                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
 155                         err = PTR_ERR(khugepaged_thread);
 156                         khugepaged_thread = NULL;
 157                         goto fail;
 158                 }
 159
 160                 if (!list_empty(&khugepaged_scan.mm_head))
 161                         wake_up_interruptible(&khugepaged_wait);
 162
 163                 set_recommended_min_free_kbytes();
 164         } else if (khugepaged_thread) {
 165                 kthread_stop(khugepaged_thread);
 166                 khugepaged_thread = NULL;
 167         }
 168 fail:
 169         return err;
 170 }
 171
 172 static atomic_t huge_zero_refcount;
 173 struct page *huge_zero_page __read_mostly;
 174
 175 static inline bool is_huge_zero_pmd(pmd_t pmd)
 176 {
 177         return is_huge_zero_page(pmd_page(pmd));
 178 }
 179
 180 static struct page *get_huge_zero_page(void)
 181 {
 182         struct page *zero_page;
 183 retry:
 184         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
 185                 return READ_ONCE(huge_zero_page);
 186
 187         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
 188                         HPAGE_PMD_ORDER);
 189         if (!zero_page) {
 190                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
 191                 return NULL;
 192         }
 193         count_vm_event(THP_ZERO_PAGE_ALLOC);
 194         preempt_disable();
 195         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
 196                 preempt_enable();
 197                 __free_pages(zero_page, compound_order(zero_page));
 198                 goto retry;
 199         }
 200
 201         /* We take additional reference here. It will be put back by shrinker */
 202         atomic_set(&huge_zero_refcount, 2);
 203         preempt_enable();
 204         return READ_ONCE(huge_zero_page);
 205 }
 206
 207 static void put_huge_zero_page(void)
 208 {
 209         /*
 210          * Counter should never go to zero here. Only shrinker can put
 211          * last reference.
 212          */
 213         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 214 }
 215
 216 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 217                                         struct shrink_control *sc)
 218 {
 219         /* we can free zero page only if last reference remains */
 220         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 221 }
 222
 223 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 224                                        struct shrink_control *sc)
 225 {
 226         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 227                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 228                 BUG_ON(zero_page == NULL);
 229                 __free_pages(zero_page, compound_order(zero_page));
 230                 return HPAGE_PMD_NR;
 231         }
 232
 233         return 0;
 234 }
 235
 236 static struct shrinker huge_zero_page_shrinker = {
 237         .count_objects = shrink_huge_zero_page_count,
 238         .scan_objects = shrink_huge_zero_page_scan,
 239         .seeks = DEFAULT_SEEKS,
 240 };
 241
 242 #ifdef CONFIG_SYSFS
 243
 244 static ssize_t double_flag_show(struct kobject *kobj,
 245                                 struct kobj_attribute *attr, char *buf,
 246                                 enum transparent_hugepage_flag enabled,
 247                                 enum transparent_hugepage_flag req_madv)
 248 {
 249         if (test_bit(enabled, &transparent_hugepage_flags)) {
 250                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
 251                 return sprintf(buf, "[always] madvise never\n");
 252         } else if (test_bit(req_madv, &transparent_hugepage_flags))
 253                 return sprintf(buf, "always [madvise] never\n");
 254         else
 255                 return sprintf(buf, "always madvise [never]\n");
 256 }
 257 static ssize_t double_flag_store(struct kobject *kobj,
 258                                  struct kobj_attribute *attr,
 259                                  const char *buf, size_t count,
 260                                  enum transparent_hugepage_flag enabled,
 261                                  enum transparent_hugepage_flag req_madv)
 262 {
 263         if (!memcmp("always", buf,
 264                     min(sizeof("always")-1, count))) {
 265                 set_bit(enabled, &transparent_hugepage_flags);
 266                 clear_bit(req_madv, &transparent_hugepage_flags);
 267         } else if (!memcmp("madvise", buf,
 268                            min(sizeof("madvise")-1, count))) {
 269                 clear_bit(enabled, &transparent_hugepage_flags);
 270                 set_bit(req_madv, &transparent_hugepage_flags);
 271         } else if (!memcmp("never", buf,
 272                            min(sizeof("never")-1, count))) {
 273                 clear_bit(enabled, &transparent_hugepage_flags);
 274                 clear_bit(req_madv, &transparent_hugepage_flags);
 275         } else
 276                 return -EINVAL;
 277
 278         return count;
 279 }
 280
 281 static ssize_t enabled_show(struct kobject *kobj,
 282                             struct kobj_attribute *attr, char *buf)
 283 {
 284         return double_flag_show(kobj, attr, buf,
 285                                 TRANSPARENT_HUGEPAGE_FLAG,
 286                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 287 }
 288 static ssize_t enabled_store(struct kobject *kobj,
 289                              struct kobj_attribute *attr,
 290                              const char *buf, size_t count)
 291 {
 292         ssize_t ret;
 293
 294         ret = double_flag_store(kobj, attr, buf, count,
 295                                 TRANSPARENT_HUGEPAGE_FLAG,
 296                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 297
 298         if (ret > 0) {
 299                 int err;
 300
 301                 mutex_lock(&khugepaged_mutex);
 302                 err = start_stop_khugepaged();
 303                 mutex_unlock(&khugepaged_mutex);
 304
 305                 if (err)
 306                         ret = err;
 307         }
 308
 309         return ret;
 310 }
 311 static struct kobj_attribute enabled_attr =
 312         __ATTR(enabled, 0644, enabled_show, enabled_store);
 313
 314 static ssize_t single_flag_show(struct kobject *kobj,
 315                                 struct kobj_attribute *attr, char *buf,
 316                                 enum transparent_hugepage_flag flag)
 317 {
 318         return sprintf(buf, "%d\n",
 319                        !!test_bit(flag, &transparent_hugepage_flags));
 320 }
 321
 322 static ssize_t single_flag_store(struct kobject *kobj,
 323                                  struct kobj_attribute *attr,
 324                                  const char *buf, size_t count,
 325                                  enum transparent_hugepage_flag flag)
 326 {
 327         unsigned long value;
 328         int ret;
 329
 330         ret = kstrtoul(buf, 10, &value);
 331         if (ret < 0)
 332                 return ret;
 333         if (value > 1)
 334                 return -EINVAL;
 335
 336         if (value)
 337                 set_bit(flag, &transparent_hugepage_flags);
 338         else
 339                 clear_bit(flag, &transparent_hugepage_flags);
 340
 341         return count;
 342 }
 343
 344 /*
 345  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 346  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 347  * memory just to allocate one more hugepage.
 348  */
 349 static ssize_t defrag_show(struct kobject *kobj,
 350                            struct kobj_attribute *attr, char *buf)
 351 {
 352         return double_flag_show(kobj, attr, buf,
 353                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 354                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 355 }
 356 static ssize_t defrag_store(struct kobject *kobj,
 357                             struct kobj_attribute *attr,
 358                             const char *buf, size_t count)
 359 {
 360         return double_flag_store(kobj, attr, buf, count,
 361                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 362                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 363 }
 364 static struct kobj_attribute defrag_attr =
 365         __ATTR(defrag, 0644, defrag_show, defrag_store);
 366
 367 static ssize_t use_zero_page_show(struct kobject *kobj,
 368                 struct kobj_attribute *attr, char *buf)
 369 {
 370         return single_flag_show(kobj, attr, buf,
 371                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 372 }
 373 static ssize_t use_zero_page_store(struct kobject *kobj,
 374                 struct kobj_attribute *attr, const char *buf, size_t count)
 375 {
 376         return single_flag_store(kobj, attr, buf, count,
 377                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 378 }
 379 static struct kobj_attribute use_zero_page_attr =
 380         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 381 #ifdef CONFIG_DEBUG_VM
 382 static ssize_t debug_cow_show(struct kobject *kobj,
 383                                 struct kobj_attribute *attr, char *buf)
 384 {
 385         return single_flag_show(kobj, attr, buf,
 386                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 387 }
 388 static ssize_t debug_cow_store(struct kobject *kobj,
 389                                struct kobj_attribute *attr,
 390                                const char *buf, size_t count)
 391 {
 392         return single_flag_store(kobj, attr, buf, count,
 393                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 394 }
 395 static struct kobj_attribute debug_cow_attr =
 396         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 397 #endif /* CONFIG_DEBUG_VM */
 398
 399 static struct attribute *hugepage_attr[] = {
 400         &enabled_attr.attr,
 401         &defrag_attr.attr,
 402         &use_zero_page_attr.attr,
 403 #ifdef CONFIG_DEBUG_VM
 404         &debug_cow_attr.attr,
 405 #endif
 406         NULL,
 407 };
 408
 409 static struct attribute_group hugepage_attr_group = {
 410         .attrs = hugepage_attr,
 411 };
 412
 413 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 414                                          struct kobj_attribute *attr,
 415                                          char *buf)
 416 {
 417         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 418 }
 419
 420 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 421                                           struct kobj_attribute *attr,
 422                                           const char *buf, size_t count)
 423 {
 424         unsigned long msecs;
 425         int err;
 426
 427         err = kstrtoul(buf, 10, &msecs);
 428         if (err || msecs > UINT_MAX)
 429                 return -EINVAL;
 430
 431         khugepaged_scan_sleep_millisecs = msecs;
 432         wake_up_interruptible(&khugepaged_wait);
 433
 434         return count;
 435 }
 436 static struct kobj_attribute scan_sleep_millisecs_attr =
 437         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 438                scan_sleep_millisecs_store);
 439
 440 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 441                                           struct kobj_attribute *attr,
 442                                           char *buf)
 443 {
 444         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 445 }
 446
 447 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 448                                            struct kobj_attribute *attr,
 449                                            const char *buf, size_t count)
 450 {
 451         unsigned long msecs;
 452         int err;
 453
 454         err = kstrtoul(buf, 10, &msecs);
 455         if (err || msecs > UINT_MAX)
 456                 return -EINVAL;
 457
 458         khugepaged_alloc_sleep_millisecs = msecs;
 459         wake_up_interruptible(&khugepaged_wait);
 460
 461         return count;
 462 }
 463 static struct kobj_attribute alloc_sleep_millisecs_attr =
 464         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 465                alloc_sleep_millisecs_store);
 466
 467 static ssize_t pages_to_scan_show(struct kobject *kobj,
 468                                   struct kobj_attribute *attr,
 469                                   char *buf)
 470 {
 471         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 472 }
 473 static ssize_t pages_to_scan_store(struct kobject *kobj,
 474                                    struct kobj_attribute *attr,
 475                                    const char *buf, size_t count)
 476 {
 477         int err;
 478         unsigned long pages;
 479
 480         err = kstrtoul(buf, 10, &pages);
 481         if (err || !pages || pages > UINT_MAX)
 482                 return -EINVAL;
 483
 484         khugepaged_pages_to_scan = pages;
 485
 486         return count;
 487 }
 488 static struct kobj_attribute pages_to_scan_attr =
 489         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 490                pages_to_scan_store);
 491
 492 static ssize_t pages_collapsed_show(struct kobject *kobj,
 493                                     struct kobj_attribute *attr,
 494                                     char *buf)
 495 {
 496         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 497 }
 498 static struct kobj_attribute pages_collapsed_attr =
 499         __ATTR_RO(pages_collapsed);
 500
 501 static ssize_t full_scans_show(struct kobject *kobj,
 502                                struct kobj_attribute *attr,
 503                                char *buf)
 504 {
 505         return sprintf(buf, "%u\n", khugepaged_full_scans);
 506 }
 507 static struct kobj_attribute full_scans_attr =
 508         __ATTR_RO(full_scans);
 509
 510 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 511                                       struct kobj_attribute *attr, char *buf)
 512 {
 513         return single_flag_show(kobj, attr, buf,
 514                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 515 }
 516 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 517                                        struct kobj_attribute *attr,
 518                                        const char *buf, size_t count)
 519 {
 520         return single_flag_store(kobj, attr, buf, count,
 521                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 522 }
 523 static struct kobj_attribute khugepaged_defrag_attr =
 524         __ATTR(defrag, 0644, khugepaged_defrag_show,
 525                khugepaged_defrag_store);
 526
 527 /*
 528  * max_ptes_none controls if khugepaged should collapse hugepages over
 529  * any unmapped ptes in turn potentially increasing the memory
 530  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 531  * reduce the available free memory in the system as it
 532  * runs. Increasing max_ptes_none will instead potentially reduce the
 533  * free memory in the system during the khugepaged scan.
 534  */
 535 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 536                                              struct kobj_attribute *attr,
 537                                              char *buf)
 538 {
 539         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 540 }
 541 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 542                                               struct kobj_attribute *attr,
 543                                               const char *buf, size_t count)
 544 {
 545         int err;
 546         unsigned long max_ptes_none;
 547
 548         err = kstrtoul(buf, 10, &max_ptes_none);
 549         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 550                 return -EINVAL;
 551
 552         khugepaged_max_ptes_none = max_ptes_none;
 553
 554         return count;
 555 }
 556 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 557         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 558                khugepaged_max_ptes_none_store);
 559
 560 static struct attribute *khugepaged_attr[] = {
 561         &khugepaged_defrag_attr.attr,
 562         &khugepaged_max_ptes_none_attr.attr,
 563         &pages_to_scan_attr.attr,
 564         &pages_collapsed_attr.attr,
 565         &full_scans_attr.attr,
 566         &scan_sleep_millisecs_attr.attr,
 567         &alloc_sleep_millisecs_attr.attr,
 568         NULL,
 569 };
 570
 571 static struct attribute_group khugepaged_attr_group = {
 572         .attrs = khugepaged_attr,
 573         .name = "khugepaged",
 574 };
 575
 576 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 577 {
 578         int err;
 579
 580         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 581         if (unlikely(!*hugepage_kobj)) {
 582                 pr_err("failed to create transparent hugepage kobject\n");
 583                 return -ENOMEM;
 584         }
 585
 586         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 587         if (err) {
 588                 pr_err("failed to register transparent hugepage group\n");
 589                 goto delete_obj;
 590         }
 591
 592         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 593         if (err) {
 594                 pr_err("failed to register transparent hugepage group\n");
 595                 goto remove_hp_group;
 596         }
 597
 598         return 0;
 599
 600 remove_hp_group:
 601         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 602 delete_obj:
 603         kobject_put(*hugepage_kobj);
 604         return err;
 605 }
 606
 607 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 608 {
 609         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 610         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 611         kobject_put(hugepage_kobj);
 612 }
 613 #else
 614 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 615 {
 616         return 0;
 617 }
 618
 619 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 620 {
 621 }
 622 #endif /* CONFIG_SYSFS */
 623
 624 static int __init hugepage_init(void)
 625 {
 626         int err;
 627         struct kobject *hugepage_kobj;
 628
 629         if (!has_transparent_hugepage()) {
 630                 transparent_hugepage_flags = 0;
 631                 return -EINVAL;
 632         }
 633
 634         err = hugepage_init_sysfs(&hugepage_kobj);
 635         if (err)
 636                 goto err_sysfs;
 637
 638         err = khugepaged_slab_init();
 639         if (err)
 640                 goto err_slab;
 641
 642         err = register_shrinker(&huge_zero_page_shrinker);
 643         if (err)
 644                 goto err_hzp_shrinker;
 645
 646         /*
 647          * By default disable transparent hugepages on smaller systems,
 648          * where the extra memory used could hurt more than TLB overhead
 649          * is likely to save.  The admin can still enable it through /sys.
 650          */
 651         if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
 652                 transparent_hugepage_flags = 0;
 653                 return 0;
 654         }
 655
 656         err = start_stop_khugepaged();
 657         if (err)
 658                 goto err_khugepaged;
 659
 660         return 0;
 661 err_khugepaged:
 662         unregister_shrinker(&huge_zero_page_shrinker);
 663 err_hzp_shrinker:
 664         khugepaged_slab_exit();
 665 err_slab:
 666         hugepage_exit_sysfs(hugepage_kobj);
 667 err_sysfs:
 668         return err;
 669 }
 670 subsys_initcall(hugepage_init);
 671
 672 static int __init setup_transparent_hugepage(char *str)
 673 {
 674         int ret = 0;
 675         if (!str)
 676                 goto out;
 677         if (!strcmp(str, "always")) {
 678                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 679                         &transparent_hugepage_flags);
 680                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 681                           &transparent_hugepage_flags);
 682                 ret = 1;
 683         } else if (!strcmp(str, "madvise")) {
 684                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 685                           &transparent_hugepage_flags);
 686                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 687                         &transparent_hugepage_flags);
 688                 ret = 1;
 689         } else if (!strcmp(str, "never")) {
 690                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 691                           &transparent_hugepage_flags);
 692                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 693                           &transparent_hugepage_flags);
 694                 ret = 1;
 695         }
 696 out:
 697         if (!ret)
 698                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 699         return ret;
 700 }
 701 __setup("transparent_hugepage=", setup_transparent_hugepage);
 702
 703 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 704 {
 705         if (likely(vma->vm_flags & VM_WRITE))
 706                 pmd = pmd_mkwrite(pmd);
 707         return pmd;
 708 }
 709
 710 static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
 711 {
 712         pmd_t entry;
 713         entry = mk_pmd(page, prot);
 714         entry = pmd_mkhuge(entry);
 715         return entry;
 716 }
 717
 718 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
 719                                         struct vm_area_struct *vma,
 720                                         unsigned long address, pmd_t *pmd,
 721                                         struct page *page, gfp_t gfp,
 722                                         unsigned int flags)
 723 {
 724         struct mem_cgroup *memcg;
 725         pgtable_t pgtable;
 726         spinlock_t *ptl;
 727         unsigned long haddr = address & HPAGE_PMD_MASK;
 728
 729         VM_BUG_ON_PAGE(!PageCompound(page), page);
 730
 731         if (mem_cgroup_try_charge(page, mm, gfp, &memcg)) {
 732                 put_page(page);
 733                 count_vm_event(THP_FAULT_FALLBACK);
 734                 return VM_FAULT_FALLBACK;
 735         }
 736
 737         pgtable = pte_alloc_one(mm, haddr);
 738         if (unlikely(!pgtable)) {
 739                 mem_cgroup_cancel_charge(page, memcg);
 740                 put_page(page);
 741                 return VM_FAULT_OOM;
 742         }
 743
 744         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 745         /*
 746          * The memory barrier inside __SetPageUptodate makes sure that
 747          * clear_huge_page writes become visible before the set_pmd_at()
 748          * write.
 749          */
 750         __SetPageUptodate(page);
 751
 752         ptl = pmd_lock(mm, pmd);
 753         if (unlikely(!pmd_none(*pmd))) {
 754                 spin_unlock(ptl);
 755                 mem_cgroup_cancel_charge(page, memcg);
 756                 put_page(page);
 757                 pte_free(mm, pgtable);
 758         } else {
 759                 pmd_t entry;
 760
 761                 /* Deliver the page fault to userland */
 762                 if (userfaultfd_missing(vma)) {
 763                         int ret;
 764
 765                         spin_unlock(ptl);
 766                         mem_cgroup_cancel_charge(page, memcg);
 767                         put_page(page);
 768                         pte_free(mm, pgtable);
 769                         ret = handle_userfault(vma, address, flags,
 770                                                VM_UFFD_MISSING);
 771                         VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 772                         return ret;
 773                 }
 774
 775                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 776                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 777                 page_add_new_anon_rmap(page, vma, haddr);
 778                 mem_cgroup_commit_charge(page, memcg, false);
 779                 lru_cache_add_active_or_unevictable(page, vma);
 780                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 781                 set_pmd_at(mm, haddr, pmd, entry);
 782                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 783                 atomic_long_inc(&mm->nr_ptes);
 784                 spin_unlock(ptl);
 785                 count_vm_event(THP_FAULT_ALLOC);
 786         }
 787
 788         return 0;
 789 }
 790
 791 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 792 {
 793         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
 794 }
 795
 796 /* Caller must hold page table lock. */
 797 static void set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 798                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 799                 struct page *zero_page)
 800 {
 801         pmd_t entry;
 802         entry = mk_pmd(zero_page, vma->vm_page_prot);
 803         entry = pmd_mkhuge(entry);
 804         pgtable_trans_huge_deposit(mm, pmd, pgtable);
 805         set_pmd_at(mm, haddr, pmd, entry);
 806         atomic_long_inc(&mm->nr_ptes);
 807 }
 808
 809 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 810                                unsigned long address, pmd_t *pmd,
 811                                unsigned int flags)
 812 {
 813         gfp_t gfp;
 814         struct page *page;
 815         unsigned long haddr = address & HPAGE_PMD_MASK;
 816
 817         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 818                 return VM_FAULT_FALLBACK;
 819         if (unlikely(anon_vma_prepare(vma)))
 820                 return VM_FAULT_OOM;
 821         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 822                 return VM_FAULT_OOM;
 823         if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm) &&
 824                         transparent_hugepage_use_zero_page()) {
 825                 spinlock_t *ptl;
 826                 pgtable_t pgtable;
 827                 struct page *zero_page;
 828                 bool set;
 829                 int ret;
 830                 pgtable = pte_alloc_one(mm, haddr);
 831                 if (unlikely(!pgtable))
 832                         return VM_FAULT_OOM;
 833                 zero_page = get_huge_zero_page();
 834                 if (unlikely(!zero_page)) {
 835                         pte_free(mm, pgtable);
 836                         count_vm_event(THP_FAULT_FALLBACK);
 837                         return VM_FAULT_FALLBACK;
 838                 }
 839                 ptl = pmd_lock(mm, pmd);
 840                 ret = 0;
 841                 set = false;
 842                 if (pmd_none(*pmd)) {
 843                         if (userfaultfd_missing(vma)) {
 844                                 spin_unlock(ptl);
 845                                 ret = handle_userfault(vma, address, flags,
 846                                                        VM_UFFD_MISSING);
 847                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 848                         } else {
 849                                 set_huge_zero_page(pgtable, mm, vma,
 850                                                    haddr, pmd,
 851                                                    zero_page);
 852                                 spin_unlock(ptl);
 853                                 set = true;
 854                         }
 855                 } else
 856                         spin_unlock(ptl);
 857                 if (!set) {
 858                         pte_free(mm, pgtable);
 859                         put_huge_zero_page();
 860                 }
 861                 return ret;
 862         }
 863         gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
 864         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 865         if (unlikely(!page)) {
 866                 count_vm_event(THP_FAULT_FALLBACK);
 867                 return VM_FAULT_FALLBACK;
 868         }
 869         return __do_huge_pmd_anonymous_page(mm, vma, address, pmd, page, gfp,
 870                                             flags);
 871 }
 872
 873 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 874                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 875                   struct vm_area_struct *vma)
 876 {
 877         spinlock_t *dst_ptl, *src_ptl;
 878         struct page *src_page;
 879         pmd_t pmd;
 880         pgtable_t pgtable;
 881         int ret;
 882
 883         ret = -ENOMEM;
 884         pgtable = pte_alloc_one(dst_mm, addr);
 885         if (unlikely(!pgtable))
 886                 goto out;
 887
 888         dst_ptl = pmd_lock(dst_mm, dst_pmd);
 889         src_ptl = pmd_lockptr(src_mm, src_pmd);
 890         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 891
 892         ret = -EAGAIN;
 893         pmd = *src_pmd;
 894         if (unlikely(!pmd_trans_huge(pmd))) {
 895                 pte_free(dst_mm, pgtable);
 896                 goto out_unlock;
 897         }
 898         /*
 899          * When page table lock is held, the huge zero pmd should not be
 900          * under splitting since we don't split the page itself, only pmd to
 901          * a page table.
 902          */
 903         if (is_huge_zero_pmd(pmd)) {
 904                 struct page *zero_page;
 905                 /*
 906                  * get_huge_zero_page() will never allocate a new page here,
 907                  * since we already have a zero page to copy. It just takes a
 908                  * reference.
 909                  */
 910                 zero_page = get_huge_zero_page();
 911                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
 912                                 zero_page);
 913                 ret = 0;
 914                 goto out_unlock;
 915         }
 916
 917         if (unlikely(pmd_trans_splitting(pmd))) {
 918                 /* split huge page running from under us */
 919                 spin_unlock(src_ptl);
 920                 spin_unlock(dst_ptl);
 921                 pte_free(dst_mm, pgtable);
 922
 923                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 924                 goto out;
 925         }
 926         src_page = pmd_page(pmd);
 927         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
 928         get_page(src_page);
 929         page_dup_rmap(src_page);
 930         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 931
 932         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 933         pmd = pmd_mkold(pmd_wrprotect(pmd));
 934         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
 935         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 936         atomic_long_inc(&dst_mm->nr_ptes);
 937
 938         ret = 0;
 939 out_unlock:
 940         spin_unlock(src_ptl);
 941         spin_unlock(dst_ptl);
 942 out:
 943         return ret;
 944 }
 945
 946 void huge_pmd_set_accessed(struct mm_struct *mm,
 947                            struct vm_area_struct *vma,
 948                            unsigned long address,
 949                            pmd_t *pmd, pmd_t orig_pmd,
 950                            int dirty)
 951 {
 952         spinlock_t *ptl;
 953         pmd_t entry;
 954         unsigned long haddr;
 955
 956         ptl = pmd_lock(mm, pmd);
 957         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 958                 goto unlock;
 959
 960         entry = pmd_mkyoung(orig_pmd);
 961         haddr = address & HPAGE_PMD_MASK;
 962         if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
 963                 update_mmu_cache_pmd(vma, address, pmd);
 964
 965 unlock:
 966         spin_unlock(ptl);
 967 }
 968
 969 /*
 970  * Save CONFIG_DEBUG_PAGEALLOC from faulting falsely on tail pages
 971  * during copy_user_huge_page()'s copy_page_rep(): in the case when
 972  * the source page gets split and a tail freed before copy completes.
 973  * Called under pmd_lock of checked pmd, so safe from splitting itself.
 974  */
 975 static void get_user_huge_page(struct page *page)
 976 {
 977         if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
 978                 struct page *endpage = page + HPAGE_PMD_NR;
 979
 980                 atomic_add(HPAGE_PMD_NR, &page->_count);
 981                 while (++page < endpage)
 982                         get_huge_page_tail(page);
 983         } else {
 984                 get_page(page);
 985         }
 986 }
 987
 988 static void put_user_huge_page(struct page *page)
 989 {
 990         if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
 991                 struct page *endpage = page + HPAGE_PMD_NR;
 992
 993                 while (page < endpage)
 994                         put_page(page++);
 995         } else {
 996                 put_page(page);
 997         }
 998 }
 999
1000 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1001                                         struct vm_area_struct *vma,
1002                                         unsigned long address,
1003                                         pmd_t *pmd, pmd_t orig_pmd,
1004                                         struct page *page,
1005                                         unsigned long haddr)
1006 {
1007         struct mem_cgroup *memcg;
1008         spinlock_t *ptl;
1009         pgtable_t pgtable;
1010         pmd_t _pmd;
1011         int ret = 0, i;
1012         struct page **pages;
1013         unsigned long mmun_start;       /* For mmu_notifiers */
1014         unsigned long mmun_end;         /* For mmu_notifiers */
1015
1016         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1017                         GFP_KERNEL);
1018         if (unlikely(!pages)) {
1019                 ret |= VM_FAULT_OOM;
1020                 goto out;
1021         }
1022
1023         for (i = 0; i < HPAGE_PMD_NR; i++) {
1024                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1025                                                __GFP_OTHER_NODE,
1026                                                vma, address, page_to_nid(page));
1027                 if (unlikely(!pages[i] ||
1028                              mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL,
1029                                                    &memcg))) {
1030                         if (pages[i])
1031                                 put_page(pages[i]);
1032                         while (--i >= 0) {
1033                                 memcg = (void *)page_private(pages[i]);
1034                                 set_page_private(pages[i], 0);
1035                                 mem_cgroup_cancel_charge(pages[i], memcg);
1036                                 put_page(pages[i]);
1037                         }
1038                         kfree(pages);
1039                         ret |= VM_FAULT_OOM;
1040                         goto out;
1041                 }
1042                 set_page_private(pages[i], (unsigned long)memcg);
1043         }
1044
1045         for (i = 0; i < HPAGE_PMD_NR; i++) {
1046                 copy_user_highpage(pages[i], page + i,
1047                                    haddr + PAGE_SIZE * i, vma);
1048                 __SetPageUptodate(pages[i]);
1049                 cond_resched();
1050         }
1051
1052         mmun_start = haddr;
1053         mmun_end   = haddr + HPAGE_PMD_SIZE;
1054         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1055
1056         ptl = pmd_lock(mm, pmd);
1057         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1058                 goto out_free_pages;
1059         VM_BUG_ON_PAGE(!PageHead(page), page);
1060
1061         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1062         /* leave pmd empty until pte is filled */
1063
1064         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1065         pmd_populate(mm, &_pmd, pgtable);
1066
1067         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1068                 pte_t *pte, entry;
1069                 entry = mk_pte(pages[i], vma->vm_page_prot);
1070                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1071                 memcg = (void *)page_private(pages[i]);
1072                 set_page_private(pages[i], 0);
1073                 page_add_new_anon_rmap(pages[i], vma, haddr);
1074                 mem_cgroup_commit_charge(pages[i], memcg, false);
1075                 lru_cache_add_active_or_unevictable(pages[i], vma);
1076                 pte = pte_offset_map(&_pmd, haddr);
1077                 VM_BUG_ON(!pte_none(*pte));
1078                 set_pte_at(mm, haddr, pte, entry);
1079                 pte_unmap(pte);
1080         }
1081         kfree(pages);
1082
1083         smp_wmb(); /* make pte visible before pmd */
1084         pmd_populate(mm, pmd, pgtable);
1085         page_remove_rmap(page);
1086         spin_unlock(ptl);
1087
1088         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1089
1090         ret |= VM_FAULT_WRITE;
1091         put_page(page);
1092
1093 out:
1094         return ret;
1095
1096 out_free_pages:
1097         spin_unlock(ptl);
1098         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1099         for (i = 0; i < HPAGE_PMD_NR; i++) {
1100                 memcg = (void *)page_private(pages[i]);
1101                 set_page_private(pages[i], 0);
1102                 mem_cgroup_cancel_charge(pages[i], memcg);
1103                 put_page(pages[i]);
1104         }
1105         kfree(pages);
1106         goto out;
1107 }
1108
1109 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1110                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1111 {
1112         spinlock_t *ptl;
1113         int ret = 0;
1114         struct page *page = NULL, *new_page;
1115         struct mem_cgroup *memcg;
1116         unsigned long haddr;
1117         unsigned long mmun_start;       /* For mmu_notifiers */
1118         unsigned long mmun_end;         /* For mmu_notifiers */
1119         gfp_t huge_gfp;                 /* for allocation and charge */
1120
1121         ptl = pmd_lockptr(mm, pmd);
1122         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1123         haddr = address & HPAGE_PMD_MASK;
1124         if (is_huge_zero_pmd(orig_pmd))
1125                 goto alloc;
1126         spin_lock(ptl);
1127         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1128                 goto out_unlock;
1129
1130         page = pmd_page(orig_pmd);
1131         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1132         if (page_mapcount(page) == 1) {
1133                 pmd_t entry;
1134                 entry = pmd_mkyoung(orig_pmd);
1135                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1136                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1137                         update_mmu_cache_pmd(vma, address, pmd);
1138                 ret |= VM_FAULT_WRITE;
1139                 goto out_unlock;
1140         }
1141         get_user_huge_page(page);
1142         spin_unlock(ptl);
1143 alloc:
1144         if (transparent_hugepage_enabled(vma) &&
1145             !transparent_hugepage_debug_cow()) {
1146                 huge_gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
1147                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1148         } else
1149                 new_page = NULL;
1150
1151         if (unlikely(!new_page)) {
1152                 if (!page) {
1153                         split_huge_page_pmd(vma, address, pmd);
1154                         ret |= VM_FAULT_FALLBACK;
1155                 } else {
1156                         ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1157                                         pmd, orig_pmd, page, haddr);
1158                         if (ret & VM_FAULT_OOM) {
1159                                 split_huge_page(page);
1160                                 ret |= VM_FAULT_FALLBACK;
1161                         }
1162                         put_user_huge_page(page);
1163                 }
1164                 count_vm_event(THP_FAULT_FALLBACK);
1165                 goto out;
1166         }
1167
1168         if (unlikely(mem_cgroup_try_charge(new_page, mm, huge_gfp, &memcg))) {
1169                 put_page(new_page);
1170                 if (page) {
1171                         split_huge_page(page);
1172                         put_user_huge_page(page);
1173                 } else
1174                         split_huge_page_pmd(vma, address, pmd);
1175                 ret |= VM_FAULT_FALLBACK;
1176                 count_vm_event(THP_FAULT_FALLBACK);
1177                 goto out;
1178         }
1179
1180         count_vm_event(THP_FAULT_ALLOC);
1181
1182         if (!page)
1183                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1184         else
1185                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1186         __SetPageUptodate(new_page);
1187
1188         mmun_start = haddr;
1189         mmun_end   = haddr + HPAGE_PMD_SIZE;
1190         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1191
1192         spin_lock(ptl);
1193         if (page)
1194                 put_user_huge_page(page);
1195         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1196                 spin_unlock(ptl);
1197                 mem_cgroup_cancel_charge(new_page, memcg);
1198                 put_page(new_page);
1199                 goto out_mn;
1200         } else {
1201                 pmd_t entry;
1202                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1203                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1204                 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1205                 page_add_new_anon_rmap(new_page, vma, haddr);
1206                 mem_cgroup_commit_charge(new_page, memcg, false);
1207                 lru_cache_add_active_or_unevictable(new_page, vma);
1208                 set_pmd_at(mm, haddr, pmd, entry);
1209                 update_mmu_cache_pmd(vma, address, pmd);
1210                 if (!page) {
1211                         add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1212                         put_huge_zero_page();
1213                 } else {
1214                         VM_BUG_ON_PAGE(!PageHead(page), page);
1215                         page_remove_rmap(page);
1216                         put_page(page);
1217                 }
1218                 ret |= VM_FAULT_WRITE;
1219         }
1220         spin_unlock(ptl);
1221 out_mn:
1222         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1223 out:
1224         return ret;
1225 out_unlock:
1226         spin_unlock(ptl);
1227         return ret;
1228 }
1229
1230 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1231                                    unsigned long addr,
1232                                    pmd_t *pmd,
1233                                    unsigned int flags)
1234 {
1235         struct mm_struct *mm = vma->vm_mm;
1236         struct page *page = NULL;
1237
1238         assert_spin_locked(pmd_lockptr(mm, pmd));
1239
1240         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1241                 goto out;
1242
1243         /* Avoid dumping huge zero page */
1244         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1245                 return ERR_PTR(-EFAULT);
1246
1247         /* Full NUMA hinting faults to serialise migration in fault paths */
1248         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1249                 goto out;
1250
1251         page = pmd_page(*pmd);
1252         VM_BUG_ON_PAGE(!PageHead(page), page);
1253         if (flags & FOLL_TOUCH) {
1254                 pmd_t _pmd;
1255                 /*
1256                  * We should set the dirty bit only for FOLL_WRITE but
1257                  * for now the dirty bit in the pmd is meaningless.
1258                  * And if the dirty bit will become meaningful and
1259                  * we'll only set it with FOLL_WRITE, an atomic
1260                  * set_bit will be required on the pmd to set the
1261                  * young bit, instead of the current set_pmd_at.
1262                  */
1263                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1264                 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1265                                           pmd, _pmd,  1))
1266                         update_mmu_cache_pmd(vma, addr, pmd);
1267         }
1268         if ((flags & FOLL_POPULATE) && (vma->vm_flags & VM_LOCKED)) {
1269                 if (page->mapping && trylock_page(page)) {
1270                         lru_add_drain();
1271                         if (page->mapping)
1272                                 mlock_vma_page(page);
1273                         unlock_page(page);
1274                 }
1275         }
1276         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1277         VM_BUG_ON_PAGE(!PageCompound(page), page);
1278         if (flags & FOLL_GET)
1279                 get_page_foll(page);
1280
1281 out:
1282         return page;
1283 }
1284
1285 /* NUMA hinting page fault entry point for trans huge pmds */
1286 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1287                                 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1288 {
1289         spinlock_t *ptl;
1290         struct anon_vma *anon_vma = NULL;
1291         struct page *page;
1292         unsigned long haddr = addr & HPAGE_PMD_MASK;
1293         int page_nid = -1, this_nid = numa_node_id();
1294         int target_nid, last_cpupid = -1;
1295         bool page_locked;
1296         bool migrated = false;
1297         bool was_writable;
1298         int flags = 0;
1299
1300         /* A PROT_NONE fault should not end up here */
1301         BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
1302
1303         ptl = pmd_lock(mm, pmdp);
1304         if (unlikely(!pmd_same(pmd, *pmdp)))
1305                 goto out_unlock;
1306
1307         /*
1308          * If there are potential migrations, wait for completion and retry
1309          * without disrupting NUMA hinting information. Do not relock and
1310          * check_same as the page may no longer be mapped.
1311          */
1312         if (unlikely(pmd_trans_migrating(*pmdp))) {
1313                 page = pmd_page(*pmdp);
1314                 spin_unlock(ptl);
1315                 wait_on_page_locked(page);
1316                 goto out;
1317         }
1318
1319         page = pmd_page(pmd);
1320         BUG_ON(is_huge_zero_page(page));
1321         page_nid = page_to_nid(page);
1322         last_cpupid = page_cpupid_last(page);
1323         count_vm_numa_event(NUMA_HINT_FAULTS);
1324         if (page_nid == this_nid) {
1325                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1326                 flags |= TNF_FAULT_LOCAL;
1327         }
1328
1329         /* See similar comment in do_numa_page for explanation */
1330         if (!(vma->vm_flags & VM_WRITE))
1331                 flags |= TNF_NO_GROUP;
1332
1333         /*
1334          * Acquire the page lock to serialise THP migrations but avoid dropping
1335          * page_table_lock if at all possible
1336          */
1337         page_locked = trylock_page(page);
1338         target_nid = mpol_misplaced(page, vma, haddr);
1339         if (target_nid == -1) {
1340                 /* If the page was locked, there are no parallel migrations */
1341                 if (page_locked)
1342                         goto clear_pmdnuma;
1343         }
1344
1345         /* Migration could have started since the pmd_trans_migrating check */
1346         if (!page_locked) {
1347                 spin_unlock(ptl);
1348                 wait_on_page_locked(page);
1349                 page_nid = -1;
1350                 goto out;
1351         }
1352
1353         /*
1354          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1355          * to serialises splits
1356          */
1357         get_page(page);
1358         spin_unlock(ptl);
1359         anon_vma = page_lock_anon_vma_read(page);
1360
1361         /* Confirm the PMD did not change while page_table_lock was released */
1362         spin_lock(ptl);
1363         if (unlikely(!pmd_same(pmd, *pmdp))) {
1364                 unlock_page(page);
1365                 put_page(page);
1366                 page_nid = -1;
1367                 goto out_unlock;
1368         }
1369
1370         /* Bail if we fail to protect against THP splits for any reason */
1371         if (unlikely(!anon_vma)) {
1372                 put_page(page);
1373                 page_nid = -1;
1374                 goto clear_pmdnuma;
1375         }
1376
1377         /*
1378          * Migrate the THP to the requested node, returns with page unlocked
1379          * and access rights restored.
1380          */
1381         spin_unlock(ptl);
1382         migrated = migrate_misplaced_transhuge_page(mm, vma,
1383                                 pmdp, pmd, addr, page, target_nid);
1384         if (migrated) {
1385                 flags |= TNF_MIGRATED;
1386                 page_nid = target_nid;
1387         } else
1388                 flags |= TNF_MIGRATE_FAIL;
1389
1390         goto out;
1391 clear_pmdnuma:
1392         BUG_ON(!PageLocked(page));
1393         was_writable = pmd_write(pmd);
1394         pmd = pmd_modify(pmd, vma->vm_page_prot);
1395         pmd = pmd_mkyoung(pmd);
1396         if (was_writable)
1397                 pmd = pmd_mkwrite(pmd);
1398         set_pmd_at(mm, haddr, pmdp, pmd);
1399         update_mmu_cache_pmd(vma, addr, pmdp);
1400         unlock_page(page);
1401 out_unlock:
1402         spin_unlock(ptl);
1403
1404 out:
1405         if (anon_vma)
1406                 page_unlock_anon_vma_read(anon_vma);
1407
1408         if (page_nid != -1)
1409                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
1410
1411         return 0;
1412 }
1413
1414 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1415                  pmd_t *pmd, unsigned long addr)
1416 {
1417         spinlock_t *ptl;
1418         int ret = 0;
1419
1420         if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1421                 struct page *page;
1422                 pgtable_t pgtable;
1423                 pmd_t orig_pmd;
1424                 /*
1425                  * For architectures like ppc64 we look at deposited pgtable
1426                  * when calling pmdp_huge_get_and_clear. So do the
1427                  * pgtable_trans_huge_withdraw after finishing pmdp related
1428                  * operations.
1429                  */
1430                 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1431                                                         tlb->fullmm);
1432                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1433                 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1434                 if (is_huge_zero_pmd(orig_pmd)) {
1435                         atomic_long_dec(&tlb->mm->nr_ptes);
1436                         spin_unlock(ptl);
1437                         put_huge_zero_page();
1438                 } else {
1439                         page = pmd_page(orig_pmd);
1440                         page_remove_rmap(page);
1441                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1442                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1443                         VM_BUG_ON_PAGE(!PageHead(page), page);
1444                         atomic_long_dec(&tlb->mm->nr_ptes);
1445                         spin_unlock(ptl);
1446                         tlb_remove_page(tlb, page);
1447                 }
1448                 pte_free(tlb->mm, pgtable);
1449                 ret = 1;
1450         }
1451         return ret;
1452 }
1453
1454 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1455                   unsigned long old_addr,
1456                   unsigned long new_addr, unsigned long old_end,
1457                   pmd_t *old_pmd, pmd_t *new_pmd)
1458 {
1459         spinlock_t *old_ptl, *new_ptl;
1460         int ret = 0;
1461         pmd_t pmd;
1462
1463         struct mm_struct *mm = vma->vm_mm;
1464
1465         if ((old_addr & ~HPAGE_PMD_MASK) ||
1466             (new_addr & ~HPAGE_PMD_MASK) ||
1467             old_end - old_addr < HPAGE_PMD_SIZE ||
1468             (new_vma->vm_flags & VM_NOHUGEPAGE))
1469                 goto out;
1470
1471         /*
1472          * The destination pmd shouldn't be established, free_pgtables()
1473          * should have release it.
1474          */
1475         if (WARN_ON(!pmd_none(*new_pmd))) {
1476                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1477                 goto out;
1478         }
1479
1480         /*
1481          * We don't have to worry about the ordering of src and dst
1482          * ptlocks because exclusive mmap_sem prevents deadlock.
1483          */
1484         ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl);
1485         if (ret == 1) {
1486                 new_ptl = pmd_lockptr(mm, new_pmd);
1487                 if (new_ptl != old_ptl)
1488                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1489                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1490                 VM_BUG_ON(!pmd_none(*new_pmd));
1491
1492                 if (pmd_move_must_withdraw(new_ptl, old_ptl)) {
1493                         pgtable_t pgtable;
1494                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1495                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1496                 }
1497                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1498                 if (new_ptl != old_ptl)
1499                         spin_unlock(new_ptl);
1500                 spin_unlock(old_ptl);
1501         }
1502 out:
1503         return ret;
1504 }
1505
1506 /*
1507  * Returns
1508  *  - 0 if PMD could not be locked
1509  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1510  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1511  */
1512 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1513                 unsigned long addr, pgprot_t newprot, int prot_numa)
1514 {
1515         struct mm_struct *mm = vma->vm_mm;
1516         spinlock_t *ptl;
1517         int ret = 0;
1518
1519         if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1520                 pmd_t entry;
1521                 bool preserve_write = prot_numa && pmd_write(*pmd);
1522                 ret = 1;
1523
1524                 /*
1525                  * Avoid trapping faults against the zero page. The read-only
1526                  * data is likely to be read-cached on the local CPU and
1527                  * local/remote hits to the zero page are not interesting.
1528                  */
1529                 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1530                         spin_unlock(ptl);
1531                         return ret;
1532                 }
1533
1534                 if (!prot_numa || !pmd_protnone(*pmd)) {
1535                         entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
1536                         entry = pmd_modify(entry, newprot);
1537                         if (preserve_write)
1538                                 entry = pmd_mkwrite(entry);
1539                         ret = HPAGE_PMD_NR;
1540                         set_pmd_at(mm, addr, pmd, entry);
1541                         BUG_ON(!preserve_write && pmd_write(entry));
1542                 }
1543                 spin_unlock(ptl);
1544         }
1545
1546         return ret;
1547 }
1548
1549 /*
1550  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1551  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1552  *
1553  * Note that if it returns 1, this routine returns without unlocking page
1554  * table locks. So callers must unlock them.
1555  */
1556 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1557                 spinlock_t **ptl)
1558 {
1559         *ptl = pmd_lock(vma->vm_mm, pmd);
1560         if (likely(pmd_trans_huge(*pmd))) {
1561                 if (unlikely(pmd_trans_splitting(*pmd))) {
1562                         spin_unlock(*ptl);
1563                         wait_split_huge_page(vma->anon_vma, pmd);
1564                         return -1;
1565                 } else {
1566                         /* Thp mapped by 'pmd' is stable, so we can
1567                          * handle it as it is. */
1568                         return 1;
1569                 }
1570         }
1571         spin_unlock(*ptl);
1572         return 0;
1573 }
1574
1575 /*
1576  * This function returns whether a given @page is mapped onto the @address
1577  * in the virtual space of @mm.
1578  *
1579  * When it's true, this function returns *pmd with holding the page table lock
1580  * and passing it back to the caller via @ptl.
1581  * If it's false, returns NULL without holding the page table lock.
1582  */
1583 pmd_t *page_check_address_pmd(struct page *page,
1584                               struct mm_struct *mm,
1585                               unsigned long address,
1586                               enum page_check_address_pmd_flag flag,
1587                               spinlock_t **ptl)
1588 {
1589         pgd_t *pgd;
1590         pud_t *pud;
1591         pmd_t *pmd;
1592
1593         if (address & ~HPAGE_PMD_MASK)
1594                 return NULL;
1595
1596         pgd = pgd_offset(mm, address);
1597         if (!pgd_present(*pgd))
1598                 return NULL;
1599         pud = pud_offset(pgd, address);
1600         if (!pud_present(*pud))
1601                 return NULL;
1602         pmd = pmd_offset(pud, address);
1603
1604         *ptl = pmd_lock(mm, pmd);
1605         if (!pmd_present(*pmd))
1606                 goto unlock;
1607         if (pmd_page(*pmd) != page)
1608                 goto unlock;
1609         /*
1610          * split_vma() may create temporary aliased mappings. There is
1611          * no risk as long as all huge pmd are found and have their
1612          * splitting bit set before __split_huge_page_refcount
1613          * runs. Finding the same huge pmd more than once during the
1614          * same rmap walk is not a problem.
1615          */
1616         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1617             pmd_trans_splitting(*pmd))
1618                 goto unlock;
1619         if (pmd_trans_huge(*pmd)) {
1620                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1621                           !pmd_trans_splitting(*pmd));
1622                 return pmd;
1623         }
1624 unlock:
1625         spin_unlock(*ptl);
1626         return NULL;
1627 }
1628
1629 static int __split_huge_page_splitting(struct page *page,
1630                                        struct vm_area_struct *vma,
1631                                        unsigned long address)
1632 {
1633         struct mm_struct *mm = vma->vm_mm;
1634         spinlock_t *ptl;
1635         pmd_t *pmd;
1636         int ret = 0;
1637         /* For mmu_notifiers */
1638         const unsigned long mmun_start = address;
1639         const unsigned long mmun_end   = address + HPAGE_PMD_SIZE;
1640
1641         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1642         pmd = page_check_address_pmd(page, mm, address,
1643                         PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl);
1644         if (pmd) {
1645                 /*
1646                  * We can't temporarily set the pmd to null in order
1647                  * to split it, the pmd must remain marked huge at all
1648                  * times or the VM won't take the pmd_trans_huge paths
1649                  * and it won't wait on the anon_vma->root->rwsem to
1650                  * serialize against split_huge_page*.
1651                  */
1652                 pmdp_splitting_flush(vma, address, pmd);
1653
1654                 ret = 1;
1655                 spin_unlock(ptl);
1656         }
1657         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1658
1659         return ret;
1660 }
1661
1662 static void __split_huge_page_refcount(struct page *page,
1663                                        struct list_head *list)
1664 {
1665         int i;
1666         struct zone *zone = page_zone(page);
1667         struct lruvec *lruvec;
1668         int tail_count = 0;
1669
1670         /* prevent PageLRU to go away from under us, and freeze lru stats */
1671         spin_lock_irq(&zone->lru_lock);
1672         lruvec = mem_cgroup_page_lruvec(page, zone);
1673
1674         compound_lock(page);
1675         /* complete memcg works before add pages to LRU */
1676         mem_cgroup_split_huge_fixup(page);
1677
1678         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1679                 struct page *page_tail = page + i;
1680
1681                 /* tail_page->_mapcount cannot change */
1682                 BUG_ON(page_mapcount(page_tail) < 0);
1683                 tail_count += page_mapcount(page_tail);
1684                 /* check for overflow */
1685                 BUG_ON(tail_count < 0);
1686                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1687                 /*
1688                  * tail_page->_count is zero and not changing from
1689                  * under us. But get_page_unless_zero() may be running
1690                  * from under us on the tail_page. If we used
1691                  * atomic_set() below instead of atomic_add(), we
1692                  * would then run atomic_set() concurrently with
1693                  * get_page_unless_zero(), and atomic_set() is
1694                  * implemented in C not using locked ops. spin_unlock
1695                  * on x86 sometime uses locked ops because of PPro
1696                  * errata 66, 92, so unless somebody can guarantee
1697                  * atomic_set() here would be safe on all archs (and
1698                  * not only on x86), it's safer to use atomic_add().
1699                  */
1700                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1701                            &page_tail->_count);
1702
1703                 /* after clearing PageTail the gup refcount can be released */
1704                 smp_mb__after_atomic();
1705
1706                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1707                 page_tail->flags |= (page->flags &
1708                                      ((1L << PG_referenced) |
1709                                       (1L << PG_swapbacked) |
1710                                       (1L << PG_mlocked) |
1711                                       (1L << PG_uptodate) |
1712                                       (1L << PG_active) |
1713                                       (1L << PG_unevictable)));
1714                 page_tail->flags |= (1L << PG_dirty);
1715
1716                 /* clear PageTail before overwriting first_page */
1717                 smp_wmb();
1718
1719                 /*
1720                  * __split_huge_page_splitting() already set the
1721                  * splitting bit in all pmd that could map this
1722                  * hugepage, that will ensure no CPU can alter the
1723                  * mapcount on the head page. The mapcount is only
1724                  * accounted in the head page and it has to be
1725                  * transferred to all tail pages in the below code. So
1726                  * for this code to be safe, the split the mapcount
1727                  * can't change. But that doesn't mean userland can't
1728                  * keep changing and reading the page contents while
1729                  * we transfer the mapcount, so the pmd splitting
1730                  * status is achieved setting a reserved bit in the
1731                  * pmd, not by clearing the present bit.
1732                 */
1733                 page_tail->_mapcount = page->_mapcount;
1734
1735                 BUG_ON(page_tail->mapping);
1736                 page_tail->mapping = page->mapping;
1737
1738                 page_tail->index = page->index + i;
1739                 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
1740
1741                 BUG_ON(!PageAnon(page_tail));
1742                 BUG_ON(!PageUptodate(page_tail));
1743                 BUG_ON(!PageDirty(page_tail));
1744                 BUG_ON(!PageSwapBacked(page_tail));
1745
1746                 lru_add_page_tail(page, page_tail, lruvec, list);
1747         }
1748         atomic_sub(tail_count, &page->_count);
1749         BUG_ON(atomic_read(&page->_count) <= 0);
1750
1751         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1752
1753         ClearPageCompound(page);
1754         compound_unlock(page);
1755         spin_unlock_irq(&zone->lru_lock);
1756
1757         for (i = 1; i < HPAGE_PMD_NR; i++) {
1758                 struct page *page_tail = page + i;
1759                 BUG_ON(page_count(page_tail) <= 0);
1760                 /*
1761                  * Tail pages may be freed if there wasn't any mapping
1762                  * like if add_to_swap() is running on a lru page that
1763                  * had its mapping zapped. And freeing these pages
1764                  * requires taking the lru_lock so we do the put_page
1765                  * of the tail pages after the split is complete.
1766                  */
1767                 put_page(page_tail);
1768         }
1769
1770         /*
1771          * Only the head page (now become a regular page) is required
1772          * to be pinned by the caller.
1773          */
1774         BUG_ON(page_count(page) <= 0);
1775 }
1776
1777 static int __split_huge_page_map(struct page *page,
1778                                  struct vm_area_struct *vma,
1779                                  unsigned long address)
1780 {
1781         struct mm_struct *mm = vma->vm_mm;
1782         spinlock_t *ptl;
1783         pmd_t *pmd, _pmd;
1784         int ret = 0, i;
1785         pgtable_t pgtable;
1786         unsigned long haddr;
1787
1788         pmd = page_check_address_pmd(page, mm, address,
1789                         PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl);
1790         if (pmd) {
1791                 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1792                 pmd_populate(mm, &_pmd, pgtable);
1793                 if (pmd_write(*pmd))
1794                         BUG_ON(page_mapcount(page) != 1);
1795
1796                 haddr = address;
1797                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1798                         pte_t *pte, entry;
1799                         BUG_ON(PageCompound(page+i));
1800                         /*
1801                          * Note that NUMA hinting access restrictions are not
1802                          * transferred to avoid any possibility of altering
1803                          * permissions across VMAs.
1804                          */
1805                         entry = mk_pte(page + i, vma->vm_page_prot);
1806                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1807                         if (!pmd_write(*pmd))
1808                                 entry = pte_wrprotect(entry);
1809                         if (!pmd_young(*pmd))
1810                                 entry = pte_mkold(entry);
1811                         pte = pte_offset_map(&_pmd, haddr);
1812                         BUG_ON(!pte_none(*pte));
1813                         set_pte_at(mm, haddr, pte, entry);
1814                         pte_unmap(pte);
1815                 }
1816
1817                 smp_wmb(); /* make pte visible before pmd */
1818                 /*
1819                  * Up to this point the pmd is present and huge and
1820                  * userland has the whole access to the hugepage
1821                  * during the split (which happens in place). If we
1822                  * overwrite the pmd with the not-huge version
1823                  * pointing to the pte here (which of course we could
1824                  * if all CPUs were bug free), userland could trigger
1825                  * a small page size TLB miss on the small sized TLB
1826                  * while the hugepage TLB entry is still established
1827                  * in the huge TLB. Some CPU doesn't like that. See
1828                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1829                  * Erratum 383 on page 93. Intel should be safe but is
1830                  * also warns that it's only safe if the permission
1831                  * and cache attributes of the two entries loaded in
1832                  * the two TLB is identical (which should be the case
1833                  * here). But it is generally safer to never allow
1834                  * small and huge TLB entries for the same virtual
1835                  * address to be loaded simultaneously. So instead of
1836                  * doing "pmd_populate(); flush_tlb_range();" we first
1837                  * mark the current pmd notpresent (atomically because
1838                  * here the pmd_trans_huge and pmd_trans_splitting
1839                  * must remain set at all times on the pmd until the
1840                  * split is complete for this pmd), then we flush the
1841                  * SMP TLB and finally we write the non-huge version
1842                  * of the pmd entry with pmd_populate.
1843                  */
1844                 pmdp_invalidate(vma, address, pmd);
1845                 pmd_populate(mm, pmd, pgtable);
1846                 ret = 1;
1847                 spin_unlock(ptl);
1848         }
1849
1850         return ret;
1851 }
1852
1853 /* must be called with anon_vma->root->rwsem held */
1854 static void __split_huge_page(struct page *page,
1855                               struct anon_vma *anon_vma,
1856                               struct list_head *list)
1857 {
1858         int mapcount, mapcount2;
1859         pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1860         struct anon_vma_chain *avc;
1861
1862         BUG_ON(!PageHead(page));
1863         BUG_ON(PageTail(page));
1864
1865         mapcount = 0;
1866         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1867                 struct vm_area_struct *vma = avc->vma;
1868                 unsigned long addr = vma_address(page, vma);
1869                 BUG_ON(is_vma_temporary_stack(vma));
1870                 mapcount += __split_huge_page_splitting(page, vma, addr);
1871         }
1872         /*
1873          * It is critical that new vmas are added to the tail of the
1874          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1875          * and establishes a child pmd before
1876          * __split_huge_page_splitting() freezes the parent pmd (so if
1877          * we fail to prevent copy_huge_pmd() from running until the
1878          * whole __split_huge_page() is complete), we will still see
1879          * the newly established pmd of the child later during the
1880          * walk, to be able to set it as pmd_trans_splitting too.
1881          */
1882         if (mapcount != page_mapcount(page)) {
1883                 pr_err("mapcount %d page_mapcount %d\n",
1884                         mapcount, page_mapcount(page));
1885                 BUG();
1886         }
1887
1888         __split_huge_page_refcount(page, list);
1889
1890         mapcount2 = 0;
1891         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1892                 struct vm_area_struct *vma = avc->vma;
1893                 unsigned long addr = vma_address(page, vma);
1894                 BUG_ON(is_vma_temporary_stack(vma));
1895                 mapcount2 += __split_huge_page_map(page, vma, addr);
1896         }
1897         if (mapcount != mapcount2) {
1898                 pr_err("mapcount %d mapcount2 %d page_mapcount %d\n",
1899                         mapcount, mapcount2, page_mapcount(page));
1900                 BUG();
1901         }
1902 }
1903
1904 /*
1905  * Split a hugepage into normal pages. This doesn't change the position of head
1906  * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1907  * @list. Both head page and tail pages will inherit mapping, flags, and so on
1908  * from the hugepage.
1909  * Return 0 if the hugepage is split successfully otherwise return 1.
1910  */
1911 int split_huge_page_to_list(struct page *page, struct list_head *list)
1912 {
1913         struct anon_vma *anon_vma;
1914         int ret = 1;
1915
1916         BUG_ON(is_huge_zero_page(page));
1917         BUG_ON(!PageAnon(page));
1918
1919         /*
1920          * The caller does not necessarily hold an mmap_sem that would prevent
1921          * the anon_vma disappearing so we first we take a reference to it
1922          * and then lock the anon_vma for write. This is similar to
1923          * page_lock_anon_vma_read except the write lock is taken to serialise
1924          * against parallel split or collapse operations.
1925          */
1926         anon_vma = page_get_anon_vma(page);
1927         if (!anon_vma)
1928                 goto out;
1929         anon_vma_lock_write(anon_vma);
1930
1931         ret = 0;
1932         if (!PageCompound(page))
1933                 goto out_unlock;
1934
1935         BUG_ON(!PageSwapBacked(page));
1936         __split_huge_page(page, anon_vma, list);
1937         count_vm_event(THP_SPLIT);
1938
1939         BUG_ON(PageCompound(page));
1940 out_unlock:
1941         anon_vma_unlock_write(anon_vma);
1942         put_anon_vma(anon_vma);
1943 out:
1944         return ret;
1945 }
1946
1947 #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
1948
1949 int hugepage_madvise(struct vm_area_struct *vma,
1950                      unsigned long *vm_flags, int advice)
1951 {
1952         switch (advice) {
1953         case MADV_HUGEPAGE:
1954 #ifdef CONFIG_S390
1955                 /*
1956                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1957                  * can't handle this properly after s390_enable_sie, so we simply
1958                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
1959                  */
1960                 if (mm_has_pgste(vma->vm_mm))
1961                         return 0;
1962 #endif
1963                 /*
1964                  * Be somewhat over-protective like KSM for now!
1965                  */
1966                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1967                         return -EINVAL;
1968                 *vm_flags &= ~VM_NOHUGEPAGE;
1969                 *vm_flags |= VM_HUGEPAGE;
1970                 /*
1971                  * If the vma become good for khugepaged to scan,
1972                  * register it here without waiting a page fault that
1973                  * may not happen any time soon.
1974                  */
1975                 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
1976                         return -ENOMEM;
1977                 break;
1978         case MADV_NOHUGEPAGE:
1979                 /*
1980                  * Be somewhat over-protective like KSM for now!
1981                  */
1982                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1983                         return -EINVAL;
1984                 *vm_flags &= ~VM_HUGEPAGE;
1985                 *vm_flags |= VM_NOHUGEPAGE;
1986                 /*
1987                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1988                  * this vma even if we leave the mm registered in khugepaged if
1989                  * it got registered before VM_NOHUGEPAGE was set.
1990                  */
1991                 break;
1992         }
1993
1994         return 0;
1995 }
1996
1997 static int __init khugepaged_slab_init(void)
1998 {
1999         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
2000                                           sizeof(struct mm_slot),
2001                                           __alignof__(struct mm_slot), 0, NULL);
2002         if (!mm_slot_cache)
2003                 return -ENOMEM;
2004
2005         return 0;
2006 }
2007
2008 static void __init khugepaged_slab_exit(void)
2009 {
2010         kmem_cache_destroy(mm_slot_cache);
2011 }
2012
2013 static inline struct mm_slot *alloc_mm_slot(void)
2014 {
2015         if (!mm_slot_cache)     /* initialization failed */
2016                 return NULL;
2017         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
2018 }
2019
2020 static inline void free_mm_slot(struct mm_slot *mm_slot)
2021 {
2022         kmem_cache_free(mm_slot_cache, mm_slot);
2023 }
2024
2025 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
2026 {
2027         struct mm_slot *mm_slot;
2028
2029         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
2030                 if (mm == mm_slot->mm)
2031                         return mm_slot;
2032
2033         return NULL;
2034 }
2035
2036 static void insert_to_mm_slots_hash(struct mm_struct *mm,
2037                                     struct mm_slot *mm_slot)
2038 {
2039         mm_slot->mm = mm;
2040         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
2041 }
2042
2043 static inline int khugepaged_test_exit(struct mm_struct *mm)
2044 {
2045         return atomic_read(&mm->mm_users) == 0;
2046 }
2047
2048 int __khugepaged_enter(struct mm_struct *mm)
2049 {
2050         struct mm_slot *mm_slot;
2051         int wakeup;
2052
2053         mm_slot = alloc_mm_slot();
2054         if (!mm_slot)
2055                 return -ENOMEM;
2056
2057         /* __khugepaged_exit() must not run from under us */
2058         VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
2059         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2060                 free_mm_slot(mm_slot);
2061                 return 0;
2062         }
2063
2064         spin_lock(&khugepaged_mm_lock);
2065         insert_to_mm_slots_hash(mm, mm_slot);
2066         /*
2067          * Insert just behind the scanning cursor, to let the area settle
2068          * down a little.
2069          */
2070         wakeup = list_empty(&khugepaged_scan.mm_head);
2071         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2072         spin_unlock(&khugepaged_mm_lock);
2073
2074         atomic_inc(&mm->mm_count);
2075         if (wakeup)
2076                 wake_up_interruptible(&khugepaged_wait);
2077
2078         return 0;
2079 }
2080
2081 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
2082                                unsigned long vm_flags)
2083 {
2084         unsigned long hstart, hend;
2085         if (!vma->anon_vma)
2086                 /*
2087                  * Not yet faulted in so we will register later in the
2088                  * page fault if needed.
2089                  */
2090                 return 0;
2091         if (vma->vm_ops)
2092                 /* khugepaged not yet working on file or special mappings */
2093                 return 0;
2094         VM_BUG_ON_VMA(vm_flags & VM_NO_THP, vma);
2095         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2096         hend = vma->vm_end & HPAGE_PMD_MASK;
2097         if (hstart < hend)
2098                 return khugepaged_enter(vma, vm_flags);
2099         return 0;
2100 }
2101
2102 void __khugepaged_exit(struct mm_struct *mm)
2103 {
2104         struct mm_slot *mm_slot;
2105         int free = 0;
2106
2107         spin_lock(&khugepaged_mm_lock);
2108         mm_slot = get_mm_slot(mm);
2109         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
2110                 hash_del(&mm_slot->hash);
2111                 list_del(&mm_slot->mm_node);
2112                 free = 1;
2113         }
2114         spin_unlock(&khugepaged_mm_lock);
2115
2116         if (free) {
2117                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2118                 free_mm_slot(mm_slot);
2119                 mmdrop(mm);
2120         } else if (mm_slot) {
2121                 /*
2122                  * This is required to serialize against
2123                  * khugepaged_test_exit() (which is guaranteed to run
2124                  * under mmap sem read mode). Stop here (after we
2125                  * return all pagetables will be destroyed) until
2126                  * khugepaged has finished working on the pagetables
2127                  * under the mmap_sem.
2128                  */
2129                 down_write(&mm->mmap_sem);
2130                 up_write(&mm->mmap_sem);
2131         }
2132 }
2133
2134 static void release_pte_page(struct page *page)
2135 {
2136         /* 0 stands for page_is_file_cache(page) == false */
2137         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2138         unlock_page(page);
2139         putback_lru_page(page);
2140 }
2141
2142 static void release_pte_pages(pte_t *pte, pte_t *_pte)
2143 {
2144         while (--_pte >= pte) {
2145                 pte_t pteval = *_pte;
2146                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
2147                         release_pte_page(pte_page(pteval));
2148         }
2149 }
2150
2151 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2152                                         unsigned long address,
2153                                         pte_t *pte)
2154 {
2155         struct page *page;
2156         pte_t *_pte;
2157         int none_or_zero = 0;
2158         bool referenced = false, writable = false;
2159         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2160              _pte++, address += PAGE_SIZE) {
2161                 pte_t pteval = *_pte;
2162                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2163                         if (!userfaultfd_armed(vma) &&
2164                             ++none_or_zero <= khugepaged_max_ptes_none)
2165                                 continue;
2166                         else
2167                                 goto out;
2168                 }
2169                 if (!pte_present(pteval))
2170                         goto out;
2171                 page = vm_normal_page(vma, address, pteval);
2172                 if (unlikely(!page))
2173                         goto out;
2174
2175                 VM_BUG_ON_PAGE(PageCompound(page), page);
2176                 VM_BUG_ON_PAGE(!PageAnon(page), page);
2177                 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2178
2179                 /*
2180                  * We can do it before isolate_lru_page because the
2181                  * page can't be freed from under us. NOTE: PG_lock
2182                  * is needed to serialize against split_huge_page
2183                  * when invoked from the VM.
2184                  */
2185                 if (!trylock_page(page))
2186                         goto out;
2187
2188                 /*
2189                  * cannot use mapcount: can't collapse if there's a gup pin.
2190                  * The page must only be referenced by the scanned process
2191                  * and page swap cache.
2192                  */
2193                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2194                         unlock_page(page);
2195                         goto out;
2196                 }
2197                 if (pte_write(pteval)) {
2198                         writable = true;
2199                 } else {
2200                         if (PageSwapCache(page) && !reuse_swap_page(page)) {
2201                                 unlock_page(page);
2202                                 goto out;
2203                         }
2204                         /*
2205                          * Page is not in the swap cache. It can be collapsed
2206                          * into a THP.
2207                          */
2208                 }
2209
2210                 /*
2211                  * Isolate the page to avoid collapsing an hugepage
2212                  * currently in use by the VM.
2213                  */
2214                 if (isolate_lru_page(page)) {
2215                         unlock_page(page);
2216                         goto out;
2217                 }
2218                 /* 0 stands for page_is_file_cache(page) == false */
2219                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2220                 VM_BUG_ON_PAGE(!PageLocked(page), page);
2221                 VM_BUG_ON_PAGE(PageLRU(page), page);
2222
2223                 /* If there is no mapped pte young don't collapse the page */
2224                 if (pte_young(pteval) || PageReferenced(page) ||
2225                     mmu_notifier_test_young(vma->vm_mm, address))
2226                         referenced = true;
2227         }
2228         if (likely(referenced && writable))
2229                 return 1;
2230 out:
2231         release_pte_pages(pte, _pte);
2232         return 0;
2233 }
2234
2235 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2236                                       struct vm_area_struct *vma,
2237                                       unsigned long address,
2238                                       spinlock_t *ptl)
2239 {
2240         pte_t *_pte;
2241         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2242                 pte_t pteval = *_pte;
2243                 struct page *src_page;
2244
2245                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2246                         clear_user_highpage(page, address);
2247                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2248                         if (is_zero_pfn(pte_pfn(pteval))) {
2249                                 /*
2250                                  * ptl mostly unnecessary.
2251                                  */
2252                                 spin_lock(ptl);
2253                                 /*
2254                                  * paravirt calls inside pte_clear here are
2255                                  * superfluous.
2256                                  */
2257                                 pte_clear(vma->vm_mm, address, _pte);
2258                                 spin_unlock(ptl);
2259                         }
2260                 } else {
2261                         src_page = pte_page(pteval);
2262                         copy_user_highpage(page, src_page, address, vma);
2263                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
2264                         release_pte_page(src_page);
2265                         /*
2266                          * ptl mostly unnecessary, but preempt has to
2267                          * be disabled to update the per-cpu stats
2268                          * inside page_remove_rmap().
2269                          */
2270                         spin_lock(ptl);
2271                         /*
2272                          * paravirt calls inside pte_clear here are
2273                          * superfluous.
2274                          */
2275                         pte_clear(vma->vm_mm, address, _pte);
2276                         page_remove_rmap(src_page);
2277                         spin_unlock(ptl);
2278                         free_page_and_swap_cache(src_page);
2279                 }
2280
2281                 address += PAGE_SIZE;
2282                 page++;
2283         }
2284 }
2285
2286 static void khugepaged_alloc_sleep(void)
2287 {
2288         wait_event_freezable_timeout(khugepaged_wait, false,
2289                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2290 }
2291
2292 static int khugepaged_node_load[MAX_NUMNODES];
2293
2294 static bool khugepaged_scan_abort(int nid)
2295 {
2296         int i;
2297
2298         /*
2299          * If zone_reclaim_mode is disabled, then no extra effort is made to
2300          * allocate memory locally.
2301          */
2302         if (!zone_reclaim_mode)
2303                 return false;
2304
2305         /* If there is a count for this node already, it must be acceptable */
2306         if (khugepaged_node_load[nid])
2307                 return false;
2308
2309         for (i = 0; i < MAX_NUMNODES; i++) {
2310                 if (!khugepaged_node_load[i])
2311                         continue;
2312                 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2313                         return true;
2314         }
2315         return false;
2316 }
2317
2318 #ifdef CONFIG_NUMA
2319 static int khugepaged_find_target_node(void)
2320 {
2321         static int last_khugepaged_target_node = NUMA_NO_NODE;
2322         int nid, target_node = 0, max_value = 0;
2323
2324         /* find first node with max normal pages hit */
2325         for (nid = 0; nid < MAX_NUMNODES; nid++)
2326                 if (khugepaged_node_load[nid] > max_value) {
2327                         max_value = khugepaged_node_load[nid];
2328                         target_node = nid;
2329                 }
2330
2331         /* do some balance if several nodes have the same hit record */
2332         if (target_node <= last_khugepaged_target_node)
2333                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2334                                 nid++)
2335                         if (max_value == khugepaged_node_load[nid]) {
2336                                 target_node = nid;
2337                                 break;
2338                         }
2339
2340         last_khugepaged_target_node = target_node;
2341         return target_node;
2342 }
2343
2344 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2345 {
2346         if (IS_ERR(*hpage)) {
2347                 if (!*wait)
2348                         return false;
2349
2350                 *wait = false;
2351                 *hpage = NULL;
2352                 khugepaged_alloc_sleep();
2353         } else if (*hpage) {
2354                 put_page(*hpage);
2355                 *hpage = NULL;
2356         }
2357
2358         return true;
2359 }
2360
2361 static struct page *
2362 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2363                        struct vm_area_struct *vma, unsigned long address,
2364                        int node)
2365 {
2366         VM_BUG_ON_PAGE(*hpage, *hpage);
2367
2368         /*
2369          * Before allocating the hugepage, release the mmap_sem read lock.
2370          * The allocation can take potentially a long time if it involves
2371          * sync compaction, and we do not need to hold the mmap_sem during
2372          * that. We will recheck the vma after taking it again in write mode.
2373          */
2374         up_read(&mm->mmap_sem);
2375
2376         *hpage = alloc_pages_exact_node(node, gfp, HPAGE_PMD_ORDER);
2377         if (unlikely(!*hpage)) {
2378                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2379                 *hpage = ERR_PTR(-ENOMEM);
2380                 return NULL;
2381         }
2382
2383         count_vm_event(THP_COLLAPSE_ALLOC);
2384         return *hpage;
2385 }
2386 #else
2387 static int khugepaged_find_target_node(void)
2388 {
2389         return 0;
2390 }
2391
2392 static inline struct page *alloc_hugepage(int defrag)
2393 {
2394         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2395                            HPAGE_PMD_ORDER);
2396 }
2397
2398 static struct page *khugepaged_alloc_hugepage(bool *wait)
2399 {
2400         struct page *hpage;
2401
2402         do {
2403                 hpage = alloc_hugepage(khugepaged_defrag());
2404                 if (!hpage) {
2405                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2406                         if (!*wait)
2407                                 return NULL;
2408
2409                         *wait = false;
2410                         khugepaged_alloc_sleep();
2411                 } else
2412                         count_vm_event(THP_COLLAPSE_ALLOC);
2413         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2414
2415         return hpage;
2416 }
2417
2418 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2419 {
2420         if (!*hpage)
2421                 *hpage = khugepaged_alloc_hugepage(wait);
2422
2423         if (unlikely(!*hpage))
2424                 return false;
2425
2426         return true;
2427 }
2428
2429 static struct page *
2430 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2431                        struct vm_area_struct *vma, unsigned long address,
2432                        int node)
2433 {
2434         up_read(&mm->mmap_sem);
2435         VM_BUG_ON(!*hpage);
2436
2437         return  *hpage;
2438 }
2439 #endif
2440
2441 static bool hugepage_vma_check(struct vm_area_struct *vma)
2442 {
2443         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2444             (vma->vm_flags & VM_NOHUGEPAGE))
2445                 return false;
2446
2447         if (!vma->anon_vma || vma->vm_ops)
2448                 return false;
2449         if (is_vma_temporary_stack(vma))
2450                 return false;
2451         VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma);
2452         return true;
2453 }
2454
2455 static void collapse_huge_page(struct mm_struct *mm,
2456                                    unsigned long address,
2457                                    struct page **hpage,
2458                                    struct vm_area_struct *vma,
2459                                    int node)
2460 {
2461         pmd_t *pmd, _pmd;
2462         pte_t *pte;
2463         pgtable_t pgtable;
2464         struct page *new_page;
2465         spinlock_t *pmd_ptl, *pte_ptl;
2466         int isolated;
2467         unsigned long hstart, hend;
2468         struct mem_cgroup *memcg;
2469         unsigned long mmun_start;       /* For mmu_notifiers */
2470         unsigned long mmun_end;         /* For mmu_notifiers */
2471         gfp_t gfp;
2472
2473         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2474
2475         /* Only allocate from the target node */
2476         gfp = alloc_hugepage_gfpmask(khugepaged_defrag(), __GFP_OTHER_NODE) |
2477                 __GFP_THISNODE;
2478
2479         /* release the mmap_sem read lock. */
2480         new_page = khugepaged_alloc_page(hpage, gfp, mm, vma, address, node);
2481         if (!new_page)
2482                 return;
2483
2484         if (unlikely(mem_cgroup_try_charge(new_page, mm,
2485                                            gfp, &memcg)))
2486                 return;
2487
2488         /*
2489          * Prevent all access to pagetables with the exception of
2490          * gup_fast later hanlded by the ptep_clear_flush and the VM
2491          * handled by the anon_vma lock + PG_lock.
2492          */
2493         down_write(&mm->mmap_sem);
2494         if (unlikely(khugepaged_test_exit(mm)))
2495                 goto out;
2496
2497         vma = find_vma(mm, address);
2498         if (!vma)
2499                 goto out;
2500         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2501         hend = vma->vm_end & HPAGE_PMD_MASK;
2502         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2503                 goto out;
2504         if (!hugepage_vma_check(vma))
2505                 goto out;
2506         pmd = mm_find_pmd(mm, address);
2507         if (!pmd)
2508                 goto out;
2509
2510         anon_vma_lock_write(vma->anon_vma);
2511
2512         pte = pte_offset_map(pmd, address);
2513         pte_ptl = pte_lockptr(mm, pmd);
2514
2515         mmun_start = address;
2516         mmun_end   = address + HPAGE_PMD_SIZE;
2517         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2518         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2519         /*
2520          * After this gup_fast can't run anymore. This also removes
2521          * any huge TLB entry from the CPU so we won't allow
2522          * huge and small TLB entries for the same virtual address
2523          * to avoid the risk of CPU bugs in that area.
2524          */
2525         _pmd = pmdp_collapse_flush(vma, address, pmd);
2526         spin_unlock(pmd_ptl);
2527         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2528
2529         spin_lock(pte_ptl);
2530         isolated = __collapse_huge_page_isolate(vma, address, pte);
2531         spin_unlock(pte_ptl);
2532
2533         if (unlikely(!isolated)) {
2534                 pte_unmap(pte);
2535                 spin_lock(pmd_ptl);
2536                 BUG_ON(!pmd_none(*pmd));
2537                 /*
2538                  * We can only use set_pmd_at when establishing
2539                  * hugepmds and never for establishing regular pmds that
2540                  * points to regular pagetables. Use pmd_populate for that
2541                  */
2542                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2543                 spin_unlock(pmd_ptl);
2544                 anon_vma_unlock_write(vma->anon_vma);
2545                 goto out;
2546         }
2547
2548         /*
2549          * All pages are isolated and locked so anon_vma rmap
2550          * can't run anymore.
2551          */
2552         anon_vma_unlock_write(vma->anon_vma);
2553
2554         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2555         pte_unmap(pte);
2556         __SetPageUptodate(new_page);
2557         pgtable = pmd_pgtable(_pmd);
2558
2559         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2560         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2561
2562         /*
2563          * spin_lock() below is not the equivalent of smp_wmb(), so
2564          * this is needed to avoid the copy_huge_page writes to become
2565          * visible after the set_pmd_at() write.
2566          */
2567         smp_wmb();
2568
2569         spin_lock(pmd_ptl);
2570         BUG_ON(!pmd_none(*pmd));
2571         page_add_new_anon_rmap(new_page, vma, address);
2572         mem_cgroup_commit_charge(new_page, memcg, false);
2573         lru_cache_add_active_or_unevictable(new_page, vma);
2574         pgtable_trans_huge_deposit(mm, pmd, pgtable);
2575         set_pmd_at(mm, address, pmd, _pmd);
2576         update_mmu_cache_pmd(vma, address, pmd);
2577         spin_unlock(pmd_ptl);
2578
2579         *hpage = NULL;
2580
2581         khugepaged_pages_collapsed++;
2582 out_up_write:
2583         up_write(&mm->mmap_sem);
2584         return;
2585
2586 out:
2587         mem_cgroup_cancel_charge(new_page, memcg);
2588         goto out_up_write;
2589 }
2590
2591 static int khugepaged_scan_pmd(struct mm_struct *mm,
2592                                struct vm_area_struct *vma,
2593                                unsigned long address,
2594                                struct page **hpage)
2595 {
2596         pmd_t *pmd;
2597         pte_t *pte, *_pte;
2598         int ret = 0, none_or_zero = 0;
2599         struct page *page;
2600         unsigned long _address;
2601         spinlock_t *ptl;
2602         int node = NUMA_NO_NODE;
2603         bool writable = false, referenced = false;
2604
2605         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2606
2607         pmd = mm_find_pmd(mm, address);
2608         if (!pmd)
2609                 goto out;
2610
2611         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2612         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2613         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2614              _pte++, _address += PAGE_SIZE) {
2615                 pte_t pteval = *_pte;
2616                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2617                         if (!userfaultfd_armed(vma) &&
2618                             ++none_or_zero <= khugepaged_max_ptes_none)
2619                                 continue;
2620                         else
2621                                 goto out_unmap;
2622                 }
2623                 if (!pte_present(pteval))
2624                         goto out_unmap;
2625                 if (pte_write(pteval))
2626                         writable = true;
2627
2628                 page = vm_normal_page(vma, _address, pteval);
2629                 if (unlikely(!page))
2630                         goto out_unmap;
2631                 /*
2632                  * Record which node the original page is from and save this
2633                  * information to khugepaged_node_load[].
2634                  * Khupaged will allocate hugepage from the node has the max
2635                  * hit record.
2636                  */
2637                 node = page_to_nid(page);
2638                 if (khugepaged_scan_abort(node))
2639                         goto out_unmap;
2640                 khugepaged_node_load[node]++;
2641                 VM_BUG_ON_PAGE(PageCompound(page), page);
2642                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2643                         goto out_unmap;
2644                 /*
2645                  * cannot use mapcount: can't collapse if there's a gup pin.
2646                  * The page must only be referenced by the scanned process
2647                  * and page swap cache.
2648                  */
2649                 if (page_count(page) != 1 + !!PageSwapCache(page))
2650                         goto out_unmap;
2651                 if (pte_young(pteval) || PageReferenced(page) ||
2652                     mmu_notifier_test_young(vma->vm_mm, address))
2653                         referenced = true;
2654         }
2655         if (referenced && writable)
2656                 ret = 1;
2657 out_unmap:
2658         pte_unmap_unlock(pte, ptl);
2659         if (ret) {
2660                 node = khugepaged_find_target_node();
2661                 /* collapse_huge_page will return with the mmap_sem released */
2662                 collapse_huge_page(mm, address, hpage, vma, node);
2663         }
2664 out:
2665         return ret;
2666 }
2667
2668 static void collect_mm_slot(struct mm_slot *mm_slot)
2669 {
2670         struct mm_struct *mm = mm_slot->mm;
2671
2672         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2673
2674         if (khugepaged_test_exit(mm)) {
2675                 /* free mm_slot */
2676                 hash_del(&mm_slot->hash);
2677                 list_del(&mm_slot->mm_node);
2678
2679                 /*
2680                  * Not strictly needed because the mm exited already.
2681                  *
2682                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2683                  */
2684
2685                 /* khugepaged_mm_lock actually not necessary for the below */
2686                 free_mm_slot(mm_slot);
2687                 mmdrop(mm);
2688         }
2689 }
2690
2691 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2692                                             struct page **hpage)
2693         __releases(&khugepaged_mm_lock)
2694         __acquires(&khugepaged_mm_lock)
2695 {
2696         struct mm_slot *mm_slot;
2697         struct mm_struct *mm;
2698         struct vm_area_struct *vma;
2699         int progress = 0;
2700
2701         VM_BUG_ON(!pages);
2702         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2703
2704         if (khugepaged_scan.mm_slot)
2705                 mm_slot = khugepaged_scan.mm_slot;
2706         else {
2707                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2708                                      struct mm_slot, mm_node);
2709                 khugepaged_scan.address = 0;
2710                 khugepaged_scan.mm_slot = mm_slot;
2711         }
2712         spin_unlock(&khugepaged_mm_lock);
2713
2714         mm = mm_slot->mm;
2715         down_read(&mm->mmap_sem);
2716         if (unlikely(khugepaged_test_exit(mm)))
2717                 vma = NULL;
2718         else
2719                 vma = find_vma(mm, khugepaged_scan.address);
2720
2721         progress++;
2722         for (; vma; vma = vma->vm_next) {
2723                 unsigned long hstart, hend;
2724
2725                 cond_resched();
2726                 if (unlikely(khugepaged_test_exit(mm))) {
2727                         progress++;
2728                         break;
2729                 }
2730                 if (!hugepage_vma_check(vma)) {
2731 skip:
2732                         progress++;
2733                         continue;
2734                 }
2735                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2736                 hend = vma->vm_end & HPAGE_PMD_MASK;
2737                 if (hstart >= hend)
2738                         goto skip;
2739                 if (khugepaged_scan.address > hend)
2740                         goto skip;
2741                 if (khugepaged_scan.address < hstart)
2742                         khugepaged_scan.address = hstart;
2743                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2744
2745                 while (khugepaged_scan.address < hend) {
2746                         int ret;
2747                         cond_resched();
2748                         if (unlikely(khugepaged_test_exit(mm)))
2749                                 goto breakouterloop;
2750
2751                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2752                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2753                                   hend);
2754                         ret = khugepaged_scan_pmd(mm, vma,
2755                                                   khugepaged_scan.address,
2756                                                   hpage);
2757                         /* move to next address */
2758                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2759                         progress += HPAGE_PMD_NR;
2760                         if (ret)
2761                                 /* we released mmap_sem so break loop */
2762                                 goto breakouterloop_mmap_sem;
2763                         if (progress >= pages)
2764                                 goto breakouterloop;
2765                 }
2766         }
2767 breakouterloop:
2768         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2769 breakouterloop_mmap_sem:
2770
2771         spin_lock(&khugepaged_mm_lock);
2772         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2773         /*
2774          * Release the current mm_slot if this mm is about to die, or
2775          * if we scanned all vmas of this mm.
2776          */
2777         if (khugepaged_test_exit(mm) || !vma) {
2778                 /*
2779                  * Make sure that if mm_users is reaching zero while
2780                  * khugepaged runs here, khugepaged_exit will find
2781                  * mm_slot not pointing to the exiting mm.
2782                  */
2783                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2784                         khugepaged_scan.mm_slot = list_entry(
2785                                 mm_slot->mm_node.next,
2786                                 struct mm_slot, mm_node);
2787                         khugepaged_scan.address = 0;
2788                 } else {
2789                         khugepaged_scan.mm_slot = NULL;
2790                         khugepaged_full_scans++;
2791                 }
2792
2793                 collect_mm_slot(mm_slot);
2794         }
2795
2796         return progress;
2797 }
2798
2799 static int khugepaged_has_work(void)
2800 {
2801         return !list_empty(&khugepaged_scan.mm_head) &&
2802                 khugepaged_enabled();
2803 }
2804
2805 static int khugepaged_wait_event(void)
2806 {
2807         return !list_empty(&khugepaged_scan.mm_head) ||
2808                 kthread_should_stop();
2809 }
2810
2811 static void khugepaged_do_scan(void)
2812 {
2813         struct page *hpage = NULL;
2814         unsigned int progress = 0, pass_through_head = 0;
2815         unsigned int pages = khugepaged_pages_to_scan;
2816         bool wait = true;
2817
2818         barrier(); /* write khugepaged_pages_to_scan to local stack */
2819
2820         while (progress < pages) {
2821                 if (!khugepaged_prealloc_page(&hpage, &wait))
2822                         break;
2823
2824                 cond_resched();
2825
2826                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2827                         break;
2828
2829                 spin_lock(&khugepaged_mm_lock);
2830                 if (!khugepaged_scan.mm_slot)
2831                         pass_through_head++;
2832                 if (khugepaged_has_work() &&
2833                     pass_through_head < 2)
2834                         progress += khugepaged_scan_mm_slot(pages - progress,
2835                                                             &hpage);
2836                 else
2837                         progress = pages;
2838                 spin_unlock(&khugepaged_mm_lock);
2839         }
2840
2841         if (!IS_ERR_OR_NULL(hpage))
2842                 put_page(hpage);
2843 }
2844
2845 static void khugepaged_wait_work(void)
2846 {
2847         if (khugepaged_has_work()) {
2848                 if (!khugepaged_scan_sleep_millisecs)
2849                         return;
2850
2851                 wait_event_freezable_timeout(khugepaged_wait,
2852                                              kthread_should_stop(),
2853                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2854                 return;
2855         }
2856
2857         if (khugepaged_enabled())
2858                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2859 }
2860
2861 static int khugepaged(void *none)
2862 {
2863         struct mm_slot *mm_slot;
2864
2865         set_freezable();
2866         set_user_nice(current, MAX_NICE);
2867
2868         while (!kthread_should_stop()) {
2869                 khugepaged_do_scan();
2870                 khugepaged_wait_work();
2871         }
2872
2873         spin_lock(&khugepaged_mm_lock);
2874         mm_slot = khugepaged_scan.mm_slot;
2875         khugepaged_scan.mm_slot = NULL;
2876         if (mm_slot)
2877                 collect_mm_slot(mm_slot);
2878         spin_unlock(&khugepaged_mm_lock);
2879         return 0;
2880 }
2881
2882 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2883                 unsigned long haddr, pmd_t *pmd)
2884 {
2885         struct mm_struct *mm = vma->vm_mm;
2886         pgtable_t pgtable;
2887         pmd_t _pmd;
2888         int i;
2889
2890         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2891         /* leave pmd empty until pte is filled */
2892
2893         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2894         pmd_populate(mm, &_pmd, pgtable);
2895
2896         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2897                 pte_t *pte, entry;
2898                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2899                 entry = pte_mkspecial(entry);
2900                 pte = pte_offset_map(&_pmd, haddr);
2901                 VM_BUG_ON(!pte_none(*pte));
2902                 set_pte_at(mm, haddr, pte, entry);
2903                 pte_unmap(pte);
2904         }
2905         smp_wmb(); /* make pte visible before pmd */
2906         pmd_populate(mm, pmd, pgtable);
2907         put_huge_zero_page();
2908 }
2909
2910 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2911                 pmd_t *pmd)
2912 {
2913         spinlock_t *ptl;
2914         struct page *page;
2915         struct mm_struct *mm = vma->vm_mm;
2916         unsigned long haddr = address & HPAGE_PMD_MASK;
2917         unsigned long mmun_start;       /* For mmu_notifiers */
2918         unsigned long mmun_end;         /* For mmu_notifiers */
2919
2920         BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
2921
2922         mmun_start = haddr;
2923         mmun_end   = haddr + HPAGE_PMD_SIZE;
2924 again:
2925         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2926         ptl = pmd_lock(mm, pmd);
2927         if (unlikely(!pmd_trans_huge(*pmd))) {
2928                 spin_unlock(ptl);
2929                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2930                 return;
2931         }
2932         if (is_huge_zero_pmd(*pmd)) {
2933                 __split_huge_zero_page_pmd(vma, haddr, pmd);
2934                 spin_unlock(ptl);
2935                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2936                 return;
2937         }
2938         page = pmd_page(*pmd);
2939         VM_BUG_ON_PAGE(!page_count(page), page);
2940         get_page(page);
2941         spin_unlock(ptl);
2942         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2943
2944         split_huge_page(page);
2945
2946         put_page(page);
2947
2948         /*
2949          * We don't always have down_write of mmap_sem here: a racing
2950          * do_huge_pmd_wp_page() might have copied-on-write to another
2951          * huge page before our split_huge_page() got the anon_vma lock.
2952          */
2953         if (unlikely(pmd_trans_huge(*pmd)))
2954                 goto again;
2955 }
2956
2957 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2958                 pmd_t *pmd)
2959 {
2960         struct vm_area_struct *vma;
2961
2962         vma = find_vma(mm, address);
2963         BUG_ON(vma == NULL);
2964         split_huge_page_pmd(vma, address, pmd);
2965 }
2966
2967 static void split_huge_page_address(struct mm_struct *mm,
2968                                     unsigned long address)
2969 {
2970         pgd_t *pgd;
2971         pud_t *pud;
2972         pmd_t *pmd;
2973
2974         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2975
2976         pgd = pgd_offset(mm, address);
2977         if (!pgd_present(*pgd))
2978                 return;
2979
2980         pud = pud_offset(pgd, address);
2981         if (!pud_present(*pud))
2982                 return;
2983
2984         pmd = pmd_offset(pud, address);
2985         if (!pmd_present(*pmd))
2986                 return;
2987         /*
2988          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2989          * materialize from under us.
2990          */
2991         split_huge_page_pmd_mm(mm, address, pmd);
2992 }
2993
2994 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2995                              unsigned long start,
2996                              unsigned long end,
2997                              long adjust_next)
2998 {
2999         /*
3000          * If the new start address isn't hpage aligned and it could
3001          * previously contain an hugepage: check if we need to split
3002          * an huge pmd.
3003          */
3004         if (start & ~HPAGE_PMD_MASK &&
3005             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
3006             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3007                 split_huge_page_address(vma->vm_mm, start);
3008
3009         /*
3010          * If the new end address isn't hpage aligned and it could
3011          * previously contain an hugepage: check if we need to split
3012          * an huge pmd.
3013          */
3014         if (end & ~HPAGE_PMD_MASK &&
3015             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
3016             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3017                 split_huge_page_address(vma->vm_mm, end);
3018
3019         /*
3020          * If we're also updating the vma->vm_next->vm_start, if the new
3021          * vm_next->vm_start isn't page aligned and it could previously
3022          * contain an hugepage: check if we need to split an huge pmd.
3023          */
3024         if (adjust_next > 0) {
3025                 struct vm_area_struct *next = vma->vm_next;
3026                 unsigned long nstart = next->vm_start;
3027                 nstart += adjust_next << PAGE_SHIFT;
3028                 if (nstart & ~HPAGE_PMD_MASK &&
3029                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
3030                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
3031                         split_huge_page_address(next->vm_mm, nstart);
3032         }
3033 }