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 bool 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         if (!pmd_none(*pmd))
 803                 return false;
 804         entry = mk_pmd(zero_page, vma->vm_page_prot);
 805         entry = pmd_mkhuge(entry);
 806         pgtable_trans_huge_deposit(mm, pmd, pgtable);
 807         set_pmd_at(mm, haddr, pmd, entry);
 808         atomic_long_inc(&mm->nr_ptes);
 809         return true;
 810 }
 811
 812 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 813                                unsigned long address, pmd_t *pmd,
 814                                unsigned int flags)
 815 {
 816         gfp_t gfp;
 817         struct page *page;
 818         unsigned long haddr = address & HPAGE_PMD_MASK;
 819
 820         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 821                 return VM_FAULT_FALLBACK;
 822         if (unlikely(anon_vma_prepare(vma)))
 823                 return VM_FAULT_OOM;
 824         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 825                 return VM_FAULT_OOM;
 826         if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm) &&
 827                         transparent_hugepage_use_zero_page()) {
 828                 spinlock_t *ptl;
 829                 pgtable_t pgtable;
 830                 struct page *zero_page;
 831                 bool set;
 832                 int ret;
 833                 pgtable = pte_alloc_one(mm, haddr);
 834                 if (unlikely(!pgtable))
 835                         return VM_FAULT_OOM;
 836                 zero_page = get_huge_zero_page();
 837                 if (unlikely(!zero_page)) {
 838                         pte_free(mm, pgtable);
 839                         count_vm_event(THP_FAULT_FALLBACK);
 840                         return VM_FAULT_FALLBACK;
 841                 }
 842                 ptl = pmd_lock(mm, pmd);
 843                 ret = 0;
 844                 set = false;
 845                 if (pmd_none(*pmd)) {
 846                         if (userfaultfd_missing(vma)) {
 847                                 spin_unlock(ptl);
 848                                 ret = handle_userfault(vma, address, flags,
 849                                                        VM_UFFD_MISSING);
 850                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 851                         } else {
 852                                 set_huge_zero_page(pgtable, mm, vma,
 853                                                    haddr, pmd,
 854                                                    zero_page);
 855                                 spin_unlock(ptl);
 856                                 set = true;
 857                         }
 858                 } else
 859                         spin_unlock(ptl);
 860                 if (!set) {
 861                         pte_free(mm, pgtable);
 862                         put_huge_zero_page();
 863                 }
 864                 return ret;
 865         }
 866         gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
 867         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 868         if (unlikely(!page)) {
 869                 count_vm_event(THP_FAULT_FALLBACK);
 870                 return VM_FAULT_FALLBACK;
 871         }
 872         return __do_huge_pmd_anonymous_page(mm, vma, address, pmd, page, gfp,
 873                                             flags);
 874 }
 875
 876 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 877                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 878                   struct vm_area_struct *vma)
 879 {
 880         spinlock_t *dst_ptl, *src_ptl;
 881         struct page *src_page;
 882         pmd_t pmd;
 883         pgtable_t pgtable;
 884         int ret;
 885
 886         ret = -ENOMEM;
 887         pgtable = pte_alloc_one(dst_mm, addr);
 888         if (unlikely(!pgtable))
 889                 goto out;
 890
 891         dst_ptl = pmd_lock(dst_mm, dst_pmd);
 892         src_ptl = pmd_lockptr(src_mm, src_pmd);
 893         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 894
 895         ret = -EAGAIN;
 896         pmd = *src_pmd;
 897         if (unlikely(!pmd_trans_huge(pmd))) {
 898                 pte_free(dst_mm, pgtable);
 899                 goto out_unlock;
 900         }
 901         /*
 902          * When page table lock is held, the huge zero pmd should not be
 903          * under splitting since we don't split the page itself, only pmd to
 904          * a page table.
 905          */
 906         if (is_huge_zero_pmd(pmd)) {
 907                 struct page *zero_page;
 908                 /*
 909                  * get_huge_zero_page() will never allocate a new page here,
 910                  * since we already have a zero page to copy. It just takes a
 911                  * reference.
 912                  */
 913                 zero_page = get_huge_zero_page();
 914                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
 915                                 zero_page);
 916                 ret = 0;
 917                 goto out_unlock;
 918         }
 919
 920         if (unlikely(pmd_trans_splitting(pmd))) {
 921                 /* split huge page running from under us */
 922                 spin_unlock(src_ptl);
 923                 spin_unlock(dst_ptl);
 924                 pte_free(dst_mm, pgtable);
 925
 926                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 927                 goto out;
 928         }
 929         src_page = pmd_page(pmd);
 930         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
 931         get_page(src_page);
 932         page_dup_rmap(src_page);
 933         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 934
 935         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 936         pmd = pmd_mkold(pmd_wrprotect(pmd));
 937         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
 938         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 939         atomic_long_inc(&dst_mm->nr_ptes);
 940
 941         ret = 0;
 942 out_unlock:
 943         spin_unlock(src_ptl);
 944         spin_unlock(dst_ptl);
 945 out:
 946         return ret;
 947 }
 948
 949 void huge_pmd_set_accessed(struct mm_struct *mm,
 950                            struct vm_area_struct *vma,
 951                            unsigned long address,
 952                            pmd_t *pmd, pmd_t orig_pmd,
 953                            int dirty)
 954 {
 955         spinlock_t *ptl;
 956         pmd_t entry;
 957         unsigned long haddr;
 958
 959         ptl = pmd_lock(mm, pmd);
 960         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 961                 goto unlock;
 962
 963         entry = pmd_mkyoung(orig_pmd);
 964         haddr = address & HPAGE_PMD_MASK;
 965         if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
 966                 update_mmu_cache_pmd(vma, address, pmd);
 967
 968 unlock:
 969         spin_unlock(ptl);
 970 }
 971
 972 /*
 973  * Save CONFIG_DEBUG_PAGEALLOC from faulting falsely on tail pages
 974  * during copy_user_huge_page()'s copy_page_rep(): in the case when
 975  * the source page gets split and a tail freed before copy completes.
 976  * Called under pmd_lock of checked pmd, so safe from splitting itself.
 977  */
 978 static void get_user_huge_page(struct page *page)
 979 {
 980         if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
 981                 struct page *endpage = page + HPAGE_PMD_NR;
 982
 983                 atomic_add(HPAGE_PMD_NR, &page->_count);
 984                 while (++page < endpage)
 985                         get_huge_page_tail(page);
 986         } else {
 987                 get_page(page);
 988         }
 989 }
 990
 991 static void put_user_huge_page(struct page *page)
 992 {
 993         if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
 994                 struct page *endpage = page + HPAGE_PMD_NR;
 995
 996                 while (page < endpage)
 997                         put_page(page++);
 998         } else {
 999                 put_page(page);
1000         }
1001 }
1002
1003 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1004                                         struct vm_area_struct *vma,
1005                                         unsigned long address,
1006                                         pmd_t *pmd, pmd_t orig_pmd,
1007                                         struct page *page,
1008                                         unsigned long haddr)
1009 {
1010         struct mem_cgroup *memcg;
1011         spinlock_t *ptl;
1012         pgtable_t pgtable;
1013         pmd_t _pmd;
1014         int ret = 0, i;
1015         struct page **pages;
1016         unsigned long mmun_start;       /* For mmu_notifiers */
1017         unsigned long mmun_end;         /* For mmu_notifiers */
1018
1019         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1020                         GFP_KERNEL);
1021         if (unlikely(!pages)) {
1022                 ret |= VM_FAULT_OOM;
1023                 goto out;
1024         }
1025
1026         for (i = 0; i < HPAGE_PMD_NR; i++) {
1027                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1028                                                __GFP_OTHER_NODE,
1029                                                vma, address, page_to_nid(page));
1030                 if (unlikely(!pages[i] ||
1031                              mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL,
1032                                                    &memcg))) {
1033                         if (pages[i])
1034                                 put_page(pages[i]);
1035                         while (--i >= 0) {
1036                                 memcg = (void *)page_private(pages[i]);
1037                                 set_page_private(pages[i], 0);
1038                                 mem_cgroup_cancel_charge(pages[i], memcg);
1039                                 put_page(pages[i]);
1040                         }
1041                         kfree(pages);
1042                         ret |= VM_FAULT_OOM;
1043                         goto out;
1044                 }
1045                 set_page_private(pages[i], (unsigned long)memcg);
1046         }
1047
1048         for (i = 0; i < HPAGE_PMD_NR; i++) {
1049                 copy_user_highpage(pages[i], page + i,
1050                                    haddr + PAGE_SIZE * i, vma);
1051                 __SetPageUptodate(pages[i]);
1052                 cond_resched();
1053         }
1054
1055         mmun_start = haddr;
1056         mmun_end   = haddr + HPAGE_PMD_SIZE;
1057         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1058
1059         ptl = pmd_lock(mm, pmd);
1060         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1061                 goto out_free_pages;
1062         VM_BUG_ON_PAGE(!PageHead(page), page);
1063
1064         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1065         /* leave pmd empty until pte is filled */
1066
1067         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1068         pmd_populate(mm, &_pmd, pgtable);
1069
1070         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1071                 pte_t *pte, entry;
1072                 entry = mk_pte(pages[i], vma->vm_page_prot);
1073                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1074                 memcg = (void *)page_private(pages[i]);
1075                 set_page_private(pages[i], 0);
1076                 page_add_new_anon_rmap(pages[i], vma, haddr);
1077                 mem_cgroup_commit_charge(pages[i], memcg, false);
1078                 lru_cache_add_active_or_unevictable(pages[i], vma);
1079                 pte = pte_offset_map(&_pmd, haddr);
1080                 VM_BUG_ON(!pte_none(*pte));
1081                 set_pte_at(mm, haddr, pte, entry);
1082                 pte_unmap(pte);
1083         }
1084         kfree(pages);
1085
1086         smp_wmb(); /* make pte visible before pmd */
1087         pmd_populate(mm, pmd, pgtable);
1088         page_remove_rmap(page);
1089         spin_unlock(ptl);
1090
1091         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1092
1093         ret |= VM_FAULT_WRITE;
1094         put_page(page);
1095
1096 out:
1097         return ret;
1098
1099 out_free_pages:
1100         spin_unlock(ptl);
1101         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1102         for (i = 0; i < HPAGE_PMD_NR; i++) {
1103                 memcg = (void *)page_private(pages[i]);
1104                 set_page_private(pages[i], 0);
1105                 mem_cgroup_cancel_charge(pages[i], memcg);
1106                 put_page(pages[i]);
1107         }
1108         kfree(pages);
1109         goto out;
1110 }
1111
1112 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1113                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1114 {
1115         spinlock_t *ptl;
1116         int ret = 0;
1117         struct page *page = NULL, *new_page;
1118         struct mem_cgroup *memcg;
1119         unsigned long haddr;
1120         unsigned long mmun_start;       /* For mmu_notifiers */
1121         unsigned long mmun_end;         /* For mmu_notifiers */
1122         gfp_t huge_gfp;                 /* for allocation and charge */
1123
1124         ptl = pmd_lockptr(mm, pmd);
1125         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1126         haddr = address & HPAGE_PMD_MASK;
1127         if (is_huge_zero_pmd(orig_pmd))
1128                 goto alloc;
1129         spin_lock(ptl);
1130         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1131                 goto out_unlock;
1132
1133         page = pmd_page(orig_pmd);
1134         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1135         if (page_mapcount(page) == 1) {
1136                 pmd_t entry;
1137                 entry = pmd_mkyoung(orig_pmd);
1138                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1139                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1140                         update_mmu_cache_pmd(vma, address, pmd);
1141                 ret |= VM_FAULT_WRITE;
1142                 goto out_unlock;
1143         }
1144         get_user_huge_page(page);
1145         spin_unlock(ptl);
1146 alloc:
1147         if (transparent_hugepage_enabled(vma) &&
1148             !transparent_hugepage_debug_cow()) {
1149                 huge_gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
1150                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1151         } else
1152                 new_page = NULL;
1153
1154         if (unlikely(!new_page)) {
1155                 if (!page) {
1156                         split_huge_page_pmd(vma, address, pmd);
1157                         ret |= VM_FAULT_FALLBACK;
1158                 } else {
1159                         ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1160                                         pmd, orig_pmd, page, haddr);
1161                         if (ret & VM_FAULT_OOM) {
1162                                 split_huge_page(page);
1163                                 ret |= VM_FAULT_FALLBACK;
1164                         }
1165                         put_user_huge_page(page);
1166                 }
1167                 count_vm_event(THP_FAULT_FALLBACK);
1168                 goto out;
1169         }
1170
1171         if (unlikely(mem_cgroup_try_charge(new_page, mm, huge_gfp, &memcg))) {
1172                 put_page(new_page);
1173                 if (page) {
1174                         split_huge_page(page);
1175                         put_user_huge_page(page);
1176                 } else
1177                         split_huge_page_pmd(vma, address, pmd);
1178                 ret |= VM_FAULT_FALLBACK;
1179                 count_vm_event(THP_FAULT_FALLBACK);
1180                 goto out;
1181         }
1182
1183         count_vm_event(THP_FAULT_ALLOC);
1184
1185         if (!page)
1186                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1187         else
1188                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1189         __SetPageUptodate(new_page);
1190
1191         mmun_start = haddr;
1192         mmun_end   = haddr + HPAGE_PMD_SIZE;
1193         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1194
1195         spin_lock(ptl);
1196         if (page)
1197                 put_user_huge_page(page);
1198         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1199                 spin_unlock(ptl);
1200                 mem_cgroup_cancel_charge(new_page, memcg);
1201                 put_page(new_page);
1202                 goto out_mn;
1203         } else {
1204                 pmd_t entry;
1205                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1206                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1207                 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1208                 page_add_new_anon_rmap(new_page, vma, haddr);
1209                 mem_cgroup_commit_charge(new_page, memcg, false);
1210                 lru_cache_add_active_or_unevictable(new_page, vma);
1211                 set_pmd_at(mm, haddr, pmd, entry);
1212                 update_mmu_cache_pmd(vma, address, pmd);
1213                 if (!page) {
1214                         add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1215                         put_huge_zero_page();
1216                 } else {
1217                         VM_BUG_ON_PAGE(!PageHead(page), page);
1218                         page_remove_rmap(page);
1219                         put_page(page);
1220                 }
1221                 ret |= VM_FAULT_WRITE;
1222         }
1223         spin_unlock(ptl);
1224 out_mn:
1225         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1226 out:
1227         return ret;
1228 out_unlock:
1229         spin_unlock(ptl);
1230         return ret;
1231 }
1232
1233 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1234                                    unsigned long addr,
1235                                    pmd_t *pmd,
1236                                    unsigned int flags)
1237 {
1238         struct mm_struct *mm = vma->vm_mm;
1239         struct page *page = NULL;
1240
1241         assert_spin_locked(pmd_lockptr(mm, pmd));
1242
1243         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1244                 goto out;
1245
1246         /* Avoid dumping huge zero page */
1247         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1248                 return ERR_PTR(-EFAULT);
1249
1250         /* Full NUMA hinting faults to serialise migration in fault paths */
1251         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1252                 goto out;
1253
1254         page = pmd_page(*pmd);
1255         VM_BUG_ON_PAGE(!PageHead(page), page);
1256         if (flags & FOLL_TOUCH) {
1257                 pmd_t _pmd;
1258                 /*
1259                  * We should set the dirty bit only for FOLL_WRITE but
1260                  * for now the dirty bit in the pmd is meaningless.
1261                  * And if the dirty bit will become meaningful and
1262                  * we'll only set it with FOLL_WRITE, an atomic
1263                  * set_bit will be required on the pmd to set the
1264                  * young bit, instead of the current set_pmd_at.
1265                  */
1266                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1267                 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1268                                           pmd, _pmd,  1))
1269                         update_mmu_cache_pmd(vma, addr, pmd);
1270         }
1271         if ((flags & FOLL_POPULATE) && (vma->vm_flags & VM_LOCKED)) {
1272                 if (page->mapping && trylock_page(page)) {
1273                         lru_add_drain();
1274                         if (page->mapping)
1275                                 mlock_vma_page(page);
1276                         unlock_page(page);
1277                 }
1278         }
1279         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1280         VM_BUG_ON_PAGE(!PageCompound(page), page);
1281         if (flags & FOLL_GET)
1282                 get_page_foll(page);
1283
1284 out:
1285         return page;
1286 }
1287
1288 /* NUMA hinting page fault entry point for trans huge pmds */
1289 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1290                                 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1291 {
1292         spinlock_t *ptl;
1293         struct anon_vma *anon_vma = NULL;
1294         struct page *page;
1295         unsigned long haddr = addr & HPAGE_PMD_MASK;
1296         int page_nid = -1, this_nid = numa_node_id();
1297         int target_nid, last_cpupid = -1;
1298         bool page_locked;
1299         bool migrated = false;
1300         bool was_writable;
1301         int flags = 0;
1302
1303         /* A PROT_NONE fault should not end up here */
1304         BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
1305
1306         ptl = pmd_lock(mm, pmdp);
1307         if (unlikely(!pmd_same(pmd, *pmdp)))
1308                 goto out_unlock;
1309
1310         /*
1311          * If there are potential migrations, wait for completion and retry
1312          * without disrupting NUMA hinting information. Do not relock and
1313          * check_same as the page may no longer be mapped.
1314          */
1315         if (unlikely(pmd_trans_migrating(*pmdp))) {
1316                 page = pmd_page(*pmdp);
1317                 spin_unlock(ptl);
1318                 wait_on_page_locked(page);
1319                 goto out;
1320         }
1321
1322         page = pmd_page(pmd);
1323         BUG_ON(is_huge_zero_page(page));
1324         page_nid = page_to_nid(page);
1325         last_cpupid = page_cpupid_last(page);
1326         count_vm_numa_event(NUMA_HINT_FAULTS);
1327         if (page_nid == this_nid) {
1328                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1329                 flags |= TNF_FAULT_LOCAL;
1330         }
1331
1332         /* See similar comment in do_numa_page for explanation */
1333         if (!(vma->vm_flags & VM_WRITE))
1334                 flags |= TNF_NO_GROUP;
1335
1336         /*
1337          * Acquire the page lock to serialise THP migrations but avoid dropping
1338          * page_table_lock if at all possible
1339          */
1340         page_locked = trylock_page(page);
1341         target_nid = mpol_misplaced(page, vma, haddr);
1342         if (target_nid == -1) {
1343                 /* If the page was locked, there are no parallel migrations */
1344                 if (page_locked)
1345                         goto clear_pmdnuma;
1346         }
1347
1348         /* Migration could have started since the pmd_trans_migrating check */
1349         if (!page_locked) {
1350                 spin_unlock(ptl);
1351                 wait_on_page_locked(page);
1352                 page_nid = -1;
1353                 goto out;
1354         }
1355
1356         /*
1357          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1358          * to serialises splits
1359          */
1360         get_page(page);
1361         spin_unlock(ptl);
1362         anon_vma = page_lock_anon_vma_read(page);
1363
1364         /* Confirm the PMD did not change while page_table_lock was released */
1365         spin_lock(ptl);
1366         if (unlikely(!pmd_same(pmd, *pmdp))) {
1367                 unlock_page(page);
1368                 put_page(page);
1369                 page_nid = -1;
1370                 goto out_unlock;
1371         }
1372
1373         /* Bail if we fail to protect against THP splits for any reason */
1374         if (unlikely(!anon_vma)) {
1375                 put_page(page);
1376                 page_nid = -1;
1377                 goto clear_pmdnuma;
1378         }
1379
1380         /*
1381          * Migrate the THP to the requested node, returns with page unlocked
1382          * and access rights restored.
1383          */
1384         spin_unlock(ptl);
1385         migrated = migrate_misplaced_transhuge_page(mm, vma,
1386                                 pmdp, pmd, addr, page, target_nid);
1387         if (migrated) {
1388                 flags |= TNF_MIGRATED;
1389                 page_nid = target_nid;
1390         } else
1391                 flags |= TNF_MIGRATE_FAIL;
1392
1393         goto out;
1394 clear_pmdnuma:
1395         BUG_ON(!PageLocked(page));
1396         was_writable = pmd_write(pmd);
1397         pmd = pmd_modify(pmd, vma->vm_page_prot);
1398         pmd = pmd_mkyoung(pmd);
1399         if (was_writable)
1400                 pmd = pmd_mkwrite(pmd);
1401         set_pmd_at(mm, haddr, pmdp, pmd);
1402         update_mmu_cache_pmd(vma, addr, pmdp);
1403         unlock_page(page);
1404 out_unlock:
1405         spin_unlock(ptl);
1406
1407 out:
1408         if (anon_vma)
1409                 page_unlock_anon_vma_read(anon_vma);
1410
1411         if (page_nid != -1)
1412                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
1413
1414         return 0;
1415 }
1416
1417 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1418                  pmd_t *pmd, unsigned long addr)
1419 {
1420         spinlock_t *ptl;
1421         int ret = 0;
1422
1423         if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1424                 struct page *page;
1425                 pgtable_t pgtable;
1426                 pmd_t orig_pmd;
1427                 /*
1428                  * For architectures like ppc64 we look at deposited pgtable
1429                  * when calling pmdp_huge_get_and_clear. So do the
1430                  * pgtable_trans_huge_withdraw after finishing pmdp related
1431                  * operations.
1432                  */
1433                 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1434                                                         tlb->fullmm);
1435                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1436                 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1437                 if (is_huge_zero_pmd(orig_pmd)) {
1438                         atomic_long_dec(&tlb->mm->nr_ptes);
1439                         spin_unlock(ptl);
1440                         put_huge_zero_page();
1441                 } else {
1442                         page = pmd_page(orig_pmd);
1443                         page_remove_rmap(page);
1444                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1445                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1446                         VM_BUG_ON_PAGE(!PageHead(page), page);
1447                         atomic_long_dec(&tlb->mm->nr_ptes);
1448                         spin_unlock(ptl);
1449                         tlb_remove_page(tlb, page);
1450                 }
1451                 pte_free(tlb->mm, pgtable);
1452                 ret = 1;
1453         }
1454         return ret;
1455 }
1456
1457 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1458                   unsigned long old_addr,
1459                   unsigned long new_addr, unsigned long old_end,
1460                   pmd_t *old_pmd, pmd_t *new_pmd)
1461 {
1462         spinlock_t *old_ptl, *new_ptl;
1463         int ret = 0;
1464         pmd_t pmd;
1465
1466         struct mm_struct *mm = vma->vm_mm;
1467
1468         if ((old_addr & ~HPAGE_PMD_MASK) ||
1469             (new_addr & ~HPAGE_PMD_MASK) ||
1470             old_end - old_addr < HPAGE_PMD_SIZE ||
1471             (new_vma->vm_flags & VM_NOHUGEPAGE))
1472                 goto out;
1473
1474         /*
1475          * The destination pmd shouldn't be established, free_pgtables()
1476          * should have release it.
1477          */
1478         if (WARN_ON(!pmd_none(*new_pmd))) {
1479                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1480                 goto out;
1481         }
1482
1483         /*
1484          * We don't have to worry about the ordering of src and dst
1485          * ptlocks because exclusive mmap_sem prevents deadlock.
1486          */
1487         ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl);
1488         if (ret == 1) {
1489                 new_ptl = pmd_lockptr(mm, new_pmd);
1490                 if (new_ptl != old_ptl)
1491                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1492                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1493                 VM_BUG_ON(!pmd_none(*new_pmd));
1494
1495                 if (pmd_move_must_withdraw(new_ptl, old_ptl)) {
1496                         pgtable_t pgtable;
1497                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1498                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1499                 }
1500                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1501                 if (new_ptl != old_ptl)
1502                         spin_unlock(new_ptl);
1503                 spin_unlock(old_ptl);
1504         }
1505 out:
1506         return ret;
1507 }
1508
1509 /*
1510  * Returns
1511  *  - 0 if PMD could not be locked
1512  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1513  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1514  */
1515 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1516                 unsigned long addr, pgprot_t newprot, int prot_numa)
1517 {
1518         struct mm_struct *mm = vma->vm_mm;
1519         spinlock_t *ptl;
1520         int ret = 0;
1521
1522         if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1523                 pmd_t entry;
1524                 bool preserve_write = prot_numa && pmd_write(*pmd);
1525                 ret = 1;
1526
1527                 /*
1528                  * Avoid trapping faults against the zero page. The read-only
1529                  * data is likely to be read-cached on the local CPU and
1530                  * local/remote hits to the zero page are not interesting.
1531                  */
1532                 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1533                         spin_unlock(ptl);
1534                         return ret;
1535                 }
1536
1537                 if (!prot_numa || !pmd_protnone(*pmd)) {
1538                         entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
1539                         entry = pmd_modify(entry, newprot);
1540                         if (preserve_write)
1541                                 entry = pmd_mkwrite(entry);
1542                         ret = HPAGE_PMD_NR;
1543                         set_pmd_at(mm, addr, pmd, entry);
1544                         BUG_ON(!preserve_write && pmd_write(entry));
1545                 }
1546                 spin_unlock(ptl);
1547         }
1548
1549         return ret;
1550 }
1551
1552 /*
1553  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1554  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1555  *
1556  * Note that if it returns 1, this routine returns without unlocking page
1557  * table locks. So callers must unlock them.
1558  */
1559 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1560                 spinlock_t **ptl)
1561 {
1562         *ptl = pmd_lock(vma->vm_mm, pmd);
1563         if (likely(pmd_trans_huge(*pmd))) {
1564                 if (unlikely(pmd_trans_splitting(*pmd))) {
1565                         spin_unlock(*ptl);
1566                         wait_split_huge_page(vma->anon_vma, pmd);
1567                         return -1;
1568                 } else {
1569                         /* Thp mapped by 'pmd' is stable, so we can
1570                          * handle it as it is. */
1571                         return 1;
1572                 }
1573         }
1574         spin_unlock(*ptl);
1575         return 0;
1576 }
1577
1578 /*
1579  * This function returns whether a given @page is mapped onto the @address
1580  * in the virtual space of @mm.
1581  *
1582  * When it's true, this function returns *pmd with holding the page table lock
1583  * and passing it back to the caller via @ptl.
1584  * If it's false, returns NULL without holding the page table lock.
1585  */
1586 pmd_t *page_check_address_pmd(struct page *page,
1587                               struct mm_struct *mm,
1588                               unsigned long address,
1589                               enum page_check_address_pmd_flag flag,
1590                               spinlock_t **ptl)
1591 {
1592         pgd_t *pgd;
1593         pud_t *pud;
1594         pmd_t *pmd;
1595
1596         if (address & ~HPAGE_PMD_MASK)
1597                 return NULL;
1598
1599         pgd = pgd_offset(mm, address);
1600         if (!pgd_present(*pgd))
1601                 return NULL;
1602         pud = pud_offset(pgd, address);
1603         if (!pud_present(*pud))
1604                 return NULL;
1605         pmd = pmd_offset(pud, address);
1606
1607         *ptl = pmd_lock(mm, pmd);
1608         if (!pmd_present(*pmd))
1609                 goto unlock;
1610         if (pmd_page(*pmd) != page)
1611                 goto unlock;
1612         /*
1613          * split_vma() may create temporary aliased mappings. There is
1614          * no risk as long as all huge pmd are found and have their
1615          * splitting bit set before __split_huge_page_refcount
1616          * runs. Finding the same huge pmd more than once during the
1617          * same rmap walk is not a problem.
1618          */
1619         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1620             pmd_trans_splitting(*pmd))
1621                 goto unlock;
1622         if (pmd_trans_huge(*pmd)) {
1623                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1624                           !pmd_trans_splitting(*pmd));
1625                 return pmd;
1626         }
1627 unlock:
1628         spin_unlock(*ptl);
1629         return NULL;
1630 }
1631
1632 static int __split_huge_page_splitting(struct page *page,
1633                                        struct vm_area_struct *vma,
1634                                        unsigned long address)
1635 {
1636         struct mm_struct *mm = vma->vm_mm;
1637         spinlock_t *ptl;
1638         pmd_t *pmd;
1639         int ret = 0;
1640         /* For mmu_notifiers */
1641         const unsigned long mmun_start = address;
1642         const unsigned long mmun_end   = address + HPAGE_PMD_SIZE;
1643
1644         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1645         pmd = page_check_address_pmd(page, mm, address,
1646                         PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl);
1647         if (pmd) {
1648                 /*
1649                  * We can't temporarily set the pmd to null in order
1650                  * to split it, the pmd must remain marked huge at all
1651                  * times or the VM won't take the pmd_trans_huge paths
1652                  * and it won't wait on the anon_vma->root->rwsem to
1653                  * serialize against split_huge_page*.
1654                  */
1655                 pmdp_splitting_flush(vma, address, pmd);
1656
1657                 ret = 1;
1658                 spin_unlock(ptl);
1659         }
1660         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1661
1662         return ret;
1663 }
1664
1665 static void __split_huge_page_refcount(struct page *page,
1666                                        struct list_head *list)
1667 {
1668         int i;
1669         struct zone *zone = page_zone(page);
1670         struct lruvec *lruvec;
1671         int tail_count = 0;
1672
1673         /* prevent PageLRU to go away from under us, and freeze lru stats */
1674         spin_lock_irq(&zone->lru_lock);
1675         lruvec = mem_cgroup_page_lruvec(page, zone);
1676
1677         compound_lock(page);
1678         /* complete memcg works before add pages to LRU */
1679         mem_cgroup_split_huge_fixup(page);
1680
1681         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1682                 struct page *page_tail = page + i;
1683
1684                 /* tail_page->_mapcount cannot change */
1685                 BUG_ON(page_mapcount(page_tail) < 0);
1686                 tail_count += page_mapcount(page_tail);
1687                 /* check for overflow */
1688                 BUG_ON(tail_count < 0);
1689                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1690                 /*
1691                  * tail_page->_count is zero and not changing from
1692                  * under us. But get_page_unless_zero() may be running
1693                  * from under us on the tail_page. If we used
1694                  * atomic_set() below instead of atomic_add(), we
1695                  * would then run atomic_set() concurrently with
1696                  * get_page_unless_zero(), and atomic_set() is
1697                  * implemented in C not using locked ops. spin_unlock
1698                  * on x86 sometime uses locked ops because of PPro
1699                  * errata 66, 92, so unless somebody can guarantee
1700                  * atomic_set() here would be safe on all archs (and
1701                  * not only on x86), it's safer to use atomic_add().
1702                  */
1703                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1704                            &page_tail->_count);
1705
1706                 /* after clearing PageTail the gup refcount can be released */
1707                 smp_mb__after_atomic();
1708
1709                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1710                 page_tail->flags |= (page->flags &
1711                                      ((1L << PG_referenced) |
1712                                       (1L << PG_swapbacked) |
1713                                       (1L << PG_mlocked) |
1714                                       (1L << PG_uptodate) |
1715                                       (1L << PG_active) |
1716                                       (1L << PG_unevictable)));
1717                 page_tail->flags |= (1L << PG_dirty);
1718
1719                 /* clear PageTail before overwriting first_page */
1720                 smp_wmb();
1721
1722                 /*
1723                  * __split_huge_page_splitting() already set the
1724                  * splitting bit in all pmd that could map this
1725                  * hugepage, that will ensure no CPU can alter the
1726                  * mapcount on the head page. The mapcount is only
1727                  * accounted in the head page and it has to be
1728                  * transferred to all tail pages in the below code. So
1729                  * for this code to be safe, the split the mapcount
1730                  * can't change. But that doesn't mean userland can't
1731                  * keep changing and reading the page contents while
1732                  * we transfer the mapcount, so the pmd splitting
1733                  * status is achieved setting a reserved bit in the
1734                  * pmd, not by clearing the present bit.
1735                 */
1736                 page_tail->_mapcount = page->_mapcount;
1737
1738                 BUG_ON(page_tail->mapping);
1739                 page_tail->mapping = page->mapping;
1740
1741                 page_tail->index = page->index + i;
1742                 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
1743
1744                 BUG_ON(!PageAnon(page_tail));
1745                 BUG_ON(!PageUptodate(page_tail));
1746                 BUG_ON(!PageDirty(page_tail));
1747                 BUG_ON(!PageSwapBacked(page_tail));
1748
1749                 lru_add_page_tail(page, page_tail, lruvec, list);
1750         }
1751         atomic_sub(tail_count, &page->_count);
1752         BUG_ON(atomic_read(&page->_count) <= 0);
1753
1754         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1755
1756         ClearPageCompound(page);
1757         compound_unlock(page);
1758         spin_unlock_irq(&zone->lru_lock);
1759
1760         for (i = 1; i < HPAGE_PMD_NR; i++) {
1761                 struct page *page_tail = page + i;
1762                 BUG_ON(page_count(page_tail) <= 0);
1763                 /*
1764                  * Tail pages may be freed if there wasn't any mapping
1765                  * like if add_to_swap() is running on a lru page that
1766                  * had its mapping zapped. And freeing these pages
1767                  * requires taking the lru_lock so we do the put_page
1768                  * of the tail pages after the split is complete.
1769                  */
1770                 put_page(page_tail);
1771         }
1772
1773         /*
1774          * Only the head page (now become a regular page) is required
1775          * to be pinned by the caller.
1776          */
1777         BUG_ON(page_count(page) <= 0);
1778 }
1779
1780 static int __split_huge_page_map(struct page *page,
1781                                  struct vm_area_struct *vma,
1782                                  unsigned long address)
1783 {
1784         struct mm_struct *mm = vma->vm_mm;
1785         spinlock_t *ptl;
1786         pmd_t *pmd, _pmd;
1787         int ret = 0, i;
1788         pgtable_t pgtable;
1789         unsigned long haddr;
1790
1791         pmd = page_check_address_pmd(page, mm, address,
1792                         PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl);
1793         if (pmd) {
1794                 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1795                 pmd_populate(mm, &_pmd, pgtable);
1796                 if (pmd_write(*pmd))
1797                         BUG_ON(page_mapcount(page) != 1);
1798
1799                 haddr = address;
1800                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1801                         pte_t *pte, entry;
1802                         BUG_ON(PageCompound(page+i));
1803                         /*
1804                          * Note that NUMA hinting access restrictions are not
1805                          * transferred to avoid any possibility of altering
1806                          * permissions across VMAs.
1807                          */
1808                         entry = mk_pte(page + i, vma->vm_page_prot);
1809                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1810                         if (!pmd_write(*pmd))
1811                                 entry = pte_wrprotect(entry);
1812                         if (!pmd_young(*pmd))
1813                                 entry = pte_mkold(entry);
1814                         pte = pte_offset_map(&_pmd, haddr);
1815                         BUG_ON(!pte_none(*pte));
1816                         set_pte_at(mm, haddr, pte, entry);
1817                         pte_unmap(pte);
1818                 }
1819
1820                 smp_wmb(); /* make pte visible before pmd */
1821                 /*
1822                  * Up to this point the pmd is present and huge and
1823                  * userland has the whole access to the hugepage
1824                  * during the split (which happens in place). If we
1825                  * overwrite the pmd with the not-huge version
1826                  * pointing to the pte here (which of course we could
1827                  * if all CPUs were bug free), userland could trigger
1828                  * a small page size TLB miss on the small sized TLB
1829                  * while the hugepage TLB entry is still established
1830                  * in the huge TLB. Some CPU doesn't like that. See
1831                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1832                  * Erratum 383 on page 93. Intel should be safe but is
1833                  * also warns that it's only safe if the permission
1834                  * and cache attributes of the two entries loaded in
1835                  * the two TLB is identical (which should be the case
1836                  * here). But it is generally safer to never allow
1837                  * small and huge TLB entries for the same virtual
1838                  * address to be loaded simultaneously. So instead of
1839                  * doing "pmd_populate(); flush_tlb_range();" we first
1840                  * mark the current pmd notpresent (atomically because
1841                  * here the pmd_trans_huge and pmd_trans_splitting
1842                  * must remain set at all times on the pmd until the
1843                  * split is complete for this pmd), then we flush the
1844                  * SMP TLB and finally we write the non-huge version
1845                  * of the pmd entry with pmd_populate.
1846                  */
1847                 pmdp_invalidate(vma, address, pmd);
1848                 pmd_populate(mm, pmd, pgtable);
1849                 ret = 1;
1850                 spin_unlock(ptl);
1851         }
1852
1853         return ret;
1854 }
1855
1856 /* must be called with anon_vma->root->rwsem held */
1857 static void __split_huge_page(struct page *page,
1858                               struct anon_vma *anon_vma,
1859                               struct list_head *list)
1860 {
1861         int mapcount, mapcount2;
1862         pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1863         struct anon_vma_chain *avc;
1864
1865         BUG_ON(!PageHead(page));
1866         BUG_ON(PageTail(page));
1867
1868         mapcount = 0;
1869         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1870                 struct vm_area_struct *vma = avc->vma;
1871                 unsigned long addr = vma_address(page, vma);
1872                 BUG_ON(is_vma_temporary_stack(vma));
1873                 mapcount += __split_huge_page_splitting(page, vma, addr);
1874         }
1875         /*
1876          * It is critical that new vmas are added to the tail of the
1877          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1878          * and establishes a child pmd before
1879          * __split_huge_page_splitting() freezes the parent pmd (so if
1880          * we fail to prevent copy_huge_pmd() from running until the
1881          * whole __split_huge_page() is complete), we will still see
1882          * the newly established pmd of the child later during the
1883          * walk, to be able to set it as pmd_trans_splitting too.
1884          */
1885         if (mapcount != page_mapcount(page)) {
1886                 pr_err("mapcount %d page_mapcount %d\n",
1887                         mapcount, page_mapcount(page));
1888                 BUG();
1889         }
1890
1891         __split_huge_page_refcount(page, list);
1892
1893         mapcount2 = 0;
1894         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1895                 struct vm_area_struct *vma = avc->vma;
1896                 unsigned long addr = vma_address(page, vma);
1897                 BUG_ON(is_vma_temporary_stack(vma));
1898                 mapcount2 += __split_huge_page_map(page, vma, addr);
1899         }
1900         if (mapcount != mapcount2) {
1901                 pr_err("mapcount %d mapcount2 %d page_mapcount %d\n",
1902                         mapcount, mapcount2, page_mapcount(page));
1903                 BUG();
1904         }
1905 }
1906
1907 /*
1908  * Split a hugepage into normal pages. This doesn't change the position of head
1909  * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1910  * @list. Both head page and tail pages will inherit mapping, flags, and so on
1911  * from the hugepage.
1912  * Return 0 if the hugepage is split successfully otherwise return 1.
1913  */
1914 int split_huge_page_to_list(struct page *page, struct list_head *list)
1915 {
1916         struct anon_vma *anon_vma;
1917         int ret = 1;
1918
1919         BUG_ON(is_huge_zero_page(page));
1920         BUG_ON(!PageAnon(page));
1921
1922         /*
1923          * The caller does not necessarily hold an mmap_sem that would prevent
1924          * the anon_vma disappearing so we first we take a reference to it
1925          * and then lock the anon_vma for write. This is similar to
1926          * page_lock_anon_vma_read except the write lock is taken to serialise
1927          * against parallel split or collapse operations.
1928          */
1929         anon_vma = page_get_anon_vma(page);
1930         if (!anon_vma)
1931                 goto out;
1932         anon_vma_lock_write(anon_vma);
1933
1934         ret = 0;
1935         if (!PageCompound(page))
1936                 goto out_unlock;
1937
1938         BUG_ON(!PageSwapBacked(page));
1939         __split_huge_page(page, anon_vma, list);
1940         count_vm_event(THP_SPLIT);
1941
1942         BUG_ON(PageCompound(page));
1943 out_unlock:
1944         anon_vma_unlock_write(anon_vma);
1945         put_anon_vma(anon_vma);
1946 out:
1947         return ret;
1948 }
1949
1950 #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
1951
1952 int hugepage_madvise(struct vm_area_struct *vma,
1953                      unsigned long *vm_flags, int advice)
1954 {
1955         switch (advice) {
1956         case MADV_HUGEPAGE:
1957 #ifdef CONFIG_S390
1958                 /*
1959                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1960                  * can't handle this properly after s390_enable_sie, so we simply
1961                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
1962                  */
1963                 if (mm_has_pgste(vma->vm_mm))
1964                         return 0;
1965 #endif
1966                 /*
1967                  * Be somewhat over-protective like KSM for now!
1968                  */
1969                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1970                         return -EINVAL;
1971                 *vm_flags &= ~VM_NOHUGEPAGE;
1972                 *vm_flags |= VM_HUGEPAGE;
1973                 /*
1974                  * If the vma become good for khugepaged to scan,
1975                  * register it here without waiting a page fault that
1976                  * may not happen any time soon.
1977                  */
1978                 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
1979                         return -ENOMEM;
1980                 break;
1981         case MADV_NOHUGEPAGE:
1982                 /*
1983                  * Be somewhat over-protective like KSM for now!
1984                  */
1985                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1986                         return -EINVAL;
1987                 *vm_flags &= ~VM_HUGEPAGE;
1988                 *vm_flags |= VM_NOHUGEPAGE;
1989                 /*
1990                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1991                  * this vma even if we leave the mm registered in khugepaged if
1992                  * it got registered before VM_NOHUGEPAGE was set.
1993                  */
1994                 break;
1995         }
1996
1997         return 0;
1998 }
1999
2000 static int __init khugepaged_slab_init(void)
2001 {
2002         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
2003                                           sizeof(struct mm_slot),
2004                                           __alignof__(struct mm_slot), 0, NULL);
2005         if (!mm_slot_cache)
2006                 return -ENOMEM;
2007
2008         return 0;
2009 }
2010
2011 static void __init khugepaged_slab_exit(void)
2012 {
2013         kmem_cache_destroy(mm_slot_cache);
2014 }
2015
2016 static inline struct mm_slot *alloc_mm_slot(void)
2017 {
2018         if (!mm_slot_cache)     /* initialization failed */
2019                 return NULL;
2020         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
2021 }
2022
2023 static inline void free_mm_slot(struct mm_slot *mm_slot)
2024 {
2025         kmem_cache_free(mm_slot_cache, mm_slot);
2026 }
2027
2028 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
2029 {
2030         struct mm_slot *mm_slot;
2031
2032         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
2033                 if (mm == mm_slot->mm)
2034                         return mm_slot;
2035
2036         return NULL;
2037 }
2038
2039 static void insert_to_mm_slots_hash(struct mm_struct *mm,
2040                                     struct mm_slot *mm_slot)
2041 {
2042         mm_slot->mm = mm;
2043         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
2044 }
2045
2046 static inline int khugepaged_test_exit(struct mm_struct *mm)
2047 {
2048         return atomic_read(&mm->mm_users) == 0;
2049 }
2050
2051 int __khugepaged_enter(struct mm_struct *mm)
2052 {
2053         struct mm_slot *mm_slot;
2054         int wakeup;
2055
2056         mm_slot = alloc_mm_slot();
2057         if (!mm_slot)
2058                 return -ENOMEM;
2059
2060         /* __khugepaged_exit() must not run from under us */
2061         VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
2062         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2063                 free_mm_slot(mm_slot);
2064                 return 0;
2065         }
2066
2067         spin_lock(&khugepaged_mm_lock);
2068         insert_to_mm_slots_hash(mm, mm_slot);
2069         /*
2070          * Insert just behind the scanning cursor, to let the area settle
2071          * down a little.
2072          */
2073         wakeup = list_empty(&khugepaged_scan.mm_head);
2074         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2075         spin_unlock(&khugepaged_mm_lock);
2076
2077         atomic_inc(&mm->mm_count);
2078         if (wakeup)
2079                 wake_up_interruptible(&khugepaged_wait);
2080
2081         return 0;
2082 }
2083
2084 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
2085                                unsigned long vm_flags)
2086 {
2087         unsigned long hstart, hend;
2088         if (!vma->anon_vma)
2089                 /*
2090                  * Not yet faulted in so we will register later in the
2091                  * page fault if needed.
2092                  */
2093                 return 0;
2094         if (vma->vm_ops)
2095                 /* khugepaged not yet working on file or special mappings */
2096                 return 0;
2097         VM_BUG_ON_VMA(vm_flags & VM_NO_THP, vma);
2098         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2099         hend = vma->vm_end & HPAGE_PMD_MASK;
2100         if (hstart < hend)
2101                 return khugepaged_enter(vma, vm_flags);
2102         return 0;
2103 }
2104
2105 void __khugepaged_exit(struct mm_struct *mm)
2106 {
2107         struct mm_slot *mm_slot;
2108         int free = 0;
2109
2110         spin_lock(&khugepaged_mm_lock);
2111         mm_slot = get_mm_slot(mm);
2112         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
2113                 hash_del(&mm_slot->hash);
2114                 list_del(&mm_slot->mm_node);
2115                 free = 1;
2116         }
2117         spin_unlock(&khugepaged_mm_lock);
2118
2119         if (free) {
2120                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2121                 free_mm_slot(mm_slot);
2122                 mmdrop(mm);
2123         } else if (mm_slot) {
2124                 /*
2125                  * This is required to serialize against
2126                  * khugepaged_test_exit() (which is guaranteed to run
2127                  * under mmap sem read mode). Stop here (after we
2128                  * return all pagetables will be destroyed) until
2129                  * khugepaged has finished working on the pagetables
2130                  * under the mmap_sem.
2131                  */
2132                 down_write(&mm->mmap_sem);
2133                 up_write(&mm->mmap_sem);
2134         }
2135 }
2136
2137 static void release_pte_page(struct page *page)
2138 {
2139         /* 0 stands for page_is_file_cache(page) == false */
2140         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2141         unlock_page(page);
2142         putback_lru_page(page);
2143 }
2144
2145 static void release_pte_pages(pte_t *pte, pte_t *_pte)
2146 {
2147         while (--_pte >= pte) {
2148                 pte_t pteval = *_pte;
2149                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
2150                         release_pte_page(pte_page(pteval));
2151         }
2152 }
2153
2154 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2155                                         unsigned long address,
2156                                         pte_t *pte)
2157 {
2158         struct page *page;
2159         pte_t *_pte;
2160         int none_or_zero = 0;
2161         bool referenced = false, writable = false;
2162         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2163              _pte++, address += PAGE_SIZE) {
2164                 pte_t pteval = *_pte;
2165                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2166                         if (!userfaultfd_armed(vma) &&
2167                             ++none_or_zero <= khugepaged_max_ptes_none)
2168                                 continue;
2169                         else
2170                                 goto out;
2171                 }
2172                 if (!pte_present(pteval))
2173                         goto out;
2174                 page = vm_normal_page(vma, address, pteval);
2175                 if (unlikely(!page))
2176                         goto out;
2177
2178                 VM_BUG_ON_PAGE(PageCompound(page), page);
2179                 VM_BUG_ON_PAGE(!PageAnon(page), page);
2180                 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2181
2182                 /*
2183                  * We can do it before isolate_lru_page because the
2184                  * page can't be freed from under us. NOTE: PG_lock
2185                  * is needed to serialize against split_huge_page
2186                  * when invoked from the VM.
2187                  */
2188                 if (!trylock_page(page))
2189                         goto out;
2190
2191                 /*
2192                  * cannot use mapcount: can't collapse if there's a gup pin.
2193                  * The page must only be referenced by the scanned process
2194                  * and page swap cache.
2195                  */
2196                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2197                         unlock_page(page);
2198                         goto out;
2199                 }
2200                 if (pte_write(pteval)) {
2201                         writable = true;
2202                 } else {
2203                         if (PageSwapCache(page) && !reuse_swap_page(page)) {
2204                                 unlock_page(page);
2205                                 goto out;
2206                         }
2207                         /*
2208                          * Page is not in the swap cache. It can be collapsed
2209                          * into a THP.
2210                          */
2211                 }
2212
2213                 /*
2214                  * Isolate the page to avoid collapsing an hugepage
2215                  * currently in use by the VM.
2216                  */
2217                 if (isolate_lru_page(page)) {
2218                         unlock_page(page);
2219                         goto out;
2220                 }
2221                 /* 0 stands for page_is_file_cache(page) == false */
2222                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2223                 VM_BUG_ON_PAGE(!PageLocked(page), page);
2224                 VM_BUG_ON_PAGE(PageLRU(page), page);
2225
2226                 /* If there is no mapped pte young don't collapse the page */
2227                 if (pte_young(pteval) || PageReferenced(page) ||
2228                     mmu_notifier_test_young(vma->vm_mm, address))
2229                         referenced = true;
2230         }
2231         if (likely(referenced && writable))
2232                 return 1;
2233 out:
2234         release_pte_pages(pte, _pte);
2235         return 0;
2236 }
2237
2238 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2239                                       struct vm_area_struct *vma,
2240                                       unsigned long address,
2241                                       spinlock_t *ptl)
2242 {
2243         pte_t *_pte;
2244         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2245                 pte_t pteval = *_pte;
2246                 struct page *src_page;
2247
2248                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2249                         clear_user_highpage(page, address);
2250                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2251                         if (is_zero_pfn(pte_pfn(pteval))) {
2252                                 /*
2253                                  * ptl mostly unnecessary.
2254                                  */
2255                                 spin_lock(ptl);
2256                                 /*
2257                                  * paravirt calls inside pte_clear here are
2258                                  * superfluous.
2259                                  */
2260                                 pte_clear(vma->vm_mm, address, _pte);
2261                                 spin_unlock(ptl);
2262                         }
2263                 } else {
2264                         src_page = pte_page(pteval);
2265                         copy_user_highpage(page, src_page, address, vma);
2266                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
2267                         release_pte_page(src_page);
2268                         /*
2269                          * ptl mostly unnecessary, but preempt has to
2270                          * be disabled to update the per-cpu stats
2271                          * inside page_remove_rmap().
2272                          */
2273                         spin_lock(ptl);
2274                         /*
2275                          * paravirt calls inside pte_clear here are
2276                          * superfluous.
2277                          */
2278                         pte_clear(vma->vm_mm, address, _pte);
2279                         page_remove_rmap(src_page);
2280                         spin_unlock(ptl);
2281                         free_page_and_swap_cache(src_page);
2282                 }
2283
2284                 address += PAGE_SIZE;
2285                 page++;
2286         }
2287 }
2288
2289 static void khugepaged_alloc_sleep(void)
2290 {
2291         wait_event_freezable_timeout(khugepaged_wait, false,
2292                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2293 }
2294
2295 static int khugepaged_node_load[MAX_NUMNODES];
2296
2297 static bool khugepaged_scan_abort(int nid)
2298 {
2299         int i;
2300
2301         /*
2302          * If zone_reclaim_mode is disabled, then no extra effort is made to
2303          * allocate memory locally.
2304          */
2305         if (!zone_reclaim_mode)
2306                 return false;
2307
2308         /* If there is a count for this node already, it must be acceptable */
2309         if (khugepaged_node_load[nid])
2310                 return false;
2311
2312         for (i = 0; i < MAX_NUMNODES; i++) {
2313                 if (!khugepaged_node_load[i])
2314                         continue;
2315                 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2316                         return true;
2317         }
2318         return false;
2319 }
2320
2321 #ifdef CONFIG_NUMA
2322 static int khugepaged_find_target_node(void)
2323 {
2324         static int last_khugepaged_target_node = NUMA_NO_NODE;
2325         int nid, target_node = 0, max_value = 0;
2326
2327         /* find first node with max normal pages hit */
2328         for (nid = 0; nid < MAX_NUMNODES; nid++)
2329                 if (khugepaged_node_load[nid] > max_value) {
2330                         max_value = khugepaged_node_load[nid];
2331                         target_node = nid;
2332                 }
2333
2334         /* do some balance if several nodes have the same hit record */
2335         if (target_node <= last_khugepaged_target_node)
2336                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2337                                 nid++)
2338                         if (max_value == khugepaged_node_load[nid]) {
2339                                 target_node = nid;
2340                                 break;
2341                         }
2342
2343         last_khugepaged_target_node = target_node;
2344         return target_node;
2345 }
2346
2347 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2348 {
2349         if (IS_ERR(*hpage)) {
2350                 if (!*wait)
2351                         return false;
2352
2353                 *wait = false;
2354                 *hpage = NULL;
2355                 khugepaged_alloc_sleep();
2356         } else if (*hpage) {
2357                 put_page(*hpage);
2358                 *hpage = NULL;
2359         }
2360
2361         return true;
2362 }
2363
2364 static struct page *
2365 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2366                        struct vm_area_struct *vma, unsigned long address,
2367                        int node)
2368 {
2369         VM_BUG_ON_PAGE(*hpage, *hpage);
2370
2371         /*
2372          * Before allocating the hugepage, release the mmap_sem read lock.
2373          * The allocation can take potentially a long time if it involves
2374          * sync compaction, and we do not need to hold the mmap_sem during
2375          * that. We will recheck the vma after taking it again in write mode.
2376          */
2377         up_read(&mm->mmap_sem);
2378
2379         *hpage = alloc_pages_exact_node(node, gfp, HPAGE_PMD_ORDER);
2380         if (unlikely(!*hpage)) {
2381                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2382                 *hpage = ERR_PTR(-ENOMEM);
2383                 return NULL;
2384         }
2385
2386         count_vm_event(THP_COLLAPSE_ALLOC);
2387         return *hpage;
2388 }
2389 #else
2390 static int khugepaged_find_target_node(void)
2391 {
2392         return 0;
2393 }
2394
2395 static inline struct page *alloc_hugepage(int defrag)
2396 {
2397         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2398                            HPAGE_PMD_ORDER);
2399 }
2400
2401 static struct page *khugepaged_alloc_hugepage(bool *wait)
2402 {
2403         struct page *hpage;
2404
2405         do {
2406                 hpage = alloc_hugepage(khugepaged_defrag());
2407                 if (!hpage) {
2408                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2409                         if (!*wait)
2410                                 return NULL;
2411
2412                         *wait = false;
2413                         khugepaged_alloc_sleep();
2414                 } else
2415                         count_vm_event(THP_COLLAPSE_ALLOC);
2416         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2417
2418         return hpage;
2419 }
2420
2421 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2422 {
2423         if (!*hpage)
2424                 *hpage = khugepaged_alloc_hugepage(wait);
2425
2426         if (unlikely(!*hpage))
2427                 return false;
2428
2429         return true;
2430 }
2431
2432 static struct page *
2433 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2434                        struct vm_area_struct *vma, unsigned long address,
2435                        int node)
2436 {
2437         up_read(&mm->mmap_sem);
2438         VM_BUG_ON(!*hpage);
2439
2440         return  *hpage;
2441 }
2442 #endif
2443
2444 static bool hugepage_vma_check(struct vm_area_struct *vma)
2445 {
2446         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2447             (vma->vm_flags & VM_NOHUGEPAGE))
2448                 return false;
2449
2450         if (!vma->anon_vma || vma->vm_ops)
2451                 return false;
2452         if (is_vma_temporary_stack(vma))
2453                 return false;
2454         VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma);
2455         return true;
2456 }
2457
2458 static void collapse_huge_page(struct mm_struct *mm,
2459                                    unsigned long address,
2460                                    struct page **hpage,
2461                                    struct vm_area_struct *vma,
2462                                    int node)
2463 {
2464         pmd_t *pmd, _pmd;
2465         pte_t *pte;
2466         pgtable_t pgtable;
2467         struct page *new_page;
2468         spinlock_t *pmd_ptl, *pte_ptl;
2469         int isolated;
2470         unsigned long hstart, hend;
2471         struct mem_cgroup *memcg;
2472         unsigned long mmun_start;       /* For mmu_notifiers */
2473         unsigned long mmun_end;         /* For mmu_notifiers */
2474         gfp_t gfp;
2475
2476         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2477
2478         /* Only allocate from the target node */
2479         gfp = alloc_hugepage_gfpmask(khugepaged_defrag(), __GFP_OTHER_NODE) |
2480                 __GFP_THISNODE;
2481
2482         /* release the mmap_sem read lock. */
2483         new_page = khugepaged_alloc_page(hpage, gfp, mm, vma, address, node);
2484         if (!new_page)
2485                 return;
2486
2487         if (unlikely(mem_cgroup_try_charge(new_page, mm,
2488                                            gfp, &memcg)))
2489                 return;
2490
2491         /*
2492          * Prevent all access to pagetables with the exception of
2493          * gup_fast later hanlded by the ptep_clear_flush and the VM
2494          * handled by the anon_vma lock + PG_lock.
2495          */
2496         down_write(&mm->mmap_sem);
2497         if (unlikely(khugepaged_test_exit(mm)))
2498                 goto out;
2499
2500         vma = find_vma(mm, address);
2501         if (!vma)
2502                 goto out;
2503         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2504         hend = vma->vm_end & HPAGE_PMD_MASK;
2505         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2506                 goto out;
2507         if (!hugepage_vma_check(vma))
2508                 goto out;
2509         pmd = mm_find_pmd(mm, address);
2510         if (!pmd)
2511                 goto out;
2512
2513         anon_vma_lock_write(vma->anon_vma);
2514
2515         pte = pte_offset_map(pmd, address);
2516         pte_ptl = pte_lockptr(mm, pmd);
2517
2518         mmun_start = address;
2519         mmun_end   = address + HPAGE_PMD_SIZE;
2520         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2521         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2522         /*
2523          * After this gup_fast can't run anymore. This also removes
2524          * any huge TLB entry from the CPU so we won't allow
2525          * huge and small TLB entries for the same virtual address
2526          * to avoid the risk of CPU bugs in that area.
2527          */
2528         _pmd = pmdp_collapse_flush(vma, address, pmd);
2529         spin_unlock(pmd_ptl);
2530         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2531
2532         spin_lock(pte_ptl);
2533         isolated = __collapse_huge_page_isolate(vma, address, pte);
2534         spin_unlock(pte_ptl);
2535
2536         if (unlikely(!isolated)) {
2537                 pte_unmap(pte);
2538                 spin_lock(pmd_ptl);
2539                 BUG_ON(!pmd_none(*pmd));
2540                 /*
2541                  * We can only use set_pmd_at when establishing
2542                  * hugepmds and never for establishing regular pmds that
2543                  * points to regular pagetables. Use pmd_populate for that
2544                  */
2545                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2546                 spin_unlock(pmd_ptl);
2547                 anon_vma_unlock_write(vma->anon_vma);
2548                 goto out;
2549         }
2550
2551         /*
2552          * All pages are isolated and locked so anon_vma rmap
2553          * can't run anymore.
2554          */
2555         anon_vma_unlock_write(vma->anon_vma);
2556
2557         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2558         pte_unmap(pte);
2559         __SetPageUptodate(new_page);
2560         pgtable = pmd_pgtable(_pmd);
2561
2562         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2563         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2564
2565         /*
2566          * spin_lock() below is not the equivalent of smp_wmb(), so
2567          * this is needed to avoid the copy_huge_page writes to become
2568          * visible after the set_pmd_at() write.
2569          */
2570         smp_wmb();
2571
2572         spin_lock(pmd_ptl);
2573         BUG_ON(!pmd_none(*pmd));
2574         page_add_new_anon_rmap(new_page, vma, address);
2575         mem_cgroup_commit_charge(new_page, memcg, false);
2576         lru_cache_add_active_or_unevictable(new_page, vma);
2577         pgtable_trans_huge_deposit(mm, pmd, pgtable);
2578         set_pmd_at(mm, address, pmd, _pmd);
2579         update_mmu_cache_pmd(vma, address, pmd);
2580         spin_unlock(pmd_ptl);
2581
2582         *hpage = NULL;
2583
2584         khugepaged_pages_collapsed++;
2585 out_up_write:
2586         up_write(&mm->mmap_sem);
2587         return;
2588
2589 out:
2590         mem_cgroup_cancel_charge(new_page, memcg);
2591         goto out_up_write;
2592 }
2593
2594 static int khugepaged_scan_pmd(struct mm_struct *mm,
2595                                struct vm_area_struct *vma,
2596                                unsigned long address,
2597                                struct page **hpage)
2598 {
2599         pmd_t *pmd;
2600         pte_t *pte, *_pte;
2601         int ret = 0, none_or_zero = 0;
2602         struct page *page;
2603         unsigned long _address;
2604         spinlock_t *ptl;
2605         int node = NUMA_NO_NODE;
2606         bool writable = false, referenced = false;
2607
2608         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2609
2610         pmd = mm_find_pmd(mm, address);
2611         if (!pmd)
2612                 goto out;
2613
2614         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2615         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2616         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2617              _pte++, _address += PAGE_SIZE) {
2618                 pte_t pteval = *_pte;
2619                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2620                         if (!userfaultfd_armed(vma) &&
2621                             ++none_or_zero <= khugepaged_max_ptes_none)
2622                                 continue;
2623                         else
2624                                 goto out_unmap;
2625                 }
2626                 if (!pte_present(pteval))
2627                         goto out_unmap;
2628                 if (pte_write(pteval))
2629                         writable = true;
2630
2631                 page = vm_normal_page(vma, _address, pteval);
2632                 if (unlikely(!page))
2633                         goto out_unmap;
2634                 /*
2635                  * Record which node the original page is from and save this
2636                  * information to khugepaged_node_load[].
2637                  * Khupaged will allocate hugepage from the node has the max
2638                  * hit record.
2639                  */
2640                 node = page_to_nid(page);
2641                 if (khugepaged_scan_abort(node))
2642                         goto out_unmap;
2643                 khugepaged_node_load[node]++;
2644                 VM_BUG_ON_PAGE(PageCompound(page), page);
2645                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2646                         goto out_unmap;
2647                 /*
2648                  * cannot use mapcount: can't collapse if there's a gup pin.
2649                  * The page must only be referenced by the scanned process
2650                  * and page swap cache.
2651                  */
2652                 if (page_count(page) != 1 + !!PageSwapCache(page))
2653                         goto out_unmap;
2654                 if (pte_young(pteval) || PageReferenced(page) ||
2655                     mmu_notifier_test_young(vma->vm_mm, address))
2656                         referenced = true;
2657         }
2658         if (referenced && writable)
2659                 ret = 1;
2660 out_unmap:
2661         pte_unmap_unlock(pte, ptl);
2662         if (ret) {
2663                 node = khugepaged_find_target_node();
2664                 /* collapse_huge_page will return with the mmap_sem released */
2665                 collapse_huge_page(mm, address, hpage, vma, node);
2666         }
2667 out:
2668         return ret;
2669 }
2670
2671 static void collect_mm_slot(struct mm_slot *mm_slot)
2672 {
2673         struct mm_struct *mm = mm_slot->mm;
2674
2675         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2676
2677         if (khugepaged_test_exit(mm)) {
2678                 /* free mm_slot */
2679                 hash_del(&mm_slot->hash);
2680                 list_del(&mm_slot->mm_node);
2681
2682                 /*
2683                  * Not strictly needed because the mm exited already.
2684                  *
2685                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2686                  */
2687
2688                 /* khugepaged_mm_lock actually not necessary for the below */
2689                 free_mm_slot(mm_slot);
2690                 mmdrop(mm);
2691         }
2692 }
2693
2694 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2695                                             struct page **hpage)
2696         __releases(&khugepaged_mm_lock)
2697         __acquires(&khugepaged_mm_lock)
2698 {
2699         struct mm_slot *mm_slot;
2700         struct mm_struct *mm;
2701         struct vm_area_struct *vma;
2702         int progress = 0;
2703
2704         VM_BUG_ON(!pages);
2705         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2706
2707         if (khugepaged_scan.mm_slot)
2708                 mm_slot = khugepaged_scan.mm_slot;
2709         else {
2710                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2711                                      struct mm_slot, mm_node);
2712                 khugepaged_scan.address = 0;
2713                 khugepaged_scan.mm_slot = mm_slot;
2714         }
2715         spin_unlock(&khugepaged_mm_lock);
2716
2717         mm = mm_slot->mm;
2718         down_read(&mm->mmap_sem);
2719         if (unlikely(khugepaged_test_exit(mm)))
2720                 vma = NULL;
2721         else
2722                 vma = find_vma(mm, khugepaged_scan.address);
2723
2724         progress++;
2725         for (; vma; vma = vma->vm_next) {
2726                 unsigned long hstart, hend;
2727
2728                 cond_resched();
2729                 if (unlikely(khugepaged_test_exit(mm))) {
2730                         progress++;
2731                         break;
2732                 }
2733                 if (!hugepage_vma_check(vma)) {
2734 skip:
2735                         progress++;
2736                         continue;
2737                 }
2738                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2739                 hend = vma->vm_end & HPAGE_PMD_MASK;
2740                 if (hstart >= hend)
2741                         goto skip;
2742                 if (khugepaged_scan.address > hend)
2743                         goto skip;
2744                 if (khugepaged_scan.address < hstart)
2745                         khugepaged_scan.address = hstart;
2746                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2747
2748                 while (khugepaged_scan.address < hend) {
2749                         int ret;
2750                         cond_resched();
2751                         if (unlikely(khugepaged_test_exit(mm)))
2752                                 goto breakouterloop;
2753
2754                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2755                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2756                                   hend);
2757                         ret = khugepaged_scan_pmd(mm, vma,
2758                                                   khugepaged_scan.address,
2759                                                   hpage);
2760                         /* move to next address */
2761                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2762                         progress += HPAGE_PMD_NR;
2763                         if (ret)
2764                                 /* we released mmap_sem so break loop */
2765                                 goto breakouterloop_mmap_sem;
2766                         if (progress >= pages)
2767                                 goto breakouterloop;
2768                 }
2769         }
2770 breakouterloop:
2771         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2772 breakouterloop_mmap_sem:
2773
2774         spin_lock(&khugepaged_mm_lock);
2775         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2776         /*
2777          * Release the current mm_slot if this mm is about to die, or
2778          * if we scanned all vmas of this mm.
2779          */
2780         if (khugepaged_test_exit(mm) || !vma) {
2781                 /*
2782                  * Make sure that if mm_users is reaching zero while
2783                  * khugepaged runs here, khugepaged_exit will find
2784                  * mm_slot not pointing to the exiting mm.
2785                  */
2786                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2787                         khugepaged_scan.mm_slot = list_entry(
2788                                 mm_slot->mm_node.next,
2789                                 struct mm_slot, mm_node);
2790                         khugepaged_scan.address = 0;
2791                 } else {
2792                         khugepaged_scan.mm_slot = NULL;
2793                         khugepaged_full_scans++;
2794                 }
2795
2796                 collect_mm_slot(mm_slot);
2797         }
2798
2799         return progress;
2800 }
2801
2802 static int khugepaged_has_work(void)
2803 {
2804         return !list_empty(&khugepaged_scan.mm_head) &&
2805                 khugepaged_enabled();
2806 }
2807
2808 static int khugepaged_wait_event(void)
2809 {
2810         return !list_empty(&khugepaged_scan.mm_head) ||
2811                 kthread_should_stop();
2812 }
2813
2814 static void khugepaged_do_scan(void)
2815 {
2816         struct page *hpage = NULL;
2817         unsigned int progress = 0, pass_through_head = 0;
2818         unsigned int pages = khugepaged_pages_to_scan;
2819         bool wait = true;
2820
2821         barrier(); /* write khugepaged_pages_to_scan to local stack */
2822
2823         while (progress < pages) {
2824                 if (!khugepaged_prealloc_page(&hpage, &wait))
2825                         break;
2826
2827                 cond_resched();
2828
2829                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2830                         break;
2831
2832                 spin_lock(&khugepaged_mm_lock);
2833                 if (!khugepaged_scan.mm_slot)
2834                         pass_through_head++;
2835                 if (khugepaged_has_work() &&
2836                     pass_through_head < 2)
2837                         progress += khugepaged_scan_mm_slot(pages - progress,
2838                                                             &hpage);
2839                 else
2840                         progress = pages;
2841                 spin_unlock(&khugepaged_mm_lock);
2842         }
2843
2844         if (!IS_ERR_OR_NULL(hpage))
2845                 put_page(hpage);
2846 }
2847
2848 static void khugepaged_wait_work(void)
2849 {
2850         if (khugepaged_has_work()) {
2851                 if (!khugepaged_scan_sleep_millisecs)
2852                         return;
2853
2854                 wait_event_freezable_timeout(khugepaged_wait,
2855                                              kthread_should_stop(),
2856                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2857                 return;
2858         }
2859
2860         if (khugepaged_enabled())
2861                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2862 }
2863
2864 static int khugepaged(void *none)
2865 {
2866         struct mm_slot *mm_slot;
2867
2868         set_freezable();
2869         set_user_nice(current, MAX_NICE);
2870
2871         while (!kthread_should_stop()) {
2872                 khugepaged_do_scan();
2873                 khugepaged_wait_work();
2874         }
2875
2876         spin_lock(&khugepaged_mm_lock);
2877         mm_slot = khugepaged_scan.mm_slot;
2878         khugepaged_scan.mm_slot = NULL;
2879         if (mm_slot)
2880                 collect_mm_slot(mm_slot);
2881         spin_unlock(&khugepaged_mm_lock);
2882         return 0;
2883 }
2884
2885 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2886                 unsigned long haddr, pmd_t *pmd)
2887 {
2888         struct mm_struct *mm = vma->vm_mm;
2889         pgtable_t pgtable;
2890         pmd_t _pmd;
2891         int i;
2892
2893         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2894         /* leave pmd empty until pte is filled */
2895
2896         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2897         pmd_populate(mm, &_pmd, pgtable);
2898
2899         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2900                 pte_t *pte, entry;
2901                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2902                 entry = pte_mkspecial(entry);
2903                 pte = pte_offset_map(&_pmd, haddr);
2904                 VM_BUG_ON(!pte_none(*pte));
2905                 set_pte_at(mm, haddr, pte, entry);
2906                 pte_unmap(pte);
2907         }
2908         smp_wmb(); /* make pte visible before pmd */
2909         pmd_populate(mm, pmd, pgtable);
2910         put_huge_zero_page();
2911 }
2912
2913 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2914                 pmd_t *pmd)
2915 {
2916         spinlock_t *ptl;
2917         struct page *page;
2918         struct mm_struct *mm = vma->vm_mm;
2919         unsigned long haddr = address & HPAGE_PMD_MASK;
2920         unsigned long mmun_start;       /* For mmu_notifiers */
2921         unsigned long mmun_end;         /* For mmu_notifiers */
2922
2923         BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
2924
2925         mmun_start = haddr;
2926         mmun_end   = haddr + HPAGE_PMD_SIZE;
2927 again:
2928         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2929         ptl = pmd_lock(mm, pmd);
2930         if (unlikely(!pmd_trans_huge(*pmd))) {
2931                 spin_unlock(ptl);
2932                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2933                 return;
2934         }
2935         if (is_huge_zero_pmd(*pmd)) {
2936                 __split_huge_zero_page_pmd(vma, haddr, pmd);
2937                 spin_unlock(ptl);
2938                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2939                 return;
2940         }
2941         page = pmd_page(*pmd);
2942         VM_BUG_ON_PAGE(!page_count(page), page);
2943         get_page(page);
2944         spin_unlock(ptl);
2945         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2946
2947         split_huge_page(page);
2948
2949         put_page(page);
2950
2951         /*
2952          * We don't always have down_write of mmap_sem here: a racing
2953          * do_huge_pmd_wp_page() might have copied-on-write to another
2954          * huge page before our split_huge_page() got the anon_vma lock.
2955          */
2956         if (unlikely(pmd_trans_huge(*pmd)))
2957                 goto again;
2958 }
2959
2960 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2961                 pmd_t *pmd)
2962 {
2963         struct vm_area_struct *vma;
2964
2965         vma = find_vma(mm, address);
2966         BUG_ON(vma == NULL);
2967         split_huge_page_pmd(vma, address, pmd);
2968 }
2969
2970 static void split_huge_page_address(struct mm_struct *mm,
2971                                     unsigned long address)
2972 {
2973         pgd_t *pgd;
2974         pud_t *pud;
2975         pmd_t *pmd;
2976
2977         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2978
2979         pgd = pgd_offset(mm, address);
2980         if (!pgd_present(*pgd))
2981                 return;
2982
2983         pud = pud_offset(pgd, address);
2984         if (!pud_present(*pud))
2985                 return;
2986
2987         pmd = pmd_offset(pud, address);
2988         if (!pmd_present(*pmd))
2989                 return;
2990         /*
2991          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2992          * materialize from under us.
2993          */
2994         split_huge_page_pmd_mm(mm, address, pmd);
2995 }
2996
2997 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2998                              unsigned long start,
2999                              unsigned long end,
3000                              long adjust_next)
3001 {
3002         /*
3003          * If the new start address isn't hpage aligned and it could
3004          * previously contain an hugepage: check if we need to split
3005          * an huge pmd.
3006          */
3007         if (start & ~HPAGE_PMD_MASK &&
3008             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
3009             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3010                 split_huge_page_address(vma->vm_mm, start);
3011
3012         /*
3013          * If the new end address isn't hpage aligned and it could
3014          * previously contain an hugepage: check if we need to split
3015          * an huge pmd.
3016          */
3017         if (end & ~HPAGE_PMD_MASK &&
3018             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
3019             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3020                 split_huge_page_address(vma->vm_mm, end);
3021
3022         /*
3023          * If we're also updating the vma->vm_next->vm_start, if the new
3024          * vm_next->vm_start isn't page aligned and it could previously
3025          * contain an hugepage: check if we need to split an huge pmd.
3026          */
3027         if (adjust_next > 0) {
3028                 struct vm_area_struct *next = vma->vm_next;
3029                 unsigned long nstart = next->vm_start;
3030                 nstart += adjust_next << PAGE_SHIFT;
3031                 if (nstart & ~HPAGE_PMD_MASK &&
3032                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
3033                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
3034                         split_huge_page_address(next->vm_mm, nstart);
3035         }
3036 }