mm/huge_memory.c

   1 /*
   2  *  Copyright (C) 2009  Red Hat, Inc.
   3  *
   4  *  This work is licensed under the terms of the GNU GPL, version 2. See
   5  *  the COPYING file in the top-level directory.
   6  */
   7
   8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   9
  10 #include <linux/mm.h>
  11 #include <linux/sched.h>
  12 #include <linux/highmem.h>
  13 #include <linux/hugetlb.h>
  14 #include <linux/mmu_notifier.h>
  15 #include <linux/rmap.h>
  16 #include <linux/swap.h>
  17 #include <linux/shrinker.h>
  18 #include <linux/mm_inline.h>
  19 #include <linux/dax.h>
  20 #include <linux/kthread.h>
  21 #include <linux/khugepaged.h>
  22 #include <linux/freezer.h>
  23 #include <linux/mman.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/migrate.h>
  26 #include <linux/hashtable.h>
  27 #include <linux/userfaultfd_k.h>
  28
  29 #include <asm/tlb.h>
  30 #include <asm/pgalloc.h>
  31 #include "internal.h"
  32
  33 /*
  34  * By default transparent hugepage support is disabled in order that avoid
  35  * to risk increase the memory footprint of applications without a guaranteed
  36  * benefit. When transparent hugepage support is enabled, is for all mappings,
  37  * and khugepaged scans all mappings.
  38  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  39  * for all hugepage allocations.
  40  */
  41 unsigned long transparent_hugepage_flags __read_mostly =
  42 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  43         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  44 #endif
  45 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  46         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  47 #endif
  48         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
  49         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  50         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  51
  52 /* default scan 8*512 pte (or vmas) every 30 second */
  53 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
  54 static unsigned int khugepaged_pages_collapsed;
  55 static unsigned int khugepaged_full_scans;
  56 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  57 /* during fragmentation poll the hugepage allocator once every minute */
  58 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  59 static struct task_struct *khugepaged_thread __read_mostly;
  60 static DEFINE_MUTEX(khugepaged_mutex);
  61 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  62 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  63 /*
  64  * default collapse hugepages if there is at least one pte mapped like
  65  * it would have happened if the vma was large enough during page
  66  * fault.
  67  */
  68 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
  69
  70 static int khugepaged(void *none);
  71 static int khugepaged_slab_init(void);
  72 static void khugepaged_slab_exit(void);
  73
  74 #define MM_SLOTS_HASH_BITS 10
  75 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
  76
  77 static struct kmem_cache *mm_slot_cache __read_mostly;
  78
  79 /**
  80  * struct mm_slot - hash lookup from mm to mm_slot
  81  * @hash: hash collision list
  82  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  83  * @mm: the mm that this information is valid for
  84  */
  85 struct mm_slot {
  86         struct hlist_node hash;
  87         struct list_head mm_node;
  88         struct mm_struct *mm;
  89 };
  90
  91 /**
  92  * struct khugepaged_scan - cursor for scanning
  93  * @mm_head: the head of the mm list to scan
  94  * @mm_slot: the current mm_slot we are scanning
  95  * @address: the next address inside that to be scanned
  96  *
  97  * There is only the one khugepaged_scan instance of this cursor structure.
  98  */
  99 struct khugepaged_scan {
 100         struct list_head mm_head;
 101         struct mm_slot *mm_slot;
 102         unsigned long address;
 103 };
 104 static struct khugepaged_scan khugepaged_scan = {
 105         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 106 };
 107
 108
 109 static int set_recommended_min_free_kbytes(void)
 110 {
 111         struct zone *zone;
 112         int nr_zones = 0;
 113         unsigned long recommended_min;
 114
 115         for_each_populated_zone(zone)
 116                 nr_zones++;
 117
 118         /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
 119         recommended_min = pageblock_nr_pages * nr_zones * 2;
 120
 121         /*
 122          * Make sure that on average at least two pageblocks are almost free
 123          * of another type, one for a migratetype to fall back to and a
 124          * second to avoid subsequent fallbacks of other types There are 3
 125          * MIGRATE_TYPES we care about.
 126          */
 127         recommended_min += pageblock_nr_pages * nr_zones *
 128                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
 129
 130         /* don't ever allow to reserve more than 5% of the lowmem */
 131         recommended_min = min(recommended_min,
 132                               (unsigned long) nr_free_buffer_pages() / 20);
 133         recommended_min <<= (PAGE_SHIFT-10);
 134
 135         if (recommended_min > min_free_kbytes) {
 136                 if (user_min_free_kbytes >= 0)
 137                         pr_info("raising min_free_kbytes from %d to %lu "
 138                                 "to help transparent hugepage allocations\n",
 139                                 min_free_kbytes, recommended_min);
 140
 141                 min_free_kbytes = recommended_min;
 142         }
 143         setup_per_zone_wmarks();
 144         return 0;
 145 }
 146
 147 static int start_stop_khugepaged(void)
 148 {
 149         int err = 0;
 150         if (khugepaged_enabled()) {
 151                 if (!khugepaged_thread)
 152                         khugepaged_thread = kthread_run(khugepaged, NULL,
 153                                                         "khugepaged");
 154                 if (unlikely(IS_ERR(khugepaged_thread))) {
 155                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
 156                         err = PTR_ERR(khugepaged_thread);
 157                         khugepaged_thread = NULL;
 158                         goto fail;
 159                 }
 160
 161                 if (!list_empty(&khugepaged_scan.mm_head))
 162                         wake_up_interruptible(&khugepaged_wait);
 163
 164                 set_recommended_min_free_kbytes();
 165         } else if (khugepaged_thread) {
 166                 kthread_stop(khugepaged_thread);
 167                 khugepaged_thread = NULL;
 168         }
 169 fail:
 170         return err;
 171 }
 172
 173 static atomic_t huge_zero_refcount;
 174 struct page *huge_zero_page __read_mostly;
 175
 176 struct page *get_huge_zero_page(void)
 177 {
 178         struct page *zero_page;
 179 retry:
 180         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
 181                 return READ_ONCE(huge_zero_page);
 182
 183         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
 184                         HPAGE_PMD_ORDER);
 185         if (!zero_page) {
 186                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
 187                 return NULL;
 188         }
 189         count_vm_event(THP_ZERO_PAGE_ALLOC);
 190         preempt_disable();
 191         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
 192                 preempt_enable();
 193                 __free_pages(zero_page, compound_order(zero_page));
 194                 goto retry;
 195         }
 196
 197         /* We take additional reference here. It will be put back by shrinker */
 198         atomic_set(&huge_zero_refcount, 2);
 199         preempt_enable();
 200         return READ_ONCE(huge_zero_page);
 201 }
 202
 203 static void put_huge_zero_page(void)
 204 {
 205         /*
 206          * Counter should never go to zero here. Only shrinker can put
 207          * last reference.
 208          */
 209         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 210 }
 211
 212 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 213                                         struct shrink_control *sc)
 214 {
 215         /* we can free zero page only if last reference remains */
 216         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 217 }
 218
 219 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 220                                        struct shrink_control *sc)
 221 {
 222         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 223                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 224                 BUG_ON(zero_page == NULL);
 225                 __free_pages(zero_page, compound_order(zero_page));
 226                 return HPAGE_PMD_NR;
 227         }
 228
 229         return 0;
 230 }
 231
 232 static struct shrinker huge_zero_page_shrinker = {
 233         .count_objects = shrink_huge_zero_page_count,
 234         .scan_objects = shrink_huge_zero_page_scan,
 235         .seeks = DEFAULT_SEEKS,
 236 };
 237
 238 #ifdef CONFIG_SYSFS
 239
 240 static ssize_t double_flag_show(struct kobject *kobj,
 241                                 struct kobj_attribute *attr, char *buf,
 242                                 enum transparent_hugepage_flag enabled,
 243                                 enum transparent_hugepage_flag req_madv)
 244 {
 245         if (test_bit(enabled, &transparent_hugepage_flags)) {
 246                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
 247                 return sprintf(buf, "[always] madvise never\n");
 248         } else if (test_bit(req_madv, &transparent_hugepage_flags))
 249                 return sprintf(buf, "always [madvise] never\n");
 250         else
 251                 return sprintf(buf, "always madvise [never]\n");
 252 }
 253 static ssize_t double_flag_store(struct kobject *kobj,
 254                                  struct kobj_attribute *attr,
 255                                  const char *buf, size_t count,
 256                                  enum transparent_hugepage_flag enabled,
 257                                  enum transparent_hugepage_flag req_madv)
 258 {
 259         if (!memcmp("always", buf,
 260                     min(sizeof("always")-1, count))) {
 261                 set_bit(enabled, &transparent_hugepage_flags);
 262                 clear_bit(req_madv, &transparent_hugepage_flags);
 263         } else if (!memcmp("madvise", buf,
 264                            min(sizeof("madvise")-1, count))) {
 265                 clear_bit(enabled, &transparent_hugepage_flags);
 266                 set_bit(req_madv, &transparent_hugepage_flags);
 267         } else if (!memcmp("never", buf,
 268                            min(sizeof("never")-1, count))) {
 269                 clear_bit(enabled, &transparent_hugepage_flags);
 270                 clear_bit(req_madv, &transparent_hugepage_flags);
 271         } else
 272                 return -EINVAL;
 273
 274         return count;
 275 }
 276
 277 static ssize_t enabled_show(struct kobject *kobj,
 278                             struct kobj_attribute *attr, char *buf)
 279 {
 280         return double_flag_show(kobj, attr, buf,
 281                                 TRANSPARENT_HUGEPAGE_FLAG,
 282                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 283 }
 284 static ssize_t enabled_store(struct kobject *kobj,
 285                              struct kobj_attribute *attr,
 286                              const char *buf, size_t count)
 287 {
 288         ssize_t ret;
 289
 290         ret = double_flag_store(kobj, attr, buf, count,
 291                                 TRANSPARENT_HUGEPAGE_FLAG,
 292                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 293
 294         if (ret > 0) {
 295                 int err;
 296
 297                 mutex_lock(&khugepaged_mutex);
 298                 err = start_stop_khugepaged();
 299                 mutex_unlock(&khugepaged_mutex);
 300
 301                 if (err)
 302                         ret = err;
 303         }
 304
 305         return ret;
 306 }
 307 static struct kobj_attribute enabled_attr =
 308         __ATTR(enabled, 0644, enabled_show, enabled_store);
 309
 310 static ssize_t single_flag_show(struct kobject *kobj,
 311                                 struct kobj_attribute *attr, char *buf,
 312                                 enum transparent_hugepage_flag flag)
 313 {
 314         return sprintf(buf, "%d\n",
 315                        !!test_bit(flag, &transparent_hugepage_flags));
 316 }
 317
 318 static ssize_t single_flag_store(struct kobject *kobj,
 319                                  struct kobj_attribute *attr,
 320                                  const char *buf, size_t count,
 321                                  enum transparent_hugepage_flag flag)
 322 {
 323         unsigned long value;
 324         int ret;
 325
 326         ret = kstrtoul(buf, 10, &value);
 327         if (ret < 0)
 328                 return ret;
 329         if (value > 1)
 330                 return -EINVAL;
 331
 332         if (value)
 333                 set_bit(flag, &transparent_hugepage_flags);
 334         else
 335                 clear_bit(flag, &transparent_hugepage_flags);
 336
 337         return count;
 338 }
 339
 340 /*
 341  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 342  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 343  * memory just to allocate one more hugepage.
 344  */
 345 static ssize_t defrag_show(struct kobject *kobj,
 346                            struct kobj_attribute *attr, char *buf)
 347 {
 348         return double_flag_show(kobj, attr, buf,
 349                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 350                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 351 }
 352 static ssize_t defrag_store(struct kobject *kobj,
 353                             struct kobj_attribute *attr,
 354                             const char *buf, size_t count)
 355 {
 356         return double_flag_store(kobj, attr, buf, count,
 357                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 358                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 359 }
 360 static struct kobj_attribute defrag_attr =
 361         __ATTR(defrag, 0644, defrag_show, defrag_store);
 362
 363 static ssize_t use_zero_page_show(struct kobject *kobj,
 364                 struct kobj_attribute *attr, char *buf)
 365 {
 366         return single_flag_show(kobj, attr, buf,
 367                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 368 }
 369 static ssize_t use_zero_page_store(struct kobject *kobj,
 370                 struct kobj_attribute *attr, const char *buf, size_t count)
 371 {
 372         return single_flag_store(kobj, attr, buf, count,
 373                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 374 }
 375 static struct kobj_attribute use_zero_page_attr =
 376         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 377 #ifdef CONFIG_DEBUG_VM
 378 static ssize_t debug_cow_show(struct kobject *kobj,
 379                                 struct kobj_attribute *attr, char *buf)
 380 {
 381         return single_flag_show(kobj, attr, buf,
 382                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 383 }
 384 static ssize_t debug_cow_store(struct kobject *kobj,
 385                                struct kobj_attribute *attr,
 386                                const char *buf, size_t count)
 387 {
 388         return single_flag_store(kobj, attr, buf, count,
 389                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 390 }
 391 static struct kobj_attribute debug_cow_attr =
 392         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 393 #endif /* CONFIG_DEBUG_VM */
 394
 395 static struct attribute *hugepage_attr[] = {
 396         &enabled_attr.attr,
 397         &defrag_attr.attr,
 398         &use_zero_page_attr.attr,
 399 #ifdef CONFIG_DEBUG_VM
 400         &debug_cow_attr.attr,
 401 #endif
 402         NULL,
 403 };
 404
 405 static struct attribute_group hugepage_attr_group = {
 406         .attrs = hugepage_attr,
 407 };
 408
 409 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 410                                          struct kobj_attribute *attr,
 411                                          char *buf)
 412 {
 413         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 414 }
 415
 416 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 417                                           struct kobj_attribute *attr,
 418                                           const char *buf, size_t count)
 419 {
 420         unsigned long msecs;
 421         int err;
 422
 423         err = kstrtoul(buf, 10, &msecs);
 424         if (err || msecs > UINT_MAX)
 425                 return -EINVAL;
 426
 427         khugepaged_scan_sleep_millisecs = msecs;
 428         wake_up_interruptible(&khugepaged_wait);
 429
 430         return count;
 431 }
 432 static struct kobj_attribute scan_sleep_millisecs_attr =
 433         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 434                scan_sleep_millisecs_store);
 435
 436 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 437                                           struct kobj_attribute *attr,
 438                                           char *buf)
 439 {
 440         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 441 }
 442
 443 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 444                                            struct kobj_attribute *attr,
 445                                            const char *buf, size_t count)
 446 {
 447         unsigned long msecs;
 448         int err;
 449
 450         err = kstrtoul(buf, 10, &msecs);
 451         if (err || msecs > UINT_MAX)
 452                 return -EINVAL;
 453
 454         khugepaged_alloc_sleep_millisecs = msecs;
 455         wake_up_interruptible(&khugepaged_wait);
 456
 457         return count;
 458 }
 459 static struct kobj_attribute alloc_sleep_millisecs_attr =
 460         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 461                alloc_sleep_millisecs_store);
 462
 463 static ssize_t pages_to_scan_show(struct kobject *kobj,
 464                                   struct kobj_attribute *attr,
 465                                   char *buf)
 466 {
 467         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 468 }
 469 static ssize_t pages_to_scan_store(struct kobject *kobj,
 470                                    struct kobj_attribute *attr,
 471                                    const char *buf, size_t count)
 472 {
 473         int err;
 474         unsigned long pages;
 475
 476         err = kstrtoul(buf, 10, &pages);
 477         if (err || !pages || pages > UINT_MAX)
 478                 return -EINVAL;
 479
 480         khugepaged_pages_to_scan = pages;
 481
 482         return count;
 483 }
 484 static struct kobj_attribute pages_to_scan_attr =
 485         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 486                pages_to_scan_store);
 487
 488 static ssize_t pages_collapsed_show(struct kobject *kobj,
 489                                     struct kobj_attribute *attr,
 490                                     char *buf)
 491 {
 492         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 493 }
 494 static struct kobj_attribute pages_collapsed_attr =
 495         __ATTR_RO(pages_collapsed);
 496
 497 static ssize_t full_scans_show(struct kobject *kobj,
 498                                struct kobj_attribute *attr,
 499                                char *buf)
 500 {
 501         return sprintf(buf, "%u\n", khugepaged_full_scans);
 502 }
 503 static struct kobj_attribute full_scans_attr =
 504         __ATTR_RO(full_scans);
 505
 506 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 507                                       struct kobj_attribute *attr, char *buf)
 508 {
 509         return single_flag_show(kobj, attr, buf,
 510                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 511 }
 512 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 513                                        struct kobj_attribute *attr,
 514                                        const char *buf, size_t count)
 515 {
 516         return single_flag_store(kobj, attr, buf, count,
 517                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 518 }
 519 static struct kobj_attribute khugepaged_defrag_attr =
 520         __ATTR(defrag, 0644, khugepaged_defrag_show,
 521                khugepaged_defrag_store);
 522
 523 /*
 524  * max_ptes_none controls if khugepaged should collapse hugepages over
 525  * any unmapped ptes in turn potentially increasing the memory
 526  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 527  * reduce the available free memory in the system as it
 528  * runs. Increasing max_ptes_none will instead potentially reduce the
 529  * free memory in the system during the khugepaged scan.
 530  */
 531 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 532                                              struct kobj_attribute *attr,
 533                                              char *buf)
 534 {
 535         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 536 }
 537 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 538                                               struct kobj_attribute *attr,
 539                                               const char *buf, size_t count)
 540 {
 541         int err;
 542         unsigned long max_ptes_none;
 543
 544         err = kstrtoul(buf, 10, &max_ptes_none);
 545         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 546                 return -EINVAL;
 547
 548         khugepaged_max_ptes_none = max_ptes_none;
 549
 550         return count;
 551 }
 552 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 553         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 554                khugepaged_max_ptes_none_store);
 555
 556 static struct attribute *khugepaged_attr[] = {
 557         &khugepaged_defrag_attr.attr,
 558         &khugepaged_max_ptes_none_attr.attr,
 559         &pages_to_scan_attr.attr,
 560         &pages_collapsed_attr.attr,
 561         &full_scans_attr.attr,
 562         &scan_sleep_millisecs_attr.attr,
 563         &alloc_sleep_millisecs_attr.attr,
 564         NULL,
 565 };
 566
 567 static struct attribute_group khugepaged_attr_group = {
 568         .attrs = khugepaged_attr,
 569         .name = "khugepaged",
 570 };
 571
 572 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 573 {
 574         int err;
 575
 576         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 577         if (unlikely(!*hugepage_kobj)) {
 578                 pr_err("failed to create transparent hugepage kobject\n");
 579                 return -ENOMEM;
 580         }
 581
 582         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 583         if (err) {
 584                 pr_err("failed to register transparent hugepage group\n");
 585                 goto delete_obj;
 586         }
 587
 588         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 589         if (err) {
 590                 pr_err("failed to register transparent hugepage group\n");
 591                 goto remove_hp_group;
 592         }
 593
 594         return 0;
 595
 596 remove_hp_group:
 597         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 598 delete_obj:
 599         kobject_put(*hugepage_kobj);
 600         return err;
 601 }
 602
 603 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 604 {
 605         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 606         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 607         kobject_put(hugepage_kobj);
 608 }
 609 #else
 610 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 611 {
 612         return 0;
 613 }
 614
 615 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 616 {
 617 }
 618 #endif /* CONFIG_SYSFS */
 619
 620 static int __init hugepage_init(void)
 621 {
 622         int err;
 623         struct kobject *hugepage_kobj;
 624
 625         if (!has_transparent_hugepage()) {
 626                 transparent_hugepage_flags = 0;
 627                 return -EINVAL;
 628         }
 629
 630         err = hugepage_init_sysfs(&hugepage_kobj);
 631         if (err)
 632                 goto err_sysfs;
 633
 634         err = khugepaged_slab_init();
 635         if (err)
 636                 goto err_slab;
 637
 638         err = register_shrinker(&huge_zero_page_shrinker);
 639         if (err)
 640                 goto err_hzp_shrinker;
 641
 642         /*
 643          * By default disable transparent hugepages on smaller systems,
 644          * where the extra memory used could hurt more than TLB overhead
 645          * is likely to save.  The admin can still enable it through /sys.
 646          */
 647         if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
 648                 transparent_hugepage_flags = 0;
 649                 return 0;
 650         }
 651
 652         err = start_stop_khugepaged();
 653         if (err)
 654                 goto err_khugepaged;
 655
 656         return 0;
 657 err_khugepaged:
 658         unregister_shrinker(&huge_zero_page_shrinker);
 659 err_hzp_shrinker:
 660         khugepaged_slab_exit();
 661 err_slab:
 662         hugepage_exit_sysfs(hugepage_kobj);
 663 err_sysfs:
 664         return err;
 665 }
 666 subsys_initcall(hugepage_init);
 667
 668 static int __init setup_transparent_hugepage(char *str)
 669 {
 670         int ret = 0;
 671         if (!str)
 672                 goto out;
 673         if (!strcmp(str, "always")) {
 674                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 675                         &transparent_hugepage_flags);
 676                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 677                           &transparent_hugepage_flags);
 678                 ret = 1;
 679         } else if (!strcmp(str, "madvise")) {
 680                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 681                           &transparent_hugepage_flags);
 682                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 683                         &transparent_hugepage_flags);
 684                 ret = 1;
 685         } else if (!strcmp(str, "never")) {
 686                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 687                           &transparent_hugepage_flags);
 688                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 689                           &transparent_hugepage_flags);
 690                 ret = 1;
 691         }
 692 out:
 693         if (!ret)
 694                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
 695         return ret;
 696 }
 697 __setup("transparent_hugepage=", setup_transparent_hugepage);
 698
 699 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 700 {
 701         if (likely(vma->vm_flags & VM_WRITE))
 702                 pmd = pmd_mkwrite(pmd);
 703         return pmd;
 704 }
 705
 706 static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
 707 {
 708         pmd_t entry;
 709         entry = mk_pmd(page, prot);
 710         entry = pmd_mkhuge(entry);
 711         return entry;
 712 }
 713
 714 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
 715                                         struct vm_area_struct *vma,
 716                                         unsigned long address, pmd_t *pmd,
 717                                         struct page *page, gfp_t gfp,
 718                                         unsigned int flags)
 719 {
 720         struct mem_cgroup *memcg;
 721         pgtable_t pgtable;
 722         spinlock_t *ptl;
 723         unsigned long haddr = address & HPAGE_PMD_MASK;
 724
 725         VM_BUG_ON_PAGE(!PageCompound(page), page);
 726
 727         if (mem_cgroup_try_charge(page, mm, gfp, &memcg)) {
 728                 put_page(page);
 729                 count_vm_event(THP_FAULT_FALLBACK);
 730                 return VM_FAULT_FALLBACK;
 731         }
 732
 733         pgtable = pte_alloc_one(mm, haddr);
 734         if (unlikely(!pgtable)) {
 735                 mem_cgroup_cancel_charge(page, memcg);
 736                 put_page(page);
 737                 return VM_FAULT_OOM;
 738         }
 739
 740         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 741         /*
 742          * The memory barrier inside __SetPageUptodate makes sure that
 743          * clear_huge_page writes become visible before the set_pmd_at()
 744          * write.
 745          */
 746         __SetPageUptodate(page);
 747
 748         ptl = pmd_lock(mm, pmd);
 749         if (unlikely(!pmd_none(*pmd))) {
 750                 spin_unlock(ptl);
 751                 mem_cgroup_cancel_charge(page, memcg);
 752                 put_page(page);
 753                 pte_free(mm, pgtable);
 754         } else {
 755                 pmd_t entry;
 756
 757                 /* Deliver the page fault to userland */
 758                 if (userfaultfd_missing(vma)) {
 759                         int ret;
 760
 761                         spin_unlock(ptl);
 762                         mem_cgroup_cancel_charge(page, memcg);
 763                         put_page(page);
 764                         pte_free(mm, pgtable);
 765                         ret = handle_userfault(vma, address, flags,
 766                                                VM_UFFD_MISSING);
 767                         VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 768                         return ret;
 769                 }
 770
 771                 entry = mk_huge_pmd(page, vma->vm_page_prot);
 772                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 773                 page_add_new_anon_rmap(page, vma, haddr);
 774                 mem_cgroup_commit_charge(page, memcg, false);
 775                 lru_cache_add_active_or_unevictable(page, vma);
 776                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
 777                 set_pmd_at(mm, haddr, pmd, entry);
 778                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 779                 atomic_long_inc(&mm->nr_ptes);
 780                 spin_unlock(ptl);
 781                 count_vm_event(THP_FAULT_ALLOC);
 782         }
 783
 784         return 0;
 785 }
 786
 787 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 788 {
 789         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
 790 }
 791
 792 /* Caller must hold page table lock. */
 793 bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 794                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 795                 struct page *zero_page)
 796 {
 797         pmd_t entry;
 798         if (!pmd_none(*pmd))
 799                 return false;
 800         entry = mk_pmd(zero_page, vma->vm_page_prot);
 801         entry = pmd_mkhuge(entry);
 802         pgtable_trans_huge_deposit(mm, pmd, pgtable);
 803         set_pmd_at(mm, haddr, pmd, entry);
 804         atomic_long_inc(&mm->nr_ptes);
 805         return true;
 806 }
 807
 808 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 809                                unsigned long address, pmd_t *pmd,
 810                                unsigned int flags)
 811 {
 812         gfp_t gfp;
 813         struct page *page;
 814         unsigned long haddr = address & HPAGE_PMD_MASK;
 815
 816         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 817                 return VM_FAULT_FALLBACK;
 818         if (unlikely(anon_vma_prepare(vma)))
 819                 return VM_FAULT_OOM;
 820         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 821                 return VM_FAULT_OOM;
 822         if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm) &&
 823                         transparent_hugepage_use_zero_page()) {
 824                 spinlock_t *ptl;
 825                 pgtable_t pgtable;
 826                 struct page *zero_page;
 827                 bool set;
 828                 int ret;
 829                 pgtable = pte_alloc_one(mm, haddr);
 830                 if (unlikely(!pgtable))
 831                         return VM_FAULT_OOM;
 832                 zero_page = get_huge_zero_page();
 833                 if (unlikely(!zero_page)) {
 834                         pte_free(mm, pgtable);
 835                         count_vm_event(THP_FAULT_FALLBACK);
 836                         return VM_FAULT_FALLBACK;
 837                 }
 838                 ptl = pmd_lock(mm, pmd);
 839                 ret = 0;
 840                 set = false;
 841                 if (pmd_none(*pmd)) {
 842                         if (userfaultfd_missing(vma)) {
 843                                 spin_unlock(ptl);
 844                                 ret = handle_userfault(vma, address, flags,
 845                                                        VM_UFFD_MISSING);
 846                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 847                         } else {
 848                                 set_huge_zero_page(pgtable, mm, vma,
 849                                                    haddr, pmd,
 850                                                    zero_page);
 851                                 spin_unlock(ptl);
 852                                 set = true;
 853                         }
 854                 } else
 855                         spin_unlock(ptl);
 856                 if (!set) {
 857                         pte_free(mm, pgtable);
 858                         put_huge_zero_page();
 859                 }
 860                 return ret;
 861         }
 862         gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
 863         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 864         if (unlikely(!page)) {
 865                 count_vm_event(THP_FAULT_FALLBACK);
 866                 return VM_FAULT_FALLBACK;
 867         }
 868         return __do_huge_pmd_anonymous_page(mm, vma, address, pmd, page, gfp,
 869                                             flags);
 870 }
 871
 872 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 873                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 874                   struct vm_area_struct *vma)
 875 {
 876         spinlock_t *dst_ptl, *src_ptl;
 877         struct page *src_page;
 878         pmd_t pmd;
 879         pgtable_t pgtable;
 880         int ret;
 881
 882         ret = -ENOMEM;
 883         pgtable = pte_alloc_one(dst_mm, addr);
 884         if (unlikely(!pgtable))
 885                 goto out;
 886
 887         dst_ptl = pmd_lock(dst_mm, dst_pmd);
 888         src_ptl = pmd_lockptr(src_mm, src_pmd);
 889         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 890
 891         ret = -EAGAIN;
 892         pmd = *src_pmd;
 893         if (unlikely(!pmd_trans_huge(pmd))) {
 894                 pte_free(dst_mm, pgtable);
 895                 goto out_unlock;
 896         }
 897         /*
 898          * When page table lock is held, the huge zero pmd should not be
 899          * under splitting since we don't split the page itself, only pmd to
 900          * a page table.
 901          */
 902         if (is_huge_zero_pmd(pmd)) {
 903                 struct page *zero_page;
 904                 /*
 905                  * get_huge_zero_page() will never allocate a new page here,
 906                  * since we already have a zero page to copy. It just takes a
 907                  * reference.
 908                  */
 909                 zero_page = get_huge_zero_page();
 910                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
 911                                 zero_page);
 912                 ret = 0;
 913                 goto out_unlock;
 914         }
 915
 916         if (unlikely(pmd_trans_splitting(pmd))) {
 917                 /* split huge page running from under us */
 918                 spin_unlock(src_ptl);
 919                 spin_unlock(dst_ptl);
 920                 pte_free(dst_mm, pgtable);
 921
 922                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 923                 goto out;
 924         }
 925         src_page = pmd_page(pmd);
 926         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
 927         get_page(src_page);
 928         page_dup_rmap(src_page);
 929         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 930
 931         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 932         pmd = pmd_mkold(pmd_wrprotect(pmd));
 933         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
 934         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 935         atomic_long_inc(&dst_mm->nr_ptes);
 936
 937         ret = 0;
 938 out_unlock:
 939         spin_unlock(src_ptl);
 940         spin_unlock(dst_ptl);
 941 out:
 942         return ret;
 943 }
 944
 945 void huge_pmd_set_accessed(struct mm_struct *mm,
 946                            struct vm_area_struct *vma,
 947                            unsigned long address,
 948                            pmd_t *pmd, pmd_t orig_pmd,
 949                            int dirty)
 950 {
 951         spinlock_t *ptl;
 952         pmd_t entry;
 953         unsigned long haddr;
 954
 955         ptl = pmd_lock(mm, pmd);
 956         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 957                 goto unlock;
 958
 959         entry = pmd_mkyoung(orig_pmd);
 960         haddr = address & HPAGE_PMD_MASK;
 961         if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
 962                 update_mmu_cache_pmd(vma, address, pmd);
 963
 964 unlock:
 965         spin_unlock(ptl);
 966 }
 967
 968 /*
 969  * Save CONFIG_DEBUG_PAGEALLOC from faulting falsely on tail pages
 970  * during copy_user_huge_page()'s copy_page_rep(): in the case when
 971  * the source page gets split and a tail freed before copy completes.
 972  * Called under pmd_lock of checked pmd, so safe from splitting itself.
 973  */
 974 static void get_user_huge_page(struct page *page)
 975 {
 976         if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
 977                 struct page *endpage = page + HPAGE_PMD_NR;
 978
 979                 atomic_add(HPAGE_PMD_NR, &page->_count);
 980                 while (++page < endpage)
 981                         get_huge_page_tail(page);
 982         } else {
 983                 get_page(page);
 984         }
 985 }
 986
 987 static void put_user_huge_page(struct page *page)
 988 {
 989         if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
 990                 struct page *endpage = page + HPAGE_PMD_NR;
 991
 992                 while (page < endpage)
 993                         put_page(page++);
 994         } else {
 995                 put_page(page);
 996         }
 997 }
 998
 999 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1000                                         struct vm_area_struct *vma,
1001                                         unsigned long address,
1002                                         pmd_t *pmd, pmd_t orig_pmd,
1003                                         struct page *page,
1004                                         unsigned long haddr)
1005 {
1006         struct mem_cgroup *memcg;
1007         spinlock_t *ptl;
1008         pgtable_t pgtable;
1009         pmd_t _pmd;
1010         int ret = 0, i;
1011         struct page **pages;
1012         unsigned long mmun_start;       /* For mmu_notifiers */
1013         unsigned long mmun_end;         /* For mmu_notifiers */
1014
1015         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1016                         GFP_KERNEL);
1017         if (unlikely(!pages)) {
1018                 ret |= VM_FAULT_OOM;
1019                 goto out;
1020         }
1021
1022         for (i = 0; i < HPAGE_PMD_NR; i++) {
1023                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1024                                                __GFP_OTHER_NODE,
1025                                                vma, address, page_to_nid(page));
1026                 if (unlikely(!pages[i] ||
1027                              mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL,
1028                                                    &memcg))) {
1029                         if (pages[i])
1030                                 put_page(pages[i]);
1031                         while (--i >= 0) {
1032                                 memcg = (void *)page_private(pages[i]);
1033                                 set_page_private(pages[i], 0);
1034                                 mem_cgroup_cancel_charge(pages[i], memcg);
1035                                 put_page(pages[i]);
1036                         }
1037                         kfree(pages);
1038                         ret |= VM_FAULT_OOM;
1039                         goto out;
1040                 }
1041                 set_page_private(pages[i], (unsigned long)memcg);
1042         }
1043
1044         for (i = 0; i < HPAGE_PMD_NR; i++) {
1045                 copy_user_highpage(pages[i], page + i,
1046                                    haddr + PAGE_SIZE * i, vma);
1047                 __SetPageUptodate(pages[i]);
1048                 cond_resched();
1049         }
1050
1051         mmun_start = haddr;
1052         mmun_end   = haddr + HPAGE_PMD_SIZE;
1053         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1054
1055         ptl = pmd_lock(mm, pmd);
1056         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1057                 goto out_free_pages;
1058         VM_BUG_ON_PAGE(!PageHead(page), page);
1059
1060         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1061         /* leave pmd empty until pte is filled */
1062
1063         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1064         pmd_populate(mm, &_pmd, pgtable);
1065
1066         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1067                 pte_t *pte, entry;
1068                 entry = mk_pte(pages[i], vma->vm_page_prot);
1069                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1070                 memcg = (void *)page_private(pages[i]);
1071                 set_page_private(pages[i], 0);
1072                 page_add_new_anon_rmap(pages[i], vma, haddr);
1073                 mem_cgroup_commit_charge(pages[i], memcg, false);
1074                 lru_cache_add_active_or_unevictable(pages[i], vma);
1075                 pte = pte_offset_map(&_pmd, haddr);
1076                 VM_BUG_ON(!pte_none(*pte));
1077                 set_pte_at(mm, haddr, pte, entry);
1078                 pte_unmap(pte);
1079         }
1080         kfree(pages);
1081
1082         smp_wmb(); /* make pte visible before pmd */
1083         pmd_populate(mm, pmd, pgtable);
1084         page_remove_rmap(page);
1085         spin_unlock(ptl);
1086
1087         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1088
1089         ret |= VM_FAULT_WRITE;
1090         put_page(page);
1091
1092 out:
1093         return ret;
1094
1095 out_free_pages:
1096         spin_unlock(ptl);
1097         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1098         for (i = 0; i < HPAGE_PMD_NR; i++) {
1099                 memcg = (void *)page_private(pages[i]);
1100                 set_page_private(pages[i], 0);
1101                 mem_cgroup_cancel_charge(pages[i], memcg);
1102                 put_page(pages[i]);
1103         }
1104         kfree(pages);
1105         goto out;
1106 }
1107
1108 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1109                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1110 {
1111         spinlock_t *ptl;
1112         int ret = 0;
1113         struct page *page = NULL, *new_page;
1114         struct mem_cgroup *memcg;
1115         unsigned long haddr;
1116         unsigned long mmun_start;       /* For mmu_notifiers */
1117         unsigned long mmun_end;         /* For mmu_notifiers */
1118         gfp_t huge_gfp;                 /* for allocation and charge */
1119
1120         ptl = pmd_lockptr(mm, pmd);
1121         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1122         haddr = address & HPAGE_PMD_MASK;
1123         if (is_huge_zero_pmd(orig_pmd))
1124                 goto alloc;
1125         spin_lock(ptl);
1126         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1127                 goto out_unlock;
1128
1129         page = pmd_page(orig_pmd);
1130         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1131         if (page_mapcount(page) == 1) {
1132                 pmd_t entry;
1133                 entry = pmd_mkyoung(orig_pmd);
1134                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1135                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1136                         update_mmu_cache_pmd(vma, address, pmd);
1137                 ret |= VM_FAULT_WRITE;
1138                 goto out_unlock;
1139         }
1140         get_user_huge_page(page);
1141         spin_unlock(ptl);
1142 alloc:
1143         if (transparent_hugepage_enabled(vma) &&
1144             !transparent_hugepage_debug_cow()) {
1145                 huge_gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
1146                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1147         } else
1148                 new_page = NULL;
1149
1150         if (unlikely(!new_page)) {
1151                 if (!page) {
1152                         split_huge_page_pmd(vma, address, pmd);
1153                         ret |= VM_FAULT_FALLBACK;
1154                 } else {
1155                         ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1156                                         pmd, orig_pmd, page, haddr);
1157                         if (ret & VM_FAULT_OOM) {
1158                                 split_huge_page(page);
1159                                 ret |= VM_FAULT_FALLBACK;
1160                         }
1161                         put_user_huge_page(page);
1162                 }
1163                 count_vm_event(THP_FAULT_FALLBACK);
1164                 goto out;
1165         }
1166
1167         if (unlikely(mem_cgroup_try_charge(new_page, mm, huge_gfp, &memcg))) {
1168                 put_page(new_page);
1169                 if (page) {
1170                         split_huge_page(page);
1171                         put_user_huge_page(page);
1172                 } else
1173                         split_huge_page_pmd(vma, address, pmd);
1174                 ret |= VM_FAULT_FALLBACK;
1175                 count_vm_event(THP_FAULT_FALLBACK);
1176                 goto out;
1177         }
1178
1179         count_vm_event(THP_FAULT_ALLOC);
1180
1181         if (!page)
1182                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1183         else
1184                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1185         __SetPageUptodate(new_page);
1186
1187         mmun_start = haddr;
1188         mmun_end   = haddr + HPAGE_PMD_SIZE;
1189         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1190
1191         spin_lock(ptl);
1192         if (page)
1193                 put_user_huge_page(page);
1194         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1195                 spin_unlock(ptl);
1196                 mem_cgroup_cancel_charge(new_page, memcg);
1197                 put_page(new_page);
1198                 goto out_mn;
1199         } else {
1200                 pmd_t entry;
1201                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1202                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1203                 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1204                 page_add_new_anon_rmap(new_page, vma, haddr);
1205                 mem_cgroup_commit_charge(new_page, memcg, false);
1206                 lru_cache_add_active_or_unevictable(new_page, vma);
1207                 set_pmd_at(mm, haddr, pmd, entry);
1208                 update_mmu_cache_pmd(vma, address, pmd);
1209                 if (!page) {
1210                         add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1211                         put_huge_zero_page();
1212                 } else {
1213                         VM_BUG_ON_PAGE(!PageHead(page), page);
1214                         page_remove_rmap(page);
1215                         put_page(page);
1216                 }
1217                 ret |= VM_FAULT_WRITE;
1218         }
1219         spin_unlock(ptl);
1220 out_mn:
1221         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1222 out:
1223         return ret;
1224 out_unlock:
1225         spin_unlock(ptl);
1226         return ret;
1227 }
1228
1229 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1230                                    unsigned long addr,
1231                                    pmd_t *pmd,
1232                                    unsigned int flags)
1233 {
1234         struct mm_struct *mm = vma->vm_mm;
1235         struct page *page = NULL;
1236
1237         assert_spin_locked(pmd_lockptr(mm, pmd));
1238
1239         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1240                 goto out;
1241
1242         /* Avoid dumping huge zero page */
1243         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1244                 return ERR_PTR(-EFAULT);
1245
1246         /* Full NUMA hinting faults to serialise migration in fault paths */
1247         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1248                 goto out;
1249
1250         page = pmd_page(*pmd);
1251         VM_BUG_ON_PAGE(!PageHead(page), page);
1252         if (flags & FOLL_TOUCH) {
1253                 pmd_t _pmd;
1254                 /*
1255                  * We should set the dirty bit only for FOLL_WRITE but
1256                  * for now the dirty bit in the pmd is meaningless.
1257                  * And if the dirty bit will become meaningful and
1258                  * we'll only set it with FOLL_WRITE, an atomic
1259                  * set_bit will be required on the pmd to set the
1260                  * young bit, instead of the current set_pmd_at.
1261                  */
1262                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1263                 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1264                                           pmd, _pmd,  1))
1265                         update_mmu_cache_pmd(vma, addr, pmd);
1266         }
1267         if ((flags & FOLL_POPULATE) && (vma->vm_flags & VM_LOCKED)) {
1268                 if (page->mapping && trylock_page(page)) {
1269                         lru_add_drain();
1270                         if (page->mapping)
1271                                 mlock_vma_page(page);
1272                         unlock_page(page);
1273                 }
1274         }
1275         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1276         VM_BUG_ON_PAGE(!PageCompound(page), page);
1277         if (flags & FOLL_GET)
1278                 get_page_foll(page);
1279
1280 out:
1281         return page;
1282 }
1283
1284 /* NUMA hinting page fault entry point for trans huge pmds */
1285 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1286                                 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1287 {
1288         spinlock_t *ptl;
1289         struct anon_vma *anon_vma = NULL;
1290         struct page *page;
1291         unsigned long haddr = addr & HPAGE_PMD_MASK;
1292         int page_nid = -1, this_nid = numa_node_id();
1293         int target_nid, last_cpupid = -1;
1294         bool page_locked;
1295         bool migrated = false;
1296         bool was_writable;
1297         int flags = 0;
1298
1299         /* A PROT_NONE fault should not end up here */
1300         BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
1301
1302         ptl = pmd_lock(mm, pmdp);
1303         if (unlikely(!pmd_same(pmd, *pmdp)))
1304                 goto out_unlock;
1305
1306         /*
1307          * If there are potential migrations, wait for completion and retry
1308          * without disrupting NUMA hinting information. Do not relock and
1309          * check_same as the page may no longer be mapped.
1310          */
1311         if (unlikely(pmd_trans_migrating(*pmdp))) {
1312                 page = pmd_page(*pmdp);
1313                 spin_unlock(ptl);
1314                 wait_on_page_locked(page);
1315                 goto out;
1316         }
1317
1318         page = pmd_page(pmd);
1319         BUG_ON(is_huge_zero_page(page));
1320         page_nid = page_to_nid(page);
1321         last_cpupid = page_cpupid_last(page);
1322         count_vm_numa_event(NUMA_HINT_FAULTS);
1323         if (page_nid == this_nid) {
1324                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1325                 flags |= TNF_FAULT_LOCAL;
1326         }
1327
1328         /* See similar comment in do_numa_page for explanation */
1329         if (!(vma->vm_flags & VM_WRITE))
1330                 flags |= TNF_NO_GROUP;
1331
1332         /*
1333          * Acquire the page lock to serialise THP migrations but avoid dropping
1334          * page_table_lock if at all possible
1335          */
1336         page_locked = trylock_page(page);
1337         target_nid = mpol_misplaced(page, vma, haddr);
1338         if (target_nid == -1) {
1339                 /* If the page was locked, there are no parallel migrations */
1340                 if (page_locked)
1341                         goto clear_pmdnuma;
1342         }
1343
1344         /* Migration could have started since the pmd_trans_migrating check */
1345         if (!page_locked) {
1346                 spin_unlock(ptl);
1347                 wait_on_page_locked(page);
1348                 page_nid = -1;
1349                 goto out;
1350         }
1351
1352         /*
1353          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1354          * to serialises splits
1355          */
1356         get_page(page);
1357         spin_unlock(ptl);
1358         anon_vma = page_lock_anon_vma_read(page);
1359
1360         /* Confirm the PMD did not change while page_table_lock was released */
1361         spin_lock(ptl);
1362         if (unlikely(!pmd_same(pmd, *pmdp))) {
1363                 unlock_page(page);
1364                 put_page(page);
1365                 page_nid = -1;
1366                 goto out_unlock;
1367         }
1368
1369         /* Bail if we fail to protect against THP splits for any reason */
1370         if (unlikely(!anon_vma)) {
1371                 put_page(page);
1372                 page_nid = -1;
1373                 goto clear_pmdnuma;
1374         }
1375
1376         /*
1377          * Migrate the THP to the requested node, returns with page unlocked
1378          * and access rights restored.
1379          */
1380         spin_unlock(ptl);
1381         migrated = migrate_misplaced_transhuge_page(mm, vma,
1382                                 pmdp, pmd, addr, page, target_nid);
1383         if (migrated) {
1384                 flags |= TNF_MIGRATED;
1385                 page_nid = target_nid;
1386         } else
1387                 flags |= TNF_MIGRATE_FAIL;
1388
1389         goto out;
1390 clear_pmdnuma:
1391         BUG_ON(!PageLocked(page));
1392         was_writable = pmd_write(pmd);
1393         pmd = pmd_modify(pmd, vma->vm_page_prot);
1394         pmd = pmd_mkyoung(pmd);
1395         if (was_writable)
1396                 pmd = pmd_mkwrite(pmd);
1397         set_pmd_at(mm, haddr, pmdp, pmd);
1398         update_mmu_cache_pmd(vma, addr, pmdp);
1399         unlock_page(page);
1400 out_unlock:
1401         spin_unlock(ptl);
1402
1403 out:
1404         if (anon_vma)
1405                 page_unlock_anon_vma_read(anon_vma);
1406
1407         if (page_nid != -1)
1408                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
1409
1410         return 0;
1411 }
1412
1413 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1414                  pmd_t *pmd, unsigned long addr)
1415 {
1416         spinlock_t *ptl;
1417         int ret = 0;
1418
1419         if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1420                 pgtable_t pgtable;
1421                 pmd_t orig_pmd;
1422                 /*
1423                  * For architectures like ppc64 we look at deposited pgtable
1424                  * when calling pmdp_huge_get_and_clear. So do the
1425                  * pgtable_trans_huge_withdraw after finishing pmdp related
1426                  * operations.
1427                  */
1428                 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1429                                                         tlb->fullmm);
1430                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1431                 if (vma_is_dax(vma)) {
1432                         if (is_huge_zero_pmd(orig_pmd)) {
1433                                 pgtable = NULL;
1434                         } else {
1435                                 spin_unlock(ptl);
1436                                 return 1;
1437                         }
1438                 } else {
1439                         pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1440                 }
1441                 if (is_huge_zero_pmd(orig_pmd)) {
1442                         atomic_long_dec(&tlb->mm->nr_ptes);
1443                         spin_unlock(ptl);
1444                         put_huge_zero_page();
1445                 } else {
1446                         struct page *page = pmd_page(orig_pmd);
1447                         page_remove_rmap(page);
1448                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1449                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1450                         VM_BUG_ON_PAGE(!PageHead(page), page);
1451                         atomic_long_dec(&tlb->mm->nr_ptes);
1452                         spin_unlock(ptl);
1453                         tlb_remove_page(tlb, page);
1454                 }
1455                 if (pgtable)
1456                         pte_free(tlb->mm, pgtable);
1457                 ret = 1;
1458         }
1459         return ret;
1460 }
1461
1462 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1463                   unsigned long old_addr,
1464                   unsigned long new_addr, unsigned long old_end,
1465                   pmd_t *old_pmd, pmd_t *new_pmd)
1466 {
1467         spinlock_t *old_ptl, *new_ptl;
1468         int ret = 0;
1469         pmd_t pmd;
1470
1471         struct mm_struct *mm = vma->vm_mm;
1472
1473         if ((old_addr & ~HPAGE_PMD_MASK) ||
1474             (new_addr & ~HPAGE_PMD_MASK) ||
1475             old_end - old_addr < HPAGE_PMD_SIZE ||
1476             (new_vma->vm_flags & VM_NOHUGEPAGE))
1477                 goto out;
1478
1479         /*
1480          * The destination pmd shouldn't be established, free_pgtables()
1481          * should have release it.
1482          */
1483         if (WARN_ON(!pmd_none(*new_pmd))) {
1484                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1485                 goto out;
1486         }
1487
1488         /*
1489          * We don't have to worry about the ordering of src and dst
1490          * ptlocks because exclusive mmap_sem prevents deadlock.
1491          */
1492         ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl);
1493         if (ret == 1) {
1494                 new_ptl = pmd_lockptr(mm, new_pmd);
1495                 if (new_ptl != old_ptl)
1496                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1497                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1498                 VM_BUG_ON(!pmd_none(*new_pmd));
1499
1500                 if (pmd_move_must_withdraw(new_ptl, old_ptl)) {
1501                         pgtable_t pgtable;
1502                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1503                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1504                 }
1505                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1506                 if (new_ptl != old_ptl)
1507                         spin_unlock(new_ptl);
1508                 spin_unlock(old_ptl);
1509         }
1510 out:
1511         return ret;
1512 }
1513
1514 /*
1515  * Returns
1516  *  - 0 if PMD could not be locked
1517  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1518  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1519  */
1520 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1521                 unsigned long addr, pgprot_t newprot, int prot_numa)
1522 {
1523         struct mm_struct *mm = vma->vm_mm;
1524         spinlock_t *ptl;
1525         int ret = 0;
1526
1527         if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1528                 pmd_t entry;
1529                 bool preserve_write = prot_numa && pmd_write(*pmd);
1530                 ret = 1;
1531
1532                 /*
1533                  * Avoid trapping faults against the zero page. The read-only
1534                  * data is likely to be read-cached on the local CPU and
1535                  * local/remote hits to the zero page are not interesting.
1536                  */
1537                 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1538                         spin_unlock(ptl);
1539                         return ret;
1540                 }
1541
1542                 if (!prot_numa || !pmd_protnone(*pmd)) {
1543                         entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
1544                         entry = pmd_modify(entry, newprot);
1545                         if (preserve_write)
1546                                 entry = pmd_mkwrite(entry);
1547                         ret = HPAGE_PMD_NR;
1548                         set_pmd_at(mm, addr, pmd, entry);
1549                         BUG_ON(!preserve_write && pmd_write(entry));
1550                 }
1551                 spin_unlock(ptl);
1552         }
1553
1554         return ret;
1555 }
1556
1557 /*
1558  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1559  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1560  *
1561  * Note that if it returns 1, this routine returns without unlocking page
1562  * table locks. So callers must unlock them.
1563  */
1564 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1565                 spinlock_t **ptl)
1566 {
1567         *ptl = pmd_lock(vma->vm_mm, pmd);
1568         if (likely(pmd_trans_huge(*pmd))) {
1569                 if (unlikely(pmd_trans_splitting(*pmd))) {
1570                         spin_unlock(*ptl);
1571                         wait_split_huge_page(vma->anon_vma, pmd);
1572                         return -1;
1573                 } else {
1574                         /* Thp mapped by 'pmd' is stable, so we can
1575                          * handle it as it is. */
1576                         return 1;
1577                 }
1578         }
1579         spin_unlock(*ptl);
1580         return 0;
1581 }
1582
1583 /*
1584  * This function returns whether a given @page is mapped onto the @address
1585  * in the virtual space of @mm.
1586  *
1587  * When it's true, this function returns *pmd with holding the page table lock
1588  * and passing it back to the caller via @ptl.
1589  * If it's false, returns NULL without holding the page table lock.
1590  */
1591 pmd_t *page_check_address_pmd(struct page *page,
1592                               struct mm_struct *mm,
1593                               unsigned long address,
1594                               enum page_check_address_pmd_flag flag,
1595                               spinlock_t **ptl)
1596 {
1597         pgd_t *pgd;
1598         pud_t *pud;
1599         pmd_t *pmd;
1600
1601         if (address & ~HPAGE_PMD_MASK)
1602                 return NULL;
1603
1604         pgd = pgd_offset(mm, address);
1605         if (!pgd_present(*pgd))
1606                 return NULL;
1607         pud = pud_offset(pgd, address);
1608         if (!pud_present(*pud))
1609                 return NULL;
1610         pmd = pmd_offset(pud, address);
1611
1612         *ptl = pmd_lock(mm, pmd);
1613         if (!pmd_present(*pmd))
1614                 goto unlock;
1615         if (pmd_page(*pmd) != page)
1616                 goto unlock;
1617         /*
1618          * split_vma() may create temporary aliased mappings. There is
1619          * no risk as long as all huge pmd are found and have their
1620          * splitting bit set before __split_huge_page_refcount
1621          * runs. Finding the same huge pmd more than once during the
1622          * same rmap walk is not a problem.
1623          */
1624         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1625             pmd_trans_splitting(*pmd))
1626                 goto unlock;
1627         if (pmd_trans_huge(*pmd)) {
1628                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1629                           !pmd_trans_splitting(*pmd));
1630                 return pmd;
1631         }
1632 unlock:
1633         spin_unlock(*ptl);
1634         return NULL;
1635 }
1636
1637 static int __split_huge_page_splitting(struct page *page,
1638                                        struct vm_area_struct *vma,
1639                                        unsigned long address)
1640 {
1641         struct mm_struct *mm = vma->vm_mm;
1642         spinlock_t *ptl;
1643         pmd_t *pmd;
1644         int ret = 0;
1645         /* For mmu_notifiers */
1646         const unsigned long mmun_start = address;
1647         const unsigned long mmun_end   = address + HPAGE_PMD_SIZE;
1648
1649         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1650         pmd = page_check_address_pmd(page, mm, address,
1651                         PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl);
1652         if (pmd) {
1653                 /*
1654                  * We can't temporarily set the pmd to null in order
1655                  * to split it, the pmd must remain marked huge at all
1656                  * times or the VM won't take the pmd_trans_huge paths
1657                  * and it won't wait on the anon_vma->root->rwsem to
1658                  * serialize against split_huge_page*.
1659                  */
1660                 pmdp_splitting_flush(vma, address, pmd);
1661
1662                 ret = 1;
1663                 spin_unlock(ptl);
1664         }
1665         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1666
1667         return ret;
1668 }
1669
1670 static void __split_huge_page_refcount(struct page *page,
1671                                        struct list_head *list)
1672 {
1673         int i;
1674         struct zone *zone = page_zone(page);
1675         struct lruvec *lruvec;
1676         int tail_count = 0;
1677
1678         /* prevent PageLRU to go away from under us, and freeze lru stats */
1679         spin_lock_irq(&zone->lru_lock);
1680         lruvec = mem_cgroup_page_lruvec(page, zone);
1681
1682         compound_lock(page);
1683         /* complete memcg works before add pages to LRU */
1684         mem_cgroup_split_huge_fixup(page);
1685
1686         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1687                 struct page *page_tail = page + i;
1688
1689                 /* tail_page->_mapcount cannot change */
1690                 BUG_ON(page_mapcount(page_tail) < 0);
1691                 tail_count += page_mapcount(page_tail);
1692                 /* check for overflow */
1693                 BUG_ON(tail_count < 0);
1694                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1695                 /*
1696                  * tail_page->_count is zero and not changing from
1697                  * under us. But get_page_unless_zero() may be running
1698                  * from under us on the tail_page. If we used
1699                  * atomic_set() below instead of atomic_add(), we
1700                  * would then run atomic_set() concurrently with
1701                  * get_page_unless_zero(), and atomic_set() is
1702                  * implemented in C not using locked ops. spin_unlock
1703                  * on x86 sometime uses locked ops because of PPro
1704                  * errata 66, 92, so unless somebody can guarantee
1705                  * atomic_set() here would be safe on all archs (and
1706                  * not only on x86), it's safer to use atomic_add().
1707                  */
1708                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1709                            &page_tail->_count);
1710
1711                 /* after clearing PageTail the gup refcount can be released */
1712                 smp_mb__after_atomic();
1713
1714                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1715                 page_tail->flags |= (page->flags &
1716                                      ((1L << PG_referenced) |
1717                                       (1L << PG_swapbacked) |
1718                                       (1L << PG_mlocked) |
1719                                       (1L << PG_uptodate) |
1720                                       (1L << PG_active) |
1721                                       (1L << PG_unevictable)));
1722                 page_tail->flags |= (1L << PG_dirty);
1723
1724                 /* clear PageTail before overwriting first_page */
1725                 smp_wmb();
1726
1727                 /*
1728                  * __split_huge_page_splitting() already set the
1729                  * splitting bit in all pmd that could map this
1730                  * hugepage, that will ensure no CPU can alter the
1731                  * mapcount on the head page. The mapcount is only
1732                  * accounted in the head page and it has to be
1733                  * transferred to all tail pages in the below code. So
1734                  * for this code to be safe, the split the mapcount
1735                  * can't change. But that doesn't mean userland can't
1736                  * keep changing and reading the page contents while
1737                  * we transfer the mapcount, so the pmd splitting
1738                  * status is achieved setting a reserved bit in the
1739                  * pmd, not by clearing the present bit.
1740                 */
1741                 page_tail->_mapcount = page->_mapcount;
1742
1743                 BUG_ON(page_tail->mapping);
1744                 page_tail->mapping = page->mapping;
1745
1746                 page_tail->index = page->index + i;
1747                 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
1748
1749                 BUG_ON(!PageAnon(page_tail));
1750                 BUG_ON(!PageUptodate(page_tail));
1751                 BUG_ON(!PageDirty(page_tail));
1752                 BUG_ON(!PageSwapBacked(page_tail));
1753
1754                 lru_add_page_tail(page, page_tail, lruvec, list);
1755         }
1756         atomic_sub(tail_count, &page->_count);
1757         BUG_ON(atomic_read(&page->_count) <= 0);
1758
1759         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1760
1761         ClearPageCompound(page);
1762         compound_unlock(page);
1763         spin_unlock_irq(&zone->lru_lock);
1764
1765         for (i = 1; i < HPAGE_PMD_NR; i++) {
1766                 struct page *page_tail = page + i;
1767                 BUG_ON(page_count(page_tail) <= 0);
1768                 /*
1769                  * Tail pages may be freed if there wasn't any mapping
1770                  * like if add_to_swap() is running on a lru page that
1771                  * had its mapping zapped. And freeing these pages
1772                  * requires taking the lru_lock so we do the put_page
1773                  * of the tail pages after the split is complete.
1774                  */
1775                 put_page(page_tail);
1776         }
1777
1778         /*
1779          * Only the head page (now become a regular page) is required
1780          * to be pinned by the caller.
1781          */
1782         BUG_ON(page_count(page) <= 0);
1783 }
1784
1785 static int __split_huge_page_map(struct page *page,
1786                                  struct vm_area_struct *vma,
1787                                  unsigned long address)
1788 {
1789         struct mm_struct *mm = vma->vm_mm;
1790         spinlock_t *ptl;
1791         pmd_t *pmd, _pmd;
1792         int ret = 0, i;
1793         pgtable_t pgtable;
1794         unsigned long haddr;
1795
1796         pmd = page_check_address_pmd(page, mm, address,
1797                         PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl);
1798         if (pmd) {
1799                 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1800                 pmd_populate(mm, &_pmd, pgtable);
1801                 if (pmd_write(*pmd))
1802                         BUG_ON(page_mapcount(page) != 1);
1803
1804                 haddr = address;
1805                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1806                         pte_t *pte, entry;
1807                         BUG_ON(PageCompound(page+i));
1808                         /*
1809                          * Note that NUMA hinting access restrictions are not
1810                          * transferred to avoid any possibility of altering
1811                          * permissions across VMAs.
1812                          */
1813                         entry = mk_pte(page + i, vma->vm_page_prot);
1814                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1815                         if (!pmd_write(*pmd))
1816                                 entry = pte_wrprotect(entry);
1817                         if (!pmd_young(*pmd))
1818                                 entry = pte_mkold(entry);
1819                         pte = pte_offset_map(&_pmd, haddr);
1820                         BUG_ON(!pte_none(*pte));
1821                         set_pte_at(mm, haddr, pte, entry);
1822                         pte_unmap(pte);
1823                 }
1824
1825                 smp_wmb(); /* make pte visible before pmd */
1826                 /*
1827                  * Up to this point the pmd is present and huge and
1828                  * userland has the whole access to the hugepage
1829                  * during the split (which happens in place). If we
1830                  * overwrite the pmd with the not-huge version
1831                  * pointing to the pte here (which of course we could
1832                  * if all CPUs were bug free), userland could trigger
1833                  * a small page size TLB miss on the small sized TLB
1834                  * while the hugepage TLB entry is still established
1835                  * in the huge TLB. Some CPU doesn't like that. See
1836                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1837                  * Erratum 383 on page 93. Intel should be safe but is
1838                  * also warns that it's only safe if the permission
1839                  * and cache attributes of the two entries loaded in
1840                  * the two TLB is identical (which should be the case
1841                  * here). But it is generally safer to never allow
1842                  * small and huge TLB entries for the same virtual
1843                  * address to be loaded simultaneously. So instead of
1844                  * doing "pmd_populate(); flush_tlb_range();" we first
1845                  * mark the current pmd notpresent (atomically because
1846                  * here the pmd_trans_huge and pmd_trans_splitting
1847                  * must remain set at all times on the pmd until the
1848                  * split is complete for this pmd), then we flush the
1849                  * SMP TLB and finally we write the non-huge version
1850                  * of the pmd entry with pmd_populate.
1851                  */
1852                 pmdp_invalidate(vma, address, pmd);
1853                 pmd_populate(mm, pmd, pgtable);
1854                 ret = 1;
1855                 spin_unlock(ptl);
1856         }
1857
1858         return ret;
1859 }
1860
1861 /* must be called with anon_vma->root->rwsem held */
1862 static void __split_huge_page(struct page *page,
1863                               struct anon_vma *anon_vma,
1864                               struct list_head *list)
1865 {
1866         int mapcount, mapcount2;
1867         pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1868         struct anon_vma_chain *avc;
1869
1870         BUG_ON(!PageHead(page));
1871         BUG_ON(PageTail(page));
1872
1873         mapcount = 0;
1874         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1875                 struct vm_area_struct *vma = avc->vma;
1876                 unsigned long addr = vma_address(page, vma);
1877                 BUG_ON(is_vma_temporary_stack(vma));
1878                 mapcount += __split_huge_page_splitting(page, vma, addr);
1879         }
1880         /*
1881          * It is critical that new vmas are added to the tail of the
1882          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1883          * and establishes a child pmd before
1884          * __split_huge_page_splitting() freezes the parent pmd (so if
1885          * we fail to prevent copy_huge_pmd() from running until the
1886          * whole __split_huge_page() is complete), we will still see
1887          * the newly established pmd of the child later during the
1888          * walk, to be able to set it as pmd_trans_splitting too.
1889          */
1890         if (mapcount != page_mapcount(page)) {
1891                 pr_err("mapcount %d page_mapcount %d\n",
1892                         mapcount, page_mapcount(page));
1893                 BUG();
1894         }
1895
1896         __split_huge_page_refcount(page, list);
1897
1898         mapcount2 = 0;
1899         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1900                 struct vm_area_struct *vma = avc->vma;
1901                 unsigned long addr = vma_address(page, vma);
1902                 BUG_ON(is_vma_temporary_stack(vma));
1903                 mapcount2 += __split_huge_page_map(page, vma, addr);
1904         }
1905         if (mapcount != mapcount2) {
1906                 pr_err("mapcount %d mapcount2 %d page_mapcount %d\n",
1907                         mapcount, mapcount2, page_mapcount(page));
1908                 BUG();
1909         }
1910 }
1911
1912 /*
1913  * Split a hugepage into normal pages. This doesn't change the position of head
1914  * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1915  * @list. Both head page and tail pages will inherit mapping, flags, and so on
1916  * from the hugepage.
1917  * Return 0 if the hugepage is split successfully otherwise return 1.
1918  */
1919 int split_huge_page_to_list(struct page *page, struct list_head *list)
1920 {
1921         struct anon_vma *anon_vma;
1922         int ret = 1;
1923
1924         BUG_ON(is_huge_zero_page(page));
1925         BUG_ON(!PageAnon(page));
1926
1927         /*
1928          * The caller does not necessarily hold an mmap_sem that would prevent
1929          * the anon_vma disappearing so we first we take a reference to it
1930          * and then lock the anon_vma for write. This is similar to
1931          * page_lock_anon_vma_read except the write lock is taken to serialise
1932          * against parallel split or collapse operations.
1933          */
1934         anon_vma = page_get_anon_vma(page);
1935         if (!anon_vma)
1936                 goto out;
1937         anon_vma_lock_write(anon_vma);
1938
1939         ret = 0;
1940         if (!PageCompound(page))
1941                 goto out_unlock;
1942
1943         BUG_ON(!PageSwapBacked(page));
1944         __split_huge_page(page, anon_vma, list);
1945         count_vm_event(THP_SPLIT);
1946
1947         BUG_ON(PageCompound(page));
1948 out_unlock:
1949         anon_vma_unlock_write(anon_vma);
1950         put_anon_vma(anon_vma);
1951 out:
1952         return ret;
1953 }
1954
1955 #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
1956
1957 int hugepage_madvise(struct vm_area_struct *vma,
1958                      unsigned long *vm_flags, int advice)
1959 {
1960         switch (advice) {
1961         case MADV_HUGEPAGE:
1962 #ifdef CONFIG_S390
1963                 /*
1964                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1965                  * can't handle this properly after s390_enable_sie, so we simply
1966                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
1967                  */
1968                 if (mm_has_pgste(vma->vm_mm))
1969                         return 0;
1970 #endif
1971                 /*
1972                  * Be somewhat over-protective like KSM for now!
1973                  */
1974                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1975                         return -EINVAL;
1976                 *vm_flags &= ~VM_NOHUGEPAGE;
1977                 *vm_flags |= VM_HUGEPAGE;
1978                 /*
1979                  * If the vma become good for khugepaged to scan,
1980                  * register it here without waiting a page fault that
1981                  * may not happen any time soon.
1982                  */
1983                 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
1984                         return -ENOMEM;
1985                 break;
1986         case MADV_NOHUGEPAGE:
1987                 /*
1988                  * Be somewhat over-protective like KSM for now!
1989                  */
1990                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1991                         return -EINVAL;
1992                 *vm_flags &= ~VM_HUGEPAGE;
1993                 *vm_flags |= VM_NOHUGEPAGE;
1994                 /*
1995                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1996                  * this vma even if we leave the mm registered in khugepaged if
1997                  * it got registered before VM_NOHUGEPAGE was set.
1998                  */
1999                 break;
2000         }
2001
2002         return 0;
2003 }
2004
2005 static int __init khugepaged_slab_init(void)
2006 {
2007         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
2008                                           sizeof(struct mm_slot),
2009                                           __alignof__(struct mm_slot), 0, NULL);
2010         if (!mm_slot_cache)
2011                 return -ENOMEM;
2012
2013         return 0;
2014 }
2015
2016 static void __init khugepaged_slab_exit(void)
2017 {
2018         kmem_cache_destroy(mm_slot_cache);
2019 }
2020
2021 static inline struct mm_slot *alloc_mm_slot(void)
2022 {
2023         if (!mm_slot_cache)     /* initialization failed */
2024                 return NULL;
2025         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
2026 }
2027
2028 static inline void free_mm_slot(struct mm_slot *mm_slot)
2029 {
2030         kmem_cache_free(mm_slot_cache, mm_slot);
2031 }
2032
2033 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
2034 {
2035         struct mm_slot *mm_slot;
2036
2037         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
2038                 if (mm == mm_slot->mm)
2039                         return mm_slot;
2040
2041         return NULL;
2042 }
2043
2044 static void insert_to_mm_slots_hash(struct mm_struct *mm,
2045                                     struct mm_slot *mm_slot)
2046 {
2047         mm_slot->mm = mm;
2048         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
2049 }
2050
2051 static inline int khugepaged_test_exit(struct mm_struct *mm)
2052 {
2053         return atomic_read(&mm->mm_users) == 0;
2054 }
2055
2056 int __khugepaged_enter(struct mm_struct *mm)
2057 {
2058         struct mm_slot *mm_slot;
2059         int wakeup;
2060
2061         mm_slot = alloc_mm_slot();
2062         if (!mm_slot)
2063                 return -ENOMEM;
2064
2065         /* __khugepaged_exit() must not run from under us */
2066         VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
2067         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2068                 free_mm_slot(mm_slot);
2069                 return 0;
2070         }
2071
2072         spin_lock(&khugepaged_mm_lock);
2073         insert_to_mm_slots_hash(mm, mm_slot);
2074         /*
2075          * Insert just behind the scanning cursor, to let the area settle
2076          * down a little.
2077          */
2078         wakeup = list_empty(&khugepaged_scan.mm_head);
2079         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2080         spin_unlock(&khugepaged_mm_lock);
2081
2082         atomic_inc(&mm->mm_count);
2083         if (wakeup)
2084                 wake_up_interruptible(&khugepaged_wait);
2085
2086         return 0;
2087 }
2088
2089 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
2090                                unsigned long vm_flags)
2091 {
2092         unsigned long hstart, hend;
2093         if (!vma->anon_vma)
2094                 /*
2095                  * Not yet faulted in so we will register later in the
2096                  * page fault if needed.
2097                  */
2098                 return 0;
2099         if (vma->vm_ops)
2100                 /* khugepaged not yet working on file or special mappings */
2101                 return 0;
2102         VM_BUG_ON_VMA(vm_flags & VM_NO_THP, vma);
2103         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2104         hend = vma->vm_end & HPAGE_PMD_MASK;
2105         if (hstart < hend)
2106                 return khugepaged_enter(vma, vm_flags);
2107         return 0;
2108 }
2109
2110 void __khugepaged_exit(struct mm_struct *mm)
2111 {
2112         struct mm_slot *mm_slot;
2113         int free = 0;
2114
2115         spin_lock(&khugepaged_mm_lock);
2116         mm_slot = get_mm_slot(mm);
2117         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
2118                 hash_del(&mm_slot->hash);
2119                 list_del(&mm_slot->mm_node);
2120                 free = 1;
2121         }
2122         spin_unlock(&khugepaged_mm_lock);
2123
2124         if (free) {
2125                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2126                 free_mm_slot(mm_slot);
2127                 mmdrop(mm);
2128         } else if (mm_slot) {
2129                 /*
2130                  * This is required to serialize against
2131                  * khugepaged_test_exit() (which is guaranteed to run
2132                  * under mmap sem read mode). Stop here (after we
2133                  * return all pagetables will be destroyed) until
2134                  * khugepaged has finished working on the pagetables
2135                  * under the mmap_sem.
2136                  */
2137                 down_write(&mm->mmap_sem);
2138                 up_write(&mm->mmap_sem);
2139         }
2140 }
2141
2142 static void release_pte_page(struct page *page)
2143 {
2144         /* 0 stands for page_is_file_cache(page) == false */
2145         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2146         unlock_page(page);
2147         putback_lru_page(page);
2148 }
2149
2150 static void release_pte_pages(pte_t *pte, pte_t *_pte)
2151 {
2152         while (--_pte >= pte) {
2153                 pte_t pteval = *_pte;
2154                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
2155                         release_pte_page(pte_page(pteval));
2156         }
2157 }
2158
2159 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2160                                         unsigned long address,
2161                                         pte_t *pte)
2162 {
2163         struct page *page;
2164         pte_t *_pte;
2165         int none_or_zero = 0;
2166         bool referenced = false, writable = false;
2167         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2168              _pte++, address += PAGE_SIZE) {
2169                 pte_t pteval = *_pte;
2170                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2171                         if (!userfaultfd_armed(vma) &&
2172                             ++none_or_zero <= khugepaged_max_ptes_none)
2173                                 continue;
2174                         else
2175                                 goto out;
2176                 }
2177                 if (!pte_present(pteval))
2178                         goto out;
2179                 page = vm_normal_page(vma, address, pteval);
2180                 if (unlikely(!page))
2181                         goto out;
2182
2183                 VM_BUG_ON_PAGE(PageCompound(page), page);
2184                 VM_BUG_ON_PAGE(!PageAnon(page), page);
2185                 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2186
2187                 /*
2188                  * We can do it before isolate_lru_page because the
2189                  * page can't be freed from under us. NOTE: PG_lock
2190                  * is needed to serialize against split_huge_page
2191                  * when invoked from the VM.
2192                  */
2193                 if (!trylock_page(page))
2194                         goto out;
2195
2196                 /*
2197                  * cannot use mapcount: can't collapse if there's a gup pin.
2198                  * The page must only be referenced by the scanned process
2199                  * and page swap cache.
2200                  */
2201                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2202                         unlock_page(page);
2203                         goto out;
2204                 }
2205                 if (pte_write(pteval)) {
2206                         writable = true;
2207                 } else {
2208                         if (PageSwapCache(page) && !reuse_swap_page(page)) {
2209                                 unlock_page(page);
2210                                 goto out;
2211                         }
2212                         /*
2213                          * Page is not in the swap cache. It can be collapsed
2214                          * into a THP.
2215                          */
2216                 }
2217
2218                 /*
2219                  * Isolate the page to avoid collapsing an hugepage
2220                  * currently in use by the VM.
2221                  */
2222                 if (isolate_lru_page(page)) {
2223                         unlock_page(page);
2224                         goto out;
2225                 }
2226                 /* 0 stands for page_is_file_cache(page) == false */
2227                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2228                 VM_BUG_ON_PAGE(!PageLocked(page), page);
2229                 VM_BUG_ON_PAGE(PageLRU(page), page);
2230
2231                 /* If there is no mapped pte young don't collapse the page */
2232                 if (pte_young(pteval) || PageReferenced(page) ||
2233                     mmu_notifier_test_young(vma->vm_mm, address))
2234                         referenced = true;
2235         }
2236         if (likely(referenced && writable))
2237                 return 1;
2238 out:
2239         release_pte_pages(pte, _pte);
2240         return 0;
2241 }
2242
2243 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2244                                       struct vm_area_struct *vma,
2245                                       unsigned long address,
2246                                       spinlock_t *ptl)
2247 {
2248         pte_t *_pte;
2249         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2250                 pte_t pteval = *_pte;
2251                 struct page *src_page;
2252
2253                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2254                         clear_user_highpage(page, address);
2255                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2256                         if (is_zero_pfn(pte_pfn(pteval))) {
2257                                 /*
2258                                  * ptl mostly unnecessary.
2259                                  */
2260                                 spin_lock(ptl);
2261                                 /*
2262                                  * paravirt calls inside pte_clear here are
2263                                  * superfluous.
2264                                  */
2265                                 pte_clear(vma->vm_mm, address, _pte);
2266                                 spin_unlock(ptl);
2267                         }
2268                 } else {
2269                         src_page = pte_page(pteval);
2270                         copy_user_highpage(page, src_page, address, vma);
2271                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
2272                         release_pte_page(src_page);
2273                         /*
2274                          * ptl mostly unnecessary, but preempt has to
2275                          * be disabled to update the per-cpu stats
2276                          * inside page_remove_rmap().
2277                          */
2278                         spin_lock(ptl);
2279                         /*
2280                          * paravirt calls inside pte_clear here are
2281                          * superfluous.
2282                          */
2283                         pte_clear(vma->vm_mm, address, _pte);
2284                         page_remove_rmap(src_page);
2285                         spin_unlock(ptl);
2286                         free_page_and_swap_cache(src_page);
2287                 }
2288
2289                 address += PAGE_SIZE;
2290                 page++;
2291         }
2292 }
2293
2294 static void khugepaged_alloc_sleep(void)
2295 {
2296         wait_event_freezable_timeout(khugepaged_wait, false,
2297                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2298 }
2299
2300 static int khugepaged_node_load[MAX_NUMNODES];
2301
2302 static bool khugepaged_scan_abort(int nid)
2303 {
2304         int i;
2305
2306         /*
2307          * If zone_reclaim_mode is disabled, then no extra effort is made to
2308          * allocate memory locally.
2309          */
2310         if (!zone_reclaim_mode)
2311                 return false;
2312
2313         /* If there is a count for this node already, it must be acceptable */
2314         if (khugepaged_node_load[nid])
2315                 return false;
2316
2317         for (i = 0; i < MAX_NUMNODES; i++) {
2318                 if (!khugepaged_node_load[i])
2319                         continue;
2320                 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2321                         return true;
2322         }
2323         return false;
2324 }
2325
2326 #ifdef CONFIG_NUMA
2327 static int khugepaged_find_target_node(void)
2328 {
2329         static int last_khugepaged_target_node = NUMA_NO_NODE;
2330         int nid, target_node = 0, max_value = 0;
2331
2332         /* find first node with max normal pages hit */
2333         for (nid = 0; nid < MAX_NUMNODES; nid++)
2334                 if (khugepaged_node_load[nid] > max_value) {
2335                         max_value = khugepaged_node_load[nid];
2336                         target_node = nid;
2337                 }
2338
2339         /* do some balance if several nodes have the same hit record */
2340         if (target_node <= last_khugepaged_target_node)
2341                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2342                                 nid++)
2343                         if (max_value == khugepaged_node_load[nid]) {
2344                                 target_node = nid;
2345                                 break;
2346                         }
2347
2348         last_khugepaged_target_node = target_node;
2349         return target_node;
2350 }
2351
2352 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2353 {
2354         if (IS_ERR(*hpage)) {
2355                 if (!*wait)
2356                         return false;
2357
2358                 *wait = false;
2359                 *hpage = NULL;
2360                 khugepaged_alloc_sleep();
2361         } else if (*hpage) {
2362                 put_page(*hpage);
2363                 *hpage = NULL;
2364         }
2365
2366         return true;
2367 }
2368
2369 static struct page *
2370 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2371                        struct vm_area_struct *vma, unsigned long address,
2372                        int node)
2373 {
2374         VM_BUG_ON_PAGE(*hpage, *hpage);
2375
2376         /*
2377          * Before allocating the hugepage, release the mmap_sem read lock.
2378          * The allocation can take potentially a long time if it involves
2379          * sync compaction, and we do not need to hold the mmap_sem during
2380          * that. We will recheck the vma after taking it again in write mode.
2381          */
2382         up_read(&mm->mmap_sem);
2383
2384         *hpage = alloc_pages_exact_node(node, gfp, HPAGE_PMD_ORDER);
2385         if (unlikely(!*hpage)) {
2386                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2387                 *hpage = ERR_PTR(-ENOMEM);
2388                 return NULL;
2389         }
2390
2391         count_vm_event(THP_COLLAPSE_ALLOC);
2392         return *hpage;
2393 }
2394 #else
2395 static int khugepaged_find_target_node(void)
2396 {
2397         return 0;
2398 }
2399
2400 static inline struct page *alloc_hugepage(int defrag)
2401 {
2402         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2403                            HPAGE_PMD_ORDER);
2404 }
2405
2406 static struct page *khugepaged_alloc_hugepage(bool *wait)
2407 {
2408         struct page *hpage;
2409
2410         do {
2411                 hpage = alloc_hugepage(khugepaged_defrag());
2412                 if (!hpage) {
2413                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2414                         if (!*wait)
2415                                 return NULL;
2416
2417                         *wait = false;
2418                         khugepaged_alloc_sleep();
2419                 } else
2420                         count_vm_event(THP_COLLAPSE_ALLOC);
2421         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2422
2423         return hpage;
2424 }
2425
2426 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2427 {
2428         if (!*hpage)
2429                 *hpage = khugepaged_alloc_hugepage(wait);
2430
2431         if (unlikely(!*hpage))
2432                 return false;
2433
2434         return true;
2435 }
2436
2437 static struct page *
2438 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2439                        struct vm_area_struct *vma, unsigned long address,
2440                        int node)
2441 {
2442         up_read(&mm->mmap_sem);
2443         VM_BUG_ON(!*hpage);
2444
2445         return  *hpage;
2446 }
2447 #endif
2448
2449 static bool hugepage_vma_check(struct vm_area_struct *vma)
2450 {
2451         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2452             (vma->vm_flags & VM_NOHUGEPAGE))
2453                 return false;
2454
2455         if (!vma->anon_vma || vma->vm_ops)
2456                 return false;
2457         if (is_vma_temporary_stack(vma))
2458                 return false;
2459         VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma);
2460         return true;
2461 }
2462
2463 static void collapse_huge_page(struct mm_struct *mm,
2464                                    unsigned long address,
2465                                    struct page **hpage,
2466                                    struct vm_area_struct *vma,
2467                                    int node)
2468 {
2469         pmd_t *pmd, _pmd;
2470         pte_t *pte;
2471         pgtable_t pgtable;
2472         struct page *new_page;
2473         spinlock_t *pmd_ptl, *pte_ptl;
2474         int isolated;
2475         unsigned long hstart, hend;
2476         struct mem_cgroup *memcg;
2477         unsigned long mmun_start;       /* For mmu_notifiers */
2478         unsigned long mmun_end;         /* For mmu_notifiers */
2479         gfp_t gfp;
2480
2481         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2482
2483         /* Only allocate from the target node */
2484         gfp = alloc_hugepage_gfpmask(khugepaged_defrag(), __GFP_OTHER_NODE) |
2485                 __GFP_THISNODE;
2486
2487         /* release the mmap_sem read lock. */
2488         new_page = khugepaged_alloc_page(hpage, gfp, mm, vma, address, node);
2489         if (!new_page)
2490                 return;
2491
2492         if (unlikely(mem_cgroup_try_charge(new_page, mm,
2493                                            gfp, &memcg)))
2494                 return;
2495
2496         /*
2497          * Prevent all access to pagetables with the exception of
2498          * gup_fast later hanlded by the ptep_clear_flush and the VM
2499          * handled by the anon_vma lock + PG_lock.
2500          */
2501         down_write(&mm->mmap_sem);
2502         if (unlikely(khugepaged_test_exit(mm)))
2503                 goto out;
2504
2505         vma = find_vma(mm, address);
2506         if (!vma)
2507                 goto out;
2508         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2509         hend = vma->vm_end & HPAGE_PMD_MASK;
2510         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2511                 goto out;
2512         if (!hugepage_vma_check(vma))
2513                 goto out;
2514         pmd = mm_find_pmd(mm, address);
2515         if (!pmd)
2516                 goto out;
2517
2518         anon_vma_lock_write(vma->anon_vma);
2519
2520         pte = pte_offset_map(pmd, address);
2521         pte_ptl = pte_lockptr(mm, pmd);
2522
2523         mmun_start = address;
2524         mmun_end   = address + HPAGE_PMD_SIZE;
2525         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2526         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2527         /*
2528          * After this gup_fast can't run anymore. This also removes
2529          * any huge TLB entry from the CPU so we won't allow
2530          * huge and small TLB entries for the same virtual address
2531          * to avoid the risk of CPU bugs in that area.
2532          */
2533         _pmd = pmdp_collapse_flush(vma, address, pmd);
2534         spin_unlock(pmd_ptl);
2535         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2536
2537         spin_lock(pte_ptl);
2538         isolated = __collapse_huge_page_isolate(vma, address, pte);
2539         spin_unlock(pte_ptl);
2540
2541         if (unlikely(!isolated)) {
2542                 pte_unmap(pte);
2543                 spin_lock(pmd_ptl);
2544                 BUG_ON(!pmd_none(*pmd));
2545                 /*
2546                  * We can only use set_pmd_at when establishing
2547                  * hugepmds and never for establishing regular pmds that
2548                  * points to regular pagetables. Use pmd_populate for that
2549                  */
2550                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2551                 spin_unlock(pmd_ptl);
2552                 anon_vma_unlock_write(vma->anon_vma);
2553                 goto out;
2554         }
2555
2556         /*
2557          * All pages are isolated and locked so anon_vma rmap
2558          * can't run anymore.
2559          */
2560         anon_vma_unlock_write(vma->anon_vma);
2561
2562         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2563         pte_unmap(pte);
2564         __SetPageUptodate(new_page);
2565         pgtable = pmd_pgtable(_pmd);
2566
2567         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2568         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2569
2570         /*
2571          * spin_lock() below is not the equivalent of smp_wmb(), so
2572          * this is needed to avoid the copy_huge_page writes to become
2573          * visible after the set_pmd_at() write.
2574          */
2575         smp_wmb();
2576
2577         spin_lock(pmd_ptl);
2578         BUG_ON(!pmd_none(*pmd));
2579         page_add_new_anon_rmap(new_page, vma, address);
2580         mem_cgroup_commit_charge(new_page, memcg, false);
2581         lru_cache_add_active_or_unevictable(new_page, vma);
2582         pgtable_trans_huge_deposit(mm, pmd, pgtable);
2583         set_pmd_at(mm, address, pmd, _pmd);
2584         update_mmu_cache_pmd(vma, address, pmd);
2585         spin_unlock(pmd_ptl);
2586
2587         *hpage = NULL;
2588
2589         khugepaged_pages_collapsed++;
2590 out_up_write:
2591         up_write(&mm->mmap_sem);
2592         return;
2593
2594 out:
2595         mem_cgroup_cancel_charge(new_page, memcg);
2596         goto out_up_write;
2597 }
2598
2599 static int khugepaged_scan_pmd(struct mm_struct *mm,
2600                                struct vm_area_struct *vma,
2601                                unsigned long address,
2602                                struct page **hpage)
2603 {
2604         pmd_t *pmd;
2605         pte_t *pte, *_pte;
2606         int ret = 0, none_or_zero = 0;
2607         struct page *page;
2608         unsigned long _address;
2609         spinlock_t *ptl;
2610         int node = NUMA_NO_NODE;
2611         bool writable = false, referenced = false;
2612
2613         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2614
2615         pmd = mm_find_pmd(mm, address);
2616         if (!pmd)
2617                 goto out;
2618
2619         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2620         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2621         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2622              _pte++, _address += PAGE_SIZE) {
2623                 pte_t pteval = *_pte;
2624                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2625                         if (!userfaultfd_armed(vma) &&
2626                             ++none_or_zero <= khugepaged_max_ptes_none)
2627                                 continue;
2628                         else
2629                                 goto out_unmap;
2630                 }
2631                 if (!pte_present(pteval))
2632                         goto out_unmap;
2633                 if (pte_write(pteval))
2634                         writable = true;
2635
2636                 page = vm_normal_page(vma, _address, pteval);
2637                 if (unlikely(!page))
2638                         goto out_unmap;
2639                 /*
2640                  * Record which node the original page is from and save this
2641                  * information to khugepaged_node_load[].
2642                  * Khupaged will allocate hugepage from the node has the max
2643                  * hit record.
2644                  */
2645                 node = page_to_nid(page);
2646                 if (khugepaged_scan_abort(node))
2647                         goto out_unmap;
2648                 khugepaged_node_load[node]++;
2649                 VM_BUG_ON_PAGE(PageCompound(page), page);
2650                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2651                         goto out_unmap;
2652                 /*
2653                  * cannot use mapcount: can't collapse if there's a gup pin.
2654                  * The page must only be referenced by the scanned process
2655                  * and page swap cache.
2656                  */
2657                 if (page_count(page) != 1 + !!PageSwapCache(page))
2658                         goto out_unmap;
2659                 if (pte_young(pteval) || PageReferenced(page) ||
2660                     mmu_notifier_test_young(vma->vm_mm, address))
2661                         referenced = true;
2662         }
2663         if (referenced && writable)
2664                 ret = 1;
2665 out_unmap:
2666         pte_unmap_unlock(pte, ptl);
2667         if (ret) {
2668                 node = khugepaged_find_target_node();
2669                 /* collapse_huge_page will return with the mmap_sem released */
2670                 collapse_huge_page(mm, address, hpage, vma, node);
2671         }
2672 out:
2673         return ret;
2674 }
2675
2676 static void collect_mm_slot(struct mm_slot *mm_slot)
2677 {
2678         struct mm_struct *mm = mm_slot->mm;
2679
2680         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2681
2682         if (khugepaged_test_exit(mm)) {
2683                 /* free mm_slot */
2684                 hash_del(&mm_slot->hash);
2685                 list_del(&mm_slot->mm_node);
2686
2687                 /*
2688                  * Not strictly needed because the mm exited already.
2689                  *
2690                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2691                  */
2692
2693                 /* khugepaged_mm_lock actually not necessary for the below */
2694                 free_mm_slot(mm_slot);
2695                 mmdrop(mm);
2696         }
2697 }
2698
2699 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2700                                             struct page **hpage)
2701         __releases(&khugepaged_mm_lock)
2702         __acquires(&khugepaged_mm_lock)
2703 {
2704         struct mm_slot *mm_slot;
2705         struct mm_struct *mm;
2706         struct vm_area_struct *vma;
2707         int progress = 0;
2708
2709         VM_BUG_ON(!pages);
2710         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2711
2712         if (khugepaged_scan.mm_slot)
2713                 mm_slot = khugepaged_scan.mm_slot;
2714         else {
2715                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2716                                      struct mm_slot, mm_node);
2717                 khugepaged_scan.address = 0;
2718                 khugepaged_scan.mm_slot = mm_slot;
2719         }
2720         spin_unlock(&khugepaged_mm_lock);
2721
2722         mm = mm_slot->mm;
2723         down_read(&mm->mmap_sem);
2724         if (unlikely(khugepaged_test_exit(mm)))
2725                 vma = NULL;
2726         else
2727                 vma = find_vma(mm, khugepaged_scan.address);
2728
2729         progress++;
2730         for (; vma; vma = vma->vm_next) {
2731                 unsigned long hstart, hend;
2732
2733                 cond_resched();
2734                 if (unlikely(khugepaged_test_exit(mm))) {
2735                         progress++;
2736                         break;
2737                 }
2738                 if (!hugepage_vma_check(vma)) {
2739 skip:
2740                         progress++;
2741                         continue;
2742                 }
2743                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2744                 hend = vma->vm_end & HPAGE_PMD_MASK;
2745                 if (hstart >= hend)
2746                         goto skip;
2747                 if (khugepaged_scan.address > hend)
2748                         goto skip;
2749                 if (khugepaged_scan.address < hstart)
2750                         khugepaged_scan.address = hstart;
2751                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2752
2753                 while (khugepaged_scan.address < hend) {
2754                         int ret;
2755                         cond_resched();
2756                         if (unlikely(khugepaged_test_exit(mm)))
2757                                 goto breakouterloop;
2758
2759                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2760                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2761                                   hend);
2762                         ret = khugepaged_scan_pmd(mm, vma,
2763                                                   khugepaged_scan.address,
2764                                                   hpage);
2765                         /* move to next address */
2766                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2767                         progress += HPAGE_PMD_NR;
2768                         if (ret)
2769                                 /* we released mmap_sem so break loop */
2770                                 goto breakouterloop_mmap_sem;
2771                         if (progress >= pages)
2772                                 goto breakouterloop;
2773                 }
2774         }
2775 breakouterloop:
2776         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2777 breakouterloop_mmap_sem:
2778
2779         spin_lock(&khugepaged_mm_lock);
2780         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2781         /*
2782          * Release the current mm_slot if this mm is about to die, or
2783          * if we scanned all vmas of this mm.
2784          */
2785         if (khugepaged_test_exit(mm) || !vma) {
2786                 /*
2787                  * Make sure that if mm_users is reaching zero while
2788                  * khugepaged runs here, khugepaged_exit will find
2789                  * mm_slot not pointing to the exiting mm.
2790                  */
2791                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2792                         khugepaged_scan.mm_slot = list_entry(
2793                                 mm_slot->mm_node.next,
2794                                 struct mm_slot, mm_node);
2795                         khugepaged_scan.address = 0;
2796                 } else {
2797                         khugepaged_scan.mm_slot = NULL;
2798                         khugepaged_full_scans++;
2799                 }
2800
2801                 collect_mm_slot(mm_slot);
2802         }
2803
2804         return progress;
2805 }
2806
2807 static int khugepaged_has_work(void)
2808 {
2809         return !list_empty(&khugepaged_scan.mm_head) &&
2810                 khugepaged_enabled();
2811 }
2812
2813 static int khugepaged_wait_event(void)
2814 {
2815         return !list_empty(&khugepaged_scan.mm_head) ||
2816                 kthread_should_stop();
2817 }
2818
2819 static void khugepaged_do_scan(void)
2820 {
2821         struct page *hpage = NULL;
2822         unsigned int progress = 0, pass_through_head = 0;
2823         unsigned int pages = khugepaged_pages_to_scan;
2824         bool wait = true;
2825
2826         barrier(); /* write khugepaged_pages_to_scan to local stack */
2827
2828         while (progress < pages) {
2829                 if (!khugepaged_prealloc_page(&hpage, &wait))
2830                         break;
2831
2832                 cond_resched();
2833
2834                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2835                         break;
2836
2837                 spin_lock(&khugepaged_mm_lock);
2838                 if (!khugepaged_scan.mm_slot)
2839                         pass_through_head++;
2840                 if (khugepaged_has_work() &&
2841                     pass_through_head < 2)
2842                         progress += khugepaged_scan_mm_slot(pages - progress,
2843                                                             &hpage);
2844                 else
2845                         progress = pages;
2846                 spin_unlock(&khugepaged_mm_lock);
2847         }
2848
2849         if (!IS_ERR_OR_NULL(hpage))
2850                 put_page(hpage);
2851 }
2852
2853 static void khugepaged_wait_work(void)
2854 {
2855         if (khugepaged_has_work()) {
2856                 if (!khugepaged_scan_sleep_millisecs)
2857                         return;
2858
2859                 wait_event_freezable_timeout(khugepaged_wait,
2860                                              kthread_should_stop(),
2861                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2862                 return;
2863         }
2864
2865         if (khugepaged_enabled())
2866                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2867 }
2868
2869 static int khugepaged(void *none)
2870 {
2871         struct mm_slot *mm_slot;
2872
2873         set_freezable();
2874         set_user_nice(current, MAX_NICE);
2875
2876         while (!kthread_should_stop()) {
2877                 khugepaged_do_scan();
2878                 khugepaged_wait_work();
2879         }
2880
2881         spin_lock(&khugepaged_mm_lock);
2882         mm_slot = khugepaged_scan.mm_slot;
2883         khugepaged_scan.mm_slot = NULL;
2884         if (mm_slot)
2885                 collect_mm_slot(mm_slot);
2886         spin_unlock(&khugepaged_mm_lock);
2887         return 0;
2888 }
2889
2890 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2891                 unsigned long haddr, pmd_t *pmd)
2892 {
2893         struct mm_struct *mm = vma->vm_mm;
2894         pgtable_t pgtable;
2895         pmd_t _pmd;
2896         int i;
2897
2898         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2899         /* leave pmd empty until pte is filled */
2900
2901         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2902         pmd_populate(mm, &_pmd, pgtable);
2903
2904         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2905                 pte_t *pte, entry;
2906                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2907                 entry = pte_mkspecial(entry);
2908                 pte = pte_offset_map(&_pmd, haddr);
2909                 VM_BUG_ON(!pte_none(*pte));
2910                 set_pte_at(mm, haddr, pte, entry);
2911                 pte_unmap(pte);
2912         }
2913         smp_wmb(); /* make pte visible before pmd */
2914         pmd_populate(mm, pmd, pgtable);
2915         put_huge_zero_page();
2916 }
2917
2918 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2919                 pmd_t *pmd)
2920 {
2921         spinlock_t *ptl;
2922         struct page *page = NULL;
2923         struct mm_struct *mm = vma->vm_mm;
2924         unsigned long haddr = address & HPAGE_PMD_MASK;
2925         unsigned long mmun_start;       /* For mmu_notifiers */
2926         unsigned long mmun_end;         /* For mmu_notifiers */
2927
2928         BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
2929
2930         mmun_start = haddr;
2931         mmun_end   = haddr + HPAGE_PMD_SIZE;
2932 again:
2933         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2934         ptl = pmd_lock(mm, pmd);
2935         if (unlikely(!pmd_trans_huge(*pmd)))
2936                 goto unlock;
2937         if (vma_is_dax(vma)) {
2938                 pmdp_huge_clear_flush(vma, haddr, pmd);
2939         } else if (is_huge_zero_pmd(*pmd)) {
2940                 __split_huge_zero_page_pmd(vma, haddr, pmd);
2941         } else {
2942                 page = pmd_page(*pmd);
2943                 VM_BUG_ON_PAGE(!page_count(page), page);
2944                 get_page(page);
2945         }
2946  unlock:
2947         spin_unlock(ptl);
2948         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2949
2950         if (!page)
2951                 return;
2952
2953         split_huge_page(page);
2954         put_page(page);
2955
2956         /*
2957          * We don't always have down_write of mmap_sem here: a racing
2958          * do_huge_pmd_wp_page() might have copied-on-write to another
2959          * huge page before our split_huge_page() got the anon_vma lock.
2960          */
2961         if (unlikely(pmd_trans_huge(*pmd)))
2962                 goto again;
2963 }
2964
2965 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2966                 pmd_t *pmd)
2967 {
2968         struct vm_area_struct *vma;
2969
2970         vma = find_vma(mm, address);
2971         BUG_ON(vma == NULL);
2972         split_huge_page_pmd(vma, address, pmd);
2973 }
2974
2975 static void split_huge_page_address(struct mm_struct *mm,
2976                                     unsigned long address)
2977 {
2978         pgd_t *pgd;
2979         pud_t *pud;
2980         pmd_t *pmd;
2981
2982         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2983
2984         pgd = pgd_offset(mm, address);
2985         if (!pgd_present(*pgd))
2986                 return;
2987
2988         pud = pud_offset(pgd, address);
2989         if (!pud_present(*pud))
2990                 return;
2991
2992         pmd = pmd_offset(pud, address);
2993         if (!pmd_present(*pmd))
2994                 return;
2995         /*
2996          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2997          * materialize from under us.
2998          */
2999         split_huge_page_pmd_mm(mm, address, pmd);
3000 }
3001
3002 void vma_adjust_trans_huge(struct vm_area_struct *vma,
3003                              unsigned long start,
3004                              unsigned long end,
3005                              long adjust_next)
3006 {
3007         /*
3008          * If the new start address isn't hpage aligned and it could
3009          * previously contain an hugepage: check if we need to split
3010          * an huge pmd.
3011          */
3012         if (start & ~HPAGE_PMD_MASK &&
3013             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
3014             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3015                 split_huge_page_address(vma->vm_mm, start);
3016
3017         /*
3018          * If the new end address isn't hpage aligned and it could
3019          * previously contain an hugepage: check if we need to split
3020          * an huge pmd.
3021          */
3022         if (end & ~HPAGE_PMD_MASK &&
3023             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
3024             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3025                 split_huge_page_address(vma->vm_mm, end);
3026
3027         /*
3028          * If we're also updating the vma->vm_next->vm_start, if the new
3029          * vm_next->vm_start isn't page aligned and it could previously
3030          * contain an hugepage: check if we need to split an huge pmd.
3031          */
3032         if (adjust_next > 0) {
3033                 struct vm_area_struct *next = vma->vm_next;
3034                 unsigned long nstart = next->vm_start;
3035                 nstart += adjust_next << PAGE_SHIFT;
3036                 if (nstart & ~HPAGE_PMD_MASK &&
3037                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
3038                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
3039                         split_huge_page_address(next->vm_mm, nstart);
3040         }
3041 }