mm/huge_memory.c

   1 /*
   2  *  Copyright (C) 2009  Red Hat, Inc.
   3  *
   4  *  This work is licensed under the terms of the GNU GPL, version 2. See
   5  *  the COPYING file in the top-level directory.
   6  */
   7
   8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   9
  10 #include <linux/mm.h>
  11 #include <linux/sched.h>
  12 #include <linux/highmem.h>
  13 #include <linux/hugetlb.h>
  14 #include <linux/mmu_notifier.h>
  15 #include <linux/rmap.h>
  16 #include <linux/swap.h>
  17 #include <linux/shrinker.h>
  18 #include <linux/mm_inline.h>
  19 #include <linux/dax.h>
  20 #include <linux/kthread.h>
  21 #include <linux/khugepaged.h>
  22 #include <linux/freezer.h>
  23 #include <linux/mman.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/migrate.h>
  26 #include <linux/hashtable.h>
  27 #include <linux/userfaultfd_k.h>
  28 #include <linux/page_idle.h>
  29
  30 #include <asm/tlb.h>
  31 #include <asm/pgalloc.h>
  32 #include "internal.h"
  33
  34 /*
  35  * By default transparent hugepage support is disabled in order that avoid
  36  * to risk increase the memory footprint of applications without a guaranteed
  37  * benefit. When transparent hugepage support is enabled, is for all mappings,
  38  * and khugepaged scans all mappings.
  39  * Defrag is invoked by khugepaged hugepage allocations and by page faults
  40  * for all hugepage allocations.
  41  */
  42 unsigned long transparent_hugepage_flags __read_mostly =
  43 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  44         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  45 #endif
  46 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  47         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  48 #endif
  49         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
  50         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  51         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  52
  53 /* default scan 8*512 pte (or vmas) every 30 second */
  54 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
  55 static unsigned int khugepaged_pages_collapsed;
  56 static unsigned int khugepaged_full_scans;
  57 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  58 /* during fragmentation poll the hugepage allocator once every minute */
  59 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  60 static struct task_struct *khugepaged_thread __read_mostly;
  61 static DEFINE_MUTEX(khugepaged_mutex);
  62 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  63 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  64 /*
  65  * default collapse hugepages if there is at least one pte mapped like
  66  * it would have happened if the vma was large enough during page
  67  * fault.
  68  */
  69 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
  70
  71 static int khugepaged(void *none);
  72 static int khugepaged_slab_init(void);
  73 static void khugepaged_slab_exit(void);
  74
  75 #define MM_SLOTS_HASH_BITS 10
  76 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
  77
  78 static struct kmem_cache *mm_slot_cache __read_mostly;
  79
  80 /**
  81  * struct mm_slot - hash lookup from mm to mm_slot
  82  * @hash: hash collision list
  83  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  84  * @mm: the mm that this information is valid for
  85  */
  86 struct mm_slot {
  87         struct hlist_node hash;
  88         struct list_head mm_node;
  89         struct mm_struct *mm;
  90 };
  91
  92 /**
  93  * struct khugepaged_scan - cursor for scanning
  94  * @mm_head: the head of the mm list to scan
  95  * @mm_slot: the current mm_slot we are scanning
  96  * @address: the next address inside that to be scanned
  97  *
  98  * There is only the one khugepaged_scan instance of this cursor structure.
  99  */
 100 struct khugepaged_scan {
 101         struct list_head mm_head;
 102         struct mm_slot *mm_slot;
 103         unsigned long address;
 104 };
 105 static struct khugepaged_scan khugepaged_scan = {
 106         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 107 };
 108
 109
 110 static void set_recommended_min_free_kbytes(void)
 111 {
 112         struct zone *zone;
 113         int nr_zones = 0;
 114         unsigned long recommended_min;
 115
 116         for_each_populated_zone(zone)
 117                 nr_zones++;
 118
 119         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
 120         recommended_min = pageblock_nr_pages * nr_zones * 2;
 121
 122         /*
 123          * Make sure that on average at least two pageblocks are almost free
 124          * of another type, one for a migratetype to fall back to and a
 125          * second to avoid subsequent fallbacks of other types There are 3
 126          * MIGRATE_TYPES we care about.
 127          */
 128         recommended_min += pageblock_nr_pages * nr_zones *
 129                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
 130
 131         /* don't ever allow to reserve more than 5% of the lowmem */
 132         recommended_min = min(recommended_min,
 133                               (unsigned long) nr_free_buffer_pages() / 20);
 134         recommended_min <<= (PAGE_SHIFT-10);
 135
 136         if (recommended_min > min_free_kbytes) {
 137                 if (user_min_free_kbytes >= 0)
 138                         pr_info("raising min_free_kbytes from %d to %lu "
 139                                 "to help transparent hugepage allocations\n",
 140                                 min_free_kbytes, recommended_min);
 141
 142                 min_free_kbytes = recommended_min;
 143         }
 144         setup_per_zone_wmarks();
 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 (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_RECLAIM)) | extra_gfp;
 790 }
 791
 792 /* Caller must hold page table lock. */
 793 static 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                 int ret;
 828                 pgtable = pte_alloc_one(mm, haddr);
 829                 if (unlikely(!pgtable))
 830                         return VM_FAULT_OOM;
 831                 zero_page = get_huge_zero_page();
 832                 if (unlikely(!zero_page)) {
 833                         pte_free(mm, pgtable);
 834                         count_vm_event(THP_FAULT_FALLBACK);
 835                         return VM_FAULT_FALLBACK;
 836                 }
 837                 ptl = pmd_lock(mm, pmd);
 838                 ret = 0;
 839                 if (pmd_none(*pmd)) {
 840                         if (userfaultfd_missing(vma)) {
 841                                 spin_unlock(ptl);
 842                                 pte_free(mm, pgtable);
 843                                 put_huge_zero_page();
 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                         }
 853                 } else {
 854                         spin_unlock(ptl);
 855                         pte_free(mm, pgtable);
 856                         put_huge_zero_page();
 857                 }
 858                 return ret;
 859         }
 860         gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
 861         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 862         if (unlikely(!page)) {
 863                 count_vm_event(THP_FAULT_FALLBACK);
 864                 return VM_FAULT_FALLBACK;
 865         }
 866         return __do_huge_pmd_anonymous_page(mm, vma, address, pmd, page, gfp,
 867                                             flags);
 868 }
 869
 870 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 871                 pmd_t *pmd, unsigned long pfn, pgprot_t prot, bool write)
 872 {
 873         struct mm_struct *mm = vma->vm_mm;
 874         pmd_t entry;
 875         spinlock_t *ptl;
 876
 877         ptl = pmd_lock(mm, pmd);
 878         if (pmd_none(*pmd)) {
 879                 entry = pmd_mkhuge(pfn_pmd(pfn, prot));
 880                 if (write) {
 881                         entry = pmd_mkyoung(pmd_mkdirty(entry));
 882                         entry = maybe_pmd_mkwrite(entry, vma);
 883                 }
 884                 set_pmd_at(mm, addr, pmd, entry);
 885                 update_mmu_cache_pmd(vma, addr, pmd);
 886         }
 887         spin_unlock(ptl);
 888 }
 889
 890 int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 891                         pmd_t *pmd, unsigned long pfn, bool write)
 892 {
 893         pgprot_t pgprot = vma->vm_page_prot;
 894         /*
 895          * If we had pmd_special, we could avoid all these restrictions,
 896          * but we need to be consistent with PTEs and architectures that
 897          * can't support a 'special' bit.
 898          */
 899         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
 900         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 901                                                 (VM_PFNMAP|VM_MIXEDMAP));
 902         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 903         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
 904
 905         if (addr < vma->vm_start || addr >= vma->vm_end)
 906                 return VM_FAULT_SIGBUS;
 907         if (track_pfn_insert(vma, &pgprot, pfn))
 908                 return VM_FAULT_SIGBUS;
 909         insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
 910         return VM_FAULT_NOPAGE;
 911 }
 912
 913 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 914                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 915                   struct vm_area_struct *vma)
 916 {
 917         spinlock_t *dst_ptl, *src_ptl;
 918         struct page *src_page;
 919         pmd_t pmd;
 920         pgtable_t pgtable;
 921         int ret;
 922
 923         ret = -ENOMEM;
 924         pgtable = pte_alloc_one(dst_mm, addr);
 925         if (unlikely(!pgtable))
 926                 goto out;
 927
 928         dst_ptl = pmd_lock(dst_mm, dst_pmd);
 929         src_ptl = pmd_lockptr(src_mm, src_pmd);
 930         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 931
 932         ret = -EAGAIN;
 933         pmd = *src_pmd;
 934         if (unlikely(!pmd_trans_huge(pmd))) {
 935                 pte_free(dst_mm, pgtable);
 936                 goto out_unlock;
 937         }
 938         /*
 939          * When page table lock is held, the huge zero pmd should not be
 940          * under splitting since we don't split the page itself, only pmd to
 941          * a page table.
 942          */
 943         if (is_huge_zero_pmd(pmd)) {
 944                 struct page *zero_page;
 945                 /*
 946                  * get_huge_zero_page() will never allocate a new page here,
 947                  * since we already have a zero page to copy. It just takes a
 948                  * reference.
 949                  */
 950                 zero_page = get_huge_zero_page();
 951                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
 952                                 zero_page);
 953                 ret = 0;
 954                 goto out_unlock;
 955         }
 956
 957         if (unlikely(pmd_trans_splitting(pmd))) {
 958                 /* split huge page running from under us */
 959                 spin_unlock(src_ptl);
 960                 spin_unlock(dst_ptl);
 961                 pte_free(dst_mm, pgtable);
 962
 963                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 964                 goto out;
 965         }
 966         src_page = pmd_page(pmd);
 967         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
 968         get_page(src_page);
 969         page_dup_rmap(src_page);
 970         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 971
 972         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 973         pmd = pmd_mkold(pmd_wrprotect(pmd));
 974         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
 975         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 976         atomic_long_inc(&dst_mm->nr_ptes);
 977
 978         ret = 0;
 979 out_unlock:
 980         spin_unlock(src_ptl);
 981         spin_unlock(dst_ptl);
 982 out:
 983         return ret;
 984 }
 985
 986 void huge_pmd_set_accessed(struct mm_struct *mm,
 987                            struct vm_area_struct *vma,
 988                            unsigned long address,
 989                            pmd_t *pmd, pmd_t orig_pmd,
 990                            int dirty)
 991 {
 992         spinlock_t *ptl;
 993         pmd_t entry;
 994         unsigned long haddr;
 995
 996         ptl = pmd_lock(mm, pmd);
 997         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 998                 goto unlock;
 999
1000         entry = pmd_mkyoung(orig_pmd);
1001         haddr = address & HPAGE_PMD_MASK;
1002         if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
1003                 update_mmu_cache_pmd(vma, address, pmd);
1004
1005 unlock:
1006         spin_unlock(ptl);
1007 }
1008
1009 /*
1010  * Save CONFIG_DEBUG_PAGEALLOC from faulting falsely on tail pages
1011  * during copy_user_huge_page()'s copy_page_rep(): in the case when
1012  * the source page gets split and a tail freed before copy completes.
1013  * Called under pmd_lock of checked pmd, so safe from splitting itself.
1014  */
1015 static void get_user_huge_page(struct page *page)
1016 {
1017         if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
1018                 struct page *endpage = page + HPAGE_PMD_NR;
1019
1020                 atomic_add(HPAGE_PMD_NR, &page->_count);
1021                 while (++page < endpage)
1022                         get_huge_page_tail(page);
1023         } else {
1024                 get_page(page);
1025         }
1026 }
1027
1028 static void put_user_huge_page(struct page *page)
1029 {
1030         if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
1031                 struct page *endpage = page + HPAGE_PMD_NR;
1032
1033                 while (page < endpage)
1034                         put_page(page++);
1035         } else {
1036                 put_page(page);
1037         }
1038 }
1039
1040 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1041                                         struct vm_area_struct *vma,
1042                                         unsigned long address,
1043                                         pmd_t *pmd, pmd_t orig_pmd,
1044                                         struct page *page,
1045                                         unsigned long haddr)
1046 {
1047         struct mem_cgroup *memcg;
1048         spinlock_t *ptl;
1049         pgtable_t pgtable;
1050         pmd_t _pmd;
1051         int ret = 0, i;
1052         struct page **pages;
1053         unsigned long mmun_start;       /* For mmu_notifiers */
1054         unsigned long mmun_end;         /* For mmu_notifiers */
1055
1056         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1057                         GFP_KERNEL);
1058         if (unlikely(!pages)) {
1059                 ret |= VM_FAULT_OOM;
1060                 goto out;
1061         }
1062
1063         for (i = 0; i < HPAGE_PMD_NR; i++) {
1064                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1065                                                __GFP_OTHER_NODE,
1066                                                vma, address, page_to_nid(page));
1067                 if (unlikely(!pages[i] ||
1068                              mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL,
1069                                                    &memcg))) {
1070                         if (pages[i])
1071                                 put_page(pages[i]);
1072                         while (--i >= 0) {
1073                                 memcg = (void *)page_private(pages[i]);
1074                                 set_page_private(pages[i], 0);
1075                                 mem_cgroup_cancel_charge(pages[i], memcg);
1076                                 put_page(pages[i]);
1077                         }
1078                         kfree(pages);
1079                         ret |= VM_FAULT_OOM;
1080                         goto out;
1081                 }
1082                 set_page_private(pages[i], (unsigned long)memcg);
1083         }
1084
1085         for (i = 0; i < HPAGE_PMD_NR; i++) {
1086                 copy_user_highpage(pages[i], page + i,
1087                                    haddr + PAGE_SIZE * i, vma);
1088                 __SetPageUptodate(pages[i]);
1089                 cond_resched();
1090         }
1091
1092         mmun_start = haddr;
1093         mmun_end   = haddr + HPAGE_PMD_SIZE;
1094         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1095
1096         ptl = pmd_lock(mm, pmd);
1097         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1098                 goto out_free_pages;
1099         VM_BUG_ON_PAGE(!PageHead(page), page);
1100
1101         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1102         /* leave pmd empty until pte is filled */
1103
1104         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1105         pmd_populate(mm, &_pmd, pgtable);
1106
1107         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1108                 pte_t *pte, entry;
1109                 entry = mk_pte(pages[i], vma->vm_page_prot);
1110                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1111                 memcg = (void *)page_private(pages[i]);
1112                 set_page_private(pages[i], 0);
1113                 page_add_new_anon_rmap(pages[i], vma, haddr);
1114                 mem_cgroup_commit_charge(pages[i], memcg, false);
1115                 lru_cache_add_active_or_unevictable(pages[i], vma);
1116                 pte = pte_offset_map(&_pmd, haddr);
1117                 VM_BUG_ON(!pte_none(*pte));
1118                 set_pte_at(mm, haddr, pte, entry);
1119                 pte_unmap(pte);
1120         }
1121         kfree(pages);
1122
1123         smp_wmb(); /* make pte visible before pmd */
1124         pmd_populate(mm, pmd, pgtable);
1125         page_remove_rmap(page);
1126         spin_unlock(ptl);
1127
1128         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1129
1130         ret |= VM_FAULT_WRITE;
1131         put_page(page);
1132
1133 out:
1134         return ret;
1135
1136 out_free_pages:
1137         spin_unlock(ptl);
1138         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1139         for (i = 0; i < HPAGE_PMD_NR; i++) {
1140                 memcg = (void *)page_private(pages[i]);
1141                 set_page_private(pages[i], 0);
1142                 mem_cgroup_cancel_charge(pages[i], memcg);
1143                 put_page(pages[i]);
1144         }
1145         kfree(pages);
1146         goto out;
1147 }
1148
1149 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1150                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1151 {
1152         spinlock_t *ptl;
1153         int ret = 0;
1154         struct page *page = NULL, *new_page;
1155         struct mem_cgroup *memcg;
1156         unsigned long haddr;
1157         unsigned long mmun_start;       /* For mmu_notifiers */
1158         unsigned long mmun_end;         /* For mmu_notifiers */
1159         gfp_t huge_gfp;                 /* for allocation and charge */
1160
1161         ptl = pmd_lockptr(mm, pmd);
1162         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1163         haddr = address & HPAGE_PMD_MASK;
1164         if (is_huge_zero_pmd(orig_pmd))
1165                 goto alloc;
1166         spin_lock(ptl);
1167         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1168                 goto out_unlock;
1169
1170         page = pmd_page(orig_pmd);
1171         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1172         if (page_mapcount(page) == 1) {
1173                 pmd_t entry;
1174                 entry = pmd_mkyoung(orig_pmd);
1175                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1176                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1177                         update_mmu_cache_pmd(vma, address, pmd);
1178                 ret |= VM_FAULT_WRITE;
1179                 goto out_unlock;
1180         }
1181         get_user_huge_page(page);
1182         spin_unlock(ptl);
1183 alloc:
1184         if (transparent_hugepage_enabled(vma) &&
1185             !transparent_hugepage_debug_cow()) {
1186                 huge_gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
1187                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1188         } else
1189                 new_page = NULL;
1190
1191         if (unlikely(!new_page)) {
1192                 if (!page) {
1193                         split_huge_page_pmd(vma, address, pmd);
1194                         ret |= VM_FAULT_FALLBACK;
1195                 } else {
1196                         ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1197                                         pmd, orig_pmd, page, haddr);
1198                         if (ret & VM_FAULT_OOM) {
1199                                 split_huge_page(page);
1200                                 ret |= VM_FAULT_FALLBACK;
1201                         }
1202                         put_user_huge_page(page);
1203                 }
1204                 count_vm_event(THP_FAULT_FALLBACK);
1205                 goto out;
1206         }
1207
1208         if (unlikely(mem_cgroup_try_charge(new_page, mm, huge_gfp, &memcg))) {
1209                 put_page(new_page);
1210                 if (page) {
1211                         split_huge_page(page);
1212                         put_user_huge_page(page);
1213                 } else
1214                         split_huge_page_pmd(vma, address, pmd);
1215                 ret |= VM_FAULT_FALLBACK;
1216                 count_vm_event(THP_FAULT_FALLBACK);
1217                 goto out;
1218         }
1219
1220         count_vm_event(THP_FAULT_ALLOC);
1221
1222         if (!page)
1223                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1224         else
1225                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1226         __SetPageUptodate(new_page);
1227
1228         mmun_start = haddr;
1229         mmun_end   = haddr + HPAGE_PMD_SIZE;
1230         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1231
1232         spin_lock(ptl);
1233         if (page)
1234                 put_user_huge_page(page);
1235         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1236                 spin_unlock(ptl);
1237                 mem_cgroup_cancel_charge(new_page, memcg);
1238                 put_page(new_page);
1239                 goto out_mn;
1240         } else {
1241                 pmd_t entry;
1242                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1243                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1244                 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1245                 page_add_new_anon_rmap(new_page, vma, haddr);
1246                 mem_cgroup_commit_charge(new_page, memcg, false);
1247                 lru_cache_add_active_or_unevictable(new_page, vma);
1248                 set_pmd_at(mm, haddr, pmd, entry);
1249                 update_mmu_cache_pmd(vma, address, pmd);
1250                 if (!page) {
1251                         add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1252                         put_huge_zero_page();
1253                 } else {
1254                         VM_BUG_ON_PAGE(!PageHead(page), page);
1255                         page_remove_rmap(page);
1256                         put_page(page);
1257                 }
1258                 ret |= VM_FAULT_WRITE;
1259         }
1260         spin_unlock(ptl);
1261 out_mn:
1262         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1263 out:
1264         return ret;
1265 out_unlock:
1266         spin_unlock(ptl);
1267         return ret;
1268 }
1269
1270 /*
1271  * FOLL_FORCE can write to even unwritable pmd's, but only
1272  * after we've gone through a COW cycle and they are dirty.
1273  */
1274 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1275 {
1276         return pmd_write(pmd) ||
1277                ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1278 }
1279
1280 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1281                                    unsigned long addr,
1282                                    pmd_t *pmd,
1283                                    unsigned int flags)
1284 {
1285         struct mm_struct *mm = vma->vm_mm;
1286         struct page *page = NULL;
1287
1288         assert_spin_locked(pmd_lockptr(mm, pmd));
1289
1290         if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1291                 goto out;
1292
1293         /* Avoid dumping huge zero page */
1294         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1295                 return ERR_PTR(-EFAULT);
1296
1297         /* Full NUMA hinting faults to serialise migration in fault paths */
1298         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1299                 goto out;
1300
1301         page = pmd_page(*pmd);
1302         VM_BUG_ON_PAGE(!PageHead(page), page);
1303         if (flags & FOLL_TOUCH) {
1304                 pmd_t _pmd;
1305                 _pmd = pmd_mkyoung(*pmd);
1306                 if (flags & FOLL_WRITE)
1307                         _pmd = pmd_mkdirty(_pmd);
1308                 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1309                                           pmd, _pmd, flags & FOLL_WRITE))
1310                         update_mmu_cache_pmd(vma, addr, pmd);
1311         }
1312         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1313                 if (page->mapping && trylock_page(page)) {
1314                         lru_add_drain();
1315                         if (page->mapping)
1316                                 mlock_vma_page(page);
1317                         unlock_page(page);
1318                 }
1319         }
1320         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1321         VM_BUG_ON_PAGE(!PageCompound(page), page);
1322         if (flags & FOLL_GET)
1323                 get_page_foll(page);
1324
1325 out:
1326         return page;
1327 }
1328
1329 /* NUMA hinting page fault entry point for trans huge pmds */
1330 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1331                                 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1332 {
1333         spinlock_t *ptl;
1334         struct anon_vma *anon_vma = NULL;
1335         struct page *page;
1336         unsigned long haddr = addr & HPAGE_PMD_MASK;
1337         int page_nid = -1, this_nid = numa_node_id();
1338         int target_nid, last_cpupid = -1;
1339         bool page_locked;
1340         bool migrated = false;
1341         bool was_writable;
1342         int flags = 0;
1343
1344         /* A PROT_NONE fault should not end up here */
1345         BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
1346
1347         ptl = pmd_lock(mm, pmdp);
1348         if (unlikely(!pmd_same(pmd, *pmdp)))
1349                 goto out_unlock;
1350
1351         /*
1352          * If there are potential migrations, wait for completion and retry
1353          * without disrupting NUMA hinting information. Do not relock and
1354          * check_same as the page may no longer be mapped.
1355          */
1356         if (unlikely(pmd_trans_migrating(*pmdp))) {
1357                 page = pmd_page(*pmdp);
1358                 if (!get_page_unless_zero(page))
1359                         goto out_unlock;
1360                 spin_unlock(ptl);
1361                 wait_on_page_locked(page);
1362                 put_page(page);
1363                 goto out;
1364         }
1365
1366         page = pmd_page(pmd);
1367         BUG_ON(is_huge_zero_page(page));
1368         page_nid = page_to_nid(page);
1369         last_cpupid = page_cpupid_last(page);
1370         count_vm_numa_event(NUMA_HINT_FAULTS);
1371         if (page_nid == this_nid) {
1372                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1373                 flags |= TNF_FAULT_LOCAL;
1374         }
1375
1376         /* See similar comment in do_numa_page for explanation */
1377         if (!(vma->vm_flags & VM_WRITE))
1378                 flags |= TNF_NO_GROUP;
1379
1380         /*
1381          * Acquire the page lock to serialise THP migrations but avoid dropping
1382          * page_table_lock if at all possible
1383          */
1384         page_locked = trylock_page(page);
1385         target_nid = mpol_misplaced(page, vma, haddr);
1386         if (target_nid == -1) {
1387                 /* If the page was locked, there are no parallel migrations */
1388                 if (page_locked)
1389                         goto clear_pmdnuma;
1390         }
1391
1392         /* Migration could have started since the pmd_trans_migrating check */
1393         if (!page_locked) {
1394                 page_nid = -1;
1395                 if (!get_page_unless_zero(page))
1396                         goto out_unlock;
1397                 spin_unlock(ptl);
1398                 wait_on_page_locked(page);
1399                 put_page(page);
1400                 goto out;
1401         }
1402
1403         /*
1404          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1405          * to serialises splits
1406          */
1407         get_page(page);
1408         spin_unlock(ptl);
1409         anon_vma = page_lock_anon_vma_read(page);
1410
1411         /* Confirm the PMD did not change while page_table_lock was released */
1412         spin_lock(ptl);
1413         if (unlikely(!pmd_same(pmd, *pmdp))) {
1414                 unlock_page(page);
1415                 put_page(page);
1416                 page_nid = -1;
1417                 goto out_unlock;
1418         }
1419
1420         /* Bail if we fail to protect against THP splits for any reason */
1421         if (unlikely(!anon_vma)) {
1422                 put_page(page);
1423                 page_nid = -1;
1424                 goto clear_pmdnuma;
1425         }
1426
1427         /*
1428          * Migrate the THP to the requested node, returns with page unlocked
1429          * and access rights restored.
1430          */
1431         spin_unlock(ptl);
1432         migrated = migrate_misplaced_transhuge_page(mm, vma,
1433                                 pmdp, pmd, addr, page, target_nid);
1434         if (migrated) {
1435                 flags |= TNF_MIGRATED;
1436                 page_nid = target_nid;
1437         } else
1438                 flags |= TNF_MIGRATE_FAIL;
1439
1440         goto out;
1441 clear_pmdnuma:
1442         BUG_ON(!PageLocked(page));
1443         was_writable = pmd_write(pmd);
1444         pmd = pmd_modify(pmd, vma->vm_page_prot);
1445         pmd = pmd_mkyoung(pmd);
1446         if (was_writable)
1447                 pmd = pmd_mkwrite(pmd);
1448         set_pmd_at(mm, haddr, pmdp, pmd);
1449         update_mmu_cache_pmd(vma, addr, pmdp);
1450         unlock_page(page);
1451 out_unlock:
1452         spin_unlock(ptl);
1453
1454 out:
1455         if (anon_vma)
1456                 page_unlock_anon_vma_read(anon_vma);
1457
1458         if (page_nid != -1)
1459                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
1460
1461         return 0;
1462 }
1463
1464 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1465                  pmd_t *pmd, unsigned long addr)
1466 {
1467         pmd_t orig_pmd;
1468         spinlock_t *ptl;
1469
1470         if (__pmd_trans_huge_lock(pmd, vma, &ptl) != 1)
1471                 return 0;
1472         /*
1473          * For architectures like ppc64 we look at deposited pgtable
1474          * when calling pmdp_huge_get_and_clear. So do the
1475          * pgtable_trans_huge_withdraw after finishing pmdp related
1476          * operations.
1477          */
1478         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1479                         tlb->fullmm);
1480         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1481         if (vma_is_dax(vma)) {
1482                 spin_unlock(ptl);
1483                 if (is_huge_zero_pmd(orig_pmd))
1484                         put_huge_zero_page();
1485         } else if (is_huge_zero_pmd(orig_pmd)) {
1486                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1487                 atomic_long_dec(&tlb->mm->nr_ptes);
1488                 spin_unlock(ptl);
1489                 put_huge_zero_page();
1490         } else {
1491                 struct page *page = pmd_page(orig_pmd);
1492                 page_remove_rmap(page);
1493                 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1494                 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1495                 VM_BUG_ON_PAGE(!PageHead(page), page);
1496                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1497                 atomic_long_dec(&tlb->mm->nr_ptes);
1498                 spin_unlock(ptl);
1499                 tlb_remove_page(tlb, page);
1500         }
1501         return 1;
1502 }
1503
1504 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1505                   unsigned long old_addr,
1506                   unsigned long new_addr, unsigned long old_end,
1507                   pmd_t *old_pmd, pmd_t *new_pmd)
1508 {
1509         spinlock_t *old_ptl, *new_ptl;
1510         int ret = 0;
1511         pmd_t pmd;
1512         bool force_flush = false;
1513         struct mm_struct *mm = vma->vm_mm;
1514
1515         if ((old_addr & ~HPAGE_PMD_MASK) ||
1516             (new_addr & ~HPAGE_PMD_MASK) ||
1517             old_end - old_addr < HPAGE_PMD_SIZE ||
1518             (new_vma->vm_flags & VM_NOHUGEPAGE))
1519                 goto out;
1520
1521         /*
1522          * The destination pmd shouldn't be established, free_pgtables()
1523          * should have release it.
1524          */
1525         if (WARN_ON(!pmd_none(*new_pmd))) {
1526                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1527                 goto out;
1528         }
1529
1530         /*
1531          * We don't have to worry about the ordering of src and dst
1532          * ptlocks because exclusive mmap_sem prevents deadlock.
1533          */
1534         ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl);
1535         if (ret == 1) {
1536                 new_ptl = pmd_lockptr(mm, new_pmd);
1537                 if (new_ptl != old_ptl)
1538                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1539                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1540                 if (pmd_present(pmd))
1541                         force_flush = true;
1542                 VM_BUG_ON(!pmd_none(*new_pmd));
1543
1544                 if (pmd_move_must_withdraw(new_ptl, old_ptl)) {
1545                         pgtable_t pgtable;
1546                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1547                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1548                 }
1549                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1550                 if (force_flush)
1551                         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1552                 if (new_ptl != old_ptl)
1553                         spin_unlock(new_ptl);
1554                 spin_unlock(old_ptl);
1555         }
1556 out:
1557         return ret;
1558 }
1559
1560 /*
1561  * Returns
1562  *  - 0 if PMD could not be locked
1563  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1564  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1565  */
1566 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1567                 unsigned long addr, pgprot_t newprot, int prot_numa)
1568 {
1569         struct mm_struct *mm = vma->vm_mm;
1570         spinlock_t *ptl;
1571         pmd_t entry;
1572         bool preserve_write;
1573
1574         int ret = 0;
1575
1576         if (__pmd_trans_huge_lock(pmd, vma, &ptl) != 1)
1577                 return 0;
1578
1579         preserve_write = prot_numa && pmd_write(*pmd);
1580         ret = 1;
1581
1582         /*
1583          * Avoid trapping faults against the zero page. The read-only
1584          * data is likely to be read-cached on the local CPU and
1585          * local/remote hits to the zero page are not interesting.
1586          */
1587         if (prot_numa && is_huge_zero_pmd(*pmd))
1588                 goto unlock;
1589
1590         if (prot_numa && pmd_protnone(*pmd))
1591                 goto unlock;
1592
1593         /*
1594          * In case prot_numa, we are under down_read(mmap_sem). It's critical
1595          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1596          * which is also under down_read(mmap_sem):
1597          *
1598          *      CPU0:                           CPU1:
1599          *                              change_huge_pmd(prot_numa=1)
1600          *                               pmdp_huge_get_and_clear_notify()
1601          * madvise_dontneed()
1602          *  zap_pmd_range()
1603          *   pmd_trans_huge(*pmd) == 0 (without ptl)
1604          *   // skip the pmd
1605          *                               set_pmd_at();
1606          *                               // pmd is re-established
1607          *
1608          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1609          * which may break userspace.
1610          *
1611          * pmdp_invalidate() is required to make sure we don't miss
1612          * dirty/young flags set by hardware.
1613          */
1614         entry = *pmd;
1615         pmdp_invalidate(vma, addr, pmd);
1616
1617         /*
1618          * Recover dirty/young flags.  It relies on pmdp_invalidate to not
1619          * corrupt them.
1620          */
1621         if (pmd_dirty(*pmd))
1622                 entry = pmd_mkdirty(entry);
1623         if (pmd_young(*pmd))
1624                 entry = pmd_mkyoung(entry);
1625
1626         entry = pmd_modify(entry, newprot);
1627         if (preserve_write)
1628                 entry = pmd_mkwrite(entry);
1629         ret = HPAGE_PMD_NR;
1630         set_pmd_at(mm, addr, pmd, entry);
1631         BUG_ON(!preserve_write && pmd_write(entry));
1632 unlock:
1633         spin_unlock(ptl);
1634         return ret;
1635 }
1636
1637 /*
1638  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1639  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1640  *
1641  * Note that if it returns 1, this routine returns without unlocking page
1642  * table locks. So callers must unlock them.
1643  */
1644 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1645                 spinlock_t **ptl)
1646 {
1647         *ptl = pmd_lock(vma->vm_mm, pmd);
1648         if (likely(pmd_trans_huge(*pmd))) {
1649                 if (unlikely(pmd_trans_splitting(*pmd))) {
1650                         spin_unlock(*ptl);
1651                         wait_split_huge_page(vma->anon_vma, pmd);
1652                         return -1;
1653                 } else {
1654                         /* Thp mapped by 'pmd' is stable, so we can
1655                          * handle it as it is. */
1656                         return 1;
1657                 }
1658         }
1659         spin_unlock(*ptl);
1660         return 0;
1661 }
1662
1663 /*
1664  * This function returns whether a given @page is mapped onto the @address
1665  * in the virtual space of @mm.
1666  *
1667  * When it's true, this function returns *pmd with holding the page table lock
1668  * and passing it back to the caller via @ptl.
1669  * If it's false, returns NULL without holding the page table lock.
1670  */
1671 pmd_t *page_check_address_pmd(struct page *page,
1672                               struct mm_struct *mm,
1673                               unsigned long address,
1674                               enum page_check_address_pmd_flag flag,
1675                               spinlock_t **ptl)
1676 {
1677         pgd_t *pgd;
1678         pud_t *pud;
1679         pmd_t *pmd;
1680
1681         if (address & ~HPAGE_PMD_MASK)
1682                 return NULL;
1683
1684         pgd = pgd_offset(mm, address);
1685         if (!pgd_present(*pgd))
1686                 return NULL;
1687         pud = pud_offset(pgd, address);
1688         if (!pud_present(*pud))
1689                 return NULL;
1690         pmd = pmd_offset(pud, address);
1691
1692         *ptl = pmd_lock(mm, pmd);
1693         if (!pmd_present(*pmd))
1694                 goto unlock;
1695         if (pmd_page(*pmd) != page)
1696                 goto unlock;
1697         /*
1698          * split_vma() may create temporary aliased mappings. There is
1699          * no risk as long as all huge pmd are found and have their
1700          * splitting bit set before __split_huge_page_refcount
1701          * runs. Finding the same huge pmd more than once during the
1702          * same rmap walk is not a problem.
1703          */
1704         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1705             pmd_trans_splitting(*pmd))
1706                 goto unlock;
1707         if (pmd_trans_huge(*pmd)) {
1708                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1709                           !pmd_trans_splitting(*pmd));
1710                 return pmd;
1711         }
1712 unlock:
1713         spin_unlock(*ptl);
1714         return NULL;
1715 }
1716
1717 static int __split_huge_page_splitting(struct page *page,
1718                                        struct vm_area_struct *vma,
1719                                        unsigned long address)
1720 {
1721         struct mm_struct *mm = vma->vm_mm;
1722         spinlock_t *ptl;
1723         pmd_t *pmd;
1724         int ret = 0;
1725         /* For mmu_notifiers */
1726         const unsigned long mmun_start = address;
1727         const unsigned long mmun_end   = address + HPAGE_PMD_SIZE;
1728
1729         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1730         pmd = page_check_address_pmd(page, mm, address,
1731                         PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl);
1732         if (pmd) {
1733                 /*
1734                  * We can't temporarily set the pmd to null in order
1735                  * to split it, the pmd must remain marked huge at all
1736                  * times or the VM won't take the pmd_trans_huge paths
1737                  * and it won't wait on the anon_vma->root->rwsem to
1738                  * serialize against split_huge_page*.
1739                  */
1740                 pmdp_splitting_flush(vma, address, pmd);
1741
1742                 ret = 1;
1743                 spin_unlock(ptl);
1744         }
1745         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1746
1747         return ret;
1748 }
1749
1750 static void __split_huge_page_refcount(struct page *page,
1751                                        struct list_head *list)
1752 {
1753         int i;
1754         struct zone *zone = page_zone(page);
1755         struct lruvec *lruvec;
1756         int tail_count = 0;
1757
1758         /* prevent PageLRU to go away from under us, and freeze lru stats */
1759         spin_lock_irq(&zone->lru_lock);
1760         lruvec = mem_cgroup_page_lruvec(page, zone);
1761
1762         compound_lock(page);
1763         /* complete memcg works before add pages to LRU */
1764         mem_cgroup_split_huge_fixup(page);
1765
1766         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1767                 struct page *page_tail = page + i;
1768
1769                 /* tail_page->_mapcount cannot change */
1770                 BUG_ON(page_mapcount(page_tail) < 0);
1771                 tail_count += page_mapcount(page_tail);
1772                 /* check for overflow */
1773                 BUG_ON(tail_count < 0);
1774                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1775                 /*
1776                  * tail_page->_count is zero and not changing from
1777                  * under us. But get_page_unless_zero() may be running
1778                  * from under us on the tail_page. If we used
1779                  * atomic_set() below instead of atomic_add(), we
1780                  * would then run atomic_set() concurrently with
1781                  * get_page_unless_zero(), and atomic_set() is
1782                  * implemented in C not using locked ops. spin_unlock
1783                  * on x86 sometime uses locked ops because of PPro
1784                  * errata 66, 92, so unless somebody can guarantee
1785                  * atomic_set() here would be safe on all archs (and
1786                  * not only on x86), it's safer to use atomic_add().
1787                  */
1788                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1789                            &page_tail->_count);
1790
1791                 /* after clearing PageTail the gup refcount can be released */
1792                 smp_mb__after_atomic();
1793
1794                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1795                 page_tail->flags |= (page->flags &
1796                                      ((1L << PG_referenced) |
1797                                       (1L << PG_swapbacked) |
1798                                       (1L << PG_mlocked) |
1799                                       (1L << PG_uptodate) |
1800                                       (1L << PG_active) |
1801                                       (1L << PG_unevictable)));
1802                 page_tail->flags |= (1L << PG_dirty);
1803
1804                 clear_compound_head(page_tail);
1805
1806                 if (page_is_young(page))
1807                         set_page_young(page_tail);
1808                 if (page_is_idle(page))
1809                         set_page_idle(page_tail);
1810
1811                 /*
1812                  * __split_huge_page_splitting() already set the
1813                  * splitting bit in all pmd that could map this
1814                  * hugepage, that will ensure no CPU can alter the
1815                  * mapcount on the head page. The mapcount is only
1816                  * accounted in the head page and it has to be
1817                  * transferred to all tail pages in the below code. So
1818                  * for this code to be safe, the split the mapcount
1819                  * can't change. But that doesn't mean userland can't
1820                  * keep changing and reading the page contents while
1821                  * we transfer the mapcount, so the pmd splitting
1822                  * status is achieved setting a reserved bit in the
1823                  * pmd, not by clearing the present bit.
1824                 */
1825                 page_tail->_mapcount = page->_mapcount;
1826
1827                 BUG_ON(page_tail->mapping);
1828                 page_tail->mapping = page->mapping;
1829
1830                 page_tail->index = page->index + i;
1831                 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
1832
1833                 BUG_ON(!PageAnon(page_tail));
1834                 BUG_ON(!PageUptodate(page_tail));
1835                 BUG_ON(!PageDirty(page_tail));
1836                 BUG_ON(!PageSwapBacked(page_tail));
1837
1838                 lru_add_page_tail(page, page_tail, lruvec, list);
1839         }
1840         atomic_sub(tail_count, &page->_count);
1841         BUG_ON(atomic_read(&page->_count) <= 0);
1842
1843         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1844
1845         ClearPageCompound(page);
1846         compound_unlock(page);
1847         spin_unlock_irq(&zone->lru_lock);
1848
1849         for (i = 1; i < HPAGE_PMD_NR; i++) {
1850                 struct page *page_tail = page + i;
1851                 BUG_ON(page_count(page_tail) <= 0);
1852                 /*
1853                  * Tail pages may be freed if there wasn't any mapping
1854                  * like if add_to_swap() is running on a lru page that
1855                  * had its mapping zapped. And freeing these pages
1856                  * requires taking the lru_lock so we do the put_page
1857                  * of the tail pages after the split is complete.
1858                  */
1859                 put_page(page_tail);
1860         }
1861
1862         /*
1863          * Only the head page (now become a regular page) is required
1864          * to be pinned by the caller.
1865          */
1866         BUG_ON(page_count(page) <= 0);
1867 }
1868
1869 static int __split_huge_page_map(struct page *page,
1870                                  struct vm_area_struct *vma,
1871                                  unsigned long address)
1872 {
1873         struct mm_struct *mm = vma->vm_mm;
1874         spinlock_t *ptl;
1875         pmd_t *pmd, _pmd;
1876         int ret = 0, i;
1877         pgtable_t pgtable;
1878         unsigned long haddr;
1879
1880         pmd = page_check_address_pmd(page, mm, address,
1881                         PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl);
1882         if (pmd) {
1883                 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1884                 pmd_populate(mm, &_pmd, pgtable);
1885                 if (pmd_write(*pmd))
1886                         BUG_ON(page_mapcount(page) != 1);
1887
1888                 haddr = address;
1889                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1890                         pte_t *pte, entry;
1891                         BUG_ON(PageCompound(page+i));
1892                         /*
1893                          * Note that NUMA hinting access restrictions are not
1894                          * transferred to avoid any possibility of altering
1895                          * permissions across VMAs.
1896                          */
1897                         entry = mk_pte(page + i, vma->vm_page_prot);
1898                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1899                         if (!pmd_write(*pmd))
1900                                 entry = pte_wrprotect(entry);
1901                         if (!pmd_young(*pmd))
1902                                 entry = pte_mkold(entry);
1903                         pte = pte_offset_map(&_pmd, haddr);
1904                         BUG_ON(!pte_none(*pte));
1905                         set_pte_at(mm, haddr, pte, entry);
1906                         pte_unmap(pte);
1907                 }
1908
1909                 smp_wmb(); /* make pte visible before pmd */
1910                 /*
1911                  * Up to this point the pmd is present and huge and
1912                  * userland has the whole access to the hugepage
1913                  * during the split (which happens in place). If we
1914                  * overwrite the pmd with the not-huge version
1915                  * pointing to the pte here (which of course we could
1916                  * if all CPUs were bug free), userland could trigger
1917                  * a small page size TLB miss on the small sized TLB
1918                  * while the hugepage TLB entry is still established
1919                  * in the huge TLB. Some CPU doesn't like that. See
1920                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1921                  * Erratum 383 on page 93. Intel should be safe but is
1922                  * also warns that it's only safe if the permission
1923                  * and cache attributes of the two entries loaded in
1924                  * the two TLB is identical (which should be the case
1925                  * here). But it is generally safer to never allow
1926                  * small and huge TLB entries for the same virtual
1927                  * address to be loaded simultaneously. So instead of
1928                  * doing "pmd_populate(); flush_pmd_tlb_range();" we first
1929                  * mark the current pmd notpresent (atomically because
1930                  * here the pmd_trans_huge and pmd_trans_splitting
1931                  * must remain set at all times on the pmd until the
1932                  * split is complete for this pmd), then we flush the
1933                  * SMP TLB and finally we write the non-huge version
1934                  * of the pmd entry with pmd_populate.
1935                  */
1936                 pmdp_invalidate(vma, address, pmd);
1937                 pmd_populate(mm, pmd, pgtable);
1938                 ret = 1;
1939                 spin_unlock(ptl);
1940         }
1941
1942         return ret;
1943 }
1944
1945 /* must be called with anon_vma->root->rwsem held */
1946 static void __split_huge_page(struct page *page,
1947                               struct anon_vma *anon_vma,
1948                               struct list_head *list)
1949 {
1950         int mapcount, mapcount2;
1951         pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1952         struct anon_vma_chain *avc;
1953
1954         BUG_ON(!PageHead(page));
1955         BUG_ON(PageTail(page));
1956
1957         mapcount = 0;
1958         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1959                 struct vm_area_struct *vma = avc->vma;
1960                 unsigned long addr = vma_address(page, vma);
1961                 BUG_ON(is_vma_temporary_stack(vma));
1962                 mapcount += __split_huge_page_splitting(page, vma, addr);
1963         }
1964         /*
1965          * It is critical that new vmas are added to the tail of the
1966          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1967          * and establishes a child pmd before
1968          * __split_huge_page_splitting() freezes the parent pmd (so if
1969          * we fail to prevent copy_huge_pmd() from running until the
1970          * whole __split_huge_page() is complete), we will still see
1971          * the newly established pmd of the child later during the
1972          * walk, to be able to set it as pmd_trans_splitting too.
1973          */
1974         if (mapcount != page_mapcount(page)) {
1975                 pr_err("mapcount %d page_mapcount %d\n",
1976                         mapcount, page_mapcount(page));
1977                 BUG();
1978         }
1979
1980         __split_huge_page_refcount(page, list);
1981
1982         mapcount2 = 0;
1983         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1984                 struct vm_area_struct *vma = avc->vma;
1985                 unsigned long addr = vma_address(page, vma);
1986                 BUG_ON(is_vma_temporary_stack(vma));
1987                 mapcount2 += __split_huge_page_map(page, vma, addr);
1988         }
1989         if (mapcount != mapcount2) {
1990                 pr_err("mapcount %d mapcount2 %d page_mapcount %d\n",
1991                         mapcount, mapcount2, page_mapcount(page));
1992                 BUG();
1993         }
1994 }
1995
1996 /*
1997  * Split a hugepage into normal pages. This doesn't change the position of head
1998  * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1999  * @list. Both head page and tail pages will inherit mapping, flags, and so on
2000  * from the hugepage.
2001  * Return 0 if the hugepage is split successfully otherwise return 1.
2002  */
2003 int split_huge_page_to_list(struct page *page, struct list_head *list)
2004 {
2005         struct anon_vma *anon_vma;
2006         int ret = 1;
2007
2008         BUG_ON(is_huge_zero_page(page));
2009         BUG_ON(!PageAnon(page));
2010
2011         /*
2012          * The caller does not necessarily hold an mmap_sem that would prevent
2013          * the anon_vma disappearing so we first we take a reference to it
2014          * and then lock the anon_vma for write. This is similar to
2015          * page_lock_anon_vma_read except the write lock is taken to serialise
2016          * against parallel split or collapse operations.
2017          */
2018         anon_vma = page_get_anon_vma(page);
2019         if (!anon_vma)
2020                 goto out;
2021         anon_vma_lock_write(anon_vma);
2022
2023         ret = 0;
2024         if (!PageCompound(page))
2025                 goto out_unlock;
2026
2027         BUG_ON(!PageSwapBacked(page));
2028         __split_huge_page(page, anon_vma, list);
2029         count_vm_event(THP_SPLIT);
2030
2031         BUG_ON(PageCompound(page));
2032 out_unlock:
2033         anon_vma_unlock_write(anon_vma);
2034         put_anon_vma(anon_vma);
2035 out:
2036         return ret;
2037 }
2038
2039 #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
2040
2041 int hugepage_madvise(struct vm_area_struct *vma,
2042                      unsigned long *vm_flags, int advice)
2043 {
2044         switch (advice) {
2045         case MADV_HUGEPAGE:
2046 #ifdef CONFIG_S390
2047                 /*
2048                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
2049                  * can't handle this properly after s390_enable_sie, so we simply
2050                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
2051                  */
2052                 if (mm_has_pgste(vma->vm_mm))
2053                         return 0;
2054 #endif
2055                 /*
2056                  * Be somewhat over-protective like KSM for now!
2057                  */
2058                 if (*vm_flags & VM_NO_THP)
2059                         return -EINVAL;
2060                 *vm_flags &= ~VM_NOHUGEPAGE;
2061                 *vm_flags |= VM_HUGEPAGE;
2062                 /*
2063                  * If the vma become good for khugepaged to scan,
2064                  * register it here without waiting a page fault that
2065                  * may not happen any time soon.
2066                  */
2067                 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
2068                         return -ENOMEM;
2069                 break;
2070         case MADV_NOHUGEPAGE:
2071                 /*
2072                  * Be somewhat over-protective like KSM for now!
2073                  */
2074                 if (*vm_flags & VM_NO_THP)
2075                         return -EINVAL;
2076                 *vm_flags &= ~VM_HUGEPAGE;
2077                 *vm_flags |= VM_NOHUGEPAGE;
2078                 /*
2079                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
2080                  * this vma even if we leave the mm registered in khugepaged if
2081                  * it got registered before VM_NOHUGEPAGE was set.
2082                  */
2083                 break;
2084         }
2085
2086         return 0;
2087 }
2088
2089 static int __init khugepaged_slab_init(void)
2090 {
2091         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
2092                                           sizeof(struct mm_slot),
2093                                           __alignof__(struct mm_slot), 0, NULL);
2094         if (!mm_slot_cache)
2095                 return -ENOMEM;
2096
2097         return 0;
2098 }
2099
2100 static void __init khugepaged_slab_exit(void)
2101 {
2102         kmem_cache_destroy(mm_slot_cache);
2103 }
2104
2105 static inline struct mm_slot *alloc_mm_slot(void)
2106 {
2107         if (!mm_slot_cache)     /* initialization failed */
2108                 return NULL;
2109         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
2110 }
2111
2112 static inline void free_mm_slot(struct mm_slot *mm_slot)
2113 {
2114         kmem_cache_free(mm_slot_cache, mm_slot);
2115 }
2116
2117 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
2118 {
2119         struct mm_slot *mm_slot;
2120
2121         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
2122                 if (mm == mm_slot->mm)
2123                         return mm_slot;
2124
2125         return NULL;
2126 }
2127
2128 static void insert_to_mm_slots_hash(struct mm_struct *mm,
2129                                     struct mm_slot *mm_slot)
2130 {
2131         mm_slot->mm = mm;
2132         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
2133 }
2134
2135 static inline int khugepaged_test_exit(struct mm_struct *mm)
2136 {
2137         return atomic_read(&mm->mm_users) == 0 || !mmget_still_valid(mm);
2138 }
2139
2140 int __khugepaged_enter(struct mm_struct *mm)
2141 {
2142         struct mm_slot *mm_slot;
2143         int wakeup;
2144
2145         mm_slot = alloc_mm_slot();
2146         if (!mm_slot)
2147                 return -ENOMEM;
2148
2149         /* __khugepaged_exit() must not run from under us */
2150         VM_BUG_ON_MM(atomic_read(&mm->mm_users) == 0, mm);
2151         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2152                 free_mm_slot(mm_slot);
2153                 return 0;
2154         }
2155
2156         spin_lock(&khugepaged_mm_lock);
2157         insert_to_mm_slots_hash(mm, mm_slot);
2158         /*
2159          * Insert just behind the scanning cursor, to let the area settle
2160          * down a little.
2161          */
2162         wakeup = list_empty(&khugepaged_scan.mm_head);
2163         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2164         spin_unlock(&khugepaged_mm_lock);
2165
2166         atomic_inc(&mm->mm_count);
2167         if (wakeup)
2168                 wake_up_interruptible(&khugepaged_wait);
2169
2170         return 0;
2171 }
2172
2173 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
2174                                unsigned long vm_flags)
2175 {
2176         unsigned long hstart, hend;
2177         if (!vma->anon_vma)
2178                 /*
2179                  * Not yet faulted in so we will register later in the
2180                  * page fault if needed.
2181                  */
2182                 return 0;
2183         if (vma->vm_ops || (vm_flags & VM_NO_THP))
2184                 /* khugepaged not yet working on file or special mappings */
2185                 return 0;
2186         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2187         hend = vma->vm_end & HPAGE_PMD_MASK;
2188         if (hstart < hend)
2189                 return khugepaged_enter(vma, vm_flags);
2190         return 0;
2191 }
2192
2193 void __khugepaged_exit(struct mm_struct *mm)
2194 {
2195         struct mm_slot *mm_slot;
2196         int free = 0;
2197
2198         spin_lock(&khugepaged_mm_lock);
2199         mm_slot = get_mm_slot(mm);
2200         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
2201                 hash_del(&mm_slot->hash);
2202                 list_del(&mm_slot->mm_node);
2203                 free = 1;
2204         }
2205         spin_unlock(&khugepaged_mm_lock);
2206
2207         if (free) {
2208                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2209                 free_mm_slot(mm_slot);
2210                 mmdrop(mm);
2211         } else if (mm_slot) {
2212                 /*
2213                  * This is required to serialize against
2214                  * khugepaged_test_exit() (which is guaranteed to run
2215                  * under mmap sem read mode). Stop here (after we
2216                  * return all pagetables will be destroyed) until
2217                  * khugepaged has finished working on the pagetables
2218                  * under the mmap_sem.
2219                  */
2220                 down_write(&mm->mmap_sem);
2221                 up_write(&mm->mmap_sem);
2222         }
2223 }
2224
2225 static void release_pte_page(struct page *page)
2226 {
2227         /* 0 stands for page_is_file_cache(page) == false */
2228         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2229         unlock_page(page);
2230         putback_lru_page(page);
2231 }
2232
2233 static void release_pte_pages(pte_t *pte, pte_t *_pte)
2234 {
2235         while (--_pte >= pte) {
2236                 pte_t pteval = *_pte;
2237                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
2238                         release_pte_page(pte_page(pteval));
2239         }
2240 }
2241
2242 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2243                                         unsigned long address,
2244                                         pte_t *pte)
2245 {
2246         struct page *page;
2247         pte_t *_pte;
2248         int none_or_zero = 0;
2249         bool referenced = false, writable = false;
2250         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2251              _pte++, address += PAGE_SIZE) {
2252                 pte_t pteval = *_pte;
2253                 if (pte_none(pteval) || (pte_present(pteval) &&
2254                                 is_zero_pfn(pte_pfn(pteval)))) {
2255                         if (!userfaultfd_armed(vma) &&
2256                             ++none_or_zero <= khugepaged_max_ptes_none)
2257                                 continue;
2258                         else
2259                                 goto out;
2260                 }
2261                 if (!pte_present(pteval))
2262                         goto out;
2263                 page = vm_normal_page(vma, address, pteval);
2264                 if (unlikely(!page))
2265                         goto out;
2266
2267                 VM_BUG_ON_PAGE(PageCompound(page), page);
2268                 VM_BUG_ON_PAGE(!PageAnon(page), page);
2269                 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2270
2271                 /*
2272                  * We can do it before isolate_lru_page because the
2273                  * page can't be freed from under us. NOTE: PG_lock
2274                  * is needed to serialize against split_huge_page
2275                  * when invoked from the VM.
2276                  */
2277                 if (!trylock_page(page))
2278                         goto out;
2279
2280                 /*
2281                  * cannot use mapcount: can't collapse if there's a gup pin.
2282                  * The page must only be referenced by the scanned process
2283                  * and page swap cache.
2284                  */
2285                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2286                         unlock_page(page);
2287                         goto out;
2288                 }
2289                 if (pte_write(pteval)) {
2290                         writable = true;
2291                 } else {
2292                         if (PageSwapCache(page) && !reuse_swap_page(page)) {
2293                                 unlock_page(page);
2294                                 goto out;
2295                         }
2296                         /*
2297                          * Page is not in the swap cache. It can be collapsed
2298                          * into a THP.
2299                          */
2300                 }
2301
2302                 /*
2303                  * Isolate the page to avoid collapsing an hugepage
2304                  * currently in use by the VM.
2305                  */
2306                 if (isolate_lru_page(page)) {
2307                         unlock_page(page);
2308                         goto out;
2309                 }
2310                 /* 0 stands for page_is_file_cache(page) == false */
2311                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2312                 VM_BUG_ON_PAGE(!PageLocked(page), page);
2313                 VM_BUG_ON_PAGE(PageLRU(page), page);
2314
2315                 /* If there is no mapped pte young don't collapse the page */
2316                 if (pte_young(pteval) ||
2317                     page_is_young(page) || PageReferenced(page) ||
2318                     mmu_notifier_test_young(vma->vm_mm, address))
2319                         referenced = true;
2320         }
2321         if (likely(referenced && writable))
2322                 return 1;
2323 out:
2324         release_pte_pages(pte, _pte);
2325         return 0;
2326 }
2327
2328 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2329                                       struct vm_area_struct *vma,
2330                                       unsigned long address,
2331                                       spinlock_t *ptl)
2332 {
2333         pte_t *_pte;
2334         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2335                 pte_t pteval = *_pte;
2336                 struct page *src_page;
2337
2338                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2339                         clear_user_highpage(page, address);
2340                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2341                         if (is_zero_pfn(pte_pfn(pteval))) {
2342                                 /*
2343                                  * ptl mostly unnecessary.
2344                                  */
2345                                 spin_lock(ptl);
2346                                 /*
2347                                  * paravirt calls inside pte_clear here are
2348                                  * superfluous.
2349                                  */
2350                                 pte_clear(vma->vm_mm, address, _pte);
2351                                 spin_unlock(ptl);
2352                         }
2353                 } else {
2354                         src_page = pte_page(pteval);
2355                         copy_user_highpage(page, src_page, address, vma);
2356                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
2357                         release_pte_page(src_page);
2358                         /*
2359                          * ptl mostly unnecessary, but preempt has to
2360                          * be disabled to update the per-cpu stats
2361                          * inside page_remove_rmap().
2362                          */
2363                         spin_lock(ptl);
2364                         /*
2365                          * paravirt calls inside pte_clear here are
2366                          * superfluous.
2367                          */
2368                         pte_clear(vma->vm_mm, address, _pte);
2369                         page_remove_rmap(src_page);
2370                         spin_unlock(ptl);
2371                         free_page_and_swap_cache(src_page);
2372                 }
2373
2374                 address += PAGE_SIZE;
2375                 page++;
2376         }
2377 }
2378
2379 static void khugepaged_alloc_sleep(void)
2380 {
2381         DEFINE_WAIT(wait);
2382
2383         add_wait_queue(&khugepaged_wait, &wait);
2384         freezable_schedule_timeout_interruptible(
2385                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2386         remove_wait_queue(&khugepaged_wait, &wait);
2387 }
2388
2389 static int khugepaged_node_load[MAX_NUMNODES];
2390
2391 static bool khugepaged_scan_abort(int nid)
2392 {
2393         int i;
2394
2395         /*
2396          * If zone_reclaim_mode is disabled, then no extra effort is made to
2397          * allocate memory locally.
2398          */
2399         if (!zone_reclaim_mode)
2400                 return false;
2401
2402         /* If there is a count for this node already, it must be acceptable */
2403         if (khugepaged_node_load[nid])
2404                 return false;
2405
2406         for (i = 0; i < MAX_NUMNODES; i++) {
2407                 if (!khugepaged_node_load[i])
2408                         continue;
2409                 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2410                         return true;
2411         }
2412         return false;
2413 }
2414
2415 #ifdef CONFIG_NUMA
2416 static int khugepaged_find_target_node(void)
2417 {
2418         static int last_khugepaged_target_node = NUMA_NO_NODE;
2419         int nid, target_node = 0, max_value = 0;
2420
2421         /* find first node with max normal pages hit */
2422         for (nid = 0; nid < MAX_NUMNODES; nid++)
2423                 if (khugepaged_node_load[nid] > max_value) {
2424                         max_value = khugepaged_node_load[nid];
2425                         target_node = nid;
2426                 }
2427
2428         /* do some balance if several nodes have the same hit record */
2429         if (target_node <= last_khugepaged_target_node)
2430                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2431                                 nid++)
2432                         if (max_value == khugepaged_node_load[nid]) {
2433                                 target_node = nid;
2434                                 break;
2435                         }
2436
2437         last_khugepaged_target_node = target_node;
2438         return target_node;
2439 }
2440
2441 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2442 {
2443         if (IS_ERR(*hpage)) {
2444                 if (!*wait)
2445                         return false;
2446
2447                 *wait = false;
2448                 *hpage = NULL;
2449                 khugepaged_alloc_sleep();
2450         } else if (*hpage) {
2451                 put_page(*hpage);
2452                 *hpage = NULL;
2453         }
2454
2455         return true;
2456 }
2457
2458 static struct page *
2459 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2460                        unsigned long address, int node)
2461 {
2462         VM_BUG_ON_PAGE(*hpage, *hpage);
2463
2464         /*
2465          * Before allocating the hugepage, release the mmap_sem read lock.
2466          * The allocation can take potentially a long time if it involves
2467          * sync compaction, and we do not need to hold the mmap_sem during
2468          * that. We will recheck the vma after taking it again in write mode.
2469          */
2470         up_read(&mm->mmap_sem);
2471
2472         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
2473         if (unlikely(!*hpage)) {
2474                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2475                 *hpage = ERR_PTR(-ENOMEM);
2476                 return NULL;
2477         }
2478
2479         count_vm_event(THP_COLLAPSE_ALLOC);
2480         return *hpage;
2481 }
2482 #else
2483 static int khugepaged_find_target_node(void)
2484 {
2485         return 0;
2486 }
2487
2488 static inline struct page *alloc_hugepage(int defrag)
2489 {
2490         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2491                            HPAGE_PMD_ORDER);
2492 }
2493
2494 static struct page *khugepaged_alloc_hugepage(bool *wait)
2495 {
2496         struct page *hpage;
2497
2498         do {
2499                 hpage = alloc_hugepage(khugepaged_defrag());
2500                 if (!hpage) {
2501                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2502                         if (!*wait)
2503                                 return NULL;
2504
2505                         *wait = false;
2506                         khugepaged_alloc_sleep();
2507                 } else
2508                         count_vm_event(THP_COLLAPSE_ALLOC);
2509         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2510
2511         return hpage;
2512 }
2513
2514 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2515 {
2516         if (!*hpage)
2517                 *hpage = khugepaged_alloc_hugepage(wait);
2518
2519         if (unlikely(!*hpage))
2520                 return false;
2521
2522         return true;
2523 }
2524
2525 static struct page *
2526 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2527                        unsigned long address, int node)
2528 {
2529         up_read(&mm->mmap_sem);
2530         VM_BUG_ON(!*hpage);
2531
2532         return  *hpage;
2533 }
2534 #endif
2535
2536 static bool hugepage_vma_check(struct vm_area_struct *vma)
2537 {
2538         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2539             (vma->vm_flags & VM_NOHUGEPAGE))
2540                 return false;
2541
2542         if (!vma->anon_vma || vma->vm_ops)
2543                 return false;
2544         if (is_vma_temporary_stack(vma))
2545                 return false;
2546         return !(vma->vm_flags & VM_NO_THP);
2547 }
2548
2549 static void collapse_huge_page(struct mm_struct *mm,
2550                                    unsigned long address,
2551                                    struct page **hpage,
2552                                    struct vm_area_struct *vma,
2553                                    int node)
2554 {
2555         pmd_t *pmd, _pmd;
2556         pte_t *pte;
2557         pgtable_t pgtable;
2558         struct page *new_page;
2559         spinlock_t *pmd_ptl, *pte_ptl;
2560         int isolated;
2561         unsigned long hstart, hend;
2562         struct mem_cgroup *memcg;
2563         unsigned long mmun_start;       /* For mmu_notifiers */
2564         unsigned long mmun_end;         /* For mmu_notifiers */
2565         gfp_t gfp;
2566
2567         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2568
2569         /* Only allocate from the target node */
2570         gfp = alloc_hugepage_gfpmask(khugepaged_defrag(), __GFP_OTHER_NODE) |
2571                 __GFP_THISNODE;
2572
2573         /* release the mmap_sem read lock. */
2574         new_page = khugepaged_alloc_page(hpage, gfp, mm, address, node);
2575         if (!new_page)
2576                 return;
2577
2578         if (unlikely(mem_cgroup_try_charge(new_page, mm,
2579                                            gfp, &memcg)))
2580                 return;
2581
2582         /*
2583          * Prevent all access to pagetables with the exception of
2584          * gup_fast later hanlded by the ptep_clear_flush and the VM
2585          * handled by the anon_vma lock + PG_lock.
2586          */
2587         down_write(&mm->mmap_sem);
2588         if (unlikely(khugepaged_test_exit(mm)))
2589                 goto out;
2590
2591         vma = find_vma(mm, address);
2592         if (!vma)
2593                 goto out;
2594         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2595         hend = vma->vm_end & HPAGE_PMD_MASK;
2596         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2597                 goto out;
2598         if (!hugepage_vma_check(vma))
2599                 goto out;
2600         pmd = mm_find_pmd(mm, address);
2601         if (!pmd)
2602                 goto out;
2603
2604         anon_vma_lock_write(vma->anon_vma);
2605
2606         pte = pte_offset_map(pmd, address);
2607         pte_ptl = pte_lockptr(mm, pmd);
2608
2609         mmun_start = address;
2610         mmun_end   = address + HPAGE_PMD_SIZE;
2611         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2612         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2613         /*
2614          * After this gup_fast can't run anymore. This also removes
2615          * any huge TLB entry from the CPU so we won't allow
2616          * huge and small TLB entries for the same virtual address
2617          * to avoid the risk of CPU bugs in that area.
2618          */
2619         _pmd = pmdp_collapse_flush(vma, address, pmd);
2620         spin_unlock(pmd_ptl);
2621         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2622
2623         spin_lock(pte_ptl);
2624         isolated = __collapse_huge_page_isolate(vma, address, pte);
2625         spin_unlock(pte_ptl);
2626
2627         if (unlikely(!isolated)) {
2628                 pte_unmap(pte);
2629                 spin_lock(pmd_ptl);
2630                 BUG_ON(!pmd_none(*pmd));
2631                 /*
2632                  * We can only use set_pmd_at when establishing
2633                  * hugepmds and never for establishing regular pmds that
2634                  * points to regular pagetables. Use pmd_populate for that
2635                  */
2636                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2637                 spin_unlock(pmd_ptl);
2638                 anon_vma_unlock_write(vma->anon_vma);
2639                 goto out;
2640         }
2641
2642         /*
2643          * All pages are isolated and locked so anon_vma rmap
2644          * can't run anymore.
2645          */
2646         anon_vma_unlock_write(vma->anon_vma);
2647
2648         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2649         pte_unmap(pte);
2650         __SetPageUptodate(new_page);
2651         pgtable = pmd_pgtable(_pmd);
2652
2653         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2654         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2655
2656         /*
2657          * spin_lock() below is not the equivalent of smp_wmb(), so
2658          * this is needed to avoid the copy_huge_page writes to become
2659          * visible after the set_pmd_at() write.
2660          */
2661         smp_wmb();
2662
2663         spin_lock(pmd_ptl);
2664         BUG_ON(!pmd_none(*pmd));
2665         page_add_new_anon_rmap(new_page, vma, address);
2666         mem_cgroup_commit_charge(new_page, memcg, false);
2667         lru_cache_add_active_or_unevictable(new_page, vma);
2668         pgtable_trans_huge_deposit(mm, pmd, pgtable);
2669         set_pmd_at(mm, address, pmd, _pmd);
2670         update_mmu_cache_pmd(vma, address, pmd);
2671         spin_unlock(pmd_ptl);
2672
2673         *hpage = NULL;
2674
2675         khugepaged_pages_collapsed++;
2676 out_up_write:
2677         up_write(&mm->mmap_sem);
2678         return;
2679
2680 out:
2681         mem_cgroup_cancel_charge(new_page, memcg);
2682         goto out_up_write;
2683 }
2684
2685 static int khugepaged_scan_pmd(struct mm_struct *mm,
2686                                struct vm_area_struct *vma,
2687                                unsigned long address,
2688                                struct page **hpage)
2689 {
2690         pmd_t *pmd;
2691         pte_t *pte, *_pte;
2692         int ret = 0, none_or_zero = 0;
2693         struct page *page;
2694         unsigned long _address;
2695         spinlock_t *ptl;
2696         int node = NUMA_NO_NODE;
2697         bool writable = false, referenced = false;
2698
2699         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2700
2701         pmd = mm_find_pmd(mm, address);
2702         if (!pmd)
2703                 goto out;
2704
2705         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2706         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2707         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2708              _pte++, _address += PAGE_SIZE) {
2709                 pte_t pteval = *_pte;
2710                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2711                         if (!userfaultfd_armed(vma) &&
2712                             ++none_or_zero <= khugepaged_max_ptes_none)
2713                                 continue;
2714                         else
2715                                 goto out_unmap;
2716                 }
2717                 if (!pte_present(pteval))
2718                         goto out_unmap;
2719                 if (pte_write(pteval))
2720                         writable = true;
2721
2722                 page = vm_normal_page(vma, _address, pteval);
2723                 if (unlikely(!page))
2724                         goto out_unmap;
2725                 /*
2726                  * Record which node the original page is from and save this
2727                  * information to khugepaged_node_load[].
2728                  * Khupaged will allocate hugepage from the node has the max
2729                  * hit record.
2730                  */
2731                 node = page_to_nid(page);
2732                 if (khugepaged_scan_abort(node))
2733                         goto out_unmap;
2734                 khugepaged_node_load[node]++;
2735                 VM_BUG_ON_PAGE(PageCompound(page), page);
2736                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2737                         goto out_unmap;
2738                 /*
2739                  * cannot use mapcount: can't collapse if there's a gup pin.
2740                  * The page must only be referenced by the scanned process
2741                  * and page swap cache.
2742                  */
2743                 if (page_count(page) != 1 + !!PageSwapCache(page))
2744                         goto out_unmap;
2745                 if (pte_young(pteval) ||
2746                     page_is_young(page) || PageReferenced(page) ||
2747                     mmu_notifier_test_young(vma->vm_mm, address))
2748                         referenced = true;
2749         }
2750         if (referenced && writable)
2751                 ret = 1;
2752 out_unmap:
2753         pte_unmap_unlock(pte, ptl);
2754         if (ret) {
2755                 node = khugepaged_find_target_node();
2756                 /* collapse_huge_page will return with the mmap_sem released */
2757                 collapse_huge_page(mm, address, hpage, vma, node);
2758         }
2759 out:
2760         return ret;
2761 }
2762
2763 static void collect_mm_slot(struct mm_slot *mm_slot)
2764 {
2765         struct mm_struct *mm = mm_slot->mm;
2766
2767         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2768
2769         if (khugepaged_test_exit(mm)) {
2770                 /* free mm_slot */
2771                 hash_del(&mm_slot->hash);
2772                 list_del(&mm_slot->mm_node);
2773
2774                 /*
2775                  * Not strictly needed because the mm exited already.
2776                  *
2777                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2778                  */
2779
2780                 /* khugepaged_mm_lock actually not necessary for the below */
2781                 free_mm_slot(mm_slot);
2782                 mmdrop(mm);
2783         }
2784 }
2785
2786 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2787                                             struct page **hpage)
2788         __releases(&khugepaged_mm_lock)
2789         __acquires(&khugepaged_mm_lock)
2790 {
2791         struct mm_slot *mm_slot;
2792         struct mm_struct *mm;
2793         struct vm_area_struct *vma;
2794         int progress = 0;
2795
2796         VM_BUG_ON(!pages);
2797         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2798
2799         if (khugepaged_scan.mm_slot)
2800                 mm_slot = khugepaged_scan.mm_slot;
2801         else {
2802                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2803                                      struct mm_slot, mm_node);
2804                 khugepaged_scan.address = 0;
2805                 khugepaged_scan.mm_slot = mm_slot;
2806         }
2807         spin_unlock(&khugepaged_mm_lock);
2808
2809         mm = mm_slot->mm;
2810         down_read(&mm->mmap_sem);
2811         if (unlikely(khugepaged_test_exit(mm)))
2812                 vma = NULL;
2813         else
2814                 vma = find_vma(mm, khugepaged_scan.address);
2815
2816         progress++;
2817         for (; vma; vma = vma->vm_next) {
2818                 unsigned long hstart, hend;
2819
2820                 cond_resched();
2821                 if (unlikely(khugepaged_test_exit(mm))) {
2822                         progress++;
2823                         break;
2824                 }
2825                 if (!hugepage_vma_check(vma)) {
2826 skip:
2827                         progress++;
2828                         continue;
2829                 }
2830                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2831                 hend = vma->vm_end & HPAGE_PMD_MASK;
2832                 if (hstart >= hend)
2833                         goto skip;
2834                 if (khugepaged_scan.address > hend)
2835                         goto skip;
2836                 if (khugepaged_scan.address < hstart)
2837                         khugepaged_scan.address = hstart;
2838                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2839
2840                 while (khugepaged_scan.address < hend) {
2841                         int ret;
2842                         cond_resched();
2843                         if (unlikely(khugepaged_test_exit(mm)))
2844                                 goto breakouterloop;
2845
2846                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2847                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2848                                   hend);
2849                         ret = khugepaged_scan_pmd(mm, vma,
2850                                                   khugepaged_scan.address,
2851                                                   hpage);
2852                         /* move to next address */
2853                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2854                         progress += HPAGE_PMD_NR;
2855                         if (ret)
2856                                 /* we released mmap_sem so break loop */
2857                                 goto breakouterloop_mmap_sem;
2858                         if (progress >= pages)
2859                                 goto breakouterloop;
2860                 }
2861         }
2862 breakouterloop:
2863         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2864 breakouterloop_mmap_sem:
2865
2866         spin_lock(&khugepaged_mm_lock);
2867         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2868         /*
2869          * Release the current mm_slot if this mm is about to die, or
2870          * if we scanned all vmas of this mm.
2871          */
2872         if (khugepaged_test_exit(mm) || !vma) {
2873                 /*
2874                  * Make sure that if mm_users is reaching zero while
2875                  * khugepaged runs here, khugepaged_exit will find
2876                  * mm_slot not pointing to the exiting mm.
2877                  */
2878                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2879                         khugepaged_scan.mm_slot = list_entry(
2880                                 mm_slot->mm_node.next,
2881                                 struct mm_slot, mm_node);
2882                         khugepaged_scan.address = 0;
2883                 } else {
2884                         khugepaged_scan.mm_slot = NULL;
2885                         khugepaged_full_scans++;
2886                 }
2887
2888                 collect_mm_slot(mm_slot);
2889         }
2890
2891         return progress;
2892 }
2893
2894 static int khugepaged_has_work(void)
2895 {
2896         return !list_empty(&khugepaged_scan.mm_head) &&
2897                 khugepaged_enabled();
2898 }
2899
2900 static int khugepaged_wait_event(void)
2901 {
2902         return !list_empty(&khugepaged_scan.mm_head) ||
2903                 kthread_should_stop();
2904 }
2905
2906 static void khugepaged_do_scan(void)
2907 {
2908         struct page *hpage = NULL;
2909         unsigned int progress = 0, pass_through_head = 0;
2910         unsigned int pages = khugepaged_pages_to_scan;
2911         bool wait = true;
2912
2913         barrier(); /* write khugepaged_pages_to_scan to local stack */
2914
2915         while (progress < pages) {
2916                 if (!khugepaged_prealloc_page(&hpage, &wait))
2917                         break;
2918
2919                 cond_resched();
2920
2921                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2922                         break;
2923
2924                 spin_lock(&khugepaged_mm_lock);
2925                 if (!khugepaged_scan.mm_slot)
2926                         pass_through_head++;
2927                 if (khugepaged_has_work() &&
2928                     pass_through_head < 2)
2929                         progress += khugepaged_scan_mm_slot(pages - progress,
2930                                                             &hpage);
2931                 else
2932                         progress = pages;
2933                 spin_unlock(&khugepaged_mm_lock);
2934         }
2935
2936         if (!IS_ERR_OR_NULL(hpage))
2937                 put_page(hpage);
2938 }
2939
2940 static void khugepaged_wait_work(void)
2941 {
2942         if (khugepaged_has_work()) {
2943                 if (!khugepaged_scan_sleep_millisecs)
2944                         return;
2945
2946                 wait_event_freezable_timeout(khugepaged_wait,
2947                                              kthread_should_stop(),
2948                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2949                 return;
2950         }
2951
2952         if (khugepaged_enabled())
2953                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2954 }
2955
2956 static int khugepaged(void *none)
2957 {
2958         struct mm_slot *mm_slot;
2959
2960         set_freezable();
2961         set_user_nice(current, MAX_NICE);
2962
2963         while (!kthread_should_stop()) {
2964                 khugepaged_do_scan();
2965                 khugepaged_wait_work();
2966         }
2967
2968         spin_lock(&khugepaged_mm_lock);
2969         mm_slot = khugepaged_scan.mm_slot;
2970         khugepaged_scan.mm_slot = NULL;
2971         if (mm_slot)
2972                 collect_mm_slot(mm_slot);
2973         spin_unlock(&khugepaged_mm_lock);
2974         return 0;
2975 }
2976
2977 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2978                 unsigned long haddr, pmd_t *pmd)
2979 {
2980         struct mm_struct *mm = vma->vm_mm;
2981         pgtable_t pgtable;
2982         pmd_t _pmd;
2983         int i;
2984
2985         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2986         /* leave pmd empty until pte is filled */
2987
2988         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2989         pmd_populate(mm, &_pmd, pgtable);
2990
2991         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2992                 pte_t *pte, entry;
2993                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2994                 entry = pte_mkspecial(entry);
2995                 pte = pte_offset_map(&_pmd, haddr);
2996                 VM_BUG_ON(!pte_none(*pte));
2997                 set_pte_at(mm, haddr, pte, entry);
2998                 pte_unmap(pte);
2999         }
3000         smp_wmb(); /* make pte visible before pmd */
3001         pmd_populate(mm, pmd, pgtable);
3002         put_huge_zero_page();
3003 }
3004
3005 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
3006                 pmd_t *pmd)
3007 {
3008         spinlock_t *ptl;
3009         struct page *page = NULL;
3010         struct mm_struct *mm = vma->vm_mm;
3011         unsigned long haddr = address & HPAGE_PMD_MASK;
3012         unsigned long mmun_start;       /* For mmu_notifiers */
3013         unsigned long mmun_end;         /* For mmu_notifiers */
3014
3015         BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
3016
3017         mmun_start = haddr;
3018         mmun_end   = haddr + HPAGE_PMD_SIZE;
3019 again:
3020         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3021         ptl = pmd_lock(mm, pmd);
3022         if (unlikely(!pmd_trans_huge(*pmd)))
3023                 goto unlock;
3024         if (vma_is_dax(vma)) {
3025                 pmd_t _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
3026                 if (is_huge_zero_pmd(_pmd))
3027                         put_huge_zero_page();
3028         } else if (is_huge_zero_pmd(*pmd)) {
3029                 __split_huge_zero_page_pmd(vma, haddr, pmd);
3030         } else {
3031                 page = pmd_page(*pmd);
3032                 VM_BUG_ON_PAGE(!page_count(page), page);
3033                 get_page(page);
3034         }
3035  unlock:
3036         spin_unlock(ptl);
3037         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3038
3039         if (!page)
3040                 return;
3041
3042         split_huge_page(page);
3043         put_page(page);
3044
3045         /*
3046          * We don't always have down_write of mmap_sem here: a racing
3047          * do_huge_pmd_wp_page() might have copied-on-write to another
3048          * huge page before our split_huge_page() got the anon_vma lock.
3049          */
3050         if (unlikely(pmd_trans_huge(*pmd)))
3051                 goto again;
3052 }
3053
3054 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
3055                 pmd_t *pmd)
3056 {
3057         struct vm_area_struct *vma;
3058
3059         vma = find_vma(mm, address);
3060         BUG_ON(vma == NULL);
3061         split_huge_page_pmd(vma, address, pmd);
3062 }
3063
3064 static void split_huge_page_address(struct mm_struct *mm,
3065                                     unsigned long address)
3066 {
3067         pgd_t *pgd;
3068         pud_t *pud;
3069         pmd_t *pmd;
3070
3071         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
3072
3073         pgd = pgd_offset(mm, address);
3074         if (!pgd_present(*pgd))
3075                 return;
3076
3077         pud = pud_offset(pgd, address);
3078         if (!pud_present(*pud))
3079                 return;
3080
3081         pmd = pmd_offset(pud, address);
3082         if (!pmd_present(*pmd))
3083                 return;
3084         /*
3085          * Caller holds the mmap_sem write mode, so a huge pmd cannot
3086          * materialize from under us.
3087          */
3088         split_huge_page_pmd_mm(mm, address, pmd);
3089 }
3090
3091 void vma_adjust_trans_huge(struct vm_area_struct *vma,
3092                              unsigned long start,
3093                              unsigned long end,
3094                              long adjust_next)
3095 {
3096         /*
3097          * If the new start address isn't hpage aligned and it could
3098          * previously contain an hugepage: check if we need to split
3099          * an huge pmd.
3100          */
3101         if (start & ~HPAGE_PMD_MASK &&
3102             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
3103             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3104                 split_huge_page_address(vma->vm_mm, start);
3105
3106         /*
3107          * If the new end address isn't hpage aligned and it could
3108          * previously contain an hugepage: check if we need to split
3109          * an huge pmd.
3110          */
3111         if (end & ~HPAGE_PMD_MASK &&
3112             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
3113             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3114                 split_huge_page_address(vma->vm_mm, end);
3115
3116         /*
3117          * If we're also updating the vma->vm_next->vm_start, if the new
3118          * vm_next->vm_start isn't page aligned and it could previously
3119          * contain an hugepage: check if we need to split an huge pmd.
3120          */
3121         if (adjust_next > 0) {
3122                 struct vm_area_struct *next = vma->vm_next;
3123                 unsigned long nstart = next->vm_start;
3124                 nstart += adjust_next << PAGE_SHIFT;
3125                 if (nstart & ~HPAGE_PMD_MASK &&
3126                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
3127                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
3128                         split_huge_page_address(next->vm_mm, nstart);
3129         }
3130 }