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 #include <linux/mm.h>
   9 #include <linux/sched.h>
  10 #include <linux/highmem.h>
  11 #include <linux/hugetlb.h>
  12 #include <linux/mmu_notifier.h>
  13 #include <linux/rmap.h>
  14 #include <linux/swap.h>
  15 #include <linux/mm_inline.h>
  16 #include <linux/kthread.h>
  17 #include <linux/khugepaged.h>
  18 #include <linux/freezer.h>
  19 #include <linux/mman.h>
  20 #include <asm/tlb.h>
  21 #include <asm/pgalloc.h>
  22 #include "internal.h"
  23
  24 /*
  25  * By default transparent hugepage support is enabled for all mappings
  26  * and khugepaged scans all mappings. Defrag is only invoked by
  27  * khugepaged hugepage allocations and by page faults inside
  28  * MADV_HUGEPAGE regions to avoid the risk of slowing down short lived
  29  * allocations.
  30  */
  31 unsigned long transparent_hugepage_flags __read_mostly =
  32 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  33         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  34 #endif
  35 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  36         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  37 #endif
  38         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
  39         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
  40
  41 /* default scan 8*512 pte (or vmas) every 30 second */
  42 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
  43 static unsigned int khugepaged_pages_collapsed;
  44 static unsigned int khugepaged_full_scans;
  45 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  46 /* during fragmentation poll the hugepage allocator once every minute */
  47 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  48 static struct task_struct *khugepaged_thread __read_mostly;
  49 static DEFINE_MUTEX(khugepaged_mutex);
  50 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  51 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  52 /*
  53  * default collapse hugepages if there is at least one pte mapped like
  54  * it would have happened if the vma was large enough during page
  55  * fault.
  56  */
  57 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
  58
  59 static int khugepaged(void *none);
  60 static int mm_slots_hash_init(void);
  61 static int khugepaged_slab_init(void);
  62 static void khugepaged_slab_free(void);
  63
  64 #define MM_SLOTS_HASH_HEADS 1024
  65 static struct hlist_head *mm_slots_hash __read_mostly;
  66 static struct kmem_cache *mm_slot_cache __read_mostly;
  67
  68 /**
  69  * struct mm_slot - hash lookup from mm to mm_slot
  70  * @hash: hash collision list
  71  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  72  * @mm: the mm that this information is valid for
  73  */
  74 struct mm_slot {
  75         struct hlist_node hash;
  76         struct list_head mm_node;
  77         struct mm_struct *mm;
  78 };
  79
  80 /**
  81  * struct khugepaged_scan - cursor for scanning
  82  * @mm_head: the head of the mm list to scan
  83  * @mm_slot: the current mm_slot we are scanning
  84  * @address: the next address inside that to be scanned
  85  *
  86  * There is only the one khugepaged_scan instance of this cursor structure.
  87  */
  88 struct khugepaged_scan {
  89         struct list_head mm_head;
  90         struct mm_slot *mm_slot;
  91         unsigned long address;
  92 };
  93 static struct khugepaged_scan khugepaged_scan = {
  94         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
  95 };
  96
  97
  98 static int set_recommended_min_free_kbytes(void)
  99 {
 100         struct zone *zone;
 101         int nr_zones = 0;
 102         unsigned long recommended_min;
 103         extern int min_free_kbytes;
 104
 105         if (!khugepaged_enabled())
 106                 return 0;
 107
 108         for_each_populated_zone(zone)
 109                 nr_zones++;
 110
 111         /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
 112         recommended_min = pageblock_nr_pages * nr_zones * 2;
 113
 114         /*
 115          * Make sure that on average at least two pageblocks are almost free
 116          * of another type, one for a migratetype to fall back to and a
 117          * second to avoid subsequent fallbacks of other types There are 3
 118          * MIGRATE_TYPES we care about.
 119          */
 120         recommended_min += pageblock_nr_pages * nr_zones *
 121                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
 122
 123         /* don't ever allow to reserve more than 5% of the lowmem */
 124         recommended_min = min(recommended_min,
 125                               (unsigned long) nr_free_buffer_pages() / 20);
 126         recommended_min <<= (PAGE_SHIFT-10);
 127
 128         if (recommended_min > min_free_kbytes)
 129                 min_free_kbytes = recommended_min;
 130         setup_per_zone_wmarks();
 131         return 0;
 132 }
 133 late_initcall(set_recommended_min_free_kbytes);
 134
 135 static int start_khugepaged(void)
 136 {
 137         int err = 0;
 138         if (khugepaged_enabled()) {
 139                 if (!khugepaged_thread)
 140                         khugepaged_thread = kthread_run(khugepaged, NULL,
 141                                                         "khugepaged");
 142                 if (unlikely(IS_ERR(khugepaged_thread))) {
 143                         printk(KERN_ERR
 144                                "khugepaged: kthread_run(khugepaged) failed\n");
 145                         err = PTR_ERR(khugepaged_thread);
 146                         khugepaged_thread = NULL;
 147                 }
 148
 149                 if (!list_empty(&khugepaged_scan.mm_head))
 150                         wake_up_interruptible(&khugepaged_wait);
 151
 152                 set_recommended_min_free_kbytes();
 153         } else if (khugepaged_thread) {
 154                 kthread_stop(khugepaged_thread);
 155                 khugepaged_thread = NULL;
 156         }
 157
 158         return err;
 159 }
 160
 161 #ifdef CONFIG_SYSFS
 162
 163 static ssize_t double_flag_show(struct kobject *kobj,
 164                                 struct kobj_attribute *attr, char *buf,
 165                                 enum transparent_hugepage_flag enabled,
 166                                 enum transparent_hugepage_flag req_madv)
 167 {
 168         if (test_bit(enabled, &transparent_hugepage_flags)) {
 169                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
 170                 return sprintf(buf, "[always] madvise never\n");
 171         } else if (test_bit(req_madv, &transparent_hugepage_flags))
 172                 return sprintf(buf, "always [madvise] never\n");
 173         else
 174                 return sprintf(buf, "always madvise [never]\n");
 175 }
 176 static ssize_t double_flag_store(struct kobject *kobj,
 177                                  struct kobj_attribute *attr,
 178                                  const char *buf, size_t count,
 179                                  enum transparent_hugepage_flag enabled,
 180                                  enum transparent_hugepage_flag req_madv)
 181 {
 182         if (!memcmp("always", buf,
 183                     min(sizeof("always")-1, count))) {
 184                 set_bit(enabled, &transparent_hugepage_flags);
 185                 clear_bit(req_madv, &transparent_hugepage_flags);
 186         } else if (!memcmp("madvise", buf,
 187                            min(sizeof("madvise")-1, count))) {
 188                 clear_bit(enabled, &transparent_hugepage_flags);
 189                 set_bit(req_madv, &transparent_hugepage_flags);
 190         } else if (!memcmp("never", buf,
 191                            min(sizeof("never")-1, count))) {
 192                 clear_bit(enabled, &transparent_hugepage_flags);
 193                 clear_bit(req_madv, &transparent_hugepage_flags);
 194         } else
 195                 return -EINVAL;
 196
 197         return count;
 198 }
 199
 200 static ssize_t enabled_show(struct kobject *kobj,
 201                             struct kobj_attribute *attr, char *buf)
 202 {
 203         return double_flag_show(kobj, attr, buf,
 204                                 TRANSPARENT_HUGEPAGE_FLAG,
 205                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 206 }
 207 static ssize_t enabled_store(struct kobject *kobj,
 208                              struct kobj_attribute *attr,
 209                              const char *buf, size_t count)
 210 {
 211         ssize_t ret;
 212
 213         ret = double_flag_store(kobj, attr, buf, count,
 214                                 TRANSPARENT_HUGEPAGE_FLAG,
 215                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 216
 217         if (ret > 0) {
 218                 int err;
 219
 220                 mutex_lock(&khugepaged_mutex);
 221                 err = start_khugepaged();
 222                 mutex_unlock(&khugepaged_mutex);
 223
 224                 if (err)
 225                         ret = err;
 226         }
 227
 228         return ret;
 229 }
 230 static struct kobj_attribute enabled_attr =
 231         __ATTR(enabled, 0644, enabled_show, enabled_store);
 232
 233 static ssize_t single_flag_show(struct kobject *kobj,
 234                                 struct kobj_attribute *attr, char *buf,
 235                                 enum transparent_hugepage_flag flag)
 236 {
 237         return sprintf(buf, "%d\n",
 238                        !!test_bit(flag, &transparent_hugepage_flags));
 239 }
 240
 241 static ssize_t single_flag_store(struct kobject *kobj,
 242                                  struct kobj_attribute *attr,
 243                                  const char *buf, size_t count,
 244                                  enum transparent_hugepage_flag flag)
 245 {
 246         unsigned long value;
 247         int ret;
 248
 249         ret = kstrtoul(buf, 10, &value);
 250         if (ret < 0)
 251                 return ret;
 252         if (value > 1)
 253                 return -EINVAL;
 254
 255         if (value)
 256                 set_bit(flag, &transparent_hugepage_flags);
 257         else
 258                 clear_bit(flag, &transparent_hugepage_flags);
 259
 260         return count;
 261 }
 262
 263 /*
 264  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 265  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 266  * memory just to allocate one more hugepage.
 267  */
 268 static ssize_t defrag_show(struct kobject *kobj,
 269                            struct kobj_attribute *attr, char *buf)
 270 {
 271         return double_flag_show(kobj, attr, buf,
 272                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 273                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 274 }
 275 static ssize_t defrag_store(struct kobject *kobj,
 276                             struct kobj_attribute *attr,
 277                             const char *buf, size_t count)
 278 {
 279         return double_flag_store(kobj, attr, buf, count,
 280                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 281                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 282 }
 283 static struct kobj_attribute defrag_attr =
 284         __ATTR(defrag, 0644, defrag_show, defrag_store);
 285
 286 #ifdef CONFIG_DEBUG_VM
 287 static ssize_t debug_cow_show(struct kobject *kobj,
 288                                 struct kobj_attribute *attr, char *buf)
 289 {
 290         return single_flag_show(kobj, attr, buf,
 291                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 292 }
 293 static ssize_t debug_cow_store(struct kobject *kobj,
 294                                struct kobj_attribute *attr,
 295                                const char *buf, size_t count)
 296 {
 297         return single_flag_store(kobj, attr, buf, count,
 298                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 299 }
 300 static struct kobj_attribute debug_cow_attr =
 301         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 302 #endif /* CONFIG_DEBUG_VM */
 303
 304 static struct attribute *hugepage_attr[] = {
 305         &enabled_attr.attr,
 306         &defrag_attr.attr,
 307 #ifdef CONFIG_DEBUG_VM
 308         &debug_cow_attr.attr,
 309 #endif
 310         NULL,
 311 };
 312
 313 static struct attribute_group hugepage_attr_group = {
 314         .attrs = hugepage_attr,
 315 };
 316
 317 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 318                                          struct kobj_attribute *attr,
 319                                          char *buf)
 320 {
 321         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 322 }
 323
 324 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 325                                           struct kobj_attribute *attr,
 326                                           const char *buf, size_t count)
 327 {
 328         unsigned long msecs;
 329         int err;
 330
 331         err = strict_strtoul(buf, 10, &msecs);
 332         if (err || msecs > UINT_MAX)
 333                 return -EINVAL;
 334
 335         khugepaged_scan_sleep_millisecs = msecs;
 336         wake_up_interruptible(&khugepaged_wait);
 337
 338         return count;
 339 }
 340 static struct kobj_attribute scan_sleep_millisecs_attr =
 341         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 342                scan_sleep_millisecs_store);
 343
 344 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 345                                           struct kobj_attribute *attr,
 346                                           char *buf)
 347 {
 348         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 349 }
 350
 351 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 352                                            struct kobj_attribute *attr,
 353                                            const char *buf, size_t count)
 354 {
 355         unsigned long msecs;
 356         int err;
 357
 358         err = strict_strtoul(buf, 10, &msecs);
 359         if (err || msecs > UINT_MAX)
 360                 return -EINVAL;
 361
 362         khugepaged_alloc_sleep_millisecs = msecs;
 363         wake_up_interruptible(&khugepaged_wait);
 364
 365         return count;
 366 }
 367 static struct kobj_attribute alloc_sleep_millisecs_attr =
 368         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 369                alloc_sleep_millisecs_store);
 370
 371 static ssize_t pages_to_scan_show(struct kobject *kobj,
 372                                   struct kobj_attribute *attr,
 373                                   char *buf)
 374 {
 375         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 376 }
 377 static ssize_t pages_to_scan_store(struct kobject *kobj,
 378                                    struct kobj_attribute *attr,
 379                                    const char *buf, size_t count)
 380 {
 381         int err;
 382         unsigned long pages;
 383
 384         err = strict_strtoul(buf, 10, &pages);
 385         if (err || !pages || pages > UINT_MAX)
 386                 return -EINVAL;
 387
 388         khugepaged_pages_to_scan = pages;
 389
 390         return count;
 391 }
 392 static struct kobj_attribute pages_to_scan_attr =
 393         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 394                pages_to_scan_store);
 395
 396 static ssize_t pages_collapsed_show(struct kobject *kobj,
 397                                     struct kobj_attribute *attr,
 398                                     char *buf)
 399 {
 400         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 401 }
 402 static struct kobj_attribute pages_collapsed_attr =
 403         __ATTR_RO(pages_collapsed);
 404
 405 static ssize_t full_scans_show(struct kobject *kobj,
 406                                struct kobj_attribute *attr,
 407                                char *buf)
 408 {
 409         return sprintf(buf, "%u\n", khugepaged_full_scans);
 410 }
 411 static struct kobj_attribute full_scans_attr =
 412         __ATTR_RO(full_scans);
 413
 414 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 415                                       struct kobj_attribute *attr, char *buf)
 416 {
 417         return single_flag_show(kobj, attr, buf,
 418                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 419 }
 420 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 421                                        struct kobj_attribute *attr,
 422                                        const char *buf, size_t count)
 423 {
 424         return single_flag_store(kobj, attr, buf, count,
 425                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 426 }
 427 static struct kobj_attribute khugepaged_defrag_attr =
 428         __ATTR(defrag, 0644, khugepaged_defrag_show,
 429                khugepaged_defrag_store);
 430
 431 /*
 432  * max_ptes_none controls if khugepaged should collapse hugepages over
 433  * any unmapped ptes in turn potentially increasing the memory
 434  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 435  * reduce the available free memory in the system as it
 436  * runs. Increasing max_ptes_none will instead potentially reduce the
 437  * free memory in the system during the khugepaged scan.
 438  */
 439 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 440                                              struct kobj_attribute *attr,
 441                                              char *buf)
 442 {
 443         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 444 }
 445 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 446                                               struct kobj_attribute *attr,
 447                                               const char *buf, size_t count)
 448 {
 449         int err;
 450         unsigned long max_ptes_none;
 451
 452         err = strict_strtoul(buf, 10, &max_ptes_none);
 453         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 454                 return -EINVAL;
 455
 456         khugepaged_max_ptes_none = max_ptes_none;
 457
 458         return count;
 459 }
 460 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 461         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 462                khugepaged_max_ptes_none_store);
 463
 464 static struct attribute *khugepaged_attr[] = {
 465         &khugepaged_defrag_attr.attr,
 466         &khugepaged_max_ptes_none_attr.attr,
 467         &pages_to_scan_attr.attr,
 468         &pages_collapsed_attr.attr,
 469         &full_scans_attr.attr,
 470         &scan_sleep_millisecs_attr.attr,
 471         &alloc_sleep_millisecs_attr.attr,
 472         NULL,
 473 };
 474
 475 static struct attribute_group khugepaged_attr_group = {
 476         .attrs = khugepaged_attr,
 477         .name = "khugepaged",
 478 };
 479
 480 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 481 {
 482         int err;
 483
 484         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 485         if (unlikely(!*hugepage_kobj)) {
 486                 printk(KERN_ERR "hugepage: failed kobject create\n");
 487                 return -ENOMEM;
 488         }
 489
 490         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 491         if (err) {
 492                 printk(KERN_ERR "hugepage: failed register hugeage group\n");
 493                 goto delete_obj;
 494         }
 495
 496         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 497         if (err) {
 498                 printk(KERN_ERR "hugepage: failed register hugeage group\n");
 499                 goto remove_hp_group;
 500         }
 501
 502         return 0;
 503
 504 remove_hp_group:
 505         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 506 delete_obj:
 507         kobject_put(*hugepage_kobj);
 508         return err;
 509 }
 510
 511 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 512 {
 513         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 514         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 515         kobject_put(hugepage_kobj);
 516 }
 517 #else
 518 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 519 {
 520         return 0;
 521 }
 522
 523 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 524 {
 525 }
 526 #endif /* CONFIG_SYSFS */
 527
 528 static int __init hugepage_init(void)
 529 {
 530         int err;
 531         struct kobject *hugepage_kobj;
 532
 533         if (!has_transparent_hugepage()) {
 534                 transparent_hugepage_flags = 0;
 535                 return -EINVAL;
 536         }
 537
 538         err = hugepage_init_sysfs(&hugepage_kobj);
 539         if (err)
 540                 return err;
 541
 542         err = khugepaged_slab_init();
 543         if (err)
 544                 goto out;
 545
 546         err = mm_slots_hash_init();
 547         if (err) {
 548                 khugepaged_slab_free();
 549                 goto out;
 550         }
 551
 552         /*
 553          * By default disable transparent hugepages on smaller systems,
 554          * where the extra memory used could hurt more than TLB overhead
 555          * is likely to save.  The admin can still enable it through /sys.
 556          */
 557         if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
 558                 transparent_hugepage_flags = 0;
 559
 560         start_khugepaged();
 561
 562         return 0;
 563 out:
 564         hugepage_exit_sysfs(hugepage_kobj);
 565         return err;
 566 }
 567 module_init(hugepage_init)
 568
 569 static int __init setup_transparent_hugepage(char *str)
 570 {
 571         int ret = 0;
 572         if (!str)
 573                 goto out;
 574         if (!strcmp(str, "always")) {
 575                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 576                         &transparent_hugepage_flags);
 577                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 578                           &transparent_hugepage_flags);
 579                 ret = 1;
 580         } else if (!strcmp(str, "madvise")) {
 581                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 582                           &transparent_hugepage_flags);
 583                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 584                         &transparent_hugepage_flags);
 585                 ret = 1;
 586         } else if (!strcmp(str, "never")) {
 587                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 588                           &transparent_hugepage_flags);
 589                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 590                           &transparent_hugepage_flags);
 591                 ret = 1;
 592         }
 593 out:
 594         if (!ret)
 595                 printk(KERN_WARNING
 596                        "transparent_hugepage= cannot parse, ignored\n");
 597         return ret;
 598 }
 599 __setup("transparent_hugepage=", setup_transparent_hugepage);
 600
 601 static void prepare_pmd_huge_pte(pgtable_t pgtable,
 602                                  struct mm_struct *mm)
 603 {
 604         assert_spin_locked(&mm->page_table_lock);
 605
 606         /* FIFO */
 607         if (!mm->pmd_huge_pte)
 608                 INIT_LIST_HEAD(&pgtable->lru);
 609         else
 610                 list_add(&pgtable->lru, &mm->pmd_huge_pte->lru);
 611         mm->pmd_huge_pte = pgtable;
 612 }
 613
 614 static inline pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 615 {
 616         if (likely(vma->vm_flags & VM_WRITE))
 617                 pmd = pmd_mkwrite(pmd);
 618         return pmd;
 619 }
 620
 621 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
 622                                         struct vm_area_struct *vma,
 623                                         unsigned long haddr, pmd_t *pmd,
 624                                         struct page *page)
 625 {
 626         pgtable_t pgtable;
 627
 628         VM_BUG_ON(!PageCompound(page));
 629         pgtable = pte_alloc_one(mm, haddr);
 630         if (unlikely(!pgtable))
 631                 return VM_FAULT_OOM;
 632
 633         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 634         __SetPageUptodate(page);
 635
 636         spin_lock(&mm->page_table_lock);
 637         if (unlikely(!pmd_none(*pmd))) {
 638                 spin_unlock(&mm->page_table_lock);
 639                 mem_cgroup_uncharge_page(page);
 640                 put_page(page);
 641                 pte_free(mm, pgtable);
 642         } else {
 643                 pmd_t entry;
 644                 entry = mk_pmd(page, vma->vm_page_prot);
 645                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 646                 entry = pmd_mkhuge(entry);
 647                 /*
 648                  * The spinlocking to take the lru_lock inside
 649                  * page_add_new_anon_rmap() acts as a full memory
 650                  * barrier to be sure clear_huge_page writes become
 651                  * visible after the set_pmd_at() write.
 652                  */
 653                 page_add_new_anon_rmap(page, vma, haddr);
 654                 set_pmd_at(mm, haddr, pmd, entry);
 655                 prepare_pmd_huge_pte(pgtable, mm);
 656                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 657                 mm->nr_ptes++;
 658                 spin_unlock(&mm->page_table_lock);
 659         }
 660
 661         return 0;
 662 }
 663
 664 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 665 {
 666         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
 667 }
 668
 669 static inline struct page *alloc_hugepage_vma(int defrag,
 670                                               struct vm_area_struct *vma,
 671                                               unsigned long haddr, int nd,
 672                                               gfp_t extra_gfp)
 673 {
 674         return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
 675                                HPAGE_PMD_ORDER, vma, haddr, nd);
 676 }
 677
 678 #ifndef CONFIG_NUMA
 679 static inline struct page *alloc_hugepage(int defrag)
 680 {
 681         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
 682                            HPAGE_PMD_ORDER);
 683 }
 684 #endif
 685
 686 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 687                                unsigned long address, pmd_t *pmd,
 688                                unsigned int flags)
 689 {
 690         struct page *page;
 691         unsigned long haddr = address & HPAGE_PMD_MASK;
 692         pte_t *pte;
 693
 694         if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) {
 695                 if (unlikely(anon_vma_prepare(vma)))
 696                         return VM_FAULT_OOM;
 697                 if (unlikely(khugepaged_enter(vma)))
 698                         return VM_FAULT_OOM;
 699                 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 700                                           vma, haddr, numa_node_id(), 0);
 701                 if (unlikely(!page)) {
 702                         count_vm_event(THP_FAULT_FALLBACK);
 703                         goto out;
 704                 }
 705                 count_vm_event(THP_FAULT_ALLOC);
 706                 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
 707                         put_page(page);
 708                         goto out;
 709                 }
 710                 if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd,
 711                                                           page))) {
 712                         mem_cgroup_uncharge_page(page);
 713                         put_page(page);
 714                         goto out;
 715                 }
 716
 717                 return 0;
 718         }
 719 out:
 720         /*
 721          * Use __pte_alloc instead of pte_alloc_map, because we can't
 722          * run pte_offset_map on the pmd, if an huge pmd could
 723          * materialize from under us from a different thread.
 724          */
 725         if (unlikely(__pte_alloc(mm, vma, pmd, address)))
 726                 return VM_FAULT_OOM;
 727         /* if an huge pmd materialized from under us just retry later */
 728         if (unlikely(pmd_trans_huge(*pmd)))
 729                 return 0;
 730         /*
 731          * A regular pmd is established and it can't morph into a huge pmd
 732          * from under us anymore at this point because we hold the mmap_sem
 733          * read mode and khugepaged takes it in write mode. So now it's
 734          * safe to run pte_offset_map().
 735          */
 736         pte = pte_offset_map(pmd, address);
 737         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
 738 }
 739
 740 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 741                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 742                   struct vm_area_struct *vma)
 743 {
 744         struct page *src_page;
 745         pmd_t pmd;
 746         pgtable_t pgtable;
 747         int ret;
 748
 749         ret = -ENOMEM;
 750         pgtable = pte_alloc_one(dst_mm, addr);
 751         if (unlikely(!pgtable))
 752                 goto out;
 753
 754         spin_lock(&dst_mm->page_table_lock);
 755         spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
 756
 757         ret = -EAGAIN;
 758         pmd = *src_pmd;
 759         if (unlikely(!pmd_trans_huge(pmd))) {
 760                 pte_free(dst_mm, pgtable);
 761                 goto out_unlock;
 762         }
 763         if (unlikely(pmd_trans_splitting(pmd))) {
 764                 /* split huge page running from under us */
 765                 spin_unlock(&src_mm->page_table_lock);
 766                 spin_unlock(&dst_mm->page_table_lock);
 767                 pte_free(dst_mm, pgtable);
 768
 769                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 770                 goto out;
 771         }
 772         src_page = pmd_page(pmd);
 773         VM_BUG_ON(!PageHead(src_page));
 774         get_page(src_page);
 775         page_dup_rmap(src_page);
 776         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 777
 778         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 779         pmd = pmd_mkold(pmd_wrprotect(pmd));
 780         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 781         prepare_pmd_huge_pte(pgtable, dst_mm);
 782         dst_mm->nr_ptes++;
 783
 784         ret = 0;
 785 out_unlock:
 786         spin_unlock(&src_mm->page_table_lock);
 787         spin_unlock(&dst_mm->page_table_lock);
 788 out:
 789         return ret;
 790 }
 791
 792 /* no "address" argument so destroys page coloring of some arch */
 793 pgtable_t get_pmd_huge_pte(struct mm_struct *mm)
 794 {
 795         pgtable_t pgtable;
 796
 797         assert_spin_locked(&mm->page_table_lock);
 798
 799         /* FIFO */
 800         pgtable = mm->pmd_huge_pte;
 801         if (list_empty(&pgtable->lru))
 802                 mm->pmd_huge_pte = NULL;
 803         else {
 804                 mm->pmd_huge_pte = list_entry(pgtable->lru.next,
 805                                               struct page, lru);
 806                 list_del(&pgtable->lru);
 807         }
 808         return pgtable;
 809 }
 810
 811 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
 812                                         struct vm_area_struct *vma,
 813                                         unsigned long address,
 814                                         pmd_t *pmd, pmd_t orig_pmd,
 815                                         struct page *page,
 816                                         unsigned long haddr)
 817 {
 818         pgtable_t pgtable;
 819         pmd_t _pmd;
 820         int ret = 0, i;
 821         struct page **pages;
 822
 823         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
 824                         GFP_KERNEL);
 825         if (unlikely(!pages)) {
 826                 ret |= VM_FAULT_OOM;
 827                 goto out;
 828         }
 829
 830         for (i = 0; i < HPAGE_PMD_NR; i++) {
 831                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
 832                                                __GFP_OTHER_NODE,
 833                                                vma, address, page_to_nid(page));
 834                 if (unlikely(!pages[i] ||
 835                              mem_cgroup_newpage_charge(pages[i], mm,
 836                                                        GFP_KERNEL))) {
 837                         if (pages[i])
 838                                 put_page(pages[i]);
 839                         mem_cgroup_uncharge_start();
 840                         while (--i >= 0) {
 841                                 mem_cgroup_uncharge_page(pages[i]);
 842                                 put_page(pages[i]);
 843                         }
 844                         mem_cgroup_uncharge_end();
 845                         kfree(pages);
 846                         ret |= VM_FAULT_OOM;
 847                         goto out;
 848                 }
 849         }
 850
 851         for (i = 0; i < HPAGE_PMD_NR; i++) {
 852                 copy_user_highpage(pages[i], page + i,
 853                                    haddr + PAGE_SIZE * i, vma);
 854                 __SetPageUptodate(pages[i]);
 855                 cond_resched();
 856         }
 857
 858         spin_lock(&mm->page_table_lock);
 859         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 860                 goto out_free_pages;
 861         VM_BUG_ON(!PageHead(page));
 862
 863         pmdp_clear_flush_notify(vma, haddr, pmd);
 864         /* leave pmd empty until pte is filled */
 865
 866         pgtable = get_pmd_huge_pte(mm);
 867         pmd_populate(mm, &_pmd, pgtable);
 868
 869         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
 870                 pte_t *pte, entry;
 871                 entry = mk_pte(pages[i], vma->vm_page_prot);
 872                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
 873                 page_add_new_anon_rmap(pages[i], vma, haddr);
 874                 pte = pte_offset_map(&_pmd, haddr);
 875                 VM_BUG_ON(!pte_none(*pte));
 876                 set_pte_at(mm, haddr, pte, entry);
 877                 pte_unmap(pte);
 878         }
 879         kfree(pages);
 880
 881         smp_wmb(); /* make pte visible before pmd */
 882         pmd_populate(mm, pmd, pgtable);
 883         page_remove_rmap(page);
 884         spin_unlock(&mm->page_table_lock);
 885
 886         ret |= VM_FAULT_WRITE;
 887         put_page(page);
 888
 889 out:
 890         return ret;
 891
 892 out_free_pages:
 893         spin_unlock(&mm->page_table_lock);
 894         mem_cgroup_uncharge_start();
 895         for (i = 0; i < HPAGE_PMD_NR; i++) {
 896                 mem_cgroup_uncharge_page(pages[i]);
 897                 put_page(pages[i]);
 898         }
 899         mem_cgroup_uncharge_end();
 900         kfree(pages);
 901         goto out;
 902 }
 903
 904 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
 905                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
 906 {
 907         int ret = 0;
 908         struct page *page, *new_page;
 909         unsigned long haddr;
 910
 911         VM_BUG_ON(!vma->anon_vma);
 912         spin_lock(&mm->page_table_lock);
 913         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 914                 goto out_unlock;
 915
 916         page = pmd_page(orig_pmd);
 917         VM_BUG_ON(!PageCompound(page) || !PageHead(page));
 918         haddr = address & HPAGE_PMD_MASK;
 919         if (page_mapcount(page) == 1) {
 920                 pmd_t entry;
 921                 entry = pmd_mkyoung(orig_pmd);
 922                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 923                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
 924                         update_mmu_cache(vma, address, entry);
 925                 ret |= VM_FAULT_WRITE;
 926                 goto out_unlock;
 927         }
 928         get_page(page);
 929         spin_unlock(&mm->page_table_lock);
 930
 931         if (transparent_hugepage_enabled(vma) &&
 932             !transparent_hugepage_debug_cow())
 933                 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 934                                               vma, haddr, numa_node_id(), 0);
 935         else
 936                 new_page = NULL;
 937
 938         if (unlikely(!new_page)) {
 939                 count_vm_event(THP_FAULT_FALLBACK);
 940                 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
 941                                                    pmd, orig_pmd, page, haddr);
 942                 if (ret & VM_FAULT_OOM)
 943                         split_huge_page(page);
 944                 put_page(page);
 945                 goto out;
 946         }
 947         count_vm_event(THP_FAULT_ALLOC);
 948
 949         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
 950                 put_page(new_page);
 951                 split_huge_page(page);
 952                 put_page(page);
 953                 ret |= VM_FAULT_OOM;
 954                 goto out;
 955         }
 956
 957         copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
 958         __SetPageUptodate(new_page);
 959
 960         spin_lock(&mm->page_table_lock);
 961         put_page(page);
 962         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
 963                 spin_unlock(&mm->page_table_lock);
 964                 mem_cgroup_uncharge_page(new_page);
 965                 put_page(new_page);
 966                 goto out;
 967         } else {
 968                 pmd_t entry;
 969                 VM_BUG_ON(!PageHead(page));
 970                 entry = mk_pmd(new_page, vma->vm_page_prot);
 971                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 972                 entry = pmd_mkhuge(entry);
 973                 pmdp_clear_flush_notify(vma, haddr, pmd);
 974                 page_add_new_anon_rmap(new_page, vma, haddr);
 975                 set_pmd_at(mm, haddr, pmd, entry);
 976                 update_mmu_cache(vma, address, entry);
 977                 page_remove_rmap(page);
 978                 put_page(page);
 979                 ret |= VM_FAULT_WRITE;
 980         }
 981 out_unlock:
 982         spin_unlock(&mm->page_table_lock);
 983 out:
 984         return ret;
 985 }
 986
 987 struct page *follow_trans_huge_pmd(struct mm_struct *mm,
 988                                    unsigned long addr,
 989                                    pmd_t *pmd,
 990                                    unsigned int flags)
 991 {
 992         struct page *page = NULL;
 993
 994         assert_spin_locked(&mm->page_table_lock);
 995
 996         if (flags & FOLL_WRITE && !pmd_write(*pmd))
 997                 goto out;
 998
 999         page = pmd_page(*pmd);
1000         VM_BUG_ON(!PageHead(page));
1001         if (flags & FOLL_TOUCH) {
1002                 pmd_t _pmd;
1003                 /*
1004                  * We should set the dirty bit only for FOLL_WRITE but
1005                  * for now the dirty bit in the pmd is meaningless.
1006                  * And if the dirty bit will become meaningful and
1007                  * we'll only set it with FOLL_WRITE, an atomic
1008                  * set_bit will be required on the pmd to set the
1009                  * young bit, instead of the current set_pmd_at.
1010                  */
1011                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1012                 set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd);
1013         }
1014         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1015         VM_BUG_ON(!PageCompound(page));
1016         if (flags & FOLL_GET)
1017                 get_page_foll(page);
1018
1019 out:
1020         return page;
1021 }
1022
1023 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1024                  pmd_t *pmd, unsigned long addr)
1025 {
1026         int ret = 0;
1027
1028         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1029                 struct page *page;
1030                 pgtable_t pgtable;
1031                 pgtable = get_pmd_huge_pte(tlb->mm);
1032                 page = pmd_page(*pmd);
1033                 pmd_clear(pmd);
1034                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1035                 page_remove_rmap(page);
1036                 VM_BUG_ON(page_mapcount(page) < 0);
1037                 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1038                 VM_BUG_ON(!PageHead(page));
1039                 tlb->mm->nr_ptes--;
1040                 spin_unlock(&tlb->mm->page_table_lock);
1041                 tlb_remove_page(tlb, page);
1042                 pte_free(tlb->mm, pgtable);
1043                 ret = 1;
1044         }
1045         return ret;
1046 }
1047
1048 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1049                 unsigned long addr, unsigned long end,
1050                 unsigned char *vec)
1051 {
1052         int ret = 0;
1053
1054         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1055                 /*
1056                  * All logical pages in the range are present
1057                  * if backed by a huge page.
1058                  */
1059                 spin_unlock(&vma->vm_mm->page_table_lock);
1060                 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1061                 ret = 1;
1062         }
1063
1064         return ret;
1065 }
1066
1067 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1068                   unsigned long old_addr,
1069                   unsigned long new_addr, unsigned long old_end,
1070                   pmd_t *old_pmd, pmd_t *new_pmd)
1071 {
1072         int ret = 0;
1073         pmd_t pmd;
1074
1075         struct mm_struct *mm = vma->vm_mm;
1076
1077         if ((old_addr & ~HPAGE_PMD_MASK) ||
1078             (new_addr & ~HPAGE_PMD_MASK) ||
1079             old_end - old_addr < HPAGE_PMD_SIZE ||
1080             (new_vma->vm_flags & VM_NOHUGEPAGE))
1081                 goto out;
1082
1083         /*
1084          * The destination pmd shouldn't be established, free_pgtables()
1085          * should have release it.
1086          */
1087         if (WARN_ON(!pmd_none(*new_pmd))) {
1088                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1089                 goto out;
1090         }
1091
1092         ret = __pmd_trans_huge_lock(old_pmd, vma);
1093         if (ret == 1) {
1094                 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1095                 VM_BUG_ON(!pmd_none(*new_pmd));
1096                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1097                 spin_unlock(&mm->page_table_lock);
1098         }
1099 out:
1100         return ret;
1101 }
1102
1103 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1104                 unsigned long addr, pgprot_t newprot)
1105 {
1106         struct mm_struct *mm = vma->vm_mm;
1107         int ret = 0;
1108
1109         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1110                 pmd_t entry;
1111                 entry = pmdp_get_and_clear(mm, addr, pmd);
1112                 entry = pmd_modify(entry, newprot);
1113                 set_pmd_at(mm, addr, pmd, entry);
1114                 spin_unlock(&vma->vm_mm->page_table_lock);
1115                 ret = 1;
1116         }
1117
1118         return ret;
1119 }
1120
1121 /*
1122  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1123  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1124  *
1125  * Note that if it returns 1, this routine returns without unlocking page
1126  * table locks. So callers must unlock them.
1127  */
1128 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1129 {
1130         spin_lock(&vma->vm_mm->page_table_lock);
1131         if (likely(pmd_trans_huge(*pmd))) {
1132                 if (unlikely(pmd_trans_splitting(*pmd))) {
1133                         spin_unlock(&vma->vm_mm->page_table_lock);
1134                         wait_split_huge_page(vma->anon_vma, pmd);
1135                         return -1;
1136                 } else {
1137                         /* Thp mapped by 'pmd' is stable, so we can
1138                          * handle it as it is. */
1139                         return 1;
1140                 }
1141         }
1142         spin_unlock(&vma->vm_mm->page_table_lock);
1143         return 0;
1144 }
1145
1146 pmd_t *page_check_address_pmd(struct page *page,
1147                               struct mm_struct *mm,
1148                               unsigned long address,
1149                               enum page_check_address_pmd_flag flag)
1150 {
1151         pgd_t *pgd;
1152         pud_t *pud;
1153         pmd_t *pmd, *ret = NULL;
1154
1155         if (address & ~HPAGE_PMD_MASK)
1156                 goto out;
1157
1158         pgd = pgd_offset(mm, address);
1159         if (!pgd_present(*pgd))
1160                 goto out;
1161
1162         pud = pud_offset(pgd, address);
1163         if (!pud_present(*pud))
1164                 goto out;
1165
1166         pmd = pmd_offset(pud, address);
1167         if (pmd_none(*pmd))
1168                 goto out;
1169         if (pmd_page(*pmd) != page)
1170                 goto out;
1171         /*
1172          * split_vma() may create temporary aliased mappings. There is
1173          * no risk as long as all huge pmd are found and have their
1174          * splitting bit set before __split_huge_page_refcount
1175          * runs. Finding the same huge pmd more than once during the
1176          * same rmap walk is not a problem.
1177          */
1178         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1179             pmd_trans_splitting(*pmd))
1180                 goto out;
1181         if (pmd_trans_huge(*pmd)) {
1182                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1183                           !pmd_trans_splitting(*pmd));
1184                 ret = pmd;
1185         }
1186 out:
1187         return ret;
1188 }
1189
1190 static int __split_huge_page_splitting(struct page *page,
1191                                        struct vm_area_struct *vma,
1192                                        unsigned long address)
1193 {
1194         struct mm_struct *mm = vma->vm_mm;
1195         pmd_t *pmd;
1196         int ret = 0;
1197
1198         spin_lock(&mm->page_table_lock);
1199         pmd = page_check_address_pmd(page, mm, address,
1200                                      PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1201         if (pmd) {
1202                 /*
1203                  * We can't temporarily set the pmd to null in order
1204                  * to split it, the pmd must remain marked huge at all
1205                  * times or the VM won't take the pmd_trans_huge paths
1206                  * and it won't wait on the anon_vma->root->mutex to
1207                  * serialize against split_huge_page*.
1208                  */
1209                 pmdp_splitting_flush_notify(vma, address, pmd);
1210                 ret = 1;
1211         }
1212         spin_unlock(&mm->page_table_lock);
1213
1214         return ret;
1215 }
1216
1217 static void __split_huge_page_refcount(struct page *page)
1218 {
1219         int i;
1220         struct zone *zone = page_zone(page);
1221         struct lruvec *lruvec;
1222         int tail_count = 0;
1223
1224         /* prevent PageLRU to go away from under us, and freeze lru stats */
1225         spin_lock_irq(&zone->lru_lock);
1226         lruvec = mem_cgroup_page_lruvec(page, zone);
1227
1228         compound_lock(page);
1229         /* complete memcg works before add pages to LRU */
1230         mem_cgroup_split_huge_fixup(page);
1231
1232         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1233                 struct page *page_tail = page + i;
1234
1235                 /* tail_page->_mapcount cannot change */
1236                 BUG_ON(page_mapcount(page_tail) < 0);
1237                 tail_count += page_mapcount(page_tail);
1238                 /* check for overflow */
1239                 BUG_ON(tail_count < 0);
1240                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1241                 /*
1242                  * tail_page->_count is zero and not changing from
1243                  * under us. But get_page_unless_zero() may be running
1244                  * from under us on the tail_page. If we used
1245                  * atomic_set() below instead of atomic_add(), we
1246                  * would then run atomic_set() concurrently with
1247                  * get_page_unless_zero(), and atomic_set() is
1248                  * implemented in C not using locked ops. spin_unlock
1249                  * on x86 sometime uses locked ops because of PPro
1250                  * errata 66, 92, so unless somebody can guarantee
1251                  * atomic_set() here would be safe on all archs (and
1252                  * not only on x86), it's safer to use atomic_add().
1253                  */
1254                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1255                            &page_tail->_count);
1256
1257                 /* after clearing PageTail the gup refcount can be released */
1258                 smp_mb();
1259
1260                 /*
1261                  * retain hwpoison flag of the poisoned tail page:
1262                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1263                  *   by the memory-failure.
1264                  */
1265                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1266                 page_tail->flags |= (page->flags &
1267                                      ((1L << PG_referenced) |
1268                                       (1L << PG_swapbacked) |
1269                                       (1L << PG_mlocked) |
1270                                       (1L << PG_uptodate)));
1271                 page_tail->flags |= (1L << PG_dirty);
1272
1273                 /* clear PageTail before overwriting first_page */
1274                 smp_wmb();
1275
1276                 /*
1277                  * __split_huge_page_splitting() already set the
1278                  * splitting bit in all pmd that could map this
1279                  * hugepage, that will ensure no CPU can alter the
1280                  * mapcount on the head page. The mapcount is only
1281                  * accounted in the head page and it has to be
1282                  * transferred to all tail pages in the below code. So
1283                  * for this code to be safe, the split the mapcount
1284                  * can't change. But that doesn't mean userland can't
1285                  * keep changing and reading the page contents while
1286                  * we transfer the mapcount, so the pmd splitting
1287                  * status is achieved setting a reserved bit in the
1288                  * pmd, not by clearing the present bit.
1289                 */
1290                 page_tail->_mapcount = page->_mapcount;
1291
1292                 BUG_ON(page_tail->mapping);
1293                 page_tail->mapping = page->mapping;
1294
1295                 page_tail->index = page->index + i;
1296
1297                 BUG_ON(!PageAnon(page_tail));
1298                 BUG_ON(!PageUptodate(page_tail));
1299                 BUG_ON(!PageDirty(page_tail));
1300                 BUG_ON(!PageSwapBacked(page_tail));
1301
1302                 lru_add_page_tail(page, page_tail, lruvec);
1303         }
1304         atomic_sub(tail_count, &page->_count);
1305         BUG_ON(atomic_read(&page->_count) <= 0);
1306
1307         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1308         __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1309
1310         ClearPageCompound(page);
1311         compound_unlock(page);
1312         spin_unlock_irq(&zone->lru_lock);
1313
1314         for (i = 1; i < HPAGE_PMD_NR; i++) {
1315                 struct page *page_tail = page + i;
1316                 BUG_ON(page_count(page_tail) <= 0);
1317                 /*
1318                  * Tail pages may be freed if there wasn't any mapping
1319                  * like if add_to_swap() is running on a lru page that
1320                  * had its mapping zapped. And freeing these pages
1321                  * requires taking the lru_lock so we do the put_page
1322                  * of the tail pages after the split is complete.
1323                  */
1324                 put_page(page_tail);
1325         }
1326
1327         /*
1328          * Only the head page (now become a regular page) is required
1329          * to be pinned by the caller.
1330          */
1331         BUG_ON(page_count(page) <= 0);
1332 }
1333
1334 static int __split_huge_page_map(struct page *page,
1335                                  struct vm_area_struct *vma,
1336                                  unsigned long address)
1337 {
1338         struct mm_struct *mm = vma->vm_mm;
1339         pmd_t *pmd, _pmd;
1340         int ret = 0, i;
1341         pgtable_t pgtable;
1342         unsigned long haddr;
1343
1344         spin_lock(&mm->page_table_lock);
1345         pmd = page_check_address_pmd(page, mm, address,
1346                                      PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1347         if (pmd) {
1348                 pgtable = get_pmd_huge_pte(mm);
1349                 pmd_populate(mm, &_pmd, pgtable);
1350
1351                 for (i = 0, haddr = address; i < HPAGE_PMD_NR;
1352                      i++, haddr += PAGE_SIZE) {
1353                         pte_t *pte, entry;
1354                         BUG_ON(PageCompound(page+i));
1355                         entry = mk_pte(page + i, vma->vm_page_prot);
1356                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1357                         if (!pmd_write(*pmd))
1358                                 entry = pte_wrprotect(entry);
1359                         else
1360                                 BUG_ON(page_mapcount(page) != 1);
1361                         if (!pmd_young(*pmd))
1362                                 entry = pte_mkold(entry);
1363                         pte = pte_offset_map(&_pmd, haddr);
1364                         BUG_ON(!pte_none(*pte));
1365                         set_pte_at(mm, haddr, pte, entry);
1366                         pte_unmap(pte);
1367                 }
1368
1369                 smp_wmb(); /* make pte visible before pmd */
1370                 /*
1371                  * Up to this point the pmd is present and huge and
1372                  * userland has the whole access to the hugepage
1373                  * during the split (which happens in place). If we
1374                  * overwrite the pmd with the not-huge version
1375                  * pointing to the pte here (which of course we could
1376                  * if all CPUs were bug free), userland could trigger
1377                  * a small page size TLB miss on the small sized TLB
1378                  * while the hugepage TLB entry is still established
1379                  * in the huge TLB. Some CPU doesn't like that. See
1380                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1381                  * Erratum 383 on page 93. Intel should be safe but is
1382                  * also warns that it's only safe if the permission
1383                  * and cache attributes of the two entries loaded in
1384                  * the two TLB is identical (which should be the case
1385                  * here). But it is generally safer to never allow
1386                  * small and huge TLB entries for the same virtual
1387                  * address to be loaded simultaneously. So instead of
1388                  * doing "pmd_populate(); flush_tlb_range();" we first
1389                  * mark the current pmd notpresent (atomically because
1390                  * here the pmd_trans_huge and pmd_trans_splitting
1391                  * must remain set at all times on the pmd until the
1392                  * split is complete for this pmd), then we flush the
1393                  * SMP TLB and finally we write the non-huge version
1394                  * of the pmd entry with pmd_populate.
1395                  */
1396                 set_pmd_at(mm, address, pmd, pmd_mknotpresent(*pmd));
1397                 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
1398                 pmd_populate(mm, pmd, pgtable);
1399                 ret = 1;
1400         }
1401         spin_unlock(&mm->page_table_lock);
1402
1403         return ret;
1404 }
1405
1406 /* must be called with anon_vma->root->mutex hold */
1407 static void __split_huge_page(struct page *page,
1408                               struct anon_vma *anon_vma)
1409 {
1410         int mapcount, mapcount2;
1411         struct anon_vma_chain *avc;
1412
1413         BUG_ON(!PageHead(page));
1414         BUG_ON(PageTail(page));
1415
1416         mapcount = 0;
1417         list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1418                 struct vm_area_struct *vma = avc->vma;
1419                 unsigned long addr = vma_address(page, vma);
1420                 BUG_ON(is_vma_temporary_stack(vma));
1421                 if (addr == -EFAULT)
1422                         continue;
1423                 mapcount += __split_huge_page_splitting(page, vma, addr);
1424         }
1425         /*
1426          * It is critical that new vmas are added to the tail of the
1427          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1428          * and establishes a child pmd before
1429          * __split_huge_page_splitting() freezes the parent pmd (so if
1430          * we fail to prevent copy_huge_pmd() from running until the
1431          * whole __split_huge_page() is complete), we will still see
1432          * the newly established pmd of the child later during the
1433          * walk, to be able to set it as pmd_trans_splitting too.
1434          */
1435         if (mapcount != page_mapcount(page))
1436                 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1437                        mapcount, page_mapcount(page));
1438         BUG_ON(mapcount != page_mapcount(page));
1439
1440         __split_huge_page_refcount(page);
1441
1442         mapcount2 = 0;
1443         list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1444                 struct vm_area_struct *vma = avc->vma;
1445                 unsigned long addr = vma_address(page, vma);
1446                 BUG_ON(is_vma_temporary_stack(vma));
1447                 if (addr == -EFAULT)
1448                         continue;
1449                 mapcount2 += __split_huge_page_map(page, vma, addr);
1450         }
1451         if (mapcount != mapcount2)
1452                 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1453                        mapcount, mapcount2, page_mapcount(page));
1454         BUG_ON(mapcount != mapcount2);
1455 }
1456
1457 int split_huge_page(struct page *page)
1458 {
1459         struct anon_vma *anon_vma;
1460         int ret = 1;
1461
1462         BUG_ON(!PageAnon(page));
1463         anon_vma = page_lock_anon_vma(page);
1464         if (!anon_vma)
1465                 goto out;
1466         ret = 0;
1467         if (!PageCompound(page))
1468                 goto out_unlock;
1469
1470         BUG_ON(!PageSwapBacked(page));
1471         __split_huge_page(page, anon_vma);
1472         count_vm_event(THP_SPLIT);
1473
1474         BUG_ON(PageCompound(page));
1475 out_unlock:
1476         page_unlock_anon_vma(anon_vma);
1477 out:
1478         return ret;
1479 }
1480
1481 #define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1482
1483 int hugepage_madvise(struct vm_area_struct *vma,
1484                      unsigned long *vm_flags, int advice)
1485 {
1486         switch (advice) {
1487         case MADV_HUGEPAGE:
1488                 /*
1489                  * Be somewhat over-protective like KSM for now!
1490                  */
1491                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1492                         return -EINVAL;
1493                 *vm_flags &= ~VM_NOHUGEPAGE;
1494                 *vm_flags |= VM_HUGEPAGE;
1495                 /*
1496                  * If the vma become good for khugepaged to scan,
1497                  * register it here without waiting a page fault that
1498                  * may not happen any time soon.
1499                  */
1500                 if (unlikely(khugepaged_enter_vma_merge(vma)))
1501                         return -ENOMEM;
1502                 break;
1503         case MADV_NOHUGEPAGE:
1504                 /*
1505                  * Be somewhat over-protective like KSM for now!
1506                  */
1507                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1508                         return -EINVAL;
1509                 *vm_flags &= ~VM_HUGEPAGE;
1510                 *vm_flags |= VM_NOHUGEPAGE;
1511                 /*
1512                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1513                  * this vma even if we leave the mm registered in khugepaged if
1514                  * it got registered before VM_NOHUGEPAGE was set.
1515                  */
1516                 break;
1517         }
1518
1519         return 0;
1520 }
1521
1522 static int __init khugepaged_slab_init(void)
1523 {
1524         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1525                                           sizeof(struct mm_slot),
1526                                           __alignof__(struct mm_slot), 0, NULL);
1527         if (!mm_slot_cache)
1528                 return -ENOMEM;
1529
1530         return 0;
1531 }
1532
1533 static void __init khugepaged_slab_free(void)
1534 {
1535         kmem_cache_destroy(mm_slot_cache);
1536         mm_slot_cache = NULL;
1537 }
1538
1539 static inline struct mm_slot *alloc_mm_slot(void)
1540 {
1541         if (!mm_slot_cache)     /* initialization failed */
1542                 return NULL;
1543         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1544 }
1545
1546 static inline void free_mm_slot(struct mm_slot *mm_slot)
1547 {
1548         kmem_cache_free(mm_slot_cache, mm_slot);
1549 }
1550
1551 static int __init mm_slots_hash_init(void)
1552 {
1553         mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
1554                                 GFP_KERNEL);
1555         if (!mm_slots_hash)
1556                 return -ENOMEM;
1557         return 0;
1558 }
1559
1560 #if 0
1561 static void __init mm_slots_hash_free(void)
1562 {
1563         kfree(mm_slots_hash);
1564         mm_slots_hash = NULL;
1565 }
1566 #endif
1567
1568 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1569 {
1570         struct mm_slot *mm_slot;
1571         struct hlist_head *bucket;
1572         struct hlist_node *node;
1573
1574         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1575                                 % MM_SLOTS_HASH_HEADS];
1576         hlist_for_each_entry(mm_slot, node, bucket, hash) {
1577                 if (mm == mm_slot->mm)
1578                         return mm_slot;
1579         }
1580         return NULL;
1581 }
1582
1583 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1584                                     struct mm_slot *mm_slot)
1585 {
1586         struct hlist_head *bucket;
1587
1588         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1589                                 % MM_SLOTS_HASH_HEADS];
1590         mm_slot->mm = mm;
1591         hlist_add_head(&mm_slot->hash, bucket);
1592 }
1593
1594 static inline int khugepaged_test_exit(struct mm_struct *mm)
1595 {
1596         return atomic_read(&mm->mm_users) == 0;
1597 }
1598
1599 int __khugepaged_enter(struct mm_struct *mm)
1600 {
1601         struct mm_slot *mm_slot;
1602         int wakeup;
1603
1604         mm_slot = alloc_mm_slot();
1605         if (!mm_slot)
1606                 return -ENOMEM;
1607
1608         /* __khugepaged_exit() must not run from under us */
1609         VM_BUG_ON(khugepaged_test_exit(mm));
1610         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1611                 free_mm_slot(mm_slot);
1612                 return 0;
1613         }
1614
1615         spin_lock(&khugepaged_mm_lock);
1616         insert_to_mm_slots_hash(mm, mm_slot);
1617         /*
1618          * Insert just behind the scanning cursor, to let the area settle
1619          * down a little.
1620          */
1621         wakeup = list_empty(&khugepaged_scan.mm_head);
1622         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1623         spin_unlock(&khugepaged_mm_lock);
1624
1625         atomic_inc(&mm->mm_count);
1626         if (wakeup)
1627                 wake_up_interruptible(&khugepaged_wait);
1628
1629         return 0;
1630 }
1631
1632 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
1633 {
1634         unsigned long hstart, hend;
1635         if (!vma->anon_vma)
1636                 /*
1637                  * Not yet faulted in so we will register later in the
1638                  * page fault if needed.
1639                  */
1640                 return 0;
1641         if (vma->vm_ops)
1642                 /* khugepaged not yet working on file or special mappings */
1643                 return 0;
1644         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
1645         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1646         hend = vma->vm_end & HPAGE_PMD_MASK;
1647         if (hstart < hend)
1648                 return khugepaged_enter(vma);
1649         return 0;
1650 }
1651
1652 void __khugepaged_exit(struct mm_struct *mm)
1653 {
1654         struct mm_slot *mm_slot;
1655         int free = 0;
1656
1657         spin_lock(&khugepaged_mm_lock);
1658         mm_slot = get_mm_slot(mm);
1659         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1660                 hlist_del(&mm_slot->hash);
1661                 list_del(&mm_slot->mm_node);
1662                 free = 1;
1663         }
1664         spin_unlock(&khugepaged_mm_lock);
1665
1666         if (free) {
1667                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1668                 free_mm_slot(mm_slot);
1669                 mmdrop(mm);
1670         } else if (mm_slot) {
1671                 /*
1672                  * This is required to serialize against
1673                  * khugepaged_test_exit() (which is guaranteed to run
1674                  * under mmap sem read mode). Stop here (after we
1675                  * return all pagetables will be destroyed) until
1676                  * khugepaged has finished working on the pagetables
1677                  * under the mmap_sem.
1678                  */
1679                 down_write(&mm->mmap_sem);
1680                 up_write(&mm->mmap_sem);
1681         }
1682 }
1683
1684 static void release_pte_page(struct page *page)
1685 {
1686         /* 0 stands for page_is_file_cache(page) == false */
1687         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1688         unlock_page(page);
1689         putback_lru_page(page);
1690 }
1691
1692 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1693 {
1694         while (--_pte >= pte) {
1695                 pte_t pteval = *_pte;
1696                 if (!pte_none(pteval))
1697                         release_pte_page(pte_page(pteval));
1698         }
1699 }
1700
1701 static void release_all_pte_pages(pte_t *pte)
1702 {
1703         release_pte_pages(pte, pte + HPAGE_PMD_NR);
1704 }
1705
1706 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1707                                         unsigned long address,
1708                                         pte_t *pte)
1709 {
1710         struct page *page;
1711         pte_t *_pte;
1712         int referenced = 0, isolated = 0, none = 0;
1713         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1714              _pte++, address += PAGE_SIZE) {
1715                 pte_t pteval = *_pte;
1716                 if (pte_none(pteval)) {
1717                         if (++none <= khugepaged_max_ptes_none)
1718                                 continue;
1719                         else {
1720                                 release_pte_pages(pte, _pte);
1721                                 goto out;
1722                         }
1723                 }
1724                 if (!pte_present(pteval) || !pte_write(pteval)) {
1725                         release_pte_pages(pte, _pte);
1726                         goto out;
1727                 }
1728                 page = vm_normal_page(vma, address, pteval);
1729                 if (unlikely(!page)) {
1730                         release_pte_pages(pte, _pte);
1731                         goto out;
1732                 }
1733                 VM_BUG_ON(PageCompound(page));
1734                 BUG_ON(!PageAnon(page));
1735                 VM_BUG_ON(!PageSwapBacked(page));
1736
1737                 /* cannot use mapcount: can't collapse if there's a gup pin */
1738                 if (page_count(page) != 1) {
1739                         release_pte_pages(pte, _pte);
1740                         goto out;
1741                 }
1742                 /*
1743                  * We can do it before isolate_lru_page because the
1744                  * page can't be freed from under us. NOTE: PG_lock
1745                  * is needed to serialize against split_huge_page
1746                  * when invoked from the VM.
1747                  */
1748                 if (!trylock_page(page)) {
1749                         release_pte_pages(pte, _pte);
1750                         goto out;
1751                 }
1752                 /*
1753                  * Isolate the page to avoid collapsing an hugepage
1754                  * currently in use by the VM.
1755                  */
1756                 if (isolate_lru_page(page)) {
1757                         unlock_page(page);
1758                         release_pte_pages(pte, _pte);
1759                         goto out;
1760                 }
1761                 /* 0 stands for page_is_file_cache(page) == false */
1762                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
1763                 VM_BUG_ON(!PageLocked(page));
1764                 VM_BUG_ON(PageLRU(page));
1765
1766                 /* If there is no mapped pte young don't collapse the page */
1767                 if (pte_young(pteval) || PageReferenced(page) ||
1768                     mmu_notifier_test_young(vma->vm_mm, address))
1769                         referenced = 1;
1770         }
1771         if (unlikely(!referenced))
1772                 release_all_pte_pages(pte);
1773         else
1774                 isolated = 1;
1775 out:
1776         return isolated;
1777 }
1778
1779 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
1780                                       struct vm_area_struct *vma,
1781                                       unsigned long address,
1782                                       spinlock_t *ptl)
1783 {
1784         pte_t *_pte;
1785         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
1786                 pte_t pteval = *_pte;
1787                 struct page *src_page;
1788
1789                 if (pte_none(pteval)) {
1790                         clear_user_highpage(page, address);
1791                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
1792                 } else {
1793                         src_page = pte_page(pteval);
1794                         copy_user_highpage(page, src_page, address, vma);
1795                         VM_BUG_ON(page_mapcount(src_page) != 1);
1796                         release_pte_page(src_page);
1797                         /*
1798                          * ptl mostly unnecessary, but preempt has to
1799                          * be disabled to update the per-cpu stats
1800                          * inside page_remove_rmap().
1801                          */
1802                         spin_lock(ptl);
1803                         /*
1804                          * paravirt calls inside pte_clear here are
1805                          * superfluous.
1806                          */
1807                         pte_clear(vma->vm_mm, address, _pte);
1808                         page_remove_rmap(src_page);
1809                         spin_unlock(ptl);
1810                         free_page_and_swap_cache(src_page);
1811                 }
1812
1813                 address += PAGE_SIZE;
1814                 page++;
1815         }
1816 }
1817
1818 static void khugepaged_alloc_sleep(void)
1819 {
1820         wait_event_freezable_timeout(khugepaged_wait, false,
1821                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
1822 }
1823
1824 #ifdef CONFIG_NUMA
1825 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
1826 {
1827         if (IS_ERR(*hpage)) {
1828                 if (!*wait)
1829                         return false;
1830
1831                 *wait = false;
1832                 khugepaged_alloc_sleep();
1833         } else if (*hpage) {
1834                 put_page(*hpage);
1835                 *hpage = NULL;
1836         }
1837
1838         return true;
1839 }
1840
1841 static struct page
1842 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
1843                        struct vm_area_struct *vma, unsigned long address,
1844                        int node)
1845 {
1846         VM_BUG_ON(*hpage);
1847         /*
1848          * Allocate the page while the vma is still valid and under
1849          * the mmap_sem read mode so there is no memory allocation
1850          * later when we take the mmap_sem in write mode. This is more
1851          * friendly behavior (OTOH it may actually hide bugs) to
1852          * filesystems in userland with daemons allocating memory in
1853          * the userland I/O paths.  Allocating memory with the
1854          * mmap_sem in read mode is good idea also to allow greater
1855          * scalability.
1856          */
1857         *hpage  = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
1858                                       node, __GFP_OTHER_NODE);
1859
1860         /*
1861          * After allocating the hugepage, release the mmap_sem read lock in
1862          * preparation for taking it in write mode.
1863          */
1864         up_read(&mm->mmap_sem);
1865         if (unlikely(!*hpage)) {
1866                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
1867                 *hpage = ERR_PTR(-ENOMEM);
1868                 return NULL;
1869         }
1870
1871         count_vm_event(THP_COLLAPSE_ALLOC);
1872         return *hpage;
1873 }
1874 #else
1875 static struct page *khugepaged_alloc_hugepage(bool *wait)
1876 {
1877         struct page *hpage;
1878
1879         do {
1880                 hpage = alloc_hugepage(khugepaged_defrag());
1881                 if (!hpage) {
1882                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
1883                         if (!*wait)
1884                                 return NULL;
1885
1886                         *wait = false;
1887                         khugepaged_alloc_sleep();
1888                 } else
1889                         count_vm_event(THP_COLLAPSE_ALLOC);
1890         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
1891
1892         return hpage;
1893 }
1894
1895 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
1896 {
1897         if (!*hpage)
1898                 *hpage = khugepaged_alloc_hugepage(wait);
1899
1900         if (unlikely(!*hpage))
1901                 return false;
1902
1903         return true;
1904 }
1905
1906 static struct page
1907 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
1908                        struct vm_area_struct *vma, unsigned long address,
1909                        int node)
1910 {
1911         up_read(&mm->mmap_sem);
1912         VM_BUG_ON(!*hpage);
1913         return  *hpage;
1914 }
1915 #endif
1916
1917 static void collapse_huge_page(struct mm_struct *mm,
1918                                    unsigned long address,
1919                                    struct page **hpage,
1920                                    struct vm_area_struct *vma,
1921                                    int node)
1922 {
1923         pgd_t *pgd;
1924         pud_t *pud;
1925         pmd_t *pmd, _pmd;
1926         pte_t *pte;
1927         pgtable_t pgtable;
1928         struct page *new_page;
1929         spinlock_t *ptl;
1930         int isolated;
1931         unsigned long hstart, hend;
1932
1933         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1934
1935         /* release the mmap_sem read lock. */
1936         new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
1937         if (!new_page)
1938                 return;
1939
1940         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL)))
1941                 return;
1942
1943         /*
1944          * Prevent all access to pagetables with the exception of
1945          * gup_fast later hanlded by the ptep_clear_flush and the VM
1946          * handled by the anon_vma lock + PG_lock.
1947          */
1948         down_write(&mm->mmap_sem);
1949         if (unlikely(khugepaged_test_exit(mm)))
1950                 goto out;
1951
1952         vma = find_vma(mm, address);
1953         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1954         hend = vma->vm_end & HPAGE_PMD_MASK;
1955         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
1956                 goto out;
1957
1958         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
1959             (vma->vm_flags & VM_NOHUGEPAGE))
1960                 goto out;
1961
1962         if (!vma->anon_vma || vma->vm_ops)
1963                 goto out;
1964         if (is_vma_temporary_stack(vma))
1965                 goto out;
1966         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
1967
1968         pgd = pgd_offset(mm, address);
1969         if (!pgd_present(*pgd))
1970                 goto out;
1971
1972         pud = pud_offset(pgd, address);
1973         if (!pud_present(*pud))
1974                 goto out;
1975
1976         pmd = pmd_offset(pud, address);
1977         /* pmd can't go away or become huge under us */
1978         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
1979                 goto out;
1980
1981         anon_vma_lock(vma->anon_vma);
1982
1983         pte = pte_offset_map(pmd, address);
1984         ptl = pte_lockptr(mm, pmd);
1985
1986         spin_lock(&mm->page_table_lock); /* probably unnecessary */
1987         /*
1988          * After this gup_fast can't run anymore. This also removes
1989          * any huge TLB entry from the CPU so we won't allow
1990          * huge and small TLB entries for the same virtual address
1991          * to avoid the risk of CPU bugs in that area.
1992          */
1993         _pmd = pmdp_clear_flush_notify(vma, address, pmd);
1994         spin_unlock(&mm->page_table_lock);
1995
1996         spin_lock(ptl);
1997         isolated = __collapse_huge_page_isolate(vma, address, pte);
1998         spin_unlock(ptl);
1999
2000         if (unlikely(!isolated)) {
2001                 pte_unmap(pte);
2002                 spin_lock(&mm->page_table_lock);
2003                 BUG_ON(!pmd_none(*pmd));
2004                 set_pmd_at(mm, address, pmd, _pmd);
2005                 spin_unlock(&mm->page_table_lock);
2006                 anon_vma_unlock(vma->anon_vma);
2007                 goto out;
2008         }
2009
2010         /*
2011          * All pages are isolated and locked so anon_vma rmap
2012          * can't run anymore.
2013          */
2014         anon_vma_unlock(vma->anon_vma);
2015
2016         __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
2017         pte_unmap(pte);
2018         __SetPageUptodate(new_page);
2019         pgtable = pmd_pgtable(_pmd);
2020         VM_BUG_ON(page_count(pgtable) != 1);
2021         VM_BUG_ON(page_mapcount(pgtable) != 0);
2022
2023         _pmd = mk_pmd(new_page, vma->vm_page_prot);
2024         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2025         _pmd = pmd_mkhuge(_pmd);
2026
2027         /*
2028          * spin_lock() below is not the equivalent of smp_wmb(), so
2029          * this is needed to avoid the copy_huge_page writes to become
2030          * visible after the set_pmd_at() write.
2031          */
2032         smp_wmb();
2033
2034         spin_lock(&mm->page_table_lock);
2035         BUG_ON(!pmd_none(*pmd));
2036         page_add_new_anon_rmap(new_page, vma, address);
2037         set_pmd_at(mm, address, pmd, _pmd);
2038         update_mmu_cache(vma, address, _pmd);
2039         prepare_pmd_huge_pte(pgtable, mm);
2040         spin_unlock(&mm->page_table_lock);
2041
2042         *hpage = NULL;
2043
2044         khugepaged_pages_collapsed++;
2045 out_up_write:
2046         up_write(&mm->mmap_sem);
2047         return;
2048
2049 out:
2050         mem_cgroup_uncharge_page(new_page);
2051         goto out_up_write;
2052 }
2053
2054 static int khugepaged_scan_pmd(struct mm_struct *mm,
2055                                struct vm_area_struct *vma,
2056                                unsigned long address,
2057                                struct page **hpage)
2058 {
2059         pgd_t *pgd;
2060         pud_t *pud;
2061         pmd_t *pmd;
2062         pte_t *pte, *_pte;
2063         int ret = 0, referenced = 0, none = 0;
2064         struct page *page;
2065         unsigned long _address;
2066         spinlock_t *ptl;
2067         int node = -1;
2068
2069         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2070
2071         pgd = pgd_offset(mm, address);
2072         if (!pgd_present(*pgd))
2073                 goto out;
2074
2075         pud = pud_offset(pgd, address);
2076         if (!pud_present(*pud))
2077                 goto out;
2078
2079         pmd = pmd_offset(pud, address);
2080         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
2081                 goto out;
2082
2083         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2084         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2085              _pte++, _address += PAGE_SIZE) {
2086                 pte_t pteval = *_pte;
2087                 if (pte_none(pteval)) {
2088                         if (++none <= khugepaged_max_ptes_none)
2089                                 continue;
2090                         else
2091                                 goto out_unmap;
2092                 }
2093                 if (!pte_present(pteval) || !pte_write(pteval))
2094                         goto out_unmap;
2095                 page = vm_normal_page(vma, _address, pteval);
2096                 if (unlikely(!page))
2097                         goto out_unmap;
2098                 /*
2099                  * Chose the node of the first page. This could
2100                  * be more sophisticated and look at more pages,
2101                  * but isn't for now.
2102                  */
2103                 if (node == -1)
2104                         node = page_to_nid(page);
2105                 VM_BUG_ON(PageCompound(page));
2106                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2107                         goto out_unmap;
2108                 /* cannot use mapcount: can't collapse if there's a gup pin */
2109                 if (page_count(page) != 1)
2110                         goto out_unmap;
2111                 if (pte_young(pteval) || PageReferenced(page) ||
2112                     mmu_notifier_test_young(vma->vm_mm, address))
2113                         referenced = 1;
2114         }
2115         if (referenced)
2116                 ret = 1;
2117 out_unmap:
2118         pte_unmap_unlock(pte, ptl);
2119         if (ret)
2120                 /* collapse_huge_page will return with the mmap_sem released */
2121                 collapse_huge_page(mm, address, hpage, vma, node);
2122 out:
2123         return ret;
2124 }
2125
2126 static void collect_mm_slot(struct mm_slot *mm_slot)
2127 {
2128         struct mm_struct *mm = mm_slot->mm;
2129
2130         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2131
2132         if (khugepaged_test_exit(mm)) {
2133                 /* free mm_slot */
2134                 hlist_del(&mm_slot->hash);
2135                 list_del(&mm_slot->mm_node);
2136
2137                 /*
2138                  * Not strictly needed because the mm exited already.
2139                  *
2140                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2141                  */
2142
2143                 /* khugepaged_mm_lock actually not necessary for the below */
2144                 free_mm_slot(mm_slot);
2145                 mmdrop(mm);
2146         }
2147 }
2148
2149 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2150                                             struct page **hpage)
2151         __releases(&khugepaged_mm_lock)
2152         __acquires(&khugepaged_mm_lock)
2153 {
2154         struct mm_slot *mm_slot;
2155         struct mm_struct *mm;
2156         struct vm_area_struct *vma;
2157         int progress = 0;
2158
2159         VM_BUG_ON(!pages);
2160         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2161
2162         if (khugepaged_scan.mm_slot)
2163                 mm_slot = khugepaged_scan.mm_slot;
2164         else {
2165                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2166                                      struct mm_slot, mm_node);
2167                 khugepaged_scan.address = 0;
2168                 khugepaged_scan.mm_slot = mm_slot;
2169         }
2170         spin_unlock(&khugepaged_mm_lock);
2171
2172         mm = mm_slot->mm;
2173         down_read(&mm->mmap_sem);
2174         if (unlikely(khugepaged_test_exit(mm)))
2175                 vma = NULL;
2176         else
2177                 vma = find_vma(mm, khugepaged_scan.address);
2178
2179         progress++;
2180         for (; vma; vma = vma->vm_next) {
2181                 unsigned long hstart, hend;
2182
2183                 cond_resched();
2184                 if (unlikely(khugepaged_test_exit(mm))) {
2185                         progress++;
2186                         break;
2187                 }
2188
2189                 if ((!(vma->vm_flags & VM_HUGEPAGE) &&
2190                      !khugepaged_always()) ||
2191                     (vma->vm_flags & VM_NOHUGEPAGE)) {
2192                 skip:
2193                         progress++;
2194                         continue;
2195                 }
2196                 if (!vma->anon_vma || vma->vm_ops)
2197                         goto skip;
2198                 if (is_vma_temporary_stack(vma))
2199                         goto skip;
2200                 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2201
2202                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2203                 hend = vma->vm_end & HPAGE_PMD_MASK;
2204                 if (hstart >= hend)
2205                         goto skip;
2206                 if (khugepaged_scan.address > hend)
2207                         goto skip;
2208                 if (khugepaged_scan.address < hstart)
2209                         khugepaged_scan.address = hstart;
2210                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2211
2212                 while (khugepaged_scan.address < hend) {
2213                         int ret;
2214                         cond_resched();
2215                         if (unlikely(khugepaged_test_exit(mm)))
2216                                 goto breakouterloop;
2217
2218                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2219                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2220                                   hend);
2221                         ret = khugepaged_scan_pmd(mm, vma,
2222                                                   khugepaged_scan.address,
2223                                                   hpage);
2224                         /* move to next address */
2225                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2226                         progress += HPAGE_PMD_NR;
2227                         if (ret)
2228                                 /* we released mmap_sem so break loop */
2229                                 goto breakouterloop_mmap_sem;
2230                         if (progress >= pages)
2231                                 goto breakouterloop;
2232                 }
2233         }
2234 breakouterloop:
2235         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2236 breakouterloop_mmap_sem:
2237
2238         spin_lock(&khugepaged_mm_lock);
2239         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2240         /*
2241          * Release the current mm_slot if this mm is about to die, or
2242          * if we scanned all vmas of this mm.
2243          */
2244         if (khugepaged_test_exit(mm) || !vma) {
2245                 /*
2246                  * Make sure that if mm_users is reaching zero while
2247                  * khugepaged runs here, khugepaged_exit will find
2248                  * mm_slot not pointing to the exiting mm.
2249                  */
2250                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2251                         khugepaged_scan.mm_slot = list_entry(
2252                                 mm_slot->mm_node.next,
2253                                 struct mm_slot, mm_node);
2254                         khugepaged_scan.address = 0;
2255                 } else {
2256                         khugepaged_scan.mm_slot = NULL;
2257                         khugepaged_full_scans++;
2258                 }
2259
2260                 collect_mm_slot(mm_slot);
2261         }
2262
2263         return progress;
2264 }
2265
2266 static int khugepaged_has_work(void)
2267 {
2268         return !list_empty(&khugepaged_scan.mm_head) &&
2269                 khugepaged_enabled();
2270 }
2271
2272 static int khugepaged_wait_event(void)
2273 {
2274         return !list_empty(&khugepaged_scan.mm_head) ||
2275                 kthread_should_stop();
2276 }
2277
2278 static void khugepaged_do_scan(void)
2279 {
2280         struct page *hpage = NULL;
2281         unsigned int progress = 0, pass_through_head = 0;
2282         unsigned int pages = khugepaged_pages_to_scan;
2283         bool wait = true;
2284
2285         barrier(); /* write khugepaged_pages_to_scan to local stack */
2286
2287         while (progress < pages) {
2288                 if (!khugepaged_prealloc_page(&hpage, &wait))
2289                         break;
2290
2291                 cond_resched();
2292
2293                 if (unlikely(kthread_should_stop() || freezing(current)))
2294                         break;
2295
2296                 spin_lock(&khugepaged_mm_lock);
2297                 if (!khugepaged_scan.mm_slot)
2298                         pass_through_head++;
2299                 if (khugepaged_has_work() &&
2300                     pass_through_head < 2)
2301                         progress += khugepaged_scan_mm_slot(pages - progress,
2302                                                             &hpage);
2303                 else
2304                         progress = pages;
2305                 spin_unlock(&khugepaged_mm_lock);
2306         }
2307
2308         if (!IS_ERR_OR_NULL(hpage))
2309                 put_page(hpage);
2310 }
2311
2312 static void khugepaged_wait_work(void)
2313 {
2314         try_to_freeze();
2315
2316         if (khugepaged_has_work()) {
2317                 if (!khugepaged_scan_sleep_millisecs)
2318                         return;
2319
2320                 wait_event_freezable_timeout(khugepaged_wait,
2321                                              kthread_should_stop(),
2322                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2323                 return;
2324         }
2325
2326         if (khugepaged_enabled())
2327                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2328 }
2329
2330 static int khugepaged(void *none)
2331 {
2332         struct mm_slot *mm_slot;
2333
2334         set_freezable();
2335         set_user_nice(current, 19);
2336
2337         while (!kthread_should_stop()) {
2338                 khugepaged_do_scan();
2339                 khugepaged_wait_work();
2340         }
2341
2342         spin_lock(&khugepaged_mm_lock);
2343         mm_slot = khugepaged_scan.mm_slot;
2344         khugepaged_scan.mm_slot = NULL;
2345         if (mm_slot)
2346                 collect_mm_slot(mm_slot);
2347         spin_unlock(&khugepaged_mm_lock);
2348         return 0;
2349 }
2350
2351 void __split_huge_page_pmd(struct mm_struct *mm, pmd_t *pmd)
2352 {
2353         struct page *page;
2354
2355         spin_lock(&mm->page_table_lock);
2356         if (unlikely(!pmd_trans_huge(*pmd))) {
2357                 spin_unlock(&mm->page_table_lock);
2358                 return;
2359         }
2360         page = pmd_page(*pmd);
2361         VM_BUG_ON(!page_count(page));
2362         get_page(page);
2363         spin_unlock(&mm->page_table_lock);
2364
2365         split_huge_page(page);
2366
2367         put_page(page);
2368         BUG_ON(pmd_trans_huge(*pmd));
2369 }
2370
2371 static void split_huge_page_address(struct mm_struct *mm,
2372                                     unsigned long address)
2373 {
2374         pgd_t *pgd;
2375         pud_t *pud;
2376         pmd_t *pmd;
2377
2378         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2379
2380         pgd = pgd_offset(mm, address);
2381         if (!pgd_present(*pgd))
2382                 return;
2383
2384         pud = pud_offset(pgd, address);
2385         if (!pud_present(*pud))
2386                 return;
2387
2388         pmd = pmd_offset(pud, address);
2389         if (!pmd_present(*pmd))
2390                 return;
2391         /*
2392          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2393          * materialize from under us.
2394          */
2395         split_huge_page_pmd(mm, pmd);
2396 }
2397
2398 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2399                              unsigned long start,
2400                              unsigned long end,
2401                              long adjust_next)
2402 {
2403         /*
2404          * If the new start address isn't hpage aligned and it could
2405          * previously contain an hugepage: check if we need to split
2406          * an huge pmd.
2407          */
2408         if (start & ~HPAGE_PMD_MASK &&
2409             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2410             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2411                 split_huge_page_address(vma->vm_mm, start);
2412
2413         /*
2414          * If the new end address isn't hpage aligned and it could
2415          * previously contain an hugepage: check if we need to split
2416          * an huge pmd.
2417          */
2418         if (end & ~HPAGE_PMD_MASK &&
2419             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2420             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2421                 split_huge_page_address(vma->vm_mm, end);
2422
2423         /*
2424          * If we're also updating the vma->vm_next->vm_start, if the new
2425          * vm_next->vm_start isn't page aligned and it could previously
2426          * contain an hugepage: check if we need to split an huge pmd.
2427          */
2428         if (adjust_next > 0) {
2429                 struct vm_area_struct *next = vma->vm_next;
2430                 unsigned long nstart = next->vm_start;
2431                 nstart += adjust_next << PAGE_SHIFT;
2432                 if (nstart & ~HPAGE_PMD_MASK &&
2433                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2434                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2435                         split_huge_page_address(next->vm_mm, nstart);
2436         }
2437 }