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 (!test_bit(TRANSPARENT_HUGEPAGE_FLAG,
 106                       &transparent_hugepage_flags) &&
 107             !test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 108                       &transparent_hugepage_flags))
 109                 return 0;
 110
 111         for_each_populated_zone(zone)
 112                 nr_zones++;
 113
 114         /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
 115         recommended_min = pageblock_nr_pages * nr_zones * 2;
 116
 117         /*
 118          * Make sure that on average at least two pageblocks are almost free
 119          * of another type, one for a migratetype to fall back to and a
 120          * second to avoid subsequent fallbacks of other types There are 3
 121          * MIGRATE_TYPES we care about.
 122          */
 123         recommended_min += pageblock_nr_pages * nr_zones *
 124                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
 125
 126         /* don't ever allow to reserve more than 5% of the lowmem */
 127         recommended_min = min(recommended_min,
 128                               (unsigned long) nr_free_buffer_pages() / 20);
 129         recommended_min <<= (PAGE_SHIFT-10);
 130
 131         if (recommended_min > min_free_kbytes)
 132                 min_free_kbytes = recommended_min;
 133         setup_per_zone_wmarks();
 134         return 0;
 135 }
 136 late_initcall(set_recommended_min_free_kbytes);
 137
 138 static int start_khugepaged(void)
 139 {
 140         int err = 0;
 141         if (khugepaged_enabled()) {
 142                 int wakeup;
 143                 if (unlikely(!mm_slot_cache || !mm_slots_hash)) {
 144                         err = -ENOMEM;
 145                         goto out;
 146                 }
 147                 mutex_lock(&khugepaged_mutex);
 148                 if (!khugepaged_thread)
 149                         khugepaged_thread = kthread_run(khugepaged, NULL,
 150                                                         "khugepaged");
 151                 if (unlikely(IS_ERR(khugepaged_thread))) {
 152                         printk(KERN_ERR
 153                                "khugepaged: kthread_run(khugepaged) failed\n");
 154                         err = PTR_ERR(khugepaged_thread);
 155                         khugepaged_thread = NULL;
 156                 }
 157                 wakeup = !list_empty(&khugepaged_scan.mm_head);
 158                 mutex_unlock(&khugepaged_mutex);
 159                 if (wakeup)
 160                         wake_up_interruptible(&khugepaged_wait);
 161
 162                 set_recommended_min_free_kbytes();
 163         } else
 164                 /* wakeup to exit */
 165                 wake_up_interruptible(&khugepaged_wait);
 166 out:
 167         return err;
 168 }
 169
 170 #ifdef CONFIG_SYSFS
 171
 172 static ssize_t double_flag_show(struct kobject *kobj,
 173                                 struct kobj_attribute *attr, char *buf,
 174                                 enum transparent_hugepage_flag enabled,
 175                                 enum transparent_hugepage_flag req_madv)
 176 {
 177         if (test_bit(enabled, &transparent_hugepage_flags)) {
 178                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
 179                 return sprintf(buf, "[always] madvise never\n");
 180         } else if (test_bit(req_madv, &transparent_hugepage_flags))
 181                 return sprintf(buf, "always [madvise] never\n");
 182         else
 183                 return sprintf(buf, "always madvise [never]\n");
 184 }
 185 static ssize_t double_flag_store(struct kobject *kobj,
 186                                  struct kobj_attribute *attr,
 187                                  const char *buf, size_t count,
 188                                  enum transparent_hugepage_flag enabled,
 189                                  enum transparent_hugepage_flag req_madv)
 190 {
 191         if (!memcmp("always", buf,
 192                     min(sizeof("always")-1, count))) {
 193                 set_bit(enabled, &transparent_hugepage_flags);
 194                 clear_bit(req_madv, &transparent_hugepage_flags);
 195         } else if (!memcmp("madvise", buf,
 196                            min(sizeof("madvise")-1, count))) {
 197                 clear_bit(enabled, &transparent_hugepage_flags);
 198                 set_bit(req_madv, &transparent_hugepage_flags);
 199         } else if (!memcmp("never", buf,
 200                            min(sizeof("never")-1, count))) {
 201                 clear_bit(enabled, &transparent_hugepage_flags);
 202                 clear_bit(req_madv, &transparent_hugepage_flags);
 203         } else
 204                 return -EINVAL;
 205
 206         return count;
 207 }
 208
 209 static ssize_t enabled_show(struct kobject *kobj,
 210                             struct kobj_attribute *attr, char *buf)
 211 {
 212         return double_flag_show(kobj, attr, buf,
 213                                 TRANSPARENT_HUGEPAGE_FLAG,
 214                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 215 }
 216 static ssize_t enabled_store(struct kobject *kobj,
 217                              struct kobj_attribute *attr,
 218                              const char *buf, size_t count)
 219 {
 220         ssize_t ret;
 221
 222         ret = double_flag_store(kobj, attr, buf, count,
 223                                 TRANSPARENT_HUGEPAGE_FLAG,
 224                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 225
 226         if (ret > 0) {
 227                 int err = start_khugepaged();
 228                 if (err)
 229                         ret = err;
 230         }
 231
 232         if (ret > 0 &&
 233             (test_bit(TRANSPARENT_HUGEPAGE_FLAG,
 234                       &transparent_hugepage_flags) ||
 235              test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 236                       &transparent_hugepage_flags)))
 237                 set_recommended_min_free_kbytes();
 238
 239         return ret;
 240 }
 241 static struct kobj_attribute enabled_attr =
 242         __ATTR(enabled, 0644, enabled_show, enabled_store);
 243
 244 static ssize_t single_flag_show(struct kobject *kobj,
 245                                 struct kobj_attribute *attr, char *buf,
 246                                 enum transparent_hugepage_flag flag)
 247 {
 248         return sprintf(buf, "%d\n",
 249                        !!test_bit(flag, &transparent_hugepage_flags));
 250 }
 251
 252 static ssize_t single_flag_store(struct kobject *kobj,
 253                                  struct kobj_attribute *attr,
 254                                  const char *buf, size_t count,
 255                                  enum transparent_hugepage_flag flag)
 256 {
 257         unsigned long value;
 258         int ret;
 259
 260         ret = kstrtoul(buf, 10, &value);
 261         if (ret < 0)
 262                 return ret;
 263         if (value > 1)
 264                 return -EINVAL;
 265
 266         if (value)
 267                 set_bit(flag, &transparent_hugepage_flags);
 268         else
 269                 clear_bit(flag, &transparent_hugepage_flags);
 270
 271         return count;
 272 }
 273
 274 /*
 275  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 276  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 277  * memory just to allocate one more hugepage.
 278  */
 279 static ssize_t defrag_show(struct kobject *kobj,
 280                            struct kobj_attribute *attr, char *buf)
 281 {
 282         return double_flag_show(kobj, attr, buf,
 283                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 284                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 285 }
 286 static ssize_t defrag_store(struct kobject *kobj,
 287                             struct kobj_attribute *attr,
 288                             const char *buf, size_t count)
 289 {
 290         return double_flag_store(kobj, attr, buf, count,
 291                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 292                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 293 }
 294 static struct kobj_attribute defrag_attr =
 295         __ATTR(defrag, 0644, defrag_show, defrag_store);
 296
 297 #ifdef CONFIG_DEBUG_VM
 298 static ssize_t debug_cow_show(struct kobject *kobj,
 299                                 struct kobj_attribute *attr, char *buf)
 300 {
 301         return single_flag_show(kobj, attr, buf,
 302                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 303 }
 304 static ssize_t debug_cow_store(struct kobject *kobj,
 305                                struct kobj_attribute *attr,
 306                                const char *buf, size_t count)
 307 {
 308         return single_flag_store(kobj, attr, buf, count,
 309                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 310 }
 311 static struct kobj_attribute debug_cow_attr =
 312         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 313 #endif /* CONFIG_DEBUG_VM */
 314
 315 static struct attribute *hugepage_attr[] = {
 316         &enabled_attr.attr,
 317         &defrag_attr.attr,
 318 #ifdef CONFIG_DEBUG_VM
 319         &debug_cow_attr.attr,
 320 #endif
 321         NULL,
 322 };
 323
 324 static struct attribute_group hugepage_attr_group = {
 325         .attrs = hugepage_attr,
 326 };
 327
 328 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 329                                          struct kobj_attribute *attr,
 330                                          char *buf)
 331 {
 332         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 333 }
 334
 335 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 336                                           struct kobj_attribute *attr,
 337                                           const char *buf, size_t count)
 338 {
 339         unsigned long msecs;
 340         int err;
 341
 342         err = strict_strtoul(buf, 10, &msecs);
 343         if (err || msecs > UINT_MAX)
 344                 return -EINVAL;
 345
 346         khugepaged_scan_sleep_millisecs = msecs;
 347         wake_up_interruptible(&khugepaged_wait);
 348
 349         return count;
 350 }
 351 static struct kobj_attribute scan_sleep_millisecs_attr =
 352         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 353                scan_sleep_millisecs_store);
 354
 355 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 356                                           struct kobj_attribute *attr,
 357                                           char *buf)
 358 {
 359         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 360 }
 361
 362 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 363                                            struct kobj_attribute *attr,
 364                                            const char *buf, size_t count)
 365 {
 366         unsigned long msecs;
 367         int err;
 368
 369         err = strict_strtoul(buf, 10, &msecs);
 370         if (err || msecs > UINT_MAX)
 371                 return -EINVAL;
 372
 373         khugepaged_alloc_sleep_millisecs = msecs;
 374         wake_up_interruptible(&khugepaged_wait);
 375
 376         return count;
 377 }
 378 static struct kobj_attribute alloc_sleep_millisecs_attr =
 379         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 380                alloc_sleep_millisecs_store);
 381
 382 static ssize_t pages_to_scan_show(struct kobject *kobj,
 383                                   struct kobj_attribute *attr,
 384                                   char *buf)
 385 {
 386         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 387 }
 388 static ssize_t pages_to_scan_store(struct kobject *kobj,
 389                                    struct kobj_attribute *attr,
 390                                    const char *buf, size_t count)
 391 {
 392         int err;
 393         unsigned long pages;
 394
 395         err = strict_strtoul(buf, 10, &pages);
 396         if (err || !pages || pages > UINT_MAX)
 397                 return -EINVAL;
 398
 399         khugepaged_pages_to_scan = pages;
 400
 401         return count;
 402 }
 403 static struct kobj_attribute pages_to_scan_attr =
 404         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 405                pages_to_scan_store);
 406
 407 static ssize_t pages_collapsed_show(struct kobject *kobj,
 408                                     struct kobj_attribute *attr,
 409                                     char *buf)
 410 {
 411         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 412 }
 413 static struct kobj_attribute pages_collapsed_attr =
 414         __ATTR_RO(pages_collapsed);
 415
 416 static ssize_t full_scans_show(struct kobject *kobj,
 417                                struct kobj_attribute *attr,
 418                                char *buf)
 419 {
 420         return sprintf(buf, "%u\n", khugepaged_full_scans);
 421 }
 422 static struct kobj_attribute full_scans_attr =
 423         __ATTR_RO(full_scans);
 424
 425 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 426                                       struct kobj_attribute *attr, char *buf)
 427 {
 428         return single_flag_show(kobj, attr, buf,
 429                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 430 }
 431 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 432                                        struct kobj_attribute *attr,
 433                                        const char *buf, size_t count)
 434 {
 435         return single_flag_store(kobj, attr, buf, count,
 436                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 437 }
 438 static struct kobj_attribute khugepaged_defrag_attr =
 439         __ATTR(defrag, 0644, khugepaged_defrag_show,
 440                khugepaged_defrag_store);
 441
 442 /*
 443  * max_ptes_none controls if khugepaged should collapse hugepages over
 444  * any unmapped ptes in turn potentially increasing the memory
 445  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 446  * reduce the available free memory in the system as it
 447  * runs. Increasing max_ptes_none will instead potentially reduce the
 448  * free memory in the system during the khugepaged scan.
 449  */
 450 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 451                                              struct kobj_attribute *attr,
 452                                              char *buf)
 453 {
 454         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 455 }
 456 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 457                                               struct kobj_attribute *attr,
 458                                               const char *buf, size_t count)
 459 {
 460         int err;
 461         unsigned long max_ptes_none;
 462
 463         err = strict_strtoul(buf, 10, &max_ptes_none);
 464         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 465                 return -EINVAL;
 466
 467         khugepaged_max_ptes_none = max_ptes_none;
 468
 469         return count;
 470 }
 471 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 472         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 473                khugepaged_max_ptes_none_store);
 474
 475 static struct attribute *khugepaged_attr[] = {
 476         &khugepaged_defrag_attr.attr,
 477         &khugepaged_max_ptes_none_attr.attr,
 478         &pages_to_scan_attr.attr,
 479         &pages_collapsed_attr.attr,
 480         &full_scans_attr.attr,
 481         &scan_sleep_millisecs_attr.attr,
 482         &alloc_sleep_millisecs_attr.attr,
 483         NULL,
 484 };
 485
 486 static struct attribute_group khugepaged_attr_group = {
 487         .attrs = khugepaged_attr,
 488         .name = "khugepaged",
 489 };
 490 #endif /* CONFIG_SYSFS */
 491
 492 static int __init hugepage_init(void)
 493 {
 494         int err;
 495 #ifdef CONFIG_SYSFS
 496         static struct kobject *hugepage_kobj;
 497 #endif
 498
 499         err = -EINVAL;
 500         if (!has_transparent_hugepage()) {
 501                 transparent_hugepage_flags = 0;
 502                 goto out;
 503         }
 504
 505 #ifdef CONFIG_SYSFS
 506         err = -ENOMEM;
 507         hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 508         if (unlikely(!hugepage_kobj)) {
 509                 printk(KERN_ERR "hugepage: failed kobject create\n");
 510                 goto out;
 511         }
 512
 513         err = sysfs_create_group(hugepage_kobj, &hugepage_attr_group);
 514         if (err) {
 515                 printk(KERN_ERR "hugepage: failed register hugeage group\n");
 516                 goto out;
 517         }
 518
 519         err = sysfs_create_group(hugepage_kobj, &khugepaged_attr_group);
 520         if (err) {
 521                 printk(KERN_ERR "hugepage: failed register hugeage group\n");
 522                 goto out;
 523         }
 524 #endif
 525
 526         err = khugepaged_slab_init();
 527         if (err)
 528                 goto out;
 529
 530         err = mm_slots_hash_init();
 531         if (err) {
 532                 khugepaged_slab_free();
 533                 goto out;
 534         }
 535
 536         /*
 537          * By default disable transparent hugepages on smaller systems,
 538          * where the extra memory used could hurt more than TLB overhead
 539          * is likely to save.  The admin can still enable it through /sys.
 540          */
 541         if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
 542                 transparent_hugepage_flags = 0;
 543
 544         start_khugepaged();
 545
 546         set_recommended_min_free_kbytes();
 547
 548 out:
 549         return err;
 550 }
 551 module_init(hugepage_init)
 552
 553 static int __init setup_transparent_hugepage(char *str)
 554 {
 555         int ret = 0;
 556         if (!str)
 557                 goto out;
 558         if (!strcmp(str, "always")) {
 559                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 560                         &transparent_hugepage_flags);
 561                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 562                           &transparent_hugepage_flags);
 563                 ret = 1;
 564         } else if (!strcmp(str, "madvise")) {
 565                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 566                           &transparent_hugepage_flags);
 567                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 568                         &transparent_hugepage_flags);
 569                 ret = 1;
 570         } else if (!strcmp(str, "never")) {
 571                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 572                           &transparent_hugepage_flags);
 573                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 574                           &transparent_hugepage_flags);
 575                 ret = 1;
 576         }
 577 out:
 578         if (!ret)
 579                 printk(KERN_WARNING
 580                        "transparent_hugepage= cannot parse, ignored\n");
 581         return ret;
 582 }
 583 __setup("transparent_hugepage=", setup_transparent_hugepage);
 584
 585 static void prepare_pmd_huge_pte(pgtable_t pgtable,
 586                                  struct mm_struct *mm)
 587 {
 588         assert_spin_locked(&mm->page_table_lock);
 589
 590         /* FIFO */
 591         if (!mm->pmd_huge_pte)
 592                 INIT_LIST_HEAD(&pgtable->lru);
 593         else
 594                 list_add(&pgtable->lru, &mm->pmd_huge_pte->lru);
 595         mm->pmd_huge_pte = pgtable;
 596 }
 597
 598 static inline pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 599 {
 600         if (likely(vma->vm_flags & VM_WRITE))
 601                 pmd = pmd_mkwrite(pmd);
 602         return pmd;
 603 }
 604
 605 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
 606                                         struct vm_area_struct *vma,
 607                                         unsigned long haddr, pmd_t *pmd,
 608                                         struct page *page)
 609 {
 610         int ret = 0;
 611         pgtable_t pgtable;
 612
 613         VM_BUG_ON(!PageCompound(page));
 614         pgtable = pte_alloc_one(mm, haddr);
 615         if (unlikely(!pgtable)) {
 616                 mem_cgroup_uncharge_page(page);
 617                 put_page(page);
 618                 return VM_FAULT_OOM;
 619         }
 620
 621         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 622         __SetPageUptodate(page);
 623
 624         spin_lock(&mm->page_table_lock);
 625         if (unlikely(!pmd_none(*pmd))) {
 626                 spin_unlock(&mm->page_table_lock);
 627                 mem_cgroup_uncharge_page(page);
 628                 put_page(page);
 629                 pte_free(mm, pgtable);
 630         } else {
 631                 pmd_t entry;
 632                 entry = mk_pmd(page, vma->vm_page_prot);
 633                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 634                 entry = pmd_mkhuge(entry);
 635                 /*
 636                  * The spinlocking to take the lru_lock inside
 637                  * page_add_new_anon_rmap() acts as a full memory
 638                  * barrier to be sure clear_huge_page writes become
 639                  * visible after the set_pmd_at() write.
 640                  */
 641                 page_add_new_anon_rmap(page, vma, haddr);
 642                 set_pmd_at(mm, haddr, pmd, entry);
 643                 prepare_pmd_huge_pte(pgtable, mm);
 644                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 645                 mm->nr_ptes++;
 646                 spin_unlock(&mm->page_table_lock);
 647         }
 648
 649         return ret;
 650 }
 651
 652 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 653 {
 654         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
 655 }
 656
 657 static inline struct page *alloc_hugepage_vma(int defrag,
 658                                               struct vm_area_struct *vma,
 659                                               unsigned long haddr, int nd,
 660                                               gfp_t extra_gfp)
 661 {
 662         return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
 663                                HPAGE_PMD_ORDER, vma, haddr, nd);
 664 }
 665
 666 #ifndef CONFIG_NUMA
 667 static inline struct page *alloc_hugepage(int defrag)
 668 {
 669         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
 670                            HPAGE_PMD_ORDER);
 671 }
 672 #endif
 673
 674 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 675                                unsigned long address, pmd_t *pmd,
 676                                unsigned int flags)
 677 {
 678         struct page *page;
 679         unsigned long haddr = address & HPAGE_PMD_MASK;
 680         pte_t *pte;
 681
 682         if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) {
 683                 if (unlikely(anon_vma_prepare(vma)))
 684                         return VM_FAULT_OOM;
 685                 if (unlikely(khugepaged_enter(vma)))
 686                         return VM_FAULT_OOM;
 687                 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 688                                           vma, haddr, numa_node_id(), 0);
 689                 if (unlikely(!page)) {
 690                         count_vm_event(THP_FAULT_FALLBACK);
 691                         goto out;
 692                 }
 693                 count_vm_event(THP_FAULT_ALLOC);
 694                 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
 695                         put_page(page);
 696                         goto out;
 697                 }
 698
 699                 return __do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page);
 700         }
 701 out:
 702         /*
 703          * Use __pte_alloc instead of pte_alloc_map, because we can't
 704          * run pte_offset_map on the pmd, if an huge pmd could
 705          * materialize from under us from a different thread.
 706          */
 707         if (unlikely(__pte_alloc(mm, vma, pmd, address)))
 708                 return VM_FAULT_OOM;
 709         /* if an huge pmd materialized from under us just retry later */
 710         if (unlikely(pmd_trans_huge(*pmd)))
 711                 return 0;
 712         /*
 713          * A regular pmd is established and it can't morph into a huge pmd
 714          * from under us anymore at this point because we hold the mmap_sem
 715          * read mode and khugepaged takes it in write mode. So now it's
 716          * safe to run pte_offset_map().
 717          */
 718         pte = pte_offset_map(pmd, address);
 719         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
 720 }
 721
 722 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 723                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 724                   struct vm_area_struct *vma)
 725 {
 726         struct page *src_page;
 727         pmd_t pmd;
 728         pgtable_t pgtable;
 729         int ret;
 730
 731         ret = -ENOMEM;
 732         pgtable = pte_alloc_one(dst_mm, addr);
 733         if (unlikely(!pgtable))
 734                 goto out;
 735
 736         spin_lock(&dst_mm->page_table_lock);
 737         spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
 738
 739         ret = -EAGAIN;
 740         pmd = *src_pmd;
 741         if (unlikely(!pmd_trans_huge(pmd))) {
 742                 pte_free(dst_mm, pgtable);
 743                 goto out_unlock;
 744         }
 745         if (unlikely(pmd_trans_splitting(pmd))) {
 746                 /* split huge page running from under us */
 747                 spin_unlock(&src_mm->page_table_lock);
 748                 spin_unlock(&dst_mm->page_table_lock);
 749                 pte_free(dst_mm, pgtable);
 750
 751                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 752                 goto out;
 753         }
 754         src_page = pmd_page(pmd);
 755         VM_BUG_ON(!PageHead(src_page));
 756         get_page(src_page);
 757         page_dup_rmap(src_page);
 758         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 759
 760         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 761         pmd = pmd_mkold(pmd_wrprotect(pmd));
 762         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 763         prepare_pmd_huge_pte(pgtable, dst_mm);
 764         dst_mm->nr_ptes++;
 765
 766         ret = 0;
 767 out_unlock:
 768         spin_unlock(&src_mm->page_table_lock);
 769         spin_unlock(&dst_mm->page_table_lock);
 770 out:
 771         return ret;
 772 }
 773
 774 /* no "address" argument so destroys page coloring of some arch */
 775 pgtable_t get_pmd_huge_pte(struct mm_struct *mm)
 776 {
 777         pgtable_t pgtable;
 778
 779         assert_spin_locked(&mm->page_table_lock);
 780
 781         /* FIFO */
 782         pgtable = mm->pmd_huge_pte;
 783         if (list_empty(&pgtable->lru))
 784                 mm->pmd_huge_pte = NULL;
 785         else {
 786                 mm->pmd_huge_pte = list_entry(pgtable->lru.next,
 787                                               struct page, lru);
 788                 list_del(&pgtable->lru);
 789         }
 790         return pgtable;
 791 }
 792
 793 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
 794                                         struct vm_area_struct *vma,
 795                                         unsigned long address,
 796                                         pmd_t *pmd, pmd_t orig_pmd,
 797                                         struct page *page,
 798                                         unsigned long haddr)
 799 {
 800         pgtable_t pgtable;
 801         pmd_t _pmd;
 802         int ret = 0, i;
 803         struct page **pages;
 804
 805         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
 806                         GFP_KERNEL);
 807         if (unlikely(!pages)) {
 808                 ret |= VM_FAULT_OOM;
 809                 goto out;
 810         }
 811
 812         for (i = 0; i < HPAGE_PMD_NR; i++) {
 813                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
 814                                                __GFP_OTHER_NODE,
 815                                                vma, address, page_to_nid(page));
 816                 if (unlikely(!pages[i] ||
 817                              mem_cgroup_newpage_charge(pages[i], mm,
 818                                                        GFP_KERNEL))) {
 819                         if (pages[i])
 820                                 put_page(pages[i]);
 821                         mem_cgroup_uncharge_start();
 822                         while (--i >= 0) {
 823                                 mem_cgroup_uncharge_page(pages[i]);
 824                                 put_page(pages[i]);
 825                         }
 826                         mem_cgroup_uncharge_end();
 827                         kfree(pages);
 828                         ret |= VM_FAULT_OOM;
 829                         goto out;
 830                 }
 831         }
 832
 833         for (i = 0; i < HPAGE_PMD_NR; i++) {
 834                 copy_user_highpage(pages[i], page + i,
 835                                    haddr + PAGE_SIZE * i, vma);
 836                 __SetPageUptodate(pages[i]);
 837                 cond_resched();
 838         }
 839
 840         spin_lock(&mm->page_table_lock);
 841         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 842                 goto out_free_pages;
 843         VM_BUG_ON(!PageHead(page));
 844
 845         pmdp_clear_flush_notify(vma, haddr, pmd);
 846         /* leave pmd empty until pte is filled */
 847
 848         pgtable = get_pmd_huge_pte(mm);
 849         pmd_populate(mm, &_pmd, pgtable);
 850
 851         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
 852                 pte_t *pte, entry;
 853                 entry = mk_pte(pages[i], vma->vm_page_prot);
 854                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
 855                 page_add_new_anon_rmap(pages[i], vma, haddr);
 856                 pte = pte_offset_map(&_pmd, haddr);
 857                 VM_BUG_ON(!pte_none(*pte));
 858                 set_pte_at(mm, haddr, pte, entry);
 859                 pte_unmap(pte);
 860         }
 861         kfree(pages);
 862
 863         smp_wmb(); /* make pte visible before pmd */
 864         pmd_populate(mm, pmd, pgtable);
 865         page_remove_rmap(page);
 866         spin_unlock(&mm->page_table_lock);
 867
 868         ret |= VM_FAULT_WRITE;
 869         put_page(page);
 870
 871 out:
 872         return ret;
 873
 874 out_free_pages:
 875         spin_unlock(&mm->page_table_lock);
 876         mem_cgroup_uncharge_start();
 877         for (i = 0; i < HPAGE_PMD_NR; i++) {
 878                 mem_cgroup_uncharge_page(pages[i]);
 879                 put_page(pages[i]);
 880         }
 881         mem_cgroup_uncharge_end();
 882         kfree(pages);
 883         goto out;
 884 }
 885
 886 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
 887                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
 888 {
 889         int ret = 0;
 890         struct page *page, *new_page;
 891         unsigned long haddr;
 892
 893         VM_BUG_ON(!vma->anon_vma);
 894         spin_lock(&mm->page_table_lock);
 895         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 896                 goto out_unlock;
 897
 898         page = pmd_page(orig_pmd);
 899         VM_BUG_ON(!PageCompound(page) || !PageHead(page));
 900         haddr = address & HPAGE_PMD_MASK;
 901         if (page_mapcount(page) == 1) {
 902                 pmd_t entry;
 903                 entry = pmd_mkyoung(orig_pmd);
 904                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 905                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
 906                         update_mmu_cache(vma, address, entry);
 907                 ret |= VM_FAULT_WRITE;
 908                 goto out_unlock;
 909         }
 910         get_page(page);
 911         spin_unlock(&mm->page_table_lock);
 912
 913         if (transparent_hugepage_enabled(vma) &&
 914             !transparent_hugepage_debug_cow())
 915                 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 916                                               vma, haddr, numa_node_id(), 0);
 917         else
 918                 new_page = NULL;
 919
 920         if (unlikely(!new_page)) {
 921                 count_vm_event(THP_FAULT_FALLBACK);
 922                 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
 923                                                    pmd, orig_pmd, page, haddr);
 924                 put_page(page);
 925                 goto out;
 926         }
 927         count_vm_event(THP_FAULT_ALLOC);
 928
 929         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
 930                 put_page(new_page);
 931                 put_page(page);
 932                 ret |= VM_FAULT_OOM;
 933                 goto out;
 934         }
 935
 936         copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
 937         __SetPageUptodate(new_page);
 938
 939         spin_lock(&mm->page_table_lock);
 940         put_page(page);
 941         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
 942                 mem_cgroup_uncharge_page(new_page);
 943                 put_page(new_page);
 944         } else {
 945                 pmd_t entry;
 946                 VM_BUG_ON(!PageHead(page));
 947                 entry = mk_pmd(new_page, vma->vm_page_prot);
 948                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 949                 entry = pmd_mkhuge(entry);
 950                 pmdp_clear_flush_notify(vma, haddr, pmd);
 951                 page_add_new_anon_rmap(new_page, vma, haddr);
 952                 set_pmd_at(mm, haddr, pmd, entry);
 953                 update_mmu_cache(vma, address, entry);
 954                 page_remove_rmap(page);
 955                 put_page(page);
 956                 ret |= VM_FAULT_WRITE;
 957         }
 958 out_unlock:
 959         spin_unlock(&mm->page_table_lock);
 960 out:
 961         return ret;
 962 }
 963
 964 struct page *follow_trans_huge_pmd(struct mm_struct *mm,
 965                                    unsigned long addr,
 966                                    pmd_t *pmd,
 967                                    unsigned int flags)
 968 {
 969         struct page *page = NULL;
 970
 971         assert_spin_locked(&mm->page_table_lock);
 972
 973         if (flags & FOLL_WRITE && !pmd_write(*pmd))
 974                 goto out;
 975
 976         page = pmd_page(*pmd);
 977         VM_BUG_ON(!PageHead(page));
 978         if (flags & FOLL_TOUCH) {
 979                 pmd_t _pmd;
 980                 /*
 981                  * We should set the dirty bit only for FOLL_WRITE but
 982                  * for now the dirty bit in the pmd is meaningless.
 983                  * And if the dirty bit will become meaningful and
 984                  * we'll only set it with FOLL_WRITE, an atomic
 985                  * set_bit will be required on the pmd to set the
 986                  * young bit, instead of the current set_pmd_at.
 987                  */
 988                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
 989                 set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd);
 990         }
 991         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
 992         VM_BUG_ON(!PageCompound(page));
 993         if (flags & FOLL_GET)
 994                 get_page_foll(page);
 995
 996 out:
 997         return page;
 998 }
 999
1000 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1001                  pmd_t *pmd)
1002 {
1003         int ret = 0;
1004
1005         spin_lock(&tlb->mm->page_table_lock);
1006         if (likely(pmd_trans_huge(*pmd))) {
1007                 if (unlikely(pmd_trans_splitting(*pmd))) {
1008                         spin_unlock(&tlb->mm->page_table_lock);
1009                         wait_split_huge_page(vma->anon_vma,
1010                                              pmd);
1011                 } else {
1012                         struct page *page;
1013                         pgtable_t pgtable;
1014                         pgtable = get_pmd_huge_pte(tlb->mm);
1015                         page = pmd_page(*pmd);
1016                         pmd_clear(pmd);
1017                         page_remove_rmap(page);
1018                         VM_BUG_ON(page_mapcount(page) < 0);
1019                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1020                         VM_BUG_ON(!PageHead(page));
1021                         tlb->mm->nr_ptes--;
1022                         spin_unlock(&tlb->mm->page_table_lock);
1023                         tlb_remove_page(tlb, page);
1024                         pte_free(tlb->mm, pgtable);
1025                         ret = 1;
1026                 }
1027         } else
1028                 spin_unlock(&tlb->mm->page_table_lock);
1029
1030         return ret;
1031 }
1032
1033 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1034                 unsigned long addr, unsigned long end,
1035                 unsigned char *vec)
1036 {
1037         int ret = 0;
1038
1039         spin_lock(&vma->vm_mm->page_table_lock);
1040         if (likely(pmd_trans_huge(*pmd))) {
1041                 ret = !pmd_trans_splitting(*pmd);
1042                 spin_unlock(&vma->vm_mm->page_table_lock);
1043                 if (unlikely(!ret))
1044                         wait_split_huge_page(vma->anon_vma, pmd);
1045                 else {
1046                         /*
1047                          * All logical pages in the range are present
1048                          * if backed by a huge page.
1049                          */
1050                         memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1051                 }
1052         } else
1053                 spin_unlock(&vma->vm_mm->page_table_lock);
1054
1055         return ret;
1056 }
1057
1058 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1059                   unsigned long old_addr,
1060                   unsigned long new_addr, unsigned long old_end,
1061                   pmd_t *old_pmd, pmd_t *new_pmd)
1062 {
1063         int ret = 0;
1064         pmd_t pmd;
1065
1066         struct mm_struct *mm = vma->vm_mm;
1067
1068         if ((old_addr & ~HPAGE_PMD_MASK) ||
1069             (new_addr & ~HPAGE_PMD_MASK) ||
1070             old_end - old_addr < HPAGE_PMD_SIZE ||
1071             (new_vma->vm_flags & VM_NOHUGEPAGE))
1072                 goto out;
1073
1074         /*
1075          * The destination pmd shouldn't be established, free_pgtables()
1076          * should have release it.
1077          */
1078         if (WARN_ON(!pmd_none(*new_pmd))) {
1079                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1080                 goto out;
1081         }
1082
1083         spin_lock(&mm->page_table_lock);
1084         if (likely(pmd_trans_huge(*old_pmd))) {
1085                 if (pmd_trans_splitting(*old_pmd)) {
1086                         spin_unlock(&mm->page_table_lock);
1087                         wait_split_huge_page(vma->anon_vma, old_pmd);
1088                         ret = -1;
1089                 } else {
1090                         pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1091                         VM_BUG_ON(!pmd_none(*new_pmd));
1092                         set_pmd_at(mm, new_addr, new_pmd, pmd);
1093                         spin_unlock(&mm->page_table_lock);
1094                         ret = 1;
1095                 }
1096         } else {
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         spin_lock(&mm->page_table_lock);
1110         if (likely(pmd_trans_huge(*pmd))) {
1111                 if (unlikely(pmd_trans_splitting(*pmd))) {
1112                         spin_unlock(&mm->page_table_lock);
1113                         wait_split_huge_page(vma->anon_vma, pmd);
1114                 } else {
1115                         pmd_t entry;
1116
1117                         entry = pmdp_get_and_clear(mm, addr, pmd);
1118                         entry = pmd_modify(entry, newprot);
1119                         set_pmd_at(mm, addr, pmd, entry);
1120                         spin_unlock(&vma->vm_mm->page_table_lock);
1121                         flush_tlb_range(vma, addr, addr + HPAGE_PMD_SIZE);
1122                         ret = 1;
1123                 }
1124         } else
1125                 spin_unlock(&vma->vm_mm->page_table_lock);
1126
1127         return ret;
1128 }
1129
1130 pmd_t *page_check_address_pmd(struct page *page,
1131                               struct mm_struct *mm,
1132                               unsigned long address,
1133                               enum page_check_address_pmd_flag flag)
1134 {
1135         pgd_t *pgd;
1136         pud_t *pud;
1137         pmd_t *pmd, *ret = NULL;
1138
1139         if (address & ~HPAGE_PMD_MASK)
1140                 goto out;
1141
1142         pgd = pgd_offset(mm, address);
1143         if (!pgd_present(*pgd))
1144                 goto out;
1145
1146         pud = pud_offset(pgd, address);
1147         if (!pud_present(*pud))
1148                 goto out;
1149
1150         pmd = pmd_offset(pud, address);
1151         if (pmd_none(*pmd))
1152                 goto out;
1153         if (pmd_page(*pmd) != page)
1154                 goto out;
1155         /*
1156          * split_vma() may create temporary aliased mappings. There is
1157          * no risk as long as all huge pmd are found and have their
1158          * splitting bit set before __split_huge_page_refcount
1159          * runs. Finding the same huge pmd more than once during the
1160          * same rmap walk is not a problem.
1161          */
1162         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1163             pmd_trans_splitting(*pmd))
1164                 goto out;
1165         if (pmd_trans_huge(*pmd)) {
1166                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1167                           !pmd_trans_splitting(*pmd));
1168                 ret = pmd;
1169         }
1170 out:
1171         return ret;
1172 }
1173
1174 static int __split_huge_page_splitting(struct page *page,
1175                                        struct vm_area_struct *vma,
1176                                        unsigned long address)
1177 {
1178         struct mm_struct *mm = vma->vm_mm;
1179         pmd_t *pmd;
1180         int ret = 0;
1181
1182         spin_lock(&mm->page_table_lock);
1183         pmd = page_check_address_pmd(page, mm, address,
1184                                      PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1185         if (pmd) {
1186                 /*
1187                  * We can't temporarily set the pmd to null in order
1188                  * to split it, the pmd must remain marked huge at all
1189                  * times or the VM won't take the pmd_trans_huge paths
1190                  * and it won't wait on the anon_vma->root->mutex to
1191                  * serialize against split_huge_page*.
1192                  */
1193                 pmdp_splitting_flush_notify(vma, address, pmd);
1194                 ret = 1;
1195         }
1196         spin_unlock(&mm->page_table_lock);
1197
1198         return ret;
1199 }
1200
1201 static void __split_huge_page_refcount(struct page *page)
1202 {
1203         int i;
1204         unsigned long head_index = page->index;
1205         struct zone *zone = page_zone(page);
1206         int zonestat;
1207         int tail_count = 0;
1208
1209         /* prevent PageLRU to go away from under us, and freeze lru stats */
1210         spin_lock_irq(&zone->lru_lock);
1211         compound_lock(page);
1212
1213         for (i = 1; i < HPAGE_PMD_NR; i++) {
1214                 struct page *page_tail = page + i;
1215
1216                 /* tail_page->_mapcount cannot change */
1217                 BUG_ON(page_mapcount(page_tail) < 0);
1218                 tail_count += page_mapcount(page_tail);
1219                 /* check for overflow */
1220                 BUG_ON(tail_count < 0);
1221                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1222                 /*
1223                  * tail_page->_count is zero and not changing from
1224                  * under us. But get_page_unless_zero() may be running
1225                  * from under us on the tail_page. If we used
1226                  * atomic_set() below instead of atomic_add(), we
1227                  * would then run atomic_set() concurrently with
1228                  * get_page_unless_zero(), and atomic_set() is
1229                  * implemented in C not using locked ops. spin_unlock
1230                  * on x86 sometime uses locked ops because of PPro
1231                  * errata 66, 92, so unless somebody can guarantee
1232                  * atomic_set() here would be safe on all archs (and
1233                  * not only on x86), it's safer to use atomic_add().
1234                  */
1235                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1236                            &page_tail->_count);
1237
1238                 /* after clearing PageTail the gup refcount can be released */
1239                 smp_mb();
1240
1241                 /*
1242                  * retain hwpoison flag of the poisoned tail page:
1243                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1244                  *   by the memory-failure.
1245                  */
1246                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1247                 page_tail->flags |= (page->flags &
1248                                      ((1L << PG_referenced) |
1249                                       (1L << PG_swapbacked) |
1250                                       (1L << PG_mlocked) |
1251                                       (1L << PG_uptodate)));
1252                 page_tail->flags |= (1L << PG_dirty);
1253
1254                 /* clear PageTail before overwriting first_page */
1255                 smp_wmb();
1256
1257                 /*
1258                  * __split_huge_page_splitting() already set the
1259                  * splitting bit in all pmd that could map this
1260                  * hugepage, that will ensure no CPU can alter the
1261                  * mapcount on the head page. The mapcount is only
1262                  * accounted in the head page and it has to be
1263                  * transferred to all tail pages in the below code. So
1264                  * for this code to be safe, the split the mapcount
1265                  * can't change. But that doesn't mean userland can't
1266                  * keep changing and reading the page contents while
1267                  * we transfer the mapcount, so the pmd splitting
1268                  * status is achieved setting a reserved bit in the
1269                  * pmd, not by clearing the present bit.
1270                 */
1271                 page_tail->_mapcount = page->_mapcount;
1272
1273                 BUG_ON(page_tail->mapping);
1274                 page_tail->mapping = page->mapping;
1275
1276                 page_tail->index = ++head_index;
1277
1278                 BUG_ON(!PageAnon(page_tail));
1279                 BUG_ON(!PageUptodate(page_tail));
1280                 BUG_ON(!PageDirty(page_tail));
1281                 BUG_ON(!PageSwapBacked(page_tail));
1282
1283                 mem_cgroup_split_huge_fixup(page, page_tail);
1284
1285                 lru_add_page_tail(zone, page, page_tail);
1286         }
1287         atomic_sub(tail_count, &page->_count);
1288         BUG_ON(atomic_read(&page->_count) <= 0);
1289
1290         __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1291         __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1292
1293         /*
1294          * A hugepage counts for HPAGE_PMD_NR pages on the LRU statistics,
1295          * so adjust those appropriately if this page is on the LRU.
1296          */
1297         if (PageLRU(page)) {
1298                 zonestat = NR_LRU_BASE + page_lru(page);
1299                 __mod_zone_page_state(zone, zonestat, -(HPAGE_PMD_NR-1));
1300         }
1301
1302         ClearPageCompound(page);
1303         compound_unlock(page);
1304         spin_unlock_irq(&zone->lru_lock);
1305
1306         for (i = 1; i < HPAGE_PMD_NR; i++) {
1307                 struct page *page_tail = page + i;
1308                 BUG_ON(page_count(page_tail) <= 0);
1309                 /*
1310                  * Tail pages may be freed if there wasn't any mapping
1311                  * like if add_to_swap() is running on a lru page that
1312                  * had its mapping zapped. And freeing these pages
1313                  * requires taking the lru_lock so we do the put_page
1314                  * of the tail pages after the split is complete.
1315                  */
1316                 put_page(page_tail);
1317         }
1318
1319         /*
1320          * Only the head page (now become a regular page) is required
1321          * to be pinned by the caller.
1322          */
1323         BUG_ON(page_count(page) <= 0);
1324 }
1325
1326 static int __split_huge_page_map(struct page *page,
1327                                  struct vm_area_struct *vma,
1328                                  unsigned long address)
1329 {
1330         struct mm_struct *mm = vma->vm_mm;
1331         pmd_t *pmd, _pmd;
1332         int ret = 0, i;
1333         pgtable_t pgtable;
1334         unsigned long haddr;
1335
1336         spin_lock(&mm->page_table_lock);
1337         pmd = page_check_address_pmd(page, mm, address,
1338                                      PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1339         if (pmd) {
1340                 pgtable = get_pmd_huge_pte(mm);
1341                 pmd_populate(mm, &_pmd, pgtable);
1342
1343                 for (i = 0, haddr = address; i < HPAGE_PMD_NR;
1344                      i++, haddr += PAGE_SIZE) {
1345                         pte_t *pte, entry;
1346                         BUG_ON(PageCompound(page+i));
1347                         entry = mk_pte(page + i, vma->vm_page_prot);
1348                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1349                         if (!pmd_write(*pmd))
1350                                 entry = pte_wrprotect(entry);
1351                         else
1352                                 BUG_ON(page_mapcount(page) != 1);
1353                         if (!pmd_young(*pmd))
1354                                 entry = pte_mkold(entry);
1355                         pte = pte_offset_map(&_pmd, haddr);
1356                         BUG_ON(!pte_none(*pte));
1357                         set_pte_at(mm, haddr, pte, entry);
1358                         pte_unmap(pte);
1359                 }
1360
1361                 smp_wmb(); /* make pte visible before pmd */
1362                 /*
1363                  * Up to this point the pmd is present and huge and
1364                  * userland has the whole access to the hugepage
1365                  * during the split (which happens in place). If we
1366                  * overwrite the pmd with the not-huge version
1367                  * pointing to the pte here (which of course we could
1368                  * if all CPUs were bug free), userland could trigger
1369                  * a small page size TLB miss on the small sized TLB
1370                  * while the hugepage TLB entry is still established
1371                  * in the huge TLB. Some CPU doesn't like that. See
1372                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1373                  * Erratum 383 on page 93. Intel should be safe but is
1374                  * also warns that it's only safe if the permission
1375                  * and cache attributes of the two entries loaded in
1376                  * the two TLB is identical (which should be the case
1377                  * here). But it is generally safer to never allow
1378                  * small and huge TLB entries for the same virtual
1379                  * address to be loaded simultaneously. So instead of
1380                  * doing "pmd_populate(); flush_tlb_range();" we first
1381                  * mark the current pmd notpresent (atomically because
1382                  * here the pmd_trans_huge and pmd_trans_splitting
1383                  * must remain set at all times on the pmd until the
1384                  * split is complete for this pmd), then we flush the
1385                  * SMP TLB and finally we write the non-huge version
1386                  * of the pmd entry with pmd_populate.
1387                  */
1388                 set_pmd_at(mm, address, pmd, pmd_mknotpresent(*pmd));
1389                 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
1390                 pmd_populate(mm, pmd, pgtable);
1391                 ret = 1;
1392         }
1393         spin_unlock(&mm->page_table_lock);
1394
1395         return ret;
1396 }
1397
1398 /* must be called with anon_vma->root->mutex hold */
1399 static void __split_huge_page(struct page *page,
1400                               struct anon_vma *anon_vma)
1401 {
1402         int mapcount, mapcount2;
1403         struct anon_vma_chain *avc;
1404
1405         BUG_ON(!PageHead(page));
1406         BUG_ON(PageTail(page));
1407
1408         mapcount = 0;
1409         list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1410                 struct vm_area_struct *vma = avc->vma;
1411                 unsigned long addr = vma_address(page, vma);
1412                 BUG_ON(is_vma_temporary_stack(vma));
1413                 if (addr == -EFAULT)
1414                         continue;
1415                 mapcount += __split_huge_page_splitting(page, vma, addr);
1416         }
1417         /*
1418          * It is critical that new vmas are added to the tail of the
1419          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1420          * and establishes a child pmd before
1421          * __split_huge_page_splitting() freezes the parent pmd (so if
1422          * we fail to prevent copy_huge_pmd() from running until the
1423          * whole __split_huge_page() is complete), we will still see
1424          * the newly established pmd of the child later during the
1425          * walk, to be able to set it as pmd_trans_splitting too.
1426          */
1427         if (mapcount != page_mapcount(page))
1428                 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1429                        mapcount, page_mapcount(page));
1430         BUG_ON(mapcount != page_mapcount(page));
1431
1432         __split_huge_page_refcount(page);
1433
1434         mapcount2 = 0;
1435         list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1436                 struct vm_area_struct *vma = avc->vma;
1437                 unsigned long addr = vma_address(page, vma);
1438                 BUG_ON(is_vma_temporary_stack(vma));
1439                 if (addr == -EFAULT)
1440                         continue;
1441                 mapcount2 += __split_huge_page_map(page, vma, addr);
1442         }
1443         if (mapcount != mapcount2)
1444                 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1445                        mapcount, mapcount2, page_mapcount(page));
1446         BUG_ON(mapcount != mapcount2);
1447 }
1448
1449 int split_huge_page(struct page *page)
1450 {
1451         struct anon_vma *anon_vma;
1452         int ret = 1;
1453
1454         BUG_ON(!PageAnon(page));
1455         anon_vma = page_lock_anon_vma(page);
1456         if (!anon_vma)
1457                 goto out;
1458         ret = 0;
1459         if (!PageCompound(page))
1460                 goto out_unlock;
1461
1462         BUG_ON(!PageSwapBacked(page));
1463         __split_huge_page(page, anon_vma);
1464         count_vm_event(THP_SPLIT);
1465
1466         BUG_ON(PageCompound(page));
1467 out_unlock:
1468         page_unlock_anon_vma(anon_vma);
1469 out:
1470         return ret;
1471 }
1472
1473 #define VM_NO_THP (VM_SPECIAL|VM_INSERTPAGE|VM_MIXEDMAP|VM_SAO| \
1474                    VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1475
1476 int hugepage_madvise(struct vm_area_struct *vma,
1477                      unsigned long *vm_flags, int advice)
1478 {
1479         switch (advice) {
1480         case MADV_HUGEPAGE:
1481                 /*
1482                  * Be somewhat over-protective like KSM for now!
1483                  */
1484                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1485                         return -EINVAL;
1486                 *vm_flags &= ~VM_NOHUGEPAGE;
1487                 *vm_flags |= VM_HUGEPAGE;
1488                 /*
1489                  * If the vma become good for khugepaged to scan,
1490                  * register it here without waiting a page fault that
1491                  * may not happen any time soon.
1492                  */
1493                 if (unlikely(khugepaged_enter_vma_merge(vma)))
1494                         return -ENOMEM;
1495                 break;
1496         case MADV_NOHUGEPAGE:
1497                 /*
1498                  * Be somewhat over-protective like KSM for now!
1499                  */
1500                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1501                         return -EINVAL;
1502                 *vm_flags &= ~VM_HUGEPAGE;
1503                 *vm_flags |= VM_NOHUGEPAGE;
1504                 /*
1505                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1506                  * this vma even if we leave the mm registered in khugepaged if
1507                  * it got registered before VM_NOHUGEPAGE was set.
1508                  */
1509                 break;
1510         }
1511
1512         return 0;
1513 }
1514
1515 static int __init khugepaged_slab_init(void)
1516 {
1517         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1518                                           sizeof(struct mm_slot),
1519                                           __alignof__(struct mm_slot), 0, NULL);
1520         if (!mm_slot_cache)
1521                 return -ENOMEM;
1522
1523         return 0;
1524 }
1525
1526 static void __init khugepaged_slab_free(void)
1527 {
1528         kmem_cache_destroy(mm_slot_cache);
1529         mm_slot_cache = NULL;
1530 }
1531
1532 static inline struct mm_slot *alloc_mm_slot(void)
1533 {
1534         if (!mm_slot_cache)     /* initialization failed */
1535                 return NULL;
1536         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1537 }
1538
1539 static inline void free_mm_slot(struct mm_slot *mm_slot)
1540 {
1541         kmem_cache_free(mm_slot_cache, mm_slot);
1542 }
1543
1544 static int __init mm_slots_hash_init(void)
1545 {
1546         mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
1547                                 GFP_KERNEL);
1548         if (!mm_slots_hash)
1549                 return -ENOMEM;
1550         return 0;
1551 }
1552
1553 #if 0
1554 static void __init mm_slots_hash_free(void)
1555 {
1556         kfree(mm_slots_hash);
1557         mm_slots_hash = NULL;
1558 }
1559 #endif
1560
1561 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1562 {
1563         struct mm_slot *mm_slot;
1564         struct hlist_head *bucket;
1565         struct hlist_node *node;
1566
1567         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1568                                 % MM_SLOTS_HASH_HEADS];
1569         hlist_for_each_entry(mm_slot, node, bucket, hash) {
1570                 if (mm == mm_slot->mm)
1571                         return mm_slot;
1572         }
1573         return NULL;
1574 }
1575
1576 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1577                                     struct mm_slot *mm_slot)
1578 {
1579         struct hlist_head *bucket;
1580
1581         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1582                                 % MM_SLOTS_HASH_HEADS];
1583         mm_slot->mm = mm;
1584         hlist_add_head(&mm_slot->hash, bucket);
1585 }
1586
1587 static inline int khugepaged_test_exit(struct mm_struct *mm)
1588 {
1589         return atomic_read(&mm->mm_users) == 0;
1590 }
1591
1592 int __khugepaged_enter(struct mm_struct *mm)
1593 {
1594         struct mm_slot *mm_slot;
1595         int wakeup;
1596
1597         mm_slot = alloc_mm_slot();
1598         if (!mm_slot)
1599                 return -ENOMEM;
1600
1601         /* __khugepaged_exit() must not run from under us */
1602         VM_BUG_ON(khugepaged_test_exit(mm));
1603         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1604                 free_mm_slot(mm_slot);
1605                 return 0;
1606         }
1607
1608         spin_lock(&khugepaged_mm_lock);
1609         insert_to_mm_slots_hash(mm, mm_slot);
1610         /*
1611          * Insert just behind the scanning cursor, to let the area settle
1612          * down a little.
1613          */
1614         wakeup = list_empty(&khugepaged_scan.mm_head);
1615         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1616         spin_unlock(&khugepaged_mm_lock);
1617
1618         atomic_inc(&mm->mm_count);
1619         if (wakeup)
1620                 wake_up_interruptible(&khugepaged_wait);
1621
1622         return 0;
1623 }
1624
1625 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
1626 {
1627         unsigned long hstart, hend;
1628         if (!vma->anon_vma)
1629                 /*
1630                  * Not yet faulted in so we will register later in the
1631                  * page fault if needed.
1632                  */
1633                 return 0;
1634         if (vma->vm_ops)
1635                 /* khugepaged not yet working on file or special mappings */
1636                 return 0;
1637         /*
1638          * If is_pfn_mapping() is true is_learn_pfn_mapping() must be
1639          * true too, verify it here.
1640          */
1641         VM_BUG_ON(is_linear_pfn_mapping(vma) || vma->vm_flags & VM_NO_THP);
1642         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1643         hend = vma->vm_end & HPAGE_PMD_MASK;
1644         if (hstart < hend)
1645                 return khugepaged_enter(vma);
1646         return 0;
1647 }
1648
1649 void __khugepaged_exit(struct mm_struct *mm)
1650 {
1651         struct mm_slot *mm_slot;
1652         int free = 0;
1653
1654         spin_lock(&khugepaged_mm_lock);
1655         mm_slot = get_mm_slot(mm);
1656         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1657                 hlist_del(&mm_slot->hash);
1658                 list_del(&mm_slot->mm_node);
1659                 free = 1;
1660         }
1661         spin_unlock(&khugepaged_mm_lock);
1662
1663         if (free) {
1664                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1665                 free_mm_slot(mm_slot);
1666                 mmdrop(mm);
1667         } else if (mm_slot) {
1668                 /*
1669                  * This is required to serialize against
1670                  * khugepaged_test_exit() (which is guaranteed to run
1671                  * under mmap sem read mode). Stop here (after we
1672                  * return all pagetables will be destroyed) until
1673                  * khugepaged has finished working on the pagetables
1674                  * under the mmap_sem.
1675                  */
1676                 down_write(&mm->mmap_sem);
1677                 up_write(&mm->mmap_sem);
1678         }
1679 }
1680
1681 static void release_pte_page(struct page *page)
1682 {
1683         /* 0 stands for page_is_file_cache(page) == false */
1684         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1685         unlock_page(page);
1686         putback_lru_page(page);
1687 }
1688
1689 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1690 {
1691         while (--_pte >= pte) {
1692                 pte_t pteval = *_pte;
1693                 if (!pte_none(pteval))
1694                         release_pte_page(pte_page(pteval));
1695         }
1696 }
1697
1698 static void release_all_pte_pages(pte_t *pte)
1699 {
1700         release_pte_pages(pte, pte + HPAGE_PMD_NR);
1701 }
1702
1703 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1704                                         unsigned long address,
1705                                         pte_t *pte)
1706 {
1707         struct page *page;
1708         pte_t *_pte;
1709         int referenced = 0, isolated = 0, none = 0;
1710         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1711              _pte++, address += PAGE_SIZE) {
1712                 pte_t pteval = *_pte;
1713                 if (pte_none(pteval)) {
1714                         if (++none <= khugepaged_max_ptes_none)
1715                                 continue;
1716                         else {
1717                                 release_pte_pages(pte, _pte);
1718                                 goto out;
1719                         }
1720                 }
1721                 if (!pte_present(pteval) || !pte_write(pteval)) {
1722                         release_pte_pages(pte, _pte);
1723                         goto out;
1724                 }
1725                 page = vm_normal_page(vma, address, pteval);
1726                 if (unlikely(!page)) {
1727                         release_pte_pages(pte, _pte);
1728                         goto out;
1729                 }
1730                 VM_BUG_ON(PageCompound(page));
1731                 BUG_ON(!PageAnon(page));
1732                 VM_BUG_ON(!PageSwapBacked(page));
1733
1734                 /* cannot use mapcount: can't collapse if there's a gup pin */
1735                 if (page_count(page) != 1) {
1736                         release_pte_pages(pte, _pte);
1737                         goto out;
1738                 }
1739                 /*
1740                  * We can do it before isolate_lru_page because the
1741                  * page can't be freed from under us. NOTE: PG_lock
1742                  * is needed to serialize against split_huge_page
1743                  * when invoked from the VM.
1744                  */
1745                 if (!trylock_page(page)) {
1746                         release_pte_pages(pte, _pte);
1747                         goto out;
1748                 }
1749                 /*
1750                  * Isolate the page to avoid collapsing an hugepage
1751                  * currently in use by the VM.
1752                  */
1753                 if (isolate_lru_page(page)) {
1754                         unlock_page(page);
1755                         release_pte_pages(pte, _pte);
1756                         goto out;
1757                 }
1758                 /* 0 stands for page_is_file_cache(page) == false */
1759                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
1760                 VM_BUG_ON(!PageLocked(page));
1761                 VM_BUG_ON(PageLRU(page));
1762
1763                 /* If there is no mapped pte young don't collapse the page */
1764                 if (pte_young(pteval) || PageReferenced(page) ||
1765                     mmu_notifier_test_young(vma->vm_mm, address))
1766                         referenced = 1;
1767         }
1768         if (unlikely(!referenced))
1769                 release_all_pte_pages(pte);
1770         else
1771                 isolated = 1;
1772 out:
1773         return isolated;
1774 }
1775
1776 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
1777                                       struct vm_area_struct *vma,
1778                                       unsigned long address,
1779                                       spinlock_t *ptl)
1780 {
1781         pte_t *_pte;
1782         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
1783                 pte_t pteval = *_pte;
1784                 struct page *src_page;
1785
1786                 if (pte_none(pteval)) {
1787                         clear_user_highpage(page, address);
1788                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
1789                 } else {
1790                         src_page = pte_page(pteval);
1791                         copy_user_highpage(page, src_page, address, vma);
1792                         VM_BUG_ON(page_mapcount(src_page) != 1);
1793                         VM_BUG_ON(page_count(src_page) != 2);
1794                         release_pte_page(src_page);
1795                         /*
1796                          * ptl mostly unnecessary, but preempt has to
1797                          * be disabled to update the per-cpu stats
1798                          * inside page_remove_rmap().
1799                          */
1800                         spin_lock(ptl);
1801                         /*
1802                          * paravirt calls inside pte_clear here are
1803                          * superfluous.
1804                          */
1805                         pte_clear(vma->vm_mm, address, _pte);
1806                         page_remove_rmap(src_page);
1807                         spin_unlock(ptl);
1808                         free_page_and_swap_cache(src_page);
1809                 }
1810
1811                 address += PAGE_SIZE;
1812                 page++;
1813         }
1814 }
1815
1816 static void collapse_huge_page(struct mm_struct *mm,
1817                                unsigned long address,
1818                                struct page **hpage,
1819                                struct vm_area_struct *vma,
1820                                int node)
1821 {
1822         pgd_t *pgd;
1823         pud_t *pud;
1824         pmd_t *pmd, _pmd;
1825         pte_t *pte;
1826         pgtable_t pgtable;
1827         struct page *new_page;
1828         spinlock_t *ptl;
1829         int isolated;
1830         unsigned long hstart, hend;
1831
1832         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1833 #ifndef CONFIG_NUMA
1834         up_read(&mm->mmap_sem);
1835         VM_BUG_ON(!*hpage);
1836         new_page = *hpage;
1837 #else
1838         VM_BUG_ON(*hpage);
1839         /*
1840          * Allocate the page while the vma is still valid and under
1841          * the mmap_sem read mode so there is no memory allocation
1842          * later when we take the mmap_sem in write mode. This is more
1843          * friendly behavior (OTOH it may actually hide bugs) to
1844          * filesystems in userland with daemons allocating memory in
1845          * the userland I/O paths.  Allocating memory with the
1846          * mmap_sem in read mode is good idea also to allow greater
1847          * scalability.
1848          */
1849         new_page = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
1850                                       node, __GFP_OTHER_NODE);
1851
1852         /*
1853          * After allocating the hugepage, release the mmap_sem read lock in
1854          * preparation for taking it in write mode.
1855          */
1856         up_read(&mm->mmap_sem);
1857         if (unlikely(!new_page)) {
1858                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
1859                 *hpage = ERR_PTR(-ENOMEM);
1860                 return;
1861         }
1862 #endif
1863
1864         count_vm_event(THP_COLLAPSE_ALLOC);
1865         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1866 #ifdef CONFIG_NUMA
1867                 put_page(new_page);
1868 #endif
1869                 return;
1870         }
1871
1872         /*
1873          * Prevent all access to pagetables with the exception of
1874          * gup_fast later hanlded by the ptep_clear_flush and the VM
1875          * handled by the anon_vma lock + PG_lock.
1876          */
1877         down_write(&mm->mmap_sem);
1878         if (unlikely(khugepaged_test_exit(mm)))
1879                 goto out;
1880
1881         vma = find_vma(mm, address);
1882         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1883         hend = vma->vm_end & HPAGE_PMD_MASK;
1884         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
1885                 goto out;
1886
1887         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
1888             (vma->vm_flags & VM_NOHUGEPAGE))
1889                 goto out;
1890
1891         if (!vma->anon_vma || vma->vm_ops)
1892                 goto out;
1893         if (is_vma_temporary_stack(vma))
1894                 goto out;
1895         /*
1896          * If is_pfn_mapping() is true is_learn_pfn_mapping() must be
1897          * true too, verify it here.
1898          */
1899         VM_BUG_ON(is_linear_pfn_mapping(vma) || vma->vm_flags & VM_NO_THP);
1900
1901         pgd = pgd_offset(mm, address);
1902         if (!pgd_present(*pgd))
1903                 goto out;
1904
1905         pud = pud_offset(pgd, address);
1906         if (!pud_present(*pud))
1907                 goto out;
1908
1909         pmd = pmd_offset(pud, address);
1910         /* pmd can't go away or become huge under us */
1911         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
1912                 goto out;
1913
1914         anon_vma_lock(vma->anon_vma);
1915
1916         pte = pte_offset_map(pmd, address);
1917         ptl = pte_lockptr(mm, pmd);
1918
1919         spin_lock(&mm->page_table_lock); /* probably unnecessary */
1920         /*
1921          * After this gup_fast can't run anymore. This also removes
1922          * any huge TLB entry from the CPU so we won't allow
1923          * huge and small TLB entries for the same virtual address
1924          * to avoid the risk of CPU bugs in that area.
1925          */
1926         _pmd = pmdp_clear_flush_notify(vma, address, pmd);
1927         spin_unlock(&mm->page_table_lock);
1928
1929         spin_lock(ptl);
1930         isolated = __collapse_huge_page_isolate(vma, address, pte);
1931         spin_unlock(ptl);
1932
1933         if (unlikely(!isolated)) {
1934                 pte_unmap(pte);
1935                 spin_lock(&mm->page_table_lock);
1936                 BUG_ON(!pmd_none(*pmd));
1937                 set_pmd_at(mm, address, pmd, _pmd);
1938                 spin_unlock(&mm->page_table_lock);
1939                 anon_vma_unlock(vma->anon_vma);
1940                 goto out;
1941         }
1942
1943         /*
1944          * All pages are isolated and locked so anon_vma rmap
1945          * can't run anymore.
1946          */
1947         anon_vma_unlock(vma->anon_vma);
1948
1949         __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
1950         pte_unmap(pte);
1951         __SetPageUptodate(new_page);
1952         pgtable = pmd_pgtable(_pmd);
1953         VM_BUG_ON(page_count(pgtable) != 1);
1954         VM_BUG_ON(page_mapcount(pgtable) != 0);
1955
1956         _pmd = mk_pmd(new_page, vma->vm_page_prot);
1957         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1958         _pmd = pmd_mkhuge(_pmd);
1959
1960         /*
1961          * spin_lock() below is not the equivalent of smp_wmb(), so
1962          * this is needed to avoid the copy_huge_page writes to become
1963          * visible after the set_pmd_at() write.
1964          */
1965         smp_wmb();
1966
1967         spin_lock(&mm->page_table_lock);
1968         BUG_ON(!pmd_none(*pmd));
1969         page_add_new_anon_rmap(new_page, vma, address);
1970         set_pmd_at(mm, address, pmd, _pmd);
1971         update_mmu_cache(vma, address, _pmd);
1972         prepare_pmd_huge_pte(pgtable, mm);
1973         spin_unlock(&mm->page_table_lock);
1974
1975 #ifndef CONFIG_NUMA
1976         *hpage = NULL;
1977 #endif
1978         khugepaged_pages_collapsed++;
1979 out_up_write:
1980         up_write(&mm->mmap_sem);
1981         return;
1982
1983 out:
1984         mem_cgroup_uncharge_page(new_page);
1985 #ifdef CONFIG_NUMA
1986         put_page(new_page);
1987 #endif
1988         goto out_up_write;
1989 }
1990
1991 static int khugepaged_scan_pmd(struct mm_struct *mm,
1992                                struct vm_area_struct *vma,
1993                                unsigned long address,
1994                                struct page **hpage)
1995 {
1996         pgd_t *pgd;
1997         pud_t *pud;
1998         pmd_t *pmd;
1999         pte_t *pte, *_pte;
2000         int ret = 0, referenced = 0, none = 0;
2001         struct page *page;
2002         unsigned long _address;
2003         spinlock_t *ptl;
2004         int node = -1;
2005
2006         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2007
2008         pgd = pgd_offset(mm, address);
2009         if (!pgd_present(*pgd))
2010                 goto out;
2011
2012         pud = pud_offset(pgd, address);
2013         if (!pud_present(*pud))
2014                 goto out;
2015
2016         pmd = pmd_offset(pud, address);
2017         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
2018                 goto out;
2019
2020         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2021         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2022              _pte++, _address += PAGE_SIZE) {
2023                 pte_t pteval = *_pte;
2024                 if (pte_none(pteval)) {
2025                         if (++none <= khugepaged_max_ptes_none)
2026                                 continue;
2027                         else
2028                                 goto out_unmap;
2029                 }
2030                 if (!pte_present(pteval) || !pte_write(pteval))
2031                         goto out_unmap;
2032                 page = vm_normal_page(vma, _address, pteval);
2033                 if (unlikely(!page))
2034                         goto out_unmap;
2035                 /*
2036                  * Chose the node of the first page. This could
2037                  * be more sophisticated and look at more pages,
2038                  * but isn't for now.
2039                  */
2040                 if (node == -1)
2041                         node = page_to_nid(page);
2042                 VM_BUG_ON(PageCompound(page));
2043                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2044                         goto out_unmap;
2045                 /* cannot use mapcount: can't collapse if there's a gup pin */
2046                 if (page_count(page) != 1)
2047                         goto out_unmap;
2048                 if (pte_young(pteval) || PageReferenced(page) ||
2049                     mmu_notifier_test_young(vma->vm_mm, address))
2050                         referenced = 1;
2051         }
2052         if (referenced)
2053                 ret = 1;
2054 out_unmap:
2055         pte_unmap_unlock(pte, ptl);
2056         if (ret)
2057                 /* collapse_huge_page will return with the mmap_sem released */
2058                 collapse_huge_page(mm, address, hpage, vma, node);
2059 out:
2060         return ret;
2061 }
2062
2063 static void collect_mm_slot(struct mm_slot *mm_slot)
2064 {
2065         struct mm_struct *mm = mm_slot->mm;
2066
2067         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2068
2069         if (khugepaged_test_exit(mm)) {
2070                 /* free mm_slot */
2071                 hlist_del(&mm_slot->hash);
2072                 list_del(&mm_slot->mm_node);
2073
2074                 /*
2075                  * Not strictly needed because the mm exited already.
2076                  *
2077                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2078                  */
2079
2080                 /* khugepaged_mm_lock actually not necessary for the below */
2081                 free_mm_slot(mm_slot);
2082                 mmdrop(mm);
2083         }
2084 }
2085
2086 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2087                                             struct page **hpage)
2088         __releases(&khugepaged_mm_lock)
2089         __acquires(&khugepaged_mm_lock)
2090 {
2091         struct mm_slot *mm_slot;
2092         struct mm_struct *mm;
2093         struct vm_area_struct *vma;
2094         int progress = 0;
2095
2096         VM_BUG_ON(!pages);
2097         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2098
2099         if (khugepaged_scan.mm_slot)
2100                 mm_slot = khugepaged_scan.mm_slot;
2101         else {
2102                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2103                                      struct mm_slot, mm_node);
2104                 khugepaged_scan.address = 0;
2105                 khugepaged_scan.mm_slot = mm_slot;
2106         }
2107         spin_unlock(&khugepaged_mm_lock);
2108
2109         mm = mm_slot->mm;
2110         down_read(&mm->mmap_sem);
2111         if (unlikely(khugepaged_test_exit(mm)))
2112                 vma = NULL;
2113         else
2114                 vma = find_vma(mm, khugepaged_scan.address);
2115
2116         progress++;
2117         for (; vma; vma = vma->vm_next) {
2118                 unsigned long hstart, hend;
2119
2120                 cond_resched();
2121                 if (unlikely(khugepaged_test_exit(mm))) {
2122                         progress++;
2123                         break;
2124                 }
2125
2126                 if ((!(vma->vm_flags & VM_HUGEPAGE) &&
2127                      !khugepaged_always()) ||
2128                     (vma->vm_flags & VM_NOHUGEPAGE)) {
2129                 skip:
2130                         progress++;
2131                         continue;
2132                 }
2133                 if (!vma->anon_vma || vma->vm_ops)
2134                         goto skip;
2135                 if (is_vma_temporary_stack(vma))
2136                         goto skip;
2137                 /*
2138                  * If is_pfn_mapping() is true is_learn_pfn_mapping()
2139                  * must be true too, verify it here.
2140                  */
2141                 VM_BUG_ON(is_linear_pfn_mapping(vma) ||
2142                           vma->vm_flags & VM_NO_THP);
2143
2144                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2145                 hend = vma->vm_end & HPAGE_PMD_MASK;
2146                 if (hstart >= hend)
2147                         goto skip;
2148                 if (khugepaged_scan.address > hend)
2149                         goto skip;
2150                 if (khugepaged_scan.address < hstart)
2151                         khugepaged_scan.address = hstart;
2152                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2153
2154                 while (khugepaged_scan.address < hend) {
2155                         int ret;
2156                         cond_resched();
2157                         if (unlikely(khugepaged_test_exit(mm)))
2158                                 goto breakouterloop;
2159
2160                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2161                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2162                                   hend);
2163                         ret = khugepaged_scan_pmd(mm, vma,
2164                                                   khugepaged_scan.address,
2165                                                   hpage);
2166                         /* move to next address */
2167                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2168                         progress += HPAGE_PMD_NR;
2169                         if (ret)
2170                                 /* we released mmap_sem so break loop */
2171                                 goto breakouterloop_mmap_sem;
2172                         if (progress >= pages)
2173                                 goto breakouterloop;
2174                 }
2175         }
2176 breakouterloop:
2177         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2178 breakouterloop_mmap_sem:
2179
2180         spin_lock(&khugepaged_mm_lock);
2181         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2182         /*
2183          * Release the current mm_slot if this mm is about to die, or
2184          * if we scanned all vmas of this mm.
2185          */
2186         if (khugepaged_test_exit(mm) || !vma) {
2187                 /*
2188                  * Make sure that if mm_users is reaching zero while
2189                  * khugepaged runs here, khugepaged_exit will find
2190                  * mm_slot not pointing to the exiting mm.
2191                  */
2192                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2193                         khugepaged_scan.mm_slot = list_entry(
2194                                 mm_slot->mm_node.next,
2195                                 struct mm_slot, mm_node);
2196                         khugepaged_scan.address = 0;
2197                 } else {
2198                         khugepaged_scan.mm_slot = NULL;
2199                         khugepaged_full_scans++;
2200                 }
2201
2202                 collect_mm_slot(mm_slot);
2203         }
2204
2205         return progress;
2206 }
2207
2208 static int khugepaged_has_work(void)
2209 {
2210         return !list_empty(&khugepaged_scan.mm_head) &&
2211                 khugepaged_enabled();
2212 }
2213
2214 static int khugepaged_wait_event(void)
2215 {
2216         return !list_empty(&khugepaged_scan.mm_head) ||
2217                 !khugepaged_enabled();
2218 }
2219
2220 static void khugepaged_do_scan(struct page **hpage)
2221 {
2222         unsigned int progress = 0, pass_through_head = 0;
2223         unsigned int pages = khugepaged_pages_to_scan;
2224
2225         barrier(); /* write khugepaged_pages_to_scan to local stack */
2226
2227         while (progress < pages) {
2228                 cond_resched();
2229
2230 #ifndef CONFIG_NUMA
2231                 if (!*hpage) {
2232                         *hpage = alloc_hugepage(khugepaged_defrag());
2233                         if (unlikely(!*hpage)) {
2234                                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2235                                 break;
2236                         }
2237                         count_vm_event(THP_COLLAPSE_ALLOC);
2238                 }
2239 #else
2240                 if (IS_ERR(*hpage))
2241                         break;
2242 #endif
2243
2244                 if (unlikely(kthread_should_stop() || freezing(current)))
2245                         break;
2246
2247                 spin_lock(&khugepaged_mm_lock);
2248                 if (!khugepaged_scan.mm_slot)
2249                         pass_through_head++;
2250                 if (khugepaged_has_work() &&
2251                     pass_through_head < 2)
2252                         progress += khugepaged_scan_mm_slot(pages - progress,
2253                                                             hpage);
2254                 else
2255                         progress = pages;
2256                 spin_unlock(&khugepaged_mm_lock);
2257         }
2258 }
2259
2260 static void khugepaged_alloc_sleep(void)
2261 {
2262         wait_event_freezable_timeout(khugepaged_wait, false,
2263                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2264 }
2265
2266 #ifndef CONFIG_NUMA
2267 static struct page *khugepaged_alloc_hugepage(void)
2268 {
2269         struct page *hpage;
2270
2271         do {
2272                 hpage = alloc_hugepage(khugepaged_defrag());
2273                 if (!hpage) {
2274                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2275                         khugepaged_alloc_sleep();
2276                 } else
2277                         count_vm_event(THP_COLLAPSE_ALLOC);
2278         } while (unlikely(!hpage) &&
2279                  likely(khugepaged_enabled()));
2280         return hpage;
2281 }
2282 #endif
2283
2284 static void khugepaged_loop(void)
2285 {
2286         struct page *hpage;
2287
2288 #ifdef CONFIG_NUMA
2289         hpage = NULL;
2290 #endif
2291         while (likely(khugepaged_enabled())) {
2292 #ifndef CONFIG_NUMA
2293                 hpage = khugepaged_alloc_hugepage();
2294                 if (unlikely(!hpage))
2295                         break;
2296 #else
2297                 if (IS_ERR(hpage)) {
2298                         khugepaged_alloc_sleep();
2299                         hpage = NULL;
2300                 }
2301 #endif
2302
2303                 khugepaged_do_scan(&hpage);
2304 #ifndef CONFIG_NUMA
2305                 if (hpage)
2306                         put_page(hpage);
2307 #endif
2308                 try_to_freeze();
2309                 if (unlikely(kthread_should_stop()))
2310                         break;
2311                 if (khugepaged_has_work()) {
2312                         if (!khugepaged_scan_sleep_millisecs)
2313                                 continue;
2314                         wait_event_freezable_timeout(khugepaged_wait, false,
2315                             msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2316                 } else if (khugepaged_enabled())
2317                         wait_event_freezable(khugepaged_wait,
2318                                              khugepaged_wait_event());
2319         }
2320 }
2321
2322 static int khugepaged(void *none)
2323 {
2324         struct mm_slot *mm_slot;
2325
2326         set_freezable();
2327         set_user_nice(current, 19);
2328
2329         /* serialize with start_khugepaged() */
2330         mutex_lock(&khugepaged_mutex);
2331
2332         for (;;) {
2333                 mutex_unlock(&khugepaged_mutex);
2334                 VM_BUG_ON(khugepaged_thread != current);
2335                 khugepaged_loop();
2336                 VM_BUG_ON(khugepaged_thread != current);
2337
2338                 mutex_lock(&khugepaged_mutex);
2339                 if (!khugepaged_enabled())
2340                         break;
2341                 if (unlikely(kthread_should_stop()))
2342                         break;
2343         }
2344
2345         spin_lock(&khugepaged_mm_lock);
2346         mm_slot = khugepaged_scan.mm_slot;
2347         khugepaged_scan.mm_slot = NULL;
2348         if (mm_slot)
2349                 collect_mm_slot(mm_slot);
2350         spin_unlock(&khugepaged_mm_lock);
2351
2352         khugepaged_thread = NULL;
2353         mutex_unlock(&khugepaged_mutex);
2354
2355         return 0;
2356 }
2357
2358 void __split_huge_page_pmd(struct mm_struct *mm, pmd_t *pmd)
2359 {
2360         struct page *page;
2361
2362         spin_lock(&mm->page_table_lock);
2363         if (unlikely(!pmd_trans_huge(*pmd))) {
2364                 spin_unlock(&mm->page_table_lock);
2365                 return;
2366         }
2367         page = pmd_page(*pmd);
2368         VM_BUG_ON(!page_count(page));
2369         get_page(page);
2370         spin_unlock(&mm->page_table_lock);
2371
2372         split_huge_page(page);
2373
2374         put_page(page);
2375         BUG_ON(pmd_trans_huge(*pmd));
2376 }
2377
2378 static void split_huge_page_address(struct mm_struct *mm,
2379                                     unsigned long address)
2380 {
2381         pgd_t *pgd;
2382         pud_t *pud;
2383         pmd_t *pmd;
2384
2385         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2386
2387         pgd = pgd_offset(mm, address);
2388         if (!pgd_present(*pgd))
2389                 return;
2390
2391         pud = pud_offset(pgd, address);
2392         if (!pud_present(*pud))
2393                 return;
2394
2395         pmd = pmd_offset(pud, address);
2396         if (!pmd_present(*pmd))
2397                 return;
2398         /*
2399          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2400          * materialize from under us.
2401          */
2402         split_huge_page_pmd(mm, pmd);
2403 }
2404
2405 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2406                              unsigned long start,
2407                              unsigned long end,
2408                              long adjust_next)
2409 {
2410         /*
2411          * If the new start address isn't hpage aligned and it could
2412          * previously contain an hugepage: check if we need to split
2413          * an huge pmd.
2414          */
2415         if (start & ~HPAGE_PMD_MASK &&
2416             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2417             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2418                 split_huge_page_address(vma->vm_mm, start);
2419
2420         /*
2421          * If the new end address isn't hpage aligned and it could
2422          * previously contain an hugepage: check if we need to split
2423          * an huge pmd.
2424          */
2425         if (end & ~HPAGE_PMD_MASK &&
2426             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2427             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2428                 split_huge_page_address(vma->vm_mm, end);
2429
2430         /*
2431          * If we're also updating the vma->vm_next->vm_start, if the new
2432          * vm_next->vm_start isn't page aligned and it could previously
2433          * contain an hugepage: check if we need to split an huge pmd.
2434          */
2435         if (adjust_next > 0) {
2436                 struct vm_area_struct *next = vma->vm_next;
2437                 unsigned long nstart = next->vm_start;
2438                 nstart += adjust_next << PAGE_SHIFT;
2439                 if (nstart & ~HPAGE_PMD_MASK &&
2440                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2441                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2442                         split_huge_page_address(next->vm_mm, nstart);
2443         }
2444 }