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