mm/huge_memory.c

   1 /*
   2  *  Copyright (C) 2009  Red Hat, Inc.
   3  *
   4  *  This work is licensed under the terms of the GNU GPL, version 2. See
   5  *  the COPYING file in the top-level directory.
   6  */
   7
   8 #include <linux/mm.h>
   9 #include <linux/sched.h>
  10 #include <linux/highmem.h>
  11 #include <linux/hugetlb.h>
  12 #include <linux/mmu_notifier.h>
  13 #include <linux/rmap.h>
  14 #include <linux/swap.h>
  15 #include <linux/mm_inline.h>
  16 #include <linux/kthread.h>
  17 #include <linux/khugepaged.h>
  18 #include <linux/freezer.h>
  19 #include <linux/mman.h>
  20 #include <asm/tlb.h>
  21 #include <asm/pgalloc.h>
  22 #include "internal.h"
  23
  24 /*
  25  * By default transparent hugepage support is enabled for all mappings
  26  * and khugepaged scans all mappings. Defrag is only invoked by
  27  * khugepaged hugepage allocations and by page faults inside
  28  * MADV_HUGEPAGE regions to avoid the risk of slowing down short lived
  29  * allocations.
  30  */
  31 unsigned long transparent_hugepage_flags __read_mostly =
  32 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  33         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  34 #endif
  35 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  36         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  37 #endif
  38         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
  39         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
  40
  41 /* default scan 8*512 pte (or vmas) every 30 second */
  42 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
  43 static unsigned int khugepaged_pages_collapsed;
  44 static unsigned int khugepaged_full_scans;
  45 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  46 /* during fragmentation poll the hugepage allocator once every minute */
  47 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  48 static struct task_struct *khugepaged_thread __read_mostly;
  49 static DEFINE_MUTEX(khugepaged_mutex);
  50 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  51 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  52 /*
  53  * default collapse hugepages if there is at least one pte mapped like
  54  * it would have happened if the vma was large enough during page
  55  * fault.
  56  */
  57 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
  58
  59 static int khugepaged(void *none);
  60 static int mm_slots_hash_init(void);
  61 static int khugepaged_slab_init(void);
  62 static void khugepaged_slab_free(void);
  63
  64 #define MM_SLOTS_HASH_HEADS 1024
  65 static struct hlist_head *mm_slots_hash __read_mostly;
  66 static struct kmem_cache *mm_slot_cache __read_mostly;
  67
  68 #ifdef CONFIG_FB
  69 extern const struct file_operations fb_fops;
  70
  71 #define is_fb_vma(vma) \
  72         (vma->vm_file && vma->vm_file->f_op == &fb_fops)
  73 #else
  74 #define is_fb_vma(vma) 0
  75 #endif
  76
  77 static void split_fb_pmd(struct vm_area_struct *vma, pmd_t *pmd);
  78
  79 /**
  80  * struct mm_slot - hash lookup from mm to mm_slot
  81  * @hash: hash collision list
  82  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  83  * @mm: the mm that this information is valid for
  84  */
  85 struct mm_slot {
  86         struct hlist_node hash;
  87         struct list_head mm_node;
  88         struct mm_struct *mm;
  89 };
  90
  91 /**
  92  * struct khugepaged_scan - cursor for scanning
  93  * @mm_head: the head of the mm list to scan
  94  * @mm_slot: the current mm_slot we are scanning
  95  * @address: the next address inside that to be scanned
  96  *
  97  * There is only the one khugepaged_scan instance of this cursor structure.
  98  */
  99 struct khugepaged_scan {
 100         struct list_head mm_head;
 101         struct mm_slot *mm_slot;
 102         unsigned long address;
 103 };
 104 static struct khugepaged_scan khugepaged_scan = {
 105         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 106 };
 107
 108
 109 static int set_recommended_min_free_kbytes(void)
 110 {
 111         struct zone *zone;
 112         int nr_zones = 0;
 113         unsigned long recommended_min;
 114         extern int min_free_kbytes;
 115
 116         if (!test_bit(TRANSPARENT_HUGEPAGE_FLAG,
 117                       &transparent_hugepage_flags) &&
 118             !test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 119                       &transparent_hugepage_flags))
 120                 return 0;
 121
 122         for_each_populated_zone(zone)
 123                 nr_zones++;
 124
 125         /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
 126         recommended_min = pageblock_nr_pages * nr_zones * 2;
 127
 128         /*
 129          * Make sure that on average at least two pageblocks are almost free
 130          * of another type, one for a migratetype to fall back to and a
 131          * second to avoid subsequent fallbacks of other types There are 3
 132          * MIGRATE_TYPES we care about.
 133          */
 134         recommended_min += pageblock_nr_pages * nr_zones *
 135                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
 136
 137         /* don't ever allow to reserve more than 5% of the lowmem */
 138         recommended_min = min(recommended_min,
 139                               (unsigned long) nr_free_buffer_pages() / 20);
 140         recommended_min <<= (PAGE_SHIFT-10);
 141
 142         if (recommended_min > min_free_kbytes)
 143                 min_free_kbytes = recommended_min;
 144         setup_per_zone_wmarks();
 145         return 0;
 146 }
 147 late_initcall(set_recommended_min_free_kbytes);
 148
 149 static int start_khugepaged(void)
 150 {
 151         int err = 0;
 152         if (khugepaged_enabled()) {
 153                 int wakeup;
 154                 if (unlikely(!mm_slot_cache || !mm_slots_hash)) {
 155                         err = -ENOMEM;
 156                         goto out;
 157                 }
 158                 mutex_lock(&khugepaged_mutex);
 159                 if (!khugepaged_thread)
 160                         khugepaged_thread = kthread_run(khugepaged, NULL,
 161                                                         "khugepaged");
 162                 if (unlikely(IS_ERR(khugepaged_thread))) {
 163                         printk(KERN_ERR
 164                                "khugepaged: kthread_run(khugepaged) failed\n");
 165                         err = PTR_ERR(khugepaged_thread);
 166                         khugepaged_thread = NULL;
 167                 }
 168                 wakeup = !list_empty(&khugepaged_scan.mm_head);
 169                 mutex_unlock(&khugepaged_mutex);
 170                 if (wakeup)
 171                         wake_up_interruptible(&khugepaged_wait);
 172
 173                 set_recommended_min_free_kbytes();
 174         } else
 175                 /* wakeup to exit */
 176                 wake_up_interruptible(&khugepaged_wait);
 177 out:
 178         return err;
 179 }
 180
 181 #ifdef CONFIG_SYSFS
 182
 183 static ssize_t double_flag_show(struct kobject *kobj,
 184                                 struct kobj_attribute *attr, char *buf,
 185                                 enum transparent_hugepage_flag enabled,
 186                                 enum transparent_hugepage_flag req_madv)
 187 {
 188         if (test_bit(enabled, &transparent_hugepage_flags)) {
 189                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
 190                 return sprintf(buf, "[always] madvise never\n");
 191         } else if (test_bit(req_madv, &transparent_hugepage_flags))
 192                 return sprintf(buf, "always [madvise] never\n");
 193         else
 194                 return sprintf(buf, "always madvise [never]\n");
 195 }
 196 static ssize_t double_flag_store(struct kobject *kobj,
 197                                  struct kobj_attribute *attr,
 198                                  const char *buf, size_t count,
 199                                  enum transparent_hugepage_flag enabled,
 200                                  enum transparent_hugepage_flag req_madv)
 201 {
 202         if (!memcmp("always", buf,
 203                     min(sizeof("always")-1, count))) {
 204                 set_bit(enabled, &transparent_hugepage_flags);
 205                 clear_bit(req_madv, &transparent_hugepage_flags);
 206         } else if (!memcmp("madvise", buf,
 207                            min(sizeof("madvise")-1, count))) {
 208                 clear_bit(enabled, &transparent_hugepage_flags);
 209                 set_bit(req_madv, &transparent_hugepage_flags);
 210         } else if (!memcmp("never", buf,
 211                            min(sizeof("never")-1, count))) {
 212                 clear_bit(enabled, &transparent_hugepage_flags);
 213                 clear_bit(req_madv, &transparent_hugepage_flags);
 214         } else
 215                 return -EINVAL;
 216
 217         return count;
 218 }
 219
 220 static ssize_t enabled_show(struct kobject *kobj,
 221                             struct kobj_attribute *attr, char *buf)
 222 {
 223         return double_flag_show(kobj, attr, buf,
 224                                 TRANSPARENT_HUGEPAGE_FLAG,
 225                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 226 }
 227 static ssize_t enabled_store(struct kobject *kobj,
 228                              struct kobj_attribute *attr,
 229                              const char *buf, size_t count)
 230 {
 231         ssize_t ret;
 232
 233         ret = double_flag_store(kobj, attr, buf, count,
 234                                 TRANSPARENT_HUGEPAGE_FLAG,
 235                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 236
 237         if (ret > 0) {
 238                 int err = start_khugepaged();
 239                 if (err)
 240                         ret = err;
 241         }
 242
 243         if (ret > 0 &&
 244             (test_bit(TRANSPARENT_HUGEPAGE_FLAG,
 245                       &transparent_hugepage_flags) ||
 246              test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 247                       &transparent_hugepage_flags)))
 248                 set_recommended_min_free_kbytes();
 249
 250         return ret;
 251 }
 252 static struct kobj_attribute enabled_attr =
 253         __ATTR(enabled, 0644, enabled_show, enabled_store);
 254
 255 static ssize_t single_flag_show(struct kobject *kobj,
 256                                 struct kobj_attribute *attr, char *buf,
 257                                 enum transparent_hugepage_flag flag)
 258 {
 259         return sprintf(buf, "%d\n",
 260                        !!test_bit(flag, &transparent_hugepage_flags));
 261 }
 262
 263 static ssize_t single_flag_store(struct kobject *kobj,
 264                                  struct kobj_attribute *attr,
 265                                  const char *buf, size_t count,
 266                                  enum transparent_hugepage_flag flag)
 267 {
 268         unsigned long value;
 269         int ret;
 270
 271         ret = kstrtoul(buf, 10, &value);
 272         if (ret < 0)
 273                 return ret;
 274         if (value > 1)
 275                 return -EINVAL;
 276
 277         if (value)
 278                 set_bit(flag, &transparent_hugepage_flags);
 279         else
 280                 clear_bit(flag, &transparent_hugepage_flags);
 281
 282         return count;
 283 }
 284
 285 /*
 286  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 287  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 288  * memory just to allocate one more hugepage.
 289  */
 290 static ssize_t defrag_show(struct kobject *kobj,
 291                            struct kobj_attribute *attr, char *buf)
 292 {
 293         return double_flag_show(kobj, attr, buf,
 294                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 295                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 296 }
 297 static ssize_t defrag_store(struct kobject *kobj,
 298                             struct kobj_attribute *attr,
 299                             const char *buf, size_t count)
 300 {
 301         return double_flag_store(kobj, attr, buf, count,
 302                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 303                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 304 }
 305 static struct kobj_attribute defrag_attr =
 306         __ATTR(defrag, 0644, defrag_show, defrag_store);
 307
 308 #ifdef CONFIG_DEBUG_VM
 309 static ssize_t debug_cow_show(struct kobject *kobj,
 310                                 struct kobj_attribute *attr, char *buf)
 311 {
 312         return single_flag_show(kobj, attr, buf,
 313                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 314 }
 315 static ssize_t debug_cow_store(struct kobject *kobj,
 316                                struct kobj_attribute *attr,
 317                                const char *buf, size_t count)
 318 {
 319         return single_flag_store(kobj, attr, buf, count,
 320                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 321 }
 322 static struct kobj_attribute debug_cow_attr =
 323         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 324 #endif /* CONFIG_DEBUG_VM */
 325
 326 static struct attribute *hugepage_attr[] = {
 327         &enabled_attr.attr,
 328         &defrag_attr.attr,
 329 #ifdef CONFIG_DEBUG_VM
 330         &debug_cow_attr.attr,
 331 #endif
 332         NULL,
 333 };
 334
 335 static struct attribute_group hugepage_attr_group = {
 336         .attrs = hugepage_attr,
 337 };
 338
 339 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 340                                          struct kobj_attribute *attr,
 341                                          char *buf)
 342 {
 343         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 344 }
 345
 346 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 347                                           struct kobj_attribute *attr,
 348                                           const char *buf, size_t count)
 349 {
 350         unsigned long msecs;
 351         int err;
 352
 353         err = strict_strtoul(buf, 10, &msecs);
 354         if (err || msecs > UINT_MAX)
 355                 return -EINVAL;
 356
 357         khugepaged_scan_sleep_millisecs = msecs;
 358         wake_up_interruptible(&khugepaged_wait);
 359
 360         return count;
 361 }
 362 static struct kobj_attribute scan_sleep_millisecs_attr =
 363         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 364                scan_sleep_millisecs_store);
 365
 366 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 367                                           struct kobj_attribute *attr,
 368                                           char *buf)
 369 {
 370         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 371 }
 372
 373 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 374                                            struct kobj_attribute *attr,
 375                                            const char *buf, size_t count)
 376 {
 377         unsigned long msecs;
 378         int err;
 379
 380         err = strict_strtoul(buf, 10, &msecs);
 381         if (err || msecs > UINT_MAX)
 382                 return -EINVAL;
 383
 384         khugepaged_alloc_sleep_millisecs = msecs;
 385         wake_up_interruptible(&khugepaged_wait);
 386
 387         return count;
 388 }
 389 static struct kobj_attribute alloc_sleep_millisecs_attr =
 390         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 391                alloc_sleep_millisecs_store);
 392
 393 static ssize_t pages_to_scan_show(struct kobject *kobj,
 394                                   struct kobj_attribute *attr,
 395                                   char *buf)
 396 {
 397         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 398 }
 399 static ssize_t pages_to_scan_store(struct kobject *kobj,
 400                                    struct kobj_attribute *attr,
 401                                    const char *buf, size_t count)
 402 {
 403         int err;
 404         unsigned long pages;
 405
 406         err = strict_strtoul(buf, 10, &pages);
 407         if (err || !pages || pages > UINT_MAX)
 408                 return -EINVAL;
 409
 410         khugepaged_pages_to_scan = pages;
 411
 412         return count;
 413 }
 414 static struct kobj_attribute pages_to_scan_attr =
 415         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 416                pages_to_scan_store);
 417
 418 static ssize_t pages_collapsed_show(struct kobject *kobj,
 419                                     struct kobj_attribute *attr,
 420                                     char *buf)
 421 {
 422         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 423 }
 424 static struct kobj_attribute pages_collapsed_attr =
 425         __ATTR_RO(pages_collapsed);
 426
 427 static ssize_t full_scans_show(struct kobject *kobj,
 428                                struct kobj_attribute *attr,
 429                                char *buf)
 430 {
 431         return sprintf(buf, "%u\n", khugepaged_full_scans);
 432 }
 433 static struct kobj_attribute full_scans_attr =
 434         __ATTR_RO(full_scans);
 435
 436 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 437                                       struct kobj_attribute *attr, char *buf)
 438 {
 439         return single_flag_show(kobj, attr, buf,
 440                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 441 }
 442 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 443                                        struct kobj_attribute *attr,
 444                                        const char *buf, size_t count)
 445 {
 446         return single_flag_store(kobj, attr, buf, count,
 447                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 448 }
 449 static struct kobj_attribute khugepaged_defrag_attr =
 450         __ATTR(defrag, 0644, khugepaged_defrag_show,
 451                khugepaged_defrag_store);
 452
 453 /*
 454  * max_ptes_none controls if khugepaged should collapse hugepages over
 455  * any unmapped ptes in turn potentially increasing the memory
 456  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 457  * reduce the available free memory in the system as it
 458  * runs. Increasing max_ptes_none will instead potentially reduce the
 459  * free memory in the system during the khugepaged scan.
 460  */
 461 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 462                                              struct kobj_attribute *attr,
 463                                              char *buf)
 464 {
 465         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 466 }
 467 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 468                                               struct kobj_attribute *attr,
 469                                               const char *buf, size_t count)
 470 {
 471         int err;
 472         unsigned long max_ptes_none;
 473
 474         err = strict_strtoul(buf, 10, &max_ptes_none);
 475         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 476                 return -EINVAL;
 477
 478         khugepaged_max_ptes_none = max_ptes_none;
 479
 480         return count;
 481 }
 482 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 483         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 484                khugepaged_max_ptes_none_store);
 485
 486 static struct attribute *khugepaged_attr[] = {
 487         &khugepaged_defrag_attr.attr,
 488         &khugepaged_max_ptes_none_attr.attr,
 489         &pages_to_scan_attr.attr,
 490         &pages_collapsed_attr.attr,
 491         &full_scans_attr.attr,
 492         &scan_sleep_millisecs_attr.attr,
 493         &alloc_sleep_millisecs_attr.attr,
 494         NULL,
 495 };
 496
 497 static struct attribute_group khugepaged_attr_group = {
 498         .attrs = khugepaged_attr,
 499         .name = "khugepaged",
 500 };
 501 #endif /* CONFIG_SYSFS */
 502
 503 static int __init hugepage_init(void)
 504 {
 505         int err;
 506 #ifdef CONFIG_SYSFS
 507         static struct kobject *hugepage_kobj;
 508 #endif
 509
 510         err = -EINVAL;
 511         if (!has_transparent_hugepage()) {
 512                 transparent_hugepage_flags = 0;
 513                 goto out;
 514         }
 515
 516 #ifdef CONFIG_SYSFS
 517         err = -ENOMEM;
 518         hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 519         if (unlikely(!hugepage_kobj)) {
 520                 printk(KERN_ERR "hugepage: failed kobject create\n");
 521                 goto out;
 522         }
 523
 524         err = sysfs_create_group(hugepage_kobj, &hugepage_attr_group);
 525         if (err) {
 526                 printk(KERN_ERR "hugepage: failed register hugeage group\n");
 527                 goto out;
 528         }
 529
 530         err = sysfs_create_group(hugepage_kobj, &khugepaged_attr_group);
 531         if (err) {
 532                 printk(KERN_ERR "hugepage: failed register hugeage group\n");
 533                 goto out;
 534         }
 535 #endif
 536
 537         err = khugepaged_slab_init();
 538         if (err)
 539                 goto out;
 540
 541         err = mm_slots_hash_init();
 542         if (err) {
 543                 khugepaged_slab_free();
 544                 goto out;
 545         }
 546
 547         /*
 548          * By default disable transparent hugepages on smaller systems,
 549          * where the extra memory used could hurt more than TLB overhead
 550          * is likely to save.  The admin can still enable it through /sys.
 551          */
 552         if (totalram_pages < (200 << (20 - PAGE_SHIFT)))
 553                 transparent_hugepage_flags = 0;
 554
 555         start_khugepaged();
 556
 557         set_recommended_min_free_kbytes();
 558
 559 out:
 560         return err;
 561 }
 562 module_init(hugepage_init)
 563
 564 static int __init setup_transparent_hugepage(char *str)
 565 {
 566         int ret = 0;
 567         if (!str)
 568                 goto out;
 569         if (!strcmp(str, "always")) {
 570                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 571                         &transparent_hugepage_flags);
 572                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 573                           &transparent_hugepage_flags);
 574                 ret = 1;
 575         } else if (!strcmp(str, "madvise")) {
 576                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 577                           &transparent_hugepage_flags);
 578                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 579                         &transparent_hugepage_flags);
 580                 ret = 1;
 581         } else if (!strcmp(str, "never")) {
 582                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 583                           &transparent_hugepage_flags);
 584                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 585                           &transparent_hugepage_flags);
 586                 ret = 1;
 587         }
 588 out:
 589         if (!ret)
 590                 printk(KERN_WARNING
 591                        "transparent_hugepage= cannot parse, ignored\n");
 592         return ret;
 593 }
 594 __setup("transparent_hugepage=", setup_transparent_hugepage);
 595
 596 static void prepare_pmd_huge_pte(pgtable_t pgtable,
 597                                  struct mm_struct *mm)
 598 {
 599         assert_spin_locked(&mm->page_table_lock);
 600
 601         /* FIFO */
 602         if (!mm->pmd_huge_pte)
 603                 INIT_LIST_HEAD(&pgtable->lru);
 604         else
 605                 list_add(&pgtable->lru, &mm->pmd_huge_pte->lru);
 606         mm->pmd_huge_pte = pgtable;
 607 }
 608
 609 static inline pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 610 {
 611         if (likely(vma->vm_flags & VM_WRITE))
 612                 pmd = pmd_mkwrite(pmd);
 613         return pmd;
 614 }
 615
 616 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
 617                                         struct vm_area_struct *vma,
 618                                         unsigned long haddr, pmd_t *pmd,
 619                                         struct page *page)
 620 {
 621         int ret = 0;
 622         pgtable_t pgtable;
 623
 624         VM_BUG_ON(!PageCompound(page));
 625         pgtable = pte_alloc_one(mm, haddr);
 626         if (unlikely(!pgtable)) {
 627                 mem_cgroup_uncharge_page(page);
 628                 put_page(page);
 629                 return VM_FAULT_OOM;
 630         }
 631
 632         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 633         __SetPageUptodate(page);
 634
 635         spin_lock(&mm->page_table_lock);
 636         if (unlikely(!pmd_none(*pmd))) {
 637                 spin_unlock(&mm->page_table_lock);
 638                 mem_cgroup_uncharge_page(page);
 639                 put_page(page);
 640                 pte_free(mm, pgtable);
 641         } else {
 642                 pmd_t entry;
 643                 entry = mk_pmd(page, vma->vm_page_prot);
 644                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 645                 entry = pmd_mkhuge(entry);
 646                 /*
 647                  * The spinlocking to take the lru_lock inside
 648                  * page_add_new_anon_rmap() acts as a full memory
 649                  * barrier to be sure clear_huge_page writes become
 650                  * visible after the set_pmd_at() write.
 651                  */
 652                 page_add_new_anon_rmap(page, vma, haddr);
 653                 set_pmd_at(mm, haddr, pmd, entry);
 654                 prepare_pmd_huge_pte(pgtable, mm);
 655                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 656                 mm->nr_ptes++;
 657                 spin_unlock(&mm->page_table_lock);
 658         }
 659
 660         return ret;
 661 }
 662
 663 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 664 {
 665         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
 666 }
 667
 668 static inline struct page *alloc_hugepage_vma(int defrag,
 669                                               struct vm_area_struct *vma,
 670                                               unsigned long haddr, int nd,
 671                                               gfp_t extra_gfp)
 672 {
 673         return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
 674                                HPAGE_PMD_ORDER, vma, haddr, nd);
 675 }
 676
 677 #ifndef CONFIG_NUMA
 678 static inline struct page *alloc_hugepage(int defrag)
 679 {
 680         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
 681                            HPAGE_PMD_ORDER);
 682 }
 683 #endif
 684
 685 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 686                                unsigned long address, pmd_t *pmd,
 687                                unsigned int flags)
 688 {
 689         struct page *page;
 690         unsigned long haddr = address & HPAGE_PMD_MASK;
 691         pte_t *pte;
 692
 693         if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) {
 694                 if (unlikely(anon_vma_prepare(vma)))
 695                         return VM_FAULT_OOM;
 696                 if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 697                         return VM_FAULT_OOM;
 698                 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 699                                           vma, haddr, numa_node_id(), 0);
 700                 if (unlikely(!page)) {
 701                         count_vm_event(THP_FAULT_FALLBACK);
 702                         goto out;
 703                 }
 704                 count_vm_event(THP_FAULT_ALLOC);
 705                 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
 706                         put_page(page);
 707                         goto out;
 708                 }
 709
 710                 return __do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page);
 711         }
 712 out:
 713         /*
 714          * Use __pte_alloc instead of pte_alloc_map, because we can't
 715          * run pte_offset_map on the pmd, if an huge pmd could
 716          * materialize from under us from a different thread.
 717          */
 718         if (unlikely(__pte_alloc(mm, vma, pmd, address)))
 719                 return VM_FAULT_OOM;
 720         /* if an huge pmd materialized from under us just retry later */
 721         if (unlikely(pmd_trans_huge(*pmd)))
 722                 return 0;
 723         /*
 724          * A regular pmd is established and it can't morph into a huge pmd
 725          * from under us anymore at this point because we hold the mmap_sem
 726          * read mode and khugepaged takes it in write mode. So now it's
 727          * safe to run pte_offset_map().
 728          */
 729         pte = pte_offset_map(pmd, address);
 730         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
 731 }
 732
 733 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 734                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 735                   struct vm_area_struct *vma)
 736 {
 737         struct page *src_page;
 738         pmd_t pmd;
 739         pgtable_t pgtable;
 740         int ret;
 741
 742         ret = -ENOMEM;
 743         pgtable = pte_alloc_one(dst_mm, addr);
 744         if (unlikely(!pgtable))
 745                 goto out;
 746
 747         spin_lock(&dst_mm->page_table_lock);
 748         spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
 749
 750         ret = -EAGAIN;
 751         pmd = *src_pmd;
 752         if (unlikely(!pmd_trans_huge(pmd))) {
 753                 pte_free(dst_mm, pgtable);
 754                 goto out_unlock;
 755         }
 756         if (unlikely(pmd_trans_splitting(pmd))) {
 757                 /* split huge page running from under us */
 758                 spin_unlock(&src_mm->page_table_lock);
 759                 spin_unlock(&dst_mm->page_table_lock);
 760                 pte_free(dst_mm, pgtable);
 761
 762                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 763                 goto out;
 764         }
 765         src_page = pmd_page(pmd);
 766         VM_BUG_ON(!PageHead(src_page));
 767         get_page(src_page);
 768         page_dup_rmap(src_page);
 769         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 770
 771         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 772         pmd = pmd_mkold(pmd_wrprotect(pmd));
 773         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 774         prepare_pmd_huge_pte(pgtable, dst_mm);
 775         dst_mm->nr_ptes++;
 776
 777         ret = 0;
 778 out_unlock:
 779         spin_unlock(&src_mm->page_table_lock);
 780         spin_unlock(&dst_mm->page_table_lock);
 781 out:
 782         return ret;
 783 }
 784
 785 /* no "address" argument so destroys page coloring of some arch */
 786 pgtable_t get_pmd_huge_pte(struct mm_struct *mm)
 787 {
 788         pgtable_t pgtable;
 789
 790         assert_spin_locked(&mm->page_table_lock);
 791
 792         /* FIFO */
 793         pgtable = mm->pmd_huge_pte;
 794         if (list_empty(&pgtable->lru))
 795                 mm->pmd_huge_pte = NULL;
 796         else {
 797                 mm->pmd_huge_pte = list_entry(pgtable->lru.next,
 798                                               struct page, lru);
 799                 list_del(&pgtable->lru);
 800         }
 801         return pgtable;
 802 }
 803
 804 void huge_pmd_set_accessed(struct mm_struct *mm,
 805                            struct vm_area_struct *vma,
 806                            unsigned long address,
 807                            pmd_t *pmd, pmd_t orig_pmd,
 808                            int dirty)
 809 {
 810         pmd_t entry;
 811         unsigned long haddr;
 812
 813         spin_lock(&mm->page_table_lock);
 814         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 815                 goto unlock;
 816
 817         entry = pmd_mkyoung(orig_pmd);
 818         haddr = address & HPAGE_PMD_MASK;
 819         if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
 820                 update_mmu_cache_pmd(vma, address, pmd);
 821
 822 unlock:
 823         spin_unlock(&mm->page_table_lock);
 824 }
 825
 826 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
 827                                         struct vm_area_struct *vma,
 828                                         unsigned long address,
 829                                         pmd_t *pmd, pmd_t orig_pmd,
 830                                         struct page *page,
 831                                         unsigned long haddr)
 832 {
 833         pgtable_t pgtable;
 834         pmd_t _pmd;
 835         int ret = 0, i;
 836         struct page **pages;
 837
 838         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
 839                         GFP_KERNEL);
 840         if (unlikely(!pages)) {
 841                 ret |= VM_FAULT_OOM;
 842                 goto out;
 843         }
 844
 845         for (i = 0; i < HPAGE_PMD_NR; i++) {
 846                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
 847                                                __GFP_OTHER_NODE,
 848                                                vma, address, page_to_nid(page));
 849                 if (unlikely(!pages[i] ||
 850                              mem_cgroup_newpage_charge(pages[i], mm,
 851                                                        GFP_KERNEL))) {
 852                         if (pages[i])
 853                                 put_page(pages[i]);
 854                         mem_cgroup_uncharge_start();
 855                         while (--i >= 0) {
 856                                 mem_cgroup_uncharge_page(pages[i]);
 857                                 put_page(pages[i]);
 858                         }
 859                         mem_cgroup_uncharge_end();
 860                         kfree(pages);
 861                         ret |= VM_FAULT_OOM;
 862                         goto out;
 863                 }
 864         }
 865
 866         for (i = 0; i < HPAGE_PMD_NR; i++) {
 867                 copy_user_highpage(pages[i], page + i,
 868                                    haddr + PAGE_SIZE * i, vma);
 869                 __SetPageUptodate(pages[i]);
 870                 cond_resched();
 871         }
 872
 873         spin_lock(&mm->page_table_lock);
 874         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 875                 goto out_free_pages;
 876         VM_BUG_ON(!PageHead(page));
 877
 878         pmdp_clear_flush_notify(vma, haddr, pmd);
 879         /* leave pmd empty until pte is filled */
 880
 881         pgtable = get_pmd_huge_pte(mm);
 882         pmd_populate(mm, &_pmd, pgtable);
 883
 884         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
 885                 pte_t *pte, entry;
 886                 entry = mk_pte(pages[i], vma->vm_page_prot);
 887                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
 888                 page_add_new_anon_rmap(pages[i], vma, haddr);
 889                 pte = pte_offset_map(&_pmd, haddr);
 890                 VM_BUG_ON(!pte_none(*pte));
 891                 set_pte_at(mm, haddr, pte, entry);
 892                 pte_unmap(pte);
 893         }
 894         kfree(pages);
 895
 896         smp_wmb(); /* make pte visible before pmd */
 897         pmd_populate(mm, pmd, pgtable);
 898         page_remove_rmap(page);
 899         spin_unlock(&mm->page_table_lock);
 900
 901         ret |= VM_FAULT_WRITE;
 902         put_page(page);
 903
 904 out:
 905         return ret;
 906
 907 out_free_pages:
 908         spin_unlock(&mm->page_table_lock);
 909         mem_cgroup_uncharge_start();
 910         for (i = 0; i < HPAGE_PMD_NR; i++) {
 911                 mem_cgroup_uncharge_page(pages[i]);
 912                 put_page(pages[i]);
 913         }
 914         mem_cgroup_uncharge_end();
 915         kfree(pages);
 916         goto out;
 917 }
 918
 919 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
 920                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
 921 {
 922         int ret = 0;
 923         struct page *page, *new_page;
 924         unsigned long haddr;
 925
 926         VM_BUG_ON(!vma->anon_vma);
 927         spin_lock(&mm->page_table_lock);
 928         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 929                 goto out_unlock;
 930
 931         page = pmd_page(orig_pmd);
 932         VM_BUG_ON(!PageCompound(page) || !PageHead(page));
 933         haddr = address & HPAGE_PMD_MASK;
 934         if (page_mapcount(page) == 1) {
 935                 pmd_t entry;
 936                 entry = pmd_mkyoung(orig_pmd);
 937                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 938                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
 939                         update_mmu_cache(vma, address, pmd);
 940                 ret |= VM_FAULT_WRITE;
 941                 goto out_unlock;
 942         }
 943         get_page(page);
 944         spin_unlock(&mm->page_table_lock);
 945
 946         if (transparent_hugepage_enabled(vma) &&
 947             !transparent_hugepage_debug_cow())
 948                 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 949                                               vma, haddr, numa_node_id(), 0);
 950         else
 951                 new_page = NULL;
 952
 953         if (unlikely(!new_page)) {
 954                 count_vm_event(THP_FAULT_FALLBACK);
 955                 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
 956                                                    pmd, orig_pmd, page, haddr);
 957                 if (ret & VM_FAULT_OOM)
 958                         split_huge_page(page);
 959                 put_page(page);
 960                 goto out;
 961         }
 962         count_vm_event(THP_FAULT_ALLOC);
 963
 964         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
 965                 put_page(new_page);
 966                 split_huge_page(page);
 967                 put_page(page);
 968                 ret |= VM_FAULT_OOM;
 969                 goto out;
 970         }
 971
 972         copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
 973         __SetPageUptodate(new_page);
 974
 975         spin_lock(&mm->page_table_lock);
 976         put_page(page);
 977         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
 978                 mem_cgroup_uncharge_page(new_page);
 979                 put_page(new_page);
 980         } else {
 981                 pmd_t entry;
 982                 VM_BUG_ON(!PageHead(page));
 983                 entry = mk_pmd(new_page, vma->vm_page_prot);
 984                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 985                 entry = pmd_mkhuge(entry);
 986                 pmdp_clear_flush_notify(vma, haddr, pmd);
 987                 page_add_new_anon_rmap(new_page, vma, haddr);
 988                 set_pmd_at(mm, haddr, pmd, entry);
 989                 update_mmu_cache(vma, address, pmd);
 990                 page_remove_rmap(page);
 991                 put_page(page);
 992                 ret |= VM_FAULT_WRITE;
 993         }
 994 out_unlock:
 995         spin_unlock(&mm->page_table_lock);
 996 out:
 997         return ret;
 998 }
 999
1000 /*
1001  * FOLL_FORCE can write to even unwritable pmd's, but only
1002  * after we've gone through a COW cycle and they are dirty.
1003  */
1004 static inline bool can_follow_write_pmd(pmd_t pmd, struct page *page,
1005                                         unsigned int flags)
1006 {
1007         return pmd_write(pmd) ||
1008                 ((flags & FOLL_FORCE) && (flags & FOLL_COW) &&
1009                  page && PageAnon(page));
1010 }
1011
1012 struct page *follow_trans_huge_pmd(struct mm_struct *mm,
1013                                    unsigned long addr,
1014                                    pmd_t *pmd,
1015                                    unsigned int flags)
1016 {
1017         struct page *page = NULL;
1018
1019         assert_spin_locked(&mm->page_table_lock);
1020
1021         page = pmd_page(*pmd);
1022         VM_BUG_ON(!PageHead(page));
1023
1024         if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, page, flags))
1025                 return NULL;
1026
1027         if (flags & FOLL_TOUCH) {
1028                 pmd_t _pmd;
1029                 /*
1030                  * We should set the dirty bit only for FOLL_WRITE but
1031                  * for now the dirty bit in the pmd is meaningless.
1032                  * And if the dirty bit will become meaningful and
1033                  * we'll only set it with FOLL_WRITE, an atomic
1034                  * set_bit will be required on the pmd to set the
1035                  * young bit, instead of the current set_pmd_at.
1036                  */
1037                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1038                 set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd);
1039         }
1040         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1041         VM_BUG_ON(!PageCompound(page));
1042         if (flags & FOLL_GET)
1043                 get_page_foll(page);
1044
1045         return page;
1046 }
1047
1048 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1049                  pmd_t *pmd)
1050 {
1051         int ret = 0;
1052
1053         spin_lock(&tlb->mm->page_table_lock);
1054         if (likely(pmd_trans_huge(*pmd))) {
1055                 if (is_fb_vma(vma)) {
1056                         split_fb_pmd(vma, pmd);
1057                         return 0;
1058                 }
1059
1060                 if (unlikely(pmd_trans_splitting(*pmd))) {
1061                         spin_unlock(&tlb->mm->page_table_lock);
1062                         wait_split_huge_page(vma->anon_vma,
1063                                              pmd);
1064                 } else {
1065                         struct page *page;
1066                         pgtable_t pgtable;
1067                         pgtable = get_pmd_huge_pte(tlb->mm);
1068                         page = pmd_page(*pmd);
1069                         pmd_clear(pmd);
1070                         page_remove_rmap(page);
1071                         VM_BUG_ON(page_mapcount(page) < 0);
1072                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1073                         VM_BUG_ON(!PageHead(page));
1074                         tlb->mm->nr_ptes--;
1075                         spin_unlock(&tlb->mm->page_table_lock);
1076                         tlb_remove_page(tlb, page);
1077                         pte_free(tlb->mm, pgtable);
1078                         ret = 1;
1079                 }
1080         } else
1081                 spin_unlock(&tlb->mm->page_table_lock);
1082
1083         return ret;
1084 }
1085
1086 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1087                 unsigned long addr, unsigned long end,
1088                 unsigned char *vec)
1089 {
1090         int ret = 0;
1091
1092         spin_lock(&vma->vm_mm->page_table_lock);
1093         if (likely(pmd_trans_huge(*pmd))) {
1094                 ret = !pmd_trans_splitting(*pmd);
1095                 spin_unlock(&vma->vm_mm->page_table_lock);
1096                 if (unlikely(!ret))
1097                         wait_split_huge_page(vma->anon_vma, pmd);
1098                 else {
1099                         /*
1100                          * All logical pages in the range are present
1101                          * if backed by a huge page.
1102                          */
1103                         memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1104                 }
1105         } else
1106                 spin_unlock(&vma->vm_mm->page_table_lock);
1107
1108         return ret;
1109 }
1110
1111 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1112                   unsigned long old_addr,
1113                   unsigned long new_addr, unsigned long old_end,
1114                   pmd_t *old_pmd, pmd_t *new_pmd)
1115 {
1116         int ret = 0;
1117         pmd_t pmd;
1118
1119         struct mm_struct *mm = vma->vm_mm;
1120
1121         if ((old_addr & ~HPAGE_PMD_MASK) ||
1122             (new_addr & ~HPAGE_PMD_MASK) ||
1123             old_end - old_addr < HPAGE_PMD_SIZE ||
1124             (new_vma->vm_flags & VM_NOHUGEPAGE))
1125                 goto out;
1126
1127         /*
1128          * The destination pmd shouldn't be established, free_pgtables()
1129          * should have release it.
1130          */
1131         if (WARN_ON(!pmd_none(*new_pmd))) {
1132                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1133                 goto out;
1134         }
1135
1136         spin_lock(&mm->page_table_lock);
1137         if (likely(pmd_trans_huge(*old_pmd))) {
1138                 if (pmd_trans_splitting(*old_pmd)) {
1139                         spin_unlock(&mm->page_table_lock);
1140                         wait_split_huge_page(vma->anon_vma, old_pmd);
1141                         ret = -1;
1142                 } else {
1143                         pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1144                         VM_BUG_ON(!pmd_none(*new_pmd));
1145                         set_pmd_at(mm, new_addr, new_pmd, pmd);
1146                         spin_unlock(&mm->page_table_lock);
1147                         ret = 1;
1148                 }
1149         } else {
1150                 spin_unlock(&mm->page_table_lock);
1151         }
1152 out:
1153         return ret;
1154 }
1155
1156 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1157                 unsigned long addr, pgprot_t newprot)
1158 {
1159         struct mm_struct *mm = vma->vm_mm;
1160         int ret = 0;
1161
1162         spin_lock(&mm->page_table_lock);
1163         if (likely(pmd_trans_huge(*pmd))) {
1164                 if (unlikely(pmd_trans_splitting(*pmd))) {
1165                         spin_unlock(&mm->page_table_lock);
1166                         wait_split_huge_page(vma->anon_vma, pmd);
1167                 } else {
1168                         pmd_t entry;
1169
1170                         entry = pmdp_get_and_clear(mm, addr, pmd);
1171                         entry = pmd_modify(entry, newprot);
1172                         set_pmd_at(mm, addr, pmd, entry);
1173                         spin_unlock(&vma->vm_mm->page_table_lock);
1174                         flush_tlb_range(vma, addr, addr + HPAGE_PMD_SIZE);
1175                         ret = 1;
1176                 }
1177         } else
1178                 spin_unlock(&vma->vm_mm->page_table_lock);
1179
1180         return ret;
1181 }
1182
1183 pmd_t *page_check_address_pmd(struct page *page,
1184                               struct mm_struct *mm,
1185                               unsigned long address,
1186                               enum page_check_address_pmd_flag flag)
1187 {
1188         pgd_t *pgd;
1189         pud_t *pud;
1190         pmd_t *pmd, *ret = NULL;
1191
1192         if (address & ~HPAGE_PMD_MASK)
1193                 goto out;
1194
1195         pgd = pgd_offset(mm, address);
1196         if (!pgd_present(*pgd))
1197                 goto out;
1198
1199         pud = pud_offset(pgd, address);
1200         if (!pud_present(*pud))
1201                 goto out;
1202
1203         pmd = pmd_offset(pud, address);
1204         if (pmd_none(*pmd))
1205                 goto out;
1206         if (pmd_page(*pmd) != page)
1207                 goto out;
1208         /*
1209          * split_vma() may create temporary aliased mappings. There is
1210          * no risk as long as all huge pmd are found and have their
1211          * splitting bit set before __split_huge_page_refcount
1212          * runs. Finding the same huge pmd more than once during the
1213          * same rmap walk is not a problem.
1214          */
1215         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1216             pmd_trans_splitting(*pmd))
1217                 goto out;
1218         if (pmd_trans_huge(*pmd)) {
1219                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1220                           !pmd_trans_splitting(*pmd));
1221                 ret = pmd;
1222         }
1223 out:
1224         return ret;
1225 }
1226
1227 static int __split_huge_page_splitting(struct page *page,
1228                                        struct vm_area_struct *vma,
1229                                        unsigned long address)
1230 {
1231         struct mm_struct *mm = vma->vm_mm;
1232         pmd_t *pmd;
1233         int ret = 0;
1234
1235         spin_lock(&mm->page_table_lock);
1236         pmd = page_check_address_pmd(page, mm, address,
1237                                      PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1238         if (pmd) {
1239                 /*
1240                  * We can't temporarily set the pmd to null in order
1241                  * to split it, the pmd must remain marked huge at all
1242                  * times or the VM won't take the pmd_trans_huge paths
1243                  * and it won't wait on the anon_vma->root->mutex to
1244                  * serialize against split_huge_page*.
1245                  */
1246                 pmdp_splitting_flush_notify(vma, address, pmd);
1247                 ret = 1;
1248         }
1249         spin_unlock(&mm->page_table_lock);
1250
1251         return ret;
1252 }
1253
1254 static void __split_huge_page_refcount(struct page *page)
1255 {
1256         int i;
1257         unsigned long head_index = page->index;
1258         struct zone *zone = page_zone(page);
1259         int zonestat;
1260         int tail_count = 0;
1261
1262         /* prevent PageLRU to go away from under us, and freeze lru stats */
1263         spin_lock_irq(&zone->lru_lock);
1264         compound_lock(page);
1265
1266         for (i = 1; i < HPAGE_PMD_NR; i++) {
1267                 struct page *page_tail = page + i;
1268
1269                 /* tail_page->_mapcount cannot change */
1270                 BUG_ON(page_mapcount(page_tail) < 0);
1271                 tail_count += page_mapcount(page_tail);
1272                 /* check for overflow */
1273                 BUG_ON(tail_count < 0);
1274                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1275                 /*
1276                  * tail_page->_count is zero and not changing from
1277                  * under us. But get_page_unless_zero() may be running
1278                  * from under us on the tail_page. If we used
1279                  * atomic_set() below instead of atomic_add(), we
1280                  * would then run atomic_set() concurrently with
1281                  * get_page_unless_zero(), and atomic_set() is
1282                  * implemented in C not using locked ops. spin_unlock
1283                  * on x86 sometime uses locked ops because of PPro
1284                  * errata 66, 92, so unless somebody can guarantee
1285                  * atomic_set() here would be safe on all archs (and
1286                  * not only on x86), it's safer to use atomic_add().
1287                  */
1288                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1289                            &page_tail->_count);
1290
1291                 /* after clearing PageTail the gup refcount can be released */
1292                 smp_mb();
1293
1294                 /*
1295                  * retain hwpoison flag of the poisoned tail page:
1296                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1297                  *   by the memory-failure.
1298                  */
1299                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1300                 page_tail->flags |= (page->flags &
1301                                      ((1L << PG_referenced) |
1302                                       (1L << PG_swapbacked) |
1303                                       (1L << PG_mlocked) |
1304                                       (1L << PG_uptodate)));
1305                 page_tail->flags |= (1L << PG_dirty);
1306
1307                 /* clear PageTail before overwriting first_page */
1308                 smp_wmb();
1309
1310                 /*
1311                  * __split_huge_page_splitting() already set the
1312                  * splitting bit in all pmd that could map this
1313                  * hugepage, that will ensure no CPU can alter the
1314                  * mapcount on the head page. The mapcount is only
1315                  * accounted in the head page and it has to be
1316                  * transferred to all tail pages in the below code. So
1317                  * for this code to be safe, the split the mapcount
1318                  * can't change. But that doesn't mean userland can't
1319                  * keep changing and reading the page contents while
1320                  * we transfer the mapcount, so the pmd splitting
1321                  * status is achieved setting a reserved bit in the
1322                  * pmd, not by clearing the present bit.
1323                 */
1324                 page_tail->_mapcount = page->_mapcount;
1325
1326                 BUG_ON(page_tail->mapping);
1327                 page_tail->mapping = page->mapping;
1328
1329                 page_tail->index = ++head_index;
1330
1331                 BUG_ON(!PageAnon(page_tail));
1332                 BUG_ON(!PageUptodate(page_tail));
1333                 BUG_ON(!PageDirty(page_tail));
1334                 BUG_ON(!PageSwapBacked(page_tail));
1335
1336                 mem_cgroup_split_huge_fixup(page, page_tail);
1337
1338                 lru_add_page_tail(zone, page, page_tail);
1339         }
1340         atomic_sub(tail_count, &page->_count);
1341         BUG_ON(atomic_read(&page->_count) <= 0);
1342
1343         __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1344         __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1345
1346         /*
1347          * A hugepage counts for HPAGE_PMD_NR pages on the LRU statistics,
1348          * so adjust those appropriately if this page is on the LRU.
1349          */
1350         if (PageLRU(page)) {
1351                 zonestat = NR_LRU_BASE + page_lru(page);
1352                 __mod_zone_page_state(zone, zonestat, -(HPAGE_PMD_NR-1));
1353         }
1354
1355         ClearPageCompound(page);
1356         compound_unlock(page);
1357         spin_unlock_irq(&zone->lru_lock);
1358
1359         for (i = 1; i < HPAGE_PMD_NR; i++) {
1360                 struct page *page_tail = page + i;
1361                 BUG_ON(page_count(page_tail) <= 0);
1362                 /*
1363                  * Tail pages may be freed if there wasn't any mapping
1364                  * like if add_to_swap() is running on a lru page that
1365                  * had its mapping zapped. And freeing these pages
1366                  * requires taking the lru_lock so we do the put_page
1367                  * of the tail pages after the split is complete.
1368                  */
1369                 put_page(page_tail);
1370         }
1371
1372         /*
1373          * Only the head page (now become a regular page) is required
1374          * to be pinned by the caller.
1375          */
1376         BUG_ON(page_count(page) <= 0);
1377 }
1378
1379 static int __split_huge_page_map(struct page *page,
1380                                  struct vm_area_struct *vma,
1381                                  unsigned long address)
1382 {
1383         struct mm_struct *mm = vma->vm_mm;
1384         pmd_t *pmd, _pmd;
1385         int ret = 0, i;
1386         pgtable_t pgtable;
1387         unsigned long haddr;
1388
1389         spin_lock(&mm->page_table_lock);
1390         pmd = page_check_address_pmd(page, mm, address,
1391                                      PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1392         if (pmd) {
1393                 pgtable = get_pmd_huge_pte(mm);
1394                 pmd_populate(mm, &_pmd, pgtable);
1395
1396                 for (i = 0, haddr = address; i < HPAGE_PMD_NR;
1397                      i++, haddr += PAGE_SIZE) {
1398                         pte_t *pte, entry;
1399                         BUG_ON(PageCompound(page+i));
1400                         entry = mk_pte(page + i, vma->vm_page_prot);
1401                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1402                         if (!pmd_write(*pmd))
1403                                 entry = pte_wrprotect(entry);
1404                         else
1405                                 BUG_ON(page_mapcount(page) != 1);
1406                         if (!pmd_young(*pmd))
1407                                 entry = pte_mkold(entry);
1408                         pte = pte_offset_map(&_pmd, haddr);
1409                         BUG_ON(!pte_none(*pte));
1410                         set_pte_at(mm, haddr, pte, entry);
1411                         pte_unmap(pte);
1412                 }
1413
1414                 smp_wmb(); /* make pte visible before pmd */
1415                 /*
1416                  * Up to this point the pmd is present and huge and
1417                  * userland has the whole access to the hugepage
1418                  * during the split (which happens in place). If we
1419                  * overwrite the pmd with the not-huge version
1420                  * pointing to the pte here (which of course we could
1421                  * if all CPUs were bug free), userland could trigger
1422                  * a small page size TLB miss on the small sized TLB
1423                  * while the hugepage TLB entry is still established
1424                  * in the huge TLB. Some CPU doesn't like that. See
1425                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1426                  * Erratum 383 on page 93. Intel should be safe but is
1427                  * also warns that it's only safe if the permission
1428                  * and cache attributes of the two entries loaded in
1429                  * the two TLB is identical (which should be the case
1430                  * here). But it is generally safer to never allow
1431                  * small and huge TLB entries for the same virtual
1432                  * address to be loaded simultaneously. So instead of
1433                  * doing "pmd_populate(); flush_tlb_range();" we first
1434                  * mark the current pmd notpresent (atomically because
1435                  * here the pmd_trans_huge and pmd_trans_splitting
1436                  * must remain set at all times on the pmd until the
1437                  * split is complete for this pmd), then we flush the
1438                  * SMP TLB and finally we write the non-huge version
1439                  * of the pmd entry with pmd_populate.
1440                  */
1441                 set_pmd_at(mm, address, pmd, pmd_mknotpresent(*pmd));
1442                 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
1443                 pmd_populate(mm, pmd, pgtable);
1444                 ret = 1;
1445         }
1446         spin_unlock(&mm->page_table_lock);
1447
1448         return ret;
1449 }
1450
1451 /* must be called with anon_vma->root->mutex hold */
1452 static void __split_huge_page(struct page *page,
1453                               struct anon_vma *anon_vma)
1454 {
1455         int mapcount, mapcount2;
1456         struct anon_vma_chain *avc;
1457
1458         BUG_ON(!PageHead(page));
1459         BUG_ON(PageTail(page));
1460
1461         mapcount = 0;
1462         list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1463                 struct vm_area_struct *vma = avc->vma;
1464                 unsigned long addr = vma_address(page, vma);
1465                 BUG_ON(is_vma_temporary_stack(vma));
1466                 if (addr == -EFAULT)
1467                         continue;
1468                 mapcount += __split_huge_page_splitting(page, vma, addr);
1469         }
1470         /*
1471          * It is critical that new vmas are added to the tail of the
1472          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1473          * and establishes a child pmd before
1474          * __split_huge_page_splitting() freezes the parent pmd (so if
1475          * we fail to prevent copy_huge_pmd() from running until the
1476          * whole __split_huge_page() is complete), we will still see
1477          * the newly established pmd of the child later during the
1478          * walk, to be able to set it as pmd_trans_splitting too.
1479          */
1480         if (mapcount != page_mapcount(page))
1481                 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1482                        mapcount, page_mapcount(page));
1483         BUG_ON(mapcount != page_mapcount(page));
1484
1485         __split_huge_page_refcount(page);
1486
1487         mapcount2 = 0;
1488         list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1489                 struct vm_area_struct *vma = avc->vma;
1490                 unsigned long addr = vma_address(page, vma);
1491                 BUG_ON(is_vma_temporary_stack(vma));
1492                 if (addr == -EFAULT)
1493                         continue;
1494                 mapcount2 += __split_huge_page_map(page, vma, addr);
1495         }
1496         if (mapcount != mapcount2)
1497                 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1498                        mapcount, mapcount2, page_mapcount(page));
1499         BUG_ON(mapcount != mapcount2);
1500 }
1501
1502 int split_huge_page(struct page *page)
1503 {
1504         struct anon_vma *anon_vma;
1505         int ret = 1;
1506
1507         BUG_ON(!PageAnon(page));
1508         anon_vma = page_lock_anon_vma(page);
1509         if (!anon_vma)
1510                 goto out;
1511         ret = 0;
1512         if (!PageCompound(page))
1513                 goto out_unlock;
1514
1515         BUG_ON(!PageSwapBacked(page));
1516         __split_huge_page(page, anon_vma);
1517         count_vm_event(THP_SPLIT);
1518
1519         BUG_ON(PageCompound(page));
1520 out_unlock:
1521         page_unlock_anon_vma(anon_vma);
1522 out:
1523         return ret;
1524 }
1525
1526 /* callers must hold mmap_sem (madvise() does) */
1527 static int collapse_fb_pmd(struct mm_struct *mm, pmd_t *pmd,
1528         unsigned long addr, struct vm_area_struct *vma)
1529 {
1530         unsigned long _addr;
1531         struct page *page;
1532         pgtable_t pgtable;
1533         pte_t *pte, *_pte;
1534         pmd_t _pmd;
1535         u32 pa;
1536
1537         pte = pte_offset_map(pmd, addr);
1538         page = pte_page(*pte);
1539         pa = __pfn_to_phys(page_to_pfn(page));
1540         _pmd = pmdp_clear_flush_notify(vma, addr, pmd);
1541
1542         if ((addr | pa) & ~HPAGE_PMD_MASK) {
1543                 printk(KERN_ERR "collapse_fb: bad alignment: %08lx->%08x\n",
1544                         addr, pa);
1545                 pte_unmap(pte);
1546                 return -EINVAL;
1547         }
1548
1549         for (_pte = pte, _addr = addr; _pte < pte + HPAGE_PMD_NR; _pte++) {
1550                 pte_t pteval = *_pte;
1551                 struct page *src_page;
1552
1553                 if (!pte_none(pteval)) {
1554                         src_page = pte_page(pteval);
1555
1556                         pte_clear(vma->vm_mm, _addr, _pte);
1557                         if (pte_present(pteval))
1558                                 page_remove_rmap(src_page);
1559                 }
1560
1561                 _addr += PAGE_SIZE;
1562         }
1563
1564         pte_unmap(pte);
1565         pgtable = pmd_pgtable(_pmd);
1566         VM_BUG_ON(page_count(pgtable) != 1);
1567         VM_BUG_ON(page_mapcount(pgtable) != 0);
1568
1569         _pmd = mk_pmd(page, vma->vm_page_prot);
1570         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1571         _pmd = pmd_mkhuge(_pmd);
1572
1573         smp_wmb();
1574
1575         spin_lock(&mm->page_table_lock);
1576         BUG_ON(!pmd_none(*pmd));
1577         set_pmd_at(mm, addr, pmd, _pmd);
1578         update_mmu_cache(vma, addr, pmd);
1579         prepare_pmd_huge_pte(pgtable, mm);
1580         spin_unlock(&mm->page_table_lock);
1581
1582         return 0;
1583 }
1584
1585 static int try_collapse_fb(struct vm_area_struct *vma)
1586 {
1587         struct mm_struct *mm = vma->vm_mm;
1588         unsigned long hstart, hend, addr;
1589         int ret = 0;
1590         pgd_t *pgd;
1591         pud_t *pud;
1592         pmd_t *pmd;
1593
1594         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1595         hend = vma->vm_end & HPAGE_PMD_MASK;
1596         if (hstart >= hend)
1597                 return -EINVAL;
1598
1599         for (addr = hstart; addr < hend; addr += HPAGE_PMD_SIZE) {
1600                 pgd = pgd_offset(mm, addr);
1601                 if (!pgd_present(*pgd))
1602                         return -EINVAL;
1603
1604                 pud = pud_offset(pgd, addr);
1605                 if (!pud_present(*pud))
1606                         return -EINVAL;
1607
1608                 pmd = pmd_offset(pud, addr);
1609                 if (!pmd_present(*pmd))
1610                         return -EINVAL;
1611                 if (pmd_trans_huge(*pmd))
1612                         continue;
1613
1614                 ret = collapse_fb_pmd(mm, pmd, addr, vma);
1615                 if (ret)
1616                         break;
1617         }
1618
1619         return ret;
1620 }
1621
1622 /* undo collapse_fb_pmd(), restore pages so that mm subsys can release them
1623  * page_table_lock() should be held */
1624 static void split_fb_pmd(struct vm_area_struct *vma, pmd_t *pmd)
1625 {
1626         struct mm_struct *mm = vma->vm_mm;
1627         unsigned long addr, haddr, pfn;
1628         struct page *page;
1629         pgtable_t pgtable;
1630         pmd_t _pmd;
1631         int i;
1632
1633         page = pmd_page(*pmd);
1634         pgtable = get_pmd_huge_pte(mm);
1635         pfn = page_to_pfn(page);
1636         addr = pfn << PAGE_SHIFT;
1637
1638         pmd_populate(mm, &_pmd, pgtable);
1639
1640         for (i = 0, haddr = addr; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1641                 pte_t *pte, entry;
1642                 BUG_ON(PageCompound(page + i));
1643                 entry = mk_pte(page + i, vma->vm_page_prot);
1644                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1645                 if (!pmd_young(*pmd))
1646                         entry = pte_mkold(entry);
1647                 atomic_set(&page[i]._mapcount, 0); // hack?
1648                 pte = pte_offset_map(&_pmd, haddr);
1649                 BUG_ON(!pte_none(*pte));
1650                 set_pte_at(mm, haddr, pte, entry);
1651                 pte_unmap(pte);
1652         }
1653
1654         set_pmd_at(mm, addr, pmd, pmd_mknotpresent(*pmd));
1655         flush_tlb_range(vma, addr, addr + HPAGE_PMD_SIZE);
1656         pmd_populate(mm, pmd, pgtable);
1657 }
1658
1659 #ifndef __arm__
1660 #error arm only..
1661 #endif
1662 static u32 pmd_to_va(struct mm_struct *mm, pmd_t *pmd)
1663 {
1664         pgd_t *pgd;
1665         pud_t *pud;
1666         pmd_t *pmd0;
1667         u32 ret;
1668
1669         pgd = pgd_offset(mm, 0);
1670         pud = pud_offset(pgd, 0);
1671         pmd0 = pmd_offset(pud, 0);
1672
1673         ret = (pmd - pmd0) << SECTION_SHIFT;
1674         return ret;
1675 }
1676
1677 #define VM_NO_THP (VM_SPECIAL|VM_INSERTPAGE|VM_MIXEDMAP|VM_SAO| \
1678                    VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1679
1680 int hugepage_madvise(struct vm_area_struct *vma,
1681                      unsigned long *vm_flags, int advice)
1682 {
1683         switch (advice) {
1684         case MADV_HUGEPAGE:
1685                 if (is_fb_vma(vma))
1686                         return try_collapse_fb(vma);
1687
1688                 /*
1689                  * Be somewhat over-protective like KSM for now!
1690                  */
1691                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1692                         return -EINVAL;
1693                 *vm_flags &= ~VM_NOHUGEPAGE;
1694                 *vm_flags |= VM_HUGEPAGE;
1695                 /*
1696                  * If the vma become good for khugepaged to scan,
1697                  * register it here without waiting a page fault that
1698                  * may not happen any time soon.
1699                  */
1700                 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
1701                         return -ENOMEM;
1702                 break;
1703         case MADV_NOHUGEPAGE:
1704                 /*
1705                  * Be somewhat over-protective like KSM for now!
1706                  */
1707                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1708                         return -EINVAL;
1709                 *vm_flags &= ~VM_HUGEPAGE;
1710                 *vm_flags |= VM_NOHUGEPAGE;
1711                 /*
1712                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1713                  * this vma even if we leave the mm registered in khugepaged if
1714                  * it got registered before VM_NOHUGEPAGE was set.
1715                  */
1716                 break;
1717         }
1718
1719         return 0;
1720 }
1721
1722 static int __init khugepaged_slab_init(void)
1723 {
1724         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1725                                           sizeof(struct mm_slot),
1726                                           __alignof__(struct mm_slot), 0, NULL);
1727         if (!mm_slot_cache)
1728                 return -ENOMEM;
1729
1730         return 0;
1731 }
1732
1733 static void __init khugepaged_slab_free(void)
1734 {
1735         kmem_cache_destroy(mm_slot_cache);
1736         mm_slot_cache = NULL;
1737 }
1738
1739 static inline struct mm_slot *alloc_mm_slot(void)
1740 {
1741         if (!mm_slot_cache)     /* initialization failed */
1742                 return NULL;
1743         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1744 }
1745
1746 static inline void free_mm_slot(struct mm_slot *mm_slot)
1747 {
1748         kmem_cache_free(mm_slot_cache, mm_slot);
1749 }
1750
1751 static int __init mm_slots_hash_init(void)
1752 {
1753         mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
1754                                 GFP_KERNEL);
1755         if (!mm_slots_hash)
1756                 return -ENOMEM;
1757         return 0;
1758 }
1759
1760 #if 0
1761 static void __init mm_slots_hash_free(void)
1762 {
1763         kfree(mm_slots_hash);
1764         mm_slots_hash = NULL;
1765 }
1766 #endif
1767
1768 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1769 {
1770         struct mm_slot *mm_slot;
1771         struct hlist_head *bucket;
1772         struct hlist_node *node;
1773
1774         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1775                                 % MM_SLOTS_HASH_HEADS];
1776         hlist_for_each_entry(mm_slot, node, bucket, hash) {
1777                 if (mm == mm_slot->mm)
1778                         return mm_slot;
1779         }
1780         return NULL;
1781 }
1782
1783 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1784                                     struct mm_slot *mm_slot)
1785 {
1786         struct hlist_head *bucket;
1787
1788         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1789                                 % MM_SLOTS_HASH_HEADS];
1790         mm_slot->mm = mm;
1791         hlist_add_head(&mm_slot->hash, bucket);
1792 }
1793
1794 static inline int khugepaged_test_exit(struct mm_struct *mm)
1795 {
1796         return atomic_read(&mm->mm_users) == 0;
1797 }
1798
1799 int __khugepaged_enter(struct mm_struct *mm)
1800 {
1801         struct mm_slot *mm_slot;
1802         int wakeup;
1803
1804         mm_slot = alloc_mm_slot();
1805         if (!mm_slot)
1806                 return -ENOMEM;
1807
1808         /* __khugepaged_exit() must not run from under us */
1809         VM_BUG_ON(khugepaged_test_exit(mm));
1810         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1811                 free_mm_slot(mm_slot);
1812                 return 0;
1813         }
1814
1815         spin_lock(&khugepaged_mm_lock);
1816         insert_to_mm_slots_hash(mm, mm_slot);
1817         /*
1818          * Insert just behind the scanning cursor, to let the area settle
1819          * down a little.
1820          */
1821         wakeup = list_empty(&khugepaged_scan.mm_head);
1822         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1823         spin_unlock(&khugepaged_mm_lock);
1824
1825         atomic_inc(&mm->mm_count);
1826         if (wakeup)
1827                 wake_up_interruptible(&khugepaged_wait);
1828
1829         return 0;
1830 }
1831
1832 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
1833                                unsigned long vm_flags)
1834 {
1835         unsigned long hstart, hend;
1836         if (!vma->anon_vma)
1837                 /*
1838                  * Not yet faulted in so we will register later in the
1839                  * page fault if needed.
1840                  */
1841                 return 0;
1842         if (vma->vm_ops || (vm_flags & VM_NO_THP))
1843                 /* khugepaged not yet working on file or special mappings */
1844                 return 0;
1845         /*
1846          * If is_pfn_mapping() is true is_learn_pfn_mapping() must be
1847          * true too, verify it here.
1848          */
1849         VM_BUG_ON(is_linear_pfn_mapping(vma));
1850         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1851         hend = vma->vm_end & HPAGE_PMD_MASK;
1852         if (hstart < hend)
1853                 return khugepaged_enter(vma, vm_flags);
1854         return 0;
1855 }
1856
1857 void __khugepaged_exit(struct mm_struct *mm)
1858 {
1859         struct mm_slot *mm_slot;
1860         int free = 0;
1861
1862         spin_lock(&khugepaged_mm_lock);
1863         mm_slot = get_mm_slot(mm);
1864         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1865                 hlist_del(&mm_slot->hash);
1866                 list_del(&mm_slot->mm_node);
1867                 free = 1;
1868         }
1869         spin_unlock(&khugepaged_mm_lock);
1870
1871         if (free) {
1872                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1873                 free_mm_slot(mm_slot);
1874                 mmdrop(mm);
1875         } else if (mm_slot) {
1876                 /*
1877                  * This is required to serialize against
1878                  * khugepaged_test_exit() (which is guaranteed to run
1879                  * under mmap sem read mode). Stop here (after we
1880                  * return all pagetables will be destroyed) until
1881                  * khugepaged has finished working on the pagetables
1882                  * under the mmap_sem.
1883                  */
1884                 down_write(&mm->mmap_sem);
1885                 up_write(&mm->mmap_sem);
1886         }
1887 }
1888
1889 static void release_pte_page(struct page *page)
1890 {
1891         /* 0 stands for page_is_file_cache(page) == false */
1892         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1893         unlock_page(page);
1894         putback_lru_page(page);
1895 }
1896
1897 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1898 {
1899         while (--_pte >= pte) {
1900                 pte_t pteval = *_pte;
1901                 if (!pte_none(pteval))
1902                         release_pte_page(pte_page(pteval));
1903         }
1904 }
1905
1906 static void release_all_pte_pages(pte_t *pte)
1907 {
1908         release_pte_pages(pte, pte + HPAGE_PMD_NR);
1909 }
1910
1911 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1912                                         unsigned long address,
1913                                         pte_t *pte)
1914 {
1915         struct page *page;
1916         pte_t *_pte;
1917         int referenced = 0, isolated = 0, none = 0;
1918         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1919              _pte++, address += PAGE_SIZE) {
1920                 pte_t pteval = *_pte;
1921                 if (pte_none(pteval)) {
1922                         if (++none <= khugepaged_max_ptes_none)
1923                                 continue;
1924                         else {
1925                                 release_pte_pages(pte, _pte);
1926                                 goto out;
1927                         }
1928                 }
1929                 if (!pte_present(pteval) || !pte_write(pteval)) {
1930                         release_pte_pages(pte, _pte);
1931                         goto out;
1932                 }
1933                 page = vm_normal_page(vma, address, pteval);
1934                 if (unlikely(!page)) {
1935                         release_pte_pages(pte, _pte);
1936                         goto out;
1937                 }
1938                 VM_BUG_ON(PageCompound(page));
1939                 BUG_ON(!PageAnon(page));
1940                 VM_BUG_ON(!PageSwapBacked(page));
1941
1942                 /* cannot use mapcount: can't collapse if there's a gup pin */
1943                 if (page_count(page) != 1) {
1944                         release_pte_pages(pte, _pte);
1945                         goto out;
1946                 }
1947                 /*
1948                  * We can do it before isolate_lru_page because the
1949                  * page can't be freed from under us. NOTE: PG_lock
1950                  * is needed to serialize against split_huge_page
1951                  * when invoked from the VM.
1952                  */
1953                 if (!trylock_page(page)) {
1954                         release_pte_pages(pte, _pte);
1955                         goto out;
1956                 }
1957                 /*
1958                  * Isolate the page to avoid collapsing an hugepage
1959                  * currently in use by the VM.
1960                  */
1961                 if (isolate_lru_page(page)) {
1962                         unlock_page(page);
1963                         release_pte_pages(pte, _pte);
1964                         goto out;
1965                 }
1966                 /* 0 stands for page_is_file_cache(page) == false */
1967                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
1968                 VM_BUG_ON(!PageLocked(page));
1969                 VM_BUG_ON(PageLRU(page));
1970
1971                 /* If there is no mapped pte young don't collapse the page */
1972                 if (pte_young(pteval) || PageReferenced(page) ||
1973                     mmu_notifier_test_young(vma->vm_mm, address))
1974                         referenced = 1;
1975         }
1976         if (unlikely(!referenced))
1977                 release_all_pte_pages(pte);
1978         else
1979                 isolated = 1;
1980 out:
1981         return isolated;
1982 }
1983
1984 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
1985                                       struct vm_area_struct *vma,
1986                                       unsigned long address,
1987                                       spinlock_t *ptl)
1988 {
1989         pte_t *_pte;
1990         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
1991                 pte_t pteval = *_pte;
1992                 struct page *src_page;
1993
1994                 if (pte_none(pteval)) {
1995                         clear_user_highpage(page, address);
1996                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
1997                 } else {
1998                         src_page = pte_page(pteval);
1999                         copy_user_highpage(page, src_page, address, vma);
2000                         VM_BUG_ON(page_mapcount(src_page) != 1);
2001                         VM_BUG_ON(page_count(src_page) != 2);
2002                         release_pte_page(src_page);
2003                         /*
2004                          * ptl mostly unnecessary, but preempt has to
2005                          * be disabled to update the per-cpu stats
2006                          * inside page_remove_rmap().
2007                          */
2008                         spin_lock(ptl);
2009                         /*
2010                          * paravirt calls inside pte_clear here are
2011                          * superfluous.
2012                          */
2013                         pte_clear(vma->vm_mm, address, _pte);
2014                         page_remove_rmap(src_page);
2015                         spin_unlock(ptl);
2016                         free_page_and_swap_cache(src_page);
2017                 }
2018
2019                 address += PAGE_SIZE;
2020                 page++;
2021         }
2022 }
2023
2024 static bool hugepage_vma_check(struct vm_area_struct *vma)
2025 {
2026         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2027             (vma->vm_flags & VM_NOHUGEPAGE))
2028                 return false;
2029
2030         if (!vma->anon_vma || vma->vm_ops)
2031                 return false;
2032         if (is_vma_temporary_stack(vma))
2033                 return false;
2034         /*
2035          * If is_pfn_mapping() is true is_learn_pfn_mapping() must be
2036          * true too, verify it here.
2037          */
2038         VM_BUG_ON(is_linear_pfn_mapping(vma));
2039         return !(vma->vm_flags & VM_NO_THP);
2040 }
2041
2042 static void collapse_huge_page(struct mm_struct *mm,
2043                                unsigned long address,
2044                                struct page **hpage,
2045                                struct vm_area_struct *vma,
2046                                int node)
2047 {
2048         pgd_t *pgd;
2049         pud_t *pud;
2050         pmd_t *pmd, _pmd;
2051         pte_t *pte;
2052         pgtable_t pgtable;
2053         struct page *new_page;
2054         spinlock_t *ptl;
2055         int isolated;
2056         unsigned long hstart, hend;
2057
2058         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2059 #ifndef CONFIG_NUMA
2060         up_read(&mm->mmap_sem);
2061         VM_BUG_ON(!*hpage);
2062         new_page = *hpage;
2063 #else
2064         VM_BUG_ON(*hpage);
2065         /*
2066          * Allocate the page while the vma is still valid and under
2067          * the mmap_sem read mode so there is no memory allocation
2068          * later when we take the mmap_sem in write mode. This is more
2069          * friendly behavior (OTOH it may actually hide bugs) to
2070          * filesystems in userland with daemons allocating memory in
2071          * the userland I/O paths.  Allocating memory with the
2072          * mmap_sem in read mode is good idea also to allow greater
2073          * scalability.
2074          */
2075         new_page = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
2076                                       node, __GFP_OTHER_NODE);
2077
2078         /*
2079          * After allocating the hugepage, release the mmap_sem read lock in
2080          * preparation for taking it in write mode.
2081          */
2082         up_read(&mm->mmap_sem);
2083         if (unlikely(!new_page)) {
2084                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2085                 *hpage = ERR_PTR(-ENOMEM);
2086                 return;
2087         }
2088 #endif
2089
2090         count_vm_event(THP_COLLAPSE_ALLOC);
2091         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
2092 #ifdef CONFIG_NUMA
2093                 put_page(new_page);
2094 #endif
2095                 return;
2096         }
2097
2098         /*
2099          * Prevent all access to pagetables with the exception of
2100          * gup_fast later hanlded by the ptep_clear_flush and the VM
2101          * handled by the anon_vma lock + PG_lock.
2102          */
2103         down_write(&mm->mmap_sem);
2104         if (unlikely(khugepaged_test_exit(mm)))
2105                 goto out;
2106
2107         vma = find_vma(mm, address);
2108         if (!vma)
2109                 goto out;
2110         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2111         hend = vma->vm_end & HPAGE_PMD_MASK;
2112         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2113                 goto out;
2114         if (!hugepage_vma_check(vma))
2115                 goto out;
2116         pgd = pgd_offset(mm, address);
2117         if (!pgd_present(*pgd))
2118                 goto out;
2119
2120         pud = pud_offset(pgd, address);
2121         if (!pud_present(*pud))
2122                 goto out;
2123
2124         pmd = pmd_offset(pud, address);
2125         /* pmd can't go away or become huge under us */
2126         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
2127                 goto out;
2128
2129         anon_vma_lock(vma->anon_vma);
2130
2131         pte = pte_offset_map(pmd, address);
2132         ptl = pte_lockptr(mm, pmd);
2133
2134         spin_lock(&mm->page_table_lock); /* probably unnecessary */
2135         /*
2136          * After this gup_fast can't run anymore. This also removes
2137          * any huge TLB entry from the CPU so we won't allow
2138          * huge and small TLB entries for the same virtual address
2139          * to avoid the risk of CPU bugs in that area.
2140          */
2141         _pmd = pmdp_clear_flush_notify(vma, address, pmd);
2142         spin_unlock(&mm->page_table_lock);
2143
2144         spin_lock(ptl);
2145         isolated = __collapse_huge_page_isolate(vma, address, pte);
2146         spin_unlock(ptl);
2147
2148         if (unlikely(!isolated)) {
2149                 pte_unmap(pte);
2150                 spin_lock(&mm->page_table_lock);
2151                 BUG_ON(!pmd_none(*pmd));
2152                 /*
2153                  * We can only use set_pmd_at when establishing
2154                  * hugepmds and never for establishing regular pmds that
2155                  * points to regular pagetables. Use pmd_populate for that
2156                  */
2157                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2158                 spin_unlock(&mm->page_table_lock);
2159                 anon_vma_unlock(vma->anon_vma);
2160                 goto out;
2161         }
2162
2163         /*
2164          * All pages are isolated and locked so anon_vma rmap
2165          * can't run anymore.
2166          */
2167         anon_vma_unlock(vma->anon_vma);
2168
2169         __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
2170         pte_unmap(pte);
2171         __SetPageUptodate(new_page);
2172         pgtable = pmd_pgtable(_pmd);
2173         VM_BUG_ON(page_count(pgtable) != 1);
2174         VM_BUG_ON(page_mapcount(pgtable) != 0);
2175
2176         _pmd = mk_pmd(new_page, vma->vm_page_prot);
2177         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2178         _pmd = pmd_mkhuge(_pmd);
2179
2180         /*
2181          * spin_lock() below is not the equivalent of smp_wmb(), so
2182          * this is needed to avoid the copy_huge_page writes to become
2183          * visible after the set_pmd_at() write.
2184          */
2185         smp_wmb();
2186
2187         spin_lock(&mm->page_table_lock);
2188         BUG_ON(!pmd_none(*pmd));
2189         page_add_new_anon_rmap(new_page, vma, address);
2190         set_pmd_at(mm, address, pmd, _pmd);
2191         update_mmu_cache(vma, address, pmd);
2192         prepare_pmd_huge_pte(pgtable, mm);
2193         spin_unlock(&mm->page_table_lock);
2194
2195 #ifndef CONFIG_NUMA
2196         *hpage = NULL;
2197 #endif
2198         khugepaged_pages_collapsed++;
2199 out_up_write:
2200         up_write(&mm->mmap_sem);
2201         return;
2202
2203 out:
2204         mem_cgroup_uncharge_page(new_page);
2205 #ifdef CONFIG_NUMA
2206         put_page(new_page);
2207 #endif
2208         goto out_up_write;
2209 }
2210
2211 static int khugepaged_scan_pmd(struct mm_struct *mm,
2212                                struct vm_area_struct *vma,
2213                                unsigned long address,
2214                                struct page **hpage)
2215 {
2216         pgd_t *pgd;
2217         pud_t *pud;
2218         pmd_t *pmd;
2219         pte_t *pte, *_pte;
2220         int ret = 0, referenced = 0, none = 0;
2221         struct page *page;
2222         unsigned long _address;
2223         spinlock_t *ptl;
2224         int node = -1;
2225
2226         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2227
2228         pgd = pgd_offset(mm, address);
2229         if (!pgd_present(*pgd))
2230                 goto out;
2231
2232         pud = pud_offset(pgd, address);
2233         if (!pud_present(*pud))
2234                 goto out;
2235
2236         pmd = pmd_offset(pud, address);
2237         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
2238                 goto out;
2239
2240         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2241         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2242              _pte++, _address += PAGE_SIZE) {
2243                 pte_t pteval = *_pte;
2244                 if (pte_none(pteval)) {
2245                         if (++none <= khugepaged_max_ptes_none)
2246                                 continue;
2247                         else
2248                                 goto out_unmap;
2249                 }
2250                 if (!pte_present(pteval) || !pte_write(pteval))
2251                         goto out_unmap;
2252                 page = vm_normal_page(vma, _address, pteval);
2253                 if (unlikely(!page))
2254                         goto out_unmap;
2255                 /*
2256                  * Chose the node of the first page. This could
2257                  * be more sophisticated and look at more pages,
2258                  * but isn't for now.
2259                  */
2260                 if (node == -1)
2261                         node = page_to_nid(page);
2262                 VM_BUG_ON(PageCompound(page));
2263                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2264                         goto out_unmap;
2265                 /* cannot use mapcount: can't collapse if there's a gup pin */
2266                 if (page_count(page) != 1)
2267                         goto out_unmap;
2268                 if (pte_young(pteval) || PageReferenced(page) ||
2269                     mmu_notifier_test_young(vma->vm_mm, address))
2270                         referenced = 1;
2271         }
2272         if (referenced)
2273                 ret = 1;
2274 out_unmap:
2275         pte_unmap_unlock(pte, ptl);
2276         if (ret)
2277                 /* collapse_huge_page will return with the mmap_sem released */
2278                 collapse_huge_page(mm, address, hpage, vma, node);
2279 out:
2280         return ret;
2281 }
2282
2283 static void collect_mm_slot(struct mm_slot *mm_slot)
2284 {
2285         struct mm_struct *mm = mm_slot->mm;
2286
2287         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2288
2289         if (khugepaged_test_exit(mm)) {
2290                 /* free mm_slot */
2291                 hlist_del(&mm_slot->hash);
2292                 list_del(&mm_slot->mm_node);
2293
2294                 /*
2295                  * Not strictly needed because the mm exited already.
2296                  *
2297                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2298                  */
2299
2300                 /* khugepaged_mm_lock actually not necessary for the below */
2301                 free_mm_slot(mm_slot);
2302                 mmdrop(mm);
2303         }
2304 }
2305
2306 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2307                                             struct page **hpage)
2308         __releases(&khugepaged_mm_lock)
2309         __acquires(&khugepaged_mm_lock)
2310 {
2311         struct mm_slot *mm_slot;
2312         struct mm_struct *mm;
2313         struct vm_area_struct *vma;
2314         int progress = 0;
2315
2316         VM_BUG_ON(!pages);
2317         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2318
2319         if (khugepaged_scan.mm_slot)
2320                 mm_slot = khugepaged_scan.mm_slot;
2321         else {
2322                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2323                                      struct mm_slot, mm_node);
2324                 khugepaged_scan.address = 0;
2325                 khugepaged_scan.mm_slot = mm_slot;
2326         }
2327         spin_unlock(&khugepaged_mm_lock);
2328
2329         mm = mm_slot->mm;
2330         down_read(&mm->mmap_sem);
2331         if (unlikely(khugepaged_test_exit(mm)))
2332                 vma = NULL;
2333         else
2334                 vma = find_vma(mm, khugepaged_scan.address);
2335
2336         progress++;
2337         for (; vma; vma = vma->vm_next) {
2338                 unsigned long hstart, hend;
2339
2340                 cond_resched();
2341                 if (unlikely(khugepaged_test_exit(mm))) {
2342                         progress++;
2343                         break;
2344                 }
2345                 if (!hugepage_vma_check(vma)) {
2346 skip:
2347                         progress++;
2348                         continue;
2349                 }
2350                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2351                 hend = vma->vm_end & HPAGE_PMD_MASK;
2352                 if (hstart >= hend)
2353                         goto skip;
2354                 if (khugepaged_scan.address > hend)
2355                         goto skip;
2356                 if (khugepaged_scan.address < hstart)
2357                         khugepaged_scan.address = hstart;
2358                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2359
2360                 while (khugepaged_scan.address < hend) {
2361                         int ret;
2362                         cond_resched();
2363                         if (unlikely(khugepaged_test_exit(mm)))
2364                                 goto breakouterloop;
2365
2366                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2367                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2368                                   hend);
2369                         ret = khugepaged_scan_pmd(mm, vma,
2370                                                   khugepaged_scan.address,
2371                                                   hpage);
2372                         /* move to next address */
2373                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2374                         progress += HPAGE_PMD_NR;
2375                         if (ret)
2376                                 /* we released mmap_sem so break loop */
2377                                 goto breakouterloop_mmap_sem;
2378                         if (progress >= pages)
2379                                 goto breakouterloop;
2380                 }
2381         }
2382 breakouterloop:
2383         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2384 breakouterloop_mmap_sem:
2385
2386         spin_lock(&khugepaged_mm_lock);
2387         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2388         /*
2389          * Release the current mm_slot if this mm is about to die, or
2390          * if we scanned all vmas of this mm.
2391          */
2392         if (khugepaged_test_exit(mm) || !vma) {
2393                 /*
2394                  * Make sure that if mm_users is reaching zero while
2395                  * khugepaged runs here, khugepaged_exit will find
2396                  * mm_slot not pointing to the exiting mm.
2397                  */
2398                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2399                         khugepaged_scan.mm_slot = list_entry(
2400                                 mm_slot->mm_node.next,
2401                                 struct mm_slot, mm_node);
2402                         khugepaged_scan.address = 0;
2403                 } else {
2404                         khugepaged_scan.mm_slot = NULL;
2405                         khugepaged_full_scans++;
2406                 }
2407
2408                 collect_mm_slot(mm_slot);
2409         }
2410
2411         return progress;
2412 }
2413
2414 static int khugepaged_has_work(void)
2415 {
2416         return !list_empty(&khugepaged_scan.mm_head) &&
2417                 khugepaged_enabled();
2418 }
2419
2420 static int khugepaged_wait_event(void)
2421 {
2422         return !list_empty(&khugepaged_scan.mm_head) ||
2423                 !khugepaged_enabled();
2424 }
2425
2426 static void khugepaged_do_scan(struct page **hpage)
2427 {
2428         unsigned int progress = 0, pass_through_head = 0;
2429         unsigned int pages = khugepaged_pages_to_scan;
2430
2431         barrier(); /* write khugepaged_pages_to_scan to local stack */
2432
2433         while (progress < pages) {
2434                 cond_resched();
2435
2436 #ifndef CONFIG_NUMA
2437                 if (!*hpage) {
2438                         *hpage = alloc_hugepage(khugepaged_defrag());
2439                         if (unlikely(!*hpage)) {
2440                                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2441                                 break;
2442                         }
2443                         count_vm_event(THP_COLLAPSE_ALLOC);
2444                 }
2445 #else
2446                 if (IS_ERR(*hpage))
2447                         break;
2448 #endif
2449
2450                 if (unlikely(kthread_should_stop() || freezing(current)))
2451                         break;
2452
2453                 spin_lock(&khugepaged_mm_lock);
2454                 if (!khugepaged_scan.mm_slot)
2455                         pass_through_head++;
2456                 if (khugepaged_has_work() &&
2457                     pass_through_head < 2)
2458                         progress += khugepaged_scan_mm_slot(pages - progress,
2459                                                             hpage);
2460                 else
2461                         progress = pages;
2462                 spin_unlock(&khugepaged_mm_lock);
2463         }
2464 }
2465
2466 static void khugepaged_alloc_sleep(void)
2467 {
2468         wait_event_freezable_timeout(khugepaged_wait, false,
2469                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2470 }
2471
2472 #ifndef CONFIG_NUMA
2473 static struct page *khugepaged_alloc_hugepage(void)
2474 {
2475         struct page *hpage;
2476
2477         do {
2478                 hpage = alloc_hugepage(khugepaged_defrag());
2479                 if (!hpage) {
2480                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2481                         khugepaged_alloc_sleep();
2482                 } else
2483                         count_vm_event(THP_COLLAPSE_ALLOC);
2484         } while (unlikely(!hpage) &&
2485                  likely(khugepaged_enabled()));
2486         return hpage;
2487 }
2488 #endif
2489
2490 static void khugepaged_loop(void)
2491 {
2492         struct page *hpage;
2493
2494 #ifdef CONFIG_NUMA
2495         hpage = NULL;
2496 #endif
2497         while (likely(khugepaged_enabled())) {
2498 #ifndef CONFIG_NUMA
2499                 hpage = khugepaged_alloc_hugepage();
2500                 if (unlikely(!hpage))
2501                         break;
2502 #else
2503                 if (IS_ERR(hpage)) {
2504                         khugepaged_alloc_sleep();
2505                         hpage = NULL;
2506                 }
2507 #endif
2508
2509                 khugepaged_do_scan(&hpage);
2510 #ifndef CONFIG_NUMA
2511                 if (hpage)
2512                         put_page(hpage);
2513 #endif
2514                 try_to_freeze();
2515                 if (unlikely(kthread_should_stop()))
2516                         break;
2517                 if (khugepaged_has_work()) {
2518                         if (!khugepaged_scan_sleep_millisecs)
2519                                 continue;
2520                         wait_event_freezable_timeout(khugepaged_wait, false,
2521                             msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2522                 } else if (khugepaged_enabled())
2523                         wait_event_freezable(khugepaged_wait,
2524                                              khugepaged_wait_event());
2525         }
2526 }
2527
2528 static int khugepaged(void *none)
2529 {
2530         struct mm_slot *mm_slot;
2531
2532         set_freezable();
2533         set_user_nice(current, 19);
2534
2535         /* serialize with start_khugepaged() */
2536         mutex_lock(&khugepaged_mutex);
2537
2538         for (;;) {
2539                 mutex_unlock(&khugepaged_mutex);
2540                 VM_BUG_ON(khugepaged_thread != current);
2541                 khugepaged_loop();
2542                 VM_BUG_ON(khugepaged_thread != current);
2543
2544                 mutex_lock(&khugepaged_mutex);
2545                 if (!khugepaged_enabled())
2546                         break;
2547                 if (unlikely(kthread_should_stop()))
2548                         break;
2549         }
2550
2551         spin_lock(&khugepaged_mm_lock);
2552         mm_slot = khugepaged_scan.mm_slot;
2553         khugepaged_scan.mm_slot = NULL;
2554         if (mm_slot)
2555                 collect_mm_slot(mm_slot);
2556         spin_unlock(&khugepaged_mm_lock);
2557
2558         khugepaged_thread = NULL;
2559         mutex_unlock(&khugepaged_mutex);
2560
2561         return 0;
2562 }
2563
2564 void __split_huge_page_pmd(struct mm_struct *mm, pmd_t *pmd)
2565 {
2566         struct vm_area_struct *vma;
2567         struct page *page;
2568
2569         spin_lock(&mm->page_table_lock);
2570         if (unlikely(!pmd_trans_huge(*pmd))) {
2571                 spin_unlock(&mm->page_table_lock);
2572                 return;
2573         }
2574         vma = find_vma(mm, pmd_to_va(mm, pmd));
2575         if (vma && is_fb_vma(vma)) {
2576                 split_fb_pmd(vma, pmd);
2577                 spin_unlock(&mm->page_table_lock);
2578                 return;
2579         }
2580         page = pmd_page(*pmd);
2581         VM_BUG_ON(!page_count(page));
2582         get_page(page);
2583         spin_unlock(&mm->page_table_lock);
2584
2585         split_huge_page(page);
2586
2587         put_page(page);
2588         BUG_ON(pmd_trans_huge(*pmd));
2589 }
2590
2591 static void split_huge_page_address(struct mm_struct *mm,
2592                                     unsigned long address)
2593 {
2594         pgd_t *pgd;
2595         pud_t *pud;
2596         pmd_t *pmd;
2597
2598         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2599
2600         pgd = pgd_offset(mm, address);
2601         if (!pgd_present(*pgd))
2602                 return;
2603
2604         pud = pud_offset(pgd, address);
2605         if (!pud_present(*pud))
2606                 return;
2607
2608         pmd = pmd_offset(pud, address);
2609         if (!pmd_present(*pmd))
2610                 return;
2611         /*
2612          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2613          * materialize from under us.
2614          */
2615         split_huge_page_pmd(mm, pmd);
2616 }
2617
2618 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2619                              unsigned long start,
2620                              unsigned long end,
2621                              long adjust_next)
2622 {
2623         /*
2624          * If the new start address isn't hpage aligned and it could
2625          * previously contain an hugepage: check if we need to split
2626          * an huge pmd.
2627          */
2628         if (start & ~HPAGE_PMD_MASK &&
2629             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2630             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2631                 split_huge_page_address(vma->vm_mm, start);
2632
2633         /*
2634          * If the new end address isn't hpage aligned and it could
2635          * previously contain an hugepage: check if we need to split
2636          * an huge pmd.
2637          */
2638         if (end & ~HPAGE_PMD_MASK &&
2639             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2640             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2641                 split_huge_page_address(vma->vm_mm, end);
2642
2643         /*
2644          * If we're also updating the vma->vm_next->vm_start, if the new
2645          * vm_next->vm_start isn't page aligned and it could previously
2646          * contain an hugepage: check if we need to split an huge pmd.
2647          */
2648         if (adjust_next > 0) {
2649                 struct vm_area_struct *next = vma->vm_next;
2650                 unsigned long nstart = next->vm_start;
2651                 nstart += adjust_next << PAGE_SHIFT;
2652                 if (nstart & ~HPAGE_PMD_MASK &&
2653                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2654                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2655                         split_huge_page_address(next->vm_mm, nstart);
2656         }
2657 }