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 struct page *follow_trans_huge_pmd(struct mm_struct *mm,
1001                                    unsigned long addr,
1002                                    pmd_t *pmd,
1003                                    unsigned int flags)
1004 {
1005         struct page *page = NULL;
1006
1007         assert_spin_locked(&mm->page_table_lock);
1008
1009         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1010                 goto out;
1011
1012         page = pmd_page(*pmd);
1013         VM_BUG_ON(!PageHead(page));
1014         if (flags & FOLL_TOUCH) {
1015                 pmd_t _pmd;
1016                 /*
1017                  * We should set the dirty bit only for FOLL_WRITE but
1018                  * for now the dirty bit in the pmd is meaningless.
1019                  * And if the dirty bit will become meaningful and
1020                  * we'll only set it with FOLL_WRITE, an atomic
1021                  * set_bit will be required on the pmd to set the
1022                  * young bit, instead of the current set_pmd_at.
1023                  */
1024                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1025                 set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd);
1026         }
1027         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1028         VM_BUG_ON(!PageCompound(page));
1029         if (flags & FOLL_GET)
1030                 get_page_foll(page);
1031
1032 out:
1033         return page;
1034 }
1035
1036 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1037                  pmd_t *pmd)
1038 {
1039         int ret = 0;
1040
1041         spin_lock(&tlb->mm->page_table_lock);
1042         if (likely(pmd_trans_huge(*pmd))) {
1043                 if (is_fb_vma(vma)) {
1044                         split_fb_pmd(vma, pmd);
1045                         return 0;
1046                 }
1047
1048                 if (unlikely(pmd_trans_splitting(*pmd))) {
1049                         spin_unlock(&tlb->mm->page_table_lock);
1050                         wait_split_huge_page(vma->anon_vma,
1051                                              pmd);
1052                 } else {
1053                         struct page *page;
1054                         pgtable_t pgtable;
1055                         pgtable = get_pmd_huge_pte(tlb->mm);
1056                         page = pmd_page(*pmd);
1057                         pmd_clear(pmd);
1058                         page_remove_rmap(page);
1059                         VM_BUG_ON(page_mapcount(page) < 0);
1060                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1061                         VM_BUG_ON(!PageHead(page));
1062                         tlb->mm->nr_ptes--;
1063                         spin_unlock(&tlb->mm->page_table_lock);
1064                         tlb_remove_page(tlb, page);
1065                         pte_free(tlb->mm, pgtable);
1066                         ret = 1;
1067                 }
1068         } else
1069                 spin_unlock(&tlb->mm->page_table_lock);
1070
1071         return ret;
1072 }
1073
1074 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1075                 unsigned long addr, unsigned long end,
1076                 unsigned char *vec)
1077 {
1078         int ret = 0;
1079
1080         spin_lock(&vma->vm_mm->page_table_lock);
1081         if (likely(pmd_trans_huge(*pmd))) {
1082                 ret = !pmd_trans_splitting(*pmd);
1083                 spin_unlock(&vma->vm_mm->page_table_lock);
1084                 if (unlikely(!ret))
1085                         wait_split_huge_page(vma->anon_vma, pmd);
1086                 else {
1087                         /*
1088                          * All logical pages in the range are present
1089                          * if backed by a huge page.
1090                          */
1091                         memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1092                 }
1093         } else
1094                 spin_unlock(&vma->vm_mm->page_table_lock);
1095
1096         return ret;
1097 }
1098
1099 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1100                   unsigned long old_addr,
1101                   unsigned long new_addr, unsigned long old_end,
1102                   pmd_t *old_pmd, pmd_t *new_pmd)
1103 {
1104         int ret = 0;
1105         pmd_t pmd;
1106
1107         struct mm_struct *mm = vma->vm_mm;
1108
1109         if ((old_addr & ~HPAGE_PMD_MASK) ||
1110             (new_addr & ~HPAGE_PMD_MASK) ||
1111             old_end - old_addr < HPAGE_PMD_SIZE ||
1112             (new_vma->vm_flags & VM_NOHUGEPAGE))
1113                 goto out;
1114
1115         /*
1116          * The destination pmd shouldn't be established, free_pgtables()
1117          * should have release it.
1118          */
1119         if (WARN_ON(!pmd_none(*new_pmd))) {
1120                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1121                 goto out;
1122         }
1123
1124         spin_lock(&mm->page_table_lock);
1125         if (likely(pmd_trans_huge(*old_pmd))) {
1126                 if (pmd_trans_splitting(*old_pmd)) {
1127                         spin_unlock(&mm->page_table_lock);
1128                         wait_split_huge_page(vma->anon_vma, old_pmd);
1129                         ret = -1;
1130                 } else {
1131                         pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1132                         VM_BUG_ON(!pmd_none(*new_pmd));
1133                         set_pmd_at(mm, new_addr, new_pmd, pmd);
1134                         spin_unlock(&mm->page_table_lock);
1135                         ret = 1;
1136                 }
1137         } else {
1138                 spin_unlock(&mm->page_table_lock);
1139         }
1140 out:
1141         return ret;
1142 }
1143
1144 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1145                 unsigned long addr, pgprot_t newprot)
1146 {
1147         struct mm_struct *mm = vma->vm_mm;
1148         int ret = 0;
1149
1150         spin_lock(&mm->page_table_lock);
1151         if (likely(pmd_trans_huge(*pmd))) {
1152                 if (unlikely(pmd_trans_splitting(*pmd))) {
1153                         spin_unlock(&mm->page_table_lock);
1154                         wait_split_huge_page(vma->anon_vma, pmd);
1155                 } else {
1156                         pmd_t entry;
1157
1158                         entry = pmdp_get_and_clear(mm, addr, pmd);
1159                         entry = pmd_modify(entry, newprot);
1160                         set_pmd_at(mm, addr, pmd, entry);
1161                         spin_unlock(&vma->vm_mm->page_table_lock);
1162                         flush_tlb_range(vma, addr, addr + HPAGE_PMD_SIZE);
1163                         ret = 1;
1164                 }
1165         } else
1166                 spin_unlock(&vma->vm_mm->page_table_lock);
1167
1168         return ret;
1169 }
1170
1171 pmd_t *page_check_address_pmd(struct page *page,
1172                               struct mm_struct *mm,
1173                               unsigned long address,
1174                               enum page_check_address_pmd_flag flag)
1175 {
1176         pgd_t *pgd;
1177         pud_t *pud;
1178         pmd_t *pmd, *ret = NULL;
1179
1180         if (address & ~HPAGE_PMD_MASK)
1181                 goto out;
1182
1183         pgd = pgd_offset(mm, address);
1184         if (!pgd_present(*pgd))
1185                 goto out;
1186
1187         pud = pud_offset(pgd, address);
1188         if (!pud_present(*pud))
1189                 goto out;
1190
1191         pmd = pmd_offset(pud, address);
1192         if (pmd_none(*pmd))
1193                 goto out;
1194         if (pmd_page(*pmd) != page)
1195                 goto out;
1196         /*
1197          * split_vma() may create temporary aliased mappings. There is
1198          * no risk as long as all huge pmd are found and have their
1199          * splitting bit set before __split_huge_page_refcount
1200          * runs. Finding the same huge pmd more than once during the
1201          * same rmap walk is not a problem.
1202          */
1203         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1204             pmd_trans_splitting(*pmd))
1205                 goto out;
1206         if (pmd_trans_huge(*pmd)) {
1207                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1208                           !pmd_trans_splitting(*pmd));
1209                 ret = pmd;
1210         }
1211 out:
1212         return ret;
1213 }
1214
1215 static int __split_huge_page_splitting(struct page *page,
1216                                        struct vm_area_struct *vma,
1217                                        unsigned long address)
1218 {
1219         struct mm_struct *mm = vma->vm_mm;
1220         pmd_t *pmd;
1221         int ret = 0;
1222
1223         spin_lock(&mm->page_table_lock);
1224         pmd = page_check_address_pmd(page, mm, address,
1225                                      PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1226         if (pmd) {
1227                 /*
1228                  * We can't temporarily set the pmd to null in order
1229                  * to split it, the pmd must remain marked huge at all
1230                  * times or the VM won't take the pmd_trans_huge paths
1231                  * and it won't wait on the anon_vma->root->mutex to
1232                  * serialize against split_huge_page*.
1233                  */
1234                 pmdp_splitting_flush_notify(vma, address, pmd);
1235                 ret = 1;
1236         }
1237         spin_unlock(&mm->page_table_lock);
1238
1239         return ret;
1240 }
1241
1242 static void __split_huge_page_refcount(struct page *page)
1243 {
1244         int i;
1245         unsigned long head_index = page->index;
1246         struct zone *zone = page_zone(page);
1247         int zonestat;
1248         int tail_count = 0;
1249
1250         /* prevent PageLRU to go away from under us, and freeze lru stats */
1251         spin_lock_irq(&zone->lru_lock);
1252         compound_lock(page);
1253
1254         for (i = 1; i < HPAGE_PMD_NR; i++) {
1255                 struct page *page_tail = page + i;
1256
1257                 /* tail_page->_mapcount cannot change */
1258                 BUG_ON(page_mapcount(page_tail) < 0);
1259                 tail_count += page_mapcount(page_tail);
1260                 /* check for overflow */
1261                 BUG_ON(tail_count < 0);
1262                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1263                 /*
1264                  * tail_page->_count is zero and not changing from
1265                  * under us. But get_page_unless_zero() may be running
1266                  * from under us on the tail_page. If we used
1267                  * atomic_set() below instead of atomic_add(), we
1268                  * would then run atomic_set() concurrently with
1269                  * get_page_unless_zero(), and atomic_set() is
1270                  * implemented in C not using locked ops. spin_unlock
1271                  * on x86 sometime uses locked ops because of PPro
1272                  * errata 66, 92, so unless somebody can guarantee
1273                  * atomic_set() here would be safe on all archs (and
1274                  * not only on x86), it's safer to use atomic_add().
1275                  */
1276                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1277                            &page_tail->_count);
1278
1279                 /* after clearing PageTail the gup refcount can be released */
1280                 smp_mb();
1281
1282                 /*
1283                  * retain hwpoison flag of the poisoned tail page:
1284                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1285                  *   by the memory-failure.
1286                  */
1287                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1288                 page_tail->flags |= (page->flags &
1289                                      ((1L << PG_referenced) |
1290                                       (1L << PG_swapbacked) |
1291                                       (1L << PG_mlocked) |
1292                                       (1L << PG_uptodate)));
1293                 page_tail->flags |= (1L << PG_dirty);
1294
1295                 /* clear PageTail before overwriting first_page */
1296                 smp_wmb();
1297
1298                 /*
1299                  * __split_huge_page_splitting() already set the
1300                  * splitting bit in all pmd that could map this
1301                  * hugepage, that will ensure no CPU can alter the
1302                  * mapcount on the head page. The mapcount is only
1303                  * accounted in the head page and it has to be
1304                  * transferred to all tail pages in the below code. So
1305                  * for this code to be safe, the split the mapcount
1306                  * can't change. But that doesn't mean userland can't
1307                  * keep changing and reading the page contents while
1308                  * we transfer the mapcount, so the pmd splitting
1309                  * status is achieved setting a reserved bit in the
1310                  * pmd, not by clearing the present bit.
1311                 */
1312                 page_tail->_mapcount = page->_mapcount;
1313
1314                 BUG_ON(page_tail->mapping);
1315                 page_tail->mapping = page->mapping;
1316
1317                 page_tail->index = ++head_index;
1318
1319                 BUG_ON(!PageAnon(page_tail));
1320                 BUG_ON(!PageUptodate(page_tail));
1321                 BUG_ON(!PageDirty(page_tail));
1322                 BUG_ON(!PageSwapBacked(page_tail));
1323
1324                 mem_cgroup_split_huge_fixup(page, page_tail);
1325
1326                 lru_add_page_tail(zone, page, page_tail);
1327         }
1328         atomic_sub(tail_count, &page->_count);
1329         BUG_ON(atomic_read(&page->_count) <= 0);
1330
1331         __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1332         __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1333
1334         /*
1335          * A hugepage counts for HPAGE_PMD_NR pages on the LRU statistics,
1336          * so adjust those appropriately if this page is on the LRU.
1337          */
1338         if (PageLRU(page)) {
1339                 zonestat = NR_LRU_BASE + page_lru(page);
1340                 __mod_zone_page_state(zone, zonestat, -(HPAGE_PMD_NR-1));
1341         }
1342
1343         ClearPageCompound(page);
1344         compound_unlock(page);
1345         spin_unlock_irq(&zone->lru_lock);
1346
1347         for (i = 1; i < HPAGE_PMD_NR; i++) {
1348                 struct page *page_tail = page + i;
1349                 BUG_ON(page_count(page_tail) <= 0);
1350                 /*
1351                  * Tail pages may be freed if there wasn't any mapping
1352                  * like if add_to_swap() is running on a lru page that
1353                  * had its mapping zapped. And freeing these pages
1354                  * requires taking the lru_lock so we do the put_page
1355                  * of the tail pages after the split is complete.
1356                  */
1357                 put_page(page_tail);
1358         }
1359
1360         /*
1361          * Only the head page (now become a regular page) is required
1362          * to be pinned by the caller.
1363          */
1364         BUG_ON(page_count(page) <= 0);
1365 }
1366
1367 static int __split_huge_page_map(struct page *page,
1368                                  struct vm_area_struct *vma,
1369                                  unsigned long address)
1370 {
1371         struct mm_struct *mm = vma->vm_mm;
1372         pmd_t *pmd, _pmd;
1373         int ret = 0, i;
1374         pgtable_t pgtable;
1375         unsigned long haddr;
1376
1377         spin_lock(&mm->page_table_lock);
1378         pmd = page_check_address_pmd(page, mm, address,
1379                                      PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1380         if (pmd) {
1381                 pgtable = get_pmd_huge_pte(mm);
1382                 pmd_populate(mm, &_pmd, pgtable);
1383
1384                 for (i = 0, haddr = address; i < HPAGE_PMD_NR;
1385                      i++, haddr += PAGE_SIZE) {
1386                         pte_t *pte, entry;
1387                         BUG_ON(PageCompound(page+i));
1388                         entry = mk_pte(page + i, vma->vm_page_prot);
1389                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1390                         if (!pmd_write(*pmd))
1391                                 entry = pte_wrprotect(entry);
1392                         else
1393                                 BUG_ON(page_mapcount(page) != 1);
1394                         if (!pmd_young(*pmd))
1395                                 entry = pte_mkold(entry);
1396                         pte = pte_offset_map(&_pmd, haddr);
1397                         BUG_ON(!pte_none(*pte));
1398                         set_pte_at(mm, haddr, pte, entry);
1399                         pte_unmap(pte);
1400                 }
1401
1402                 smp_wmb(); /* make pte visible before pmd */
1403                 /*
1404                  * Up to this point the pmd is present and huge and
1405                  * userland has the whole access to the hugepage
1406                  * during the split (which happens in place). If we
1407                  * overwrite the pmd with the not-huge version
1408                  * pointing to the pte here (which of course we could
1409                  * if all CPUs were bug free), userland could trigger
1410                  * a small page size TLB miss on the small sized TLB
1411                  * while the hugepage TLB entry is still established
1412                  * in the huge TLB. Some CPU doesn't like that. See
1413                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1414                  * Erratum 383 on page 93. Intel should be safe but is
1415                  * also warns that it's only safe if the permission
1416                  * and cache attributes of the two entries loaded in
1417                  * the two TLB is identical (which should be the case
1418                  * here). But it is generally safer to never allow
1419                  * small and huge TLB entries for the same virtual
1420                  * address to be loaded simultaneously. So instead of
1421                  * doing "pmd_populate(); flush_tlb_range();" we first
1422                  * mark the current pmd notpresent (atomically because
1423                  * here the pmd_trans_huge and pmd_trans_splitting
1424                  * must remain set at all times on the pmd until the
1425                  * split is complete for this pmd), then we flush the
1426                  * SMP TLB and finally we write the non-huge version
1427                  * of the pmd entry with pmd_populate.
1428                  */
1429                 set_pmd_at(mm, address, pmd, pmd_mknotpresent(*pmd));
1430                 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
1431                 pmd_populate(mm, pmd, pgtable);
1432                 ret = 1;
1433         }
1434         spin_unlock(&mm->page_table_lock);
1435
1436         return ret;
1437 }
1438
1439 /* must be called with anon_vma->root->mutex hold */
1440 static void __split_huge_page(struct page *page,
1441                               struct anon_vma *anon_vma)
1442 {
1443         int mapcount, mapcount2;
1444         struct anon_vma_chain *avc;
1445
1446         BUG_ON(!PageHead(page));
1447         BUG_ON(PageTail(page));
1448
1449         mapcount = 0;
1450         list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1451                 struct vm_area_struct *vma = avc->vma;
1452                 unsigned long addr = vma_address(page, vma);
1453                 BUG_ON(is_vma_temporary_stack(vma));
1454                 if (addr == -EFAULT)
1455                         continue;
1456                 mapcount += __split_huge_page_splitting(page, vma, addr);
1457         }
1458         /*
1459          * It is critical that new vmas are added to the tail of the
1460          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1461          * and establishes a child pmd before
1462          * __split_huge_page_splitting() freezes the parent pmd (so if
1463          * we fail to prevent copy_huge_pmd() from running until the
1464          * whole __split_huge_page() is complete), we will still see
1465          * the newly established pmd of the child later during the
1466          * walk, to be able to set it as pmd_trans_splitting too.
1467          */
1468         if (mapcount != page_mapcount(page))
1469                 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1470                        mapcount, page_mapcount(page));
1471         BUG_ON(mapcount != page_mapcount(page));
1472
1473         __split_huge_page_refcount(page);
1474
1475         mapcount2 = 0;
1476         list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1477                 struct vm_area_struct *vma = avc->vma;
1478                 unsigned long addr = vma_address(page, vma);
1479                 BUG_ON(is_vma_temporary_stack(vma));
1480                 if (addr == -EFAULT)
1481                         continue;
1482                 mapcount2 += __split_huge_page_map(page, vma, addr);
1483         }
1484         if (mapcount != mapcount2)
1485                 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1486                        mapcount, mapcount2, page_mapcount(page));
1487         BUG_ON(mapcount != mapcount2);
1488 }
1489
1490 int split_huge_page(struct page *page)
1491 {
1492         struct anon_vma *anon_vma;
1493         int ret = 1;
1494
1495         BUG_ON(!PageAnon(page));
1496         anon_vma = page_lock_anon_vma(page);
1497         if (!anon_vma)
1498                 goto out;
1499         ret = 0;
1500         if (!PageCompound(page))
1501                 goto out_unlock;
1502
1503         BUG_ON(!PageSwapBacked(page));
1504         __split_huge_page(page, anon_vma);
1505         count_vm_event(THP_SPLIT);
1506
1507         BUG_ON(PageCompound(page));
1508 out_unlock:
1509         page_unlock_anon_vma(anon_vma);
1510 out:
1511         return ret;
1512 }
1513
1514 /* callers must hold mmap_sem (madvise() does) */
1515 static int collapse_fb_pmd(struct mm_struct *mm, pmd_t *pmd,
1516         unsigned long addr, struct vm_area_struct *vma)
1517 {
1518         unsigned long _addr;
1519         struct page *page;
1520         pgtable_t pgtable;
1521         pte_t *pte, *_pte;
1522         pmd_t _pmd;
1523         u32 pa;
1524
1525         pte = pte_offset_map(pmd, addr);
1526         page = pte_page(*pte);
1527         pa = __pfn_to_phys(page_to_pfn(page));
1528         _pmd = pmdp_clear_flush_notify(vma, addr, pmd);
1529
1530         if ((addr | pa) & ~HPAGE_PMD_MASK) {
1531                 printk(KERN_ERR "collapse_fb: bad alignment: %08lx->%08x\n",
1532                         addr, pa);
1533                 pte_unmap(pte);
1534                 return -EINVAL;
1535         }
1536
1537         for (_pte = pte, _addr = addr; _pte < pte + HPAGE_PMD_NR; _pte++) {
1538                 pte_t pteval = *_pte;
1539                 struct page *src_page;
1540
1541                 if (!pte_none(pteval)) {
1542                         src_page = pte_page(pteval);
1543
1544                         pte_clear(vma->vm_mm, _addr, _pte);
1545                         if (pte_present(pteval))
1546                                 page_remove_rmap(src_page);
1547                 }
1548
1549                 _addr += PAGE_SIZE;
1550         }
1551
1552         pte_unmap(pte);
1553         pgtable = pmd_pgtable(_pmd);
1554         VM_BUG_ON(page_count(pgtable) != 1);
1555         VM_BUG_ON(page_mapcount(pgtable) != 0);
1556
1557         _pmd = mk_pmd(page, vma->vm_page_prot);
1558         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1559         _pmd = pmd_mkhuge(_pmd);
1560
1561         smp_wmb();
1562
1563         spin_lock(&mm->page_table_lock);
1564         BUG_ON(!pmd_none(*pmd));
1565         set_pmd_at(mm, addr, pmd, _pmd);
1566         update_mmu_cache(vma, addr, pmd);
1567         prepare_pmd_huge_pte(pgtable, mm);
1568         spin_unlock(&mm->page_table_lock);
1569
1570         return 0;
1571 }
1572
1573 static int try_collapse_fb(struct vm_area_struct *vma)
1574 {
1575         struct mm_struct *mm = vma->vm_mm;
1576         unsigned long hstart, hend, addr;
1577         int ret = 0;
1578         pgd_t *pgd;
1579         pud_t *pud;
1580         pmd_t *pmd;
1581
1582         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1583         hend = vma->vm_end & HPAGE_PMD_MASK;
1584         if (hstart >= hend)
1585                 return -EINVAL;
1586
1587         for (addr = hstart; addr < hend; addr += HPAGE_PMD_SIZE) {
1588                 pgd = pgd_offset(mm, addr);
1589                 if (!pgd_present(*pgd))
1590                         return -EINVAL;
1591
1592                 pud = pud_offset(pgd, addr);
1593                 if (!pud_present(*pud))
1594                         return -EINVAL;
1595
1596                 pmd = pmd_offset(pud, addr);
1597                 if (!pmd_present(*pmd))
1598                         return -EINVAL;
1599                 if (pmd_trans_huge(*pmd))
1600                         continue;
1601
1602                 ret = collapse_fb_pmd(mm, pmd, addr, vma);
1603                 if (ret)
1604                         break;
1605         }
1606
1607         return ret;
1608 }
1609
1610 /* undo collapse_fb_pmd(), restore pages so that mm subsys can release them
1611  * page_table_lock() should be held */
1612 static void split_fb_pmd(struct vm_area_struct *vma, pmd_t *pmd)
1613 {
1614         struct mm_struct *mm = vma->vm_mm;
1615         unsigned long addr, haddr, pfn;
1616         struct page *page;
1617         pgtable_t pgtable;
1618         pmd_t _pmd;
1619         int i;
1620
1621         page = pmd_page(*pmd);
1622         pgtable = get_pmd_huge_pte(mm);
1623         pfn = page_to_pfn(page);
1624         addr = pfn << PAGE_SHIFT;
1625
1626         pmd_populate(mm, &_pmd, pgtable);
1627
1628         for (i = 0, haddr = addr; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1629                 pte_t *pte, entry;
1630                 BUG_ON(PageCompound(page + i));
1631                 entry = mk_pte(page + i, vma->vm_page_prot);
1632                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1633                 if (!pmd_young(*pmd))
1634                         entry = pte_mkold(entry);
1635                 atomic_set(&page[i]._mapcount, 0); // hack?
1636                 pte = pte_offset_map(&_pmd, haddr);
1637                 BUG_ON(!pte_none(*pte));
1638                 set_pte_at(mm, haddr, pte, entry);
1639                 pte_unmap(pte);
1640         }
1641
1642         set_pmd_at(mm, addr, pmd, pmd_mknotpresent(*pmd));
1643         flush_tlb_range(vma, addr, addr + HPAGE_PMD_SIZE);
1644         pmd_populate(mm, pmd, pgtable);
1645 }
1646
1647 #ifndef __arm__
1648 #error arm only..
1649 #endif
1650 static u32 pmd_to_va(struct mm_struct *mm, pmd_t *pmd)
1651 {
1652         pgd_t *pgd;
1653         pud_t *pud;
1654         pmd_t *pmd0;
1655         u32 ret;
1656
1657         pgd = pgd_offset(mm, 0);
1658         pud = pud_offset(pgd, 0);
1659         pmd0 = pmd_offset(pud, 0);
1660
1661         ret = (pmd - pmd0) << SECTION_SHIFT;
1662         return ret;
1663 }
1664
1665 #define VM_NO_THP (VM_SPECIAL|VM_INSERTPAGE|VM_MIXEDMAP|VM_SAO| \
1666                    VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1667
1668 int hugepage_madvise(struct vm_area_struct *vma,
1669                      unsigned long *vm_flags, int advice)
1670 {
1671         switch (advice) {
1672         case MADV_HUGEPAGE:
1673                 if (is_fb_vma(vma))
1674                         return try_collapse_fb(vma);
1675
1676                 /*
1677                  * Be somewhat over-protective like KSM for now!
1678                  */
1679                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1680                         return -EINVAL;
1681                 *vm_flags &= ~VM_NOHUGEPAGE;
1682                 *vm_flags |= VM_HUGEPAGE;
1683                 /*
1684                  * If the vma become good for khugepaged to scan,
1685                  * register it here without waiting a page fault that
1686                  * may not happen any time soon.
1687                  */
1688                 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
1689                         return -ENOMEM;
1690                 break;
1691         case MADV_NOHUGEPAGE:
1692                 /*
1693                  * Be somewhat over-protective like KSM for now!
1694                  */
1695                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1696                         return -EINVAL;
1697                 *vm_flags &= ~VM_HUGEPAGE;
1698                 *vm_flags |= VM_NOHUGEPAGE;
1699                 /*
1700                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1701                  * this vma even if we leave the mm registered in khugepaged if
1702                  * it got registered before VM_NOHUGEPAGE was set.
1703                  */
1704                 break;
1705         }
1706
1707         return 0;
1708 }
1709
1710 static int __init khugepaged_slab_init(void)
1711 {
1712         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1713                                           sizeof(struct mm_slot),
1714                                           __alignof__(struct mm_slot), 0, NULL);
1715         if (!mm_slot_cache)
1716                 return -ENOMEM;
1717
1718         return 0;
1719 }
1720
1721 static void __init khugepaged_slab_free(void)
1722 {
1723         kmem_cache_destroy(mm_slot_cache);
1724         mm_slot_cache = NULL;
1725 }
1726
1727 static inline struct mm_slot *alloc_mm_slot(void)
1728 {
1729         if (!mm_slot_cache)     /* initialization failed */
1730                 return NULL;
1731         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1732 }
1733
1734 static inline void free_mm_slot(struct mm_slot *mm_slot)
1735 {
1736         kmem_cache_free(mm_slot_cache, mm_slot);
1737 }
1738
1739 static int __init mm_slots_hash_init(void)
1740 {
1741         mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
1742                                 GFP_KERNEL);
1743         if (!mm_slots_hash)
1744                 return -ENOMEM;
1745         return 0;
1746 }
1747
1748 #if 0
1749 static void __init mm_slots_hash_free(void)
1750 {
1751         kfree(mm_slots_hash);
1752         mm_slots_hash = NULL;
1753 }
1754 #endif
1755
1756 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1757 {
1758         struct mm_slot *mm_slot;
1759         struct hlist_head *bucket;
1760         struct hlist_node *node;
1761
1762         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1763                                 % MM_SLOTS_HASH_HEADS];
1764         hlist_for_each_entry(mm_slot, node, bucket, hash) {
1765                 if (mm == mm_slot->mm)
1766                         return mm_slot;
1767         }
1768         return NULL;
1769 }
1770
1771 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1772                                     struct mm_slot *mm_slot)
1773 {
1774         struct hlist_head *bucket;
1775
1776         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1777                                 % MM_SLOTS_HASH_HEADS];
1778         mm_slot->mm = mm;
1779         hlist_add_head(&mm_slot->hash, bucket);
1780 }
1781
1782 static inline int khugepaged_test_exit(struct mm_struct *mm)
1783 {
1784         return atomic_read(&mm->mm_users) == 0;
1785 }
1786
1787 int __khugepaged_enter(struct mm_struct *mm)
1788 {
1789         struct mm_slot *mm_slot;
1790         int wakeup;
1791
1792         mm_slot = alloc_mm_slot();
1793         if (!mm_slot)
1794                 return -ENOMEM;
1795
1796         /* __khugepaged_exit() must not run from under us */
1797         VM_BUG_ON(khugepaged_test_exit(mm));
1798         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1799                 free_mm_slot(mm_slot);
1800                 return 0;
1801         }
1802
1803         spin_lock(&khugepaged_mm_lock);
1804         insert_to_mm_slots_hash(mm, mm_slot);
1805         /*
1806          * Insert just behind the scanning cursor, to let the area settle
1807          * down a little.
1808          */
1809         wakeup = list_empty(&khugepaged_scan.mm_head);
1810         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1811         spin_unlock(&khugepaged_mm_lock);
1812
1813         atomic_inc(&mm->mm_count);
1814         if (wakeup)
1815                 wake_up_interruptible(&khugepaged_wait);
1816
1817         return 0;
1818 }
1819
1820 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
1821                                unsigned long vm_flags)
1822 {
1823         unsigned long hstart, hend;
1824         if (!vma->anon_vma)
1825                 /*
1826                  * Not yet faulted in so we will register later in the
1827                  * page fault if needed.
1828                  */
1829                 return 0;
1830         if (vma->vm_ops)
1831                 /* khugepaged not yet working on file or special mappings */
1832                 return 0;
1833         /*
1834          * If is_pfn_mapping() is true is_learn_pfn_mapping() must be
1835          * true too, verify it here.
1836          */
1837         VM_BUG_ON(is_linear_pfn_mapping(vma) || vm_flags & VM_NO_THP);
1838         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1839         hend = vma->vm_end & HPAGE_PMD_MASK;
1840         if (hstart < hend)
1841                 return khugepaged_enter(vma, vm_flags);
1842         return 0;
1843 }
1844
1845 void __khugepaged_exit(struct mm_struct *mm)
1846 {
1847         struct mm_slot *mm_slot;
1848         int free = 0;
1849
1850         spin_lock(&khugepaged_mm_lock);
1851         mm_slot = get_mm_slot(mm);
1852         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1853                 hlist_del(&mm_slot->hash);
1854                 list_del(&mm_slot->mm_node);
1855                 free = 1;
1856         }
1857         spin_unlock(&khugepaged_mm_lock);
1858
1859         if (free) {
1860                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1861                 free_mm_slot(mm_slot);
1862                 mmdrop(mm);
1863         } else if (mm_slot) {
1864                 /*
1865                  * This is required to serialize against
1866                  * khugepaged_test_exit() (which is guaranteed to run
1867                  * under mmap sem read mode). Stop here (after we
1868                  * return all pagetables will be destroyed) until
1869                  * khugepaged has finished working on the pagetables
1870                  * under the mmap_sem.
1871                  */
1872                 down_write(&mm->mmap_sem);
1873                 up_write(&mm->mmap_sem);
1874         }
1875 }
1876
1877 static void release_pte_page(struct page *page)
1878 {
1879         /* 0 stands for page_is_file_cache(page) == false */
1880         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1881         unlock_page(page);
1882         putback_lru_page(page);
1883 }
1884
1885 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1886 {
1887         while (--_pte >= pte) {
1888                 pte_t pteval = *_pte;
1889                 if (!pte_none(pteval))
1890                         release_pte_page(pte_page(pteval));
1891         }
1892 }
1893
1894 static void release_all_pte_pages(pte_t *pte)
1895 {
1896         release_pte_pages(pte, pte + HPAGE_PMD_NR);
1897 }
1898
1899 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1900                                         unsigned long address,
1901                                         pte_t *pte)
1902 {
1903         struct page *page;
1904         pte_t *_pte;
1905         int referenced = 0, isolated = 0, none = 0;
1906         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1907              _pte++, address += PAGE_SIZE) {
1908                 pte_t pteval = *_pte;
1909                 if (pte_none(pteval)) {
1910                         if (++none <= khugepaged_max_ptes_none)
1911                                 continue;
1912                         else {
1913                                 release_pte_pages(pte, _pte);
1914                                 goto out;
1915                         }
1916                 }
1917                 if (!pte_present(pteval) || !pte_write(pteval)) {
1918                         release_pte_pages(pte, _pte);
1919                         goto out;
1920                 }
1921                 page = vm_normal_page(vma, address, pteval);
1922                 if (unlikely(!page)) {
1923                         release_pte_pages(pte, _pte);
1924                         goto out;
1925                 }
1926                 VM_BUG_ON(PageCompound(page));
1927                 BUG_ON(!PageAnon(page));
1928                 VM_BUG_ON(!PageSwapBacked(page));
1929
1930                 /* cannot use mapcount: can't collapse if there's a gup pin */
1931                 if (page_count(page) != 1) {
1932                         release_pte_pages(pte, _pte);
1933                         goto out;
1934                 }
1935                 /*
1936                  * We can do it before isolate_lru_page because the
1937                  * page can't be freed from under us. NOTE: PG_lock
1938                  * is needed to serialize against split_huge_page
1939                  * when invoked from the VM.
1940                  */
1941                 if (!trylock_page(page)) {
1942                         release_pte_pages(pte, _pte);
1943                         goto out;
1944                 }
1945                 /*
1946                  * Isolate the page to avoid collapsing an hugepage
1947                  * currently in use by the VM.
1948                  */
1949                 if (isolate_lru_page(page)) {
1950                         unlock_page(page);
1951                         release_pte_pages(pte, _pte);
1952                         goto out;
1953                 }
1954                 /* 0 stands for page_is_file_cache(page) == false */
1955                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
1956                 VM_BUG_ON(!PageLocked(page));
1957                 VM_BUG_ON(PageLRU(page));
1958
1959                 /* If there is no mapped pte young don't collapse the page */
1960                 if (pte_young(pteval) || PageReferenced(page) ||
1961                     mmu_notifier_test_young(vma->vm_mm, address))
1962                         referenced = 1;
1963         }
1964         if (unlikely(!referenced))
1965                 release_all_pte_pages(pte);
1966         else
1967                 isolated = 1;
1968 out:
1969         return isolated;
1970 }
1971
1972 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
1973                                       struct vm_area_struct *vma,
1974                                       unsigned long address,
1975                                       spinlock_t *ptl)
1976 {
1977         pte_t *_pte;
1978         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
1979                 pte_t pteval = *_pte;
1980                 struct page *src_page;
1981
1982                 if (pte_none(pteval)) {
1983                         clear_user_highpage(page, address);
1984                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
1985                 } else {
1986                         src_page = pte_page(pteval);
1987                         copy_user_highpage(page, src_page, address, vma);
1988                         VM_BUG_ON(page_mapcount(src_page) != 1);
1989                         VM_BUG_ON(page_count(src_page) != 2);
1990                         release_pte_page(src_page);
1991                         /*
1992                          * ptl mostly unnecessary, but preempt has to
1993                          * be disabled to update the per-cpu stats
1994                          * inside page_remove_rmap().
1995                          */
1996                         spin_lock(ptl);
1997                         /*
1998                          * paravirt calls inside pte_clear here are
1999                          * superfluous.
2000                          */
2001                         pte_clear(vma->vm_mm, address, _pte);
2002                         page_remove_rmap(src_page);
2003                         spin_unlock(ptl);
2004                         free_page_and_swap_cache(src_page);
2005                 }
2006
2007                 address += PAGE_SIZE;
2008                 page++;
2009         }
2010 }
2011
2012 static void collapse_huge_page(struct mm_struct *mm,
2013                                unsigned long address,
2014                                struct page **hpage,
2015                                struct vm_area_struct *vma,
2016                                int node)
2017 {
2018         pgd_t *pgd;
2019         pud_t *pud;
2020         pmd_t *pmd, _pmd;
2021         pte_t *pte;
2022         pgtable_t pgtable;
2023         struct page *new_page;
2024         spinlock_t *ptl;
2025         int isolated;
2026         unsigned long hstart, hend;
2027
2028         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2029 #ifndef CONFIG_NUMA
2030         up_read(&mm->mmap_sem);
2031         VM_BUG_ON(!*hpage);
2032         new_page = *hpage;
2033 #else
2034         VM_BUG_ON(*hpage);
2035         /*
2036          * Allocate the page while the vma is still valid and under
2037          * the mmap_sem read mode so there is no memory allocation
2038          * later when we take the mmap_sem in write mode. This is more
2039          * friendly behavior (OTOH it may actually hide bugs) to
2040          * filesystems in userland with daemons allocating memory in
2041          * the userland I/O paths.  Allocating memory with the
2042          * mmap_sem in read mode is good idea also to allow greater
2043          * scalability.
2044          */
2045         new_page = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
2046                                       node, __GFP_OTHER_NODE);
2047
2048         /*
2049          * After allocating the hugepage, release the mmap_sem read lock in
2050          * preparation for taking it in write mode.
2051          */
2052         up_read(&mm->mmap_sem);
2053         if (unlikely(!new_page)) {
2054                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2055                 *hpage = ERR_PTR(-ENOMEM);
2056                 return;
2057         }
2058 #endif
2059
2060         count_vm_event(THP_COLLAPSE_ALLOC);
2061         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
2062 #ifdef CONFIG_NUMA
2063                 put_page(new_page);
2064 #endif
2065                 return;
2066         }
2067
2068         /*
2069          * Prevent all access to pagetables with the exception of
2070          * gup_fast later hanlded by the ptep_clear_flush and the VM
2071          * handled by the anon_vma lock + PG_lock.
2072          */
2073         down_write(&mm->mmap_sem);
2074         if (unlikely(khugepaged_test_exit(mm)))
2075                 goto out;
2076
2077         vma = find_vma(mm, address);
2078         if (!vma)
2079                 goto out;
2080         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2081         hend = vma->vm_end & HPAGE_PMD_MASK;
2082         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2083                 goto out;
2084
2085         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2086             (vma->vm_flags & VM_NOHUGEPAGE))
2087                 goto out;
2088
2089         if (!vma->anon_vma || vma->vm_ops)
2090                 goto out;
2091         if (is_vma_temporary_stack(vma))
2092                 goto out;
2093         /*
2094          * If is_pfn_mapping() is true is_learn_pfn_mapping() must be
2095          * true too, verify it here.
2096          */
2097         VM_BUG_ON(is_linear_pfn_mapping(vma) || vma->vm_flags & VM_NO_THP);
2098
2099         pgd = pgd_offset(mm, address);
2100         if (!pgd_present(*pgd))
2101                 goto out;
2102
2103         pud = pud_offset(pgd, address);
2104         if (!pud_present(*pud))
2105                 goto out;
2106
2107         pmd = pmd_offset(pud, address);
2108         /* pmd can't go away or become huge under us */
2109         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
2110                 goto out;
2111
2112         anon_vma_lock(vma->anon_vma);
2113
2114         pte = pte_offset_map(pmd, address);
2115         ptl = pte_lockptr(mm, pmd);
2116
2117         spin_lock(&mm->page_table_lock); /* probably unnecessary */
2118         /*
2119          * After this gup_fast can't run anymore. This also removes
2120          * any huge TLB entry from the CPU so we won't allow
2121          * huge and small TLB entries for the same virtual address
2122          * to avoid the risk of CPU bugs in that area.
2123          */
2124         _pmd = pmdp_clear_flush_notify(vma, address, pmd);
2125         spin_unlock(&mm->page_table_lock);
2126
2127         spin_lock(ptl);
2128         isolated = __collapse_huge_page_isolate(vma, address, pte);
2129         spin_unlock(ptl);
2130
2131         if (unlikely(!isolated)) {
2132                 pte_unmap(pte);
2133                 spin_lock(&mm->page_table_lock);
2134                 BUG_ON(!pmd_none(*pmd));
2135                 /*
2136                  * We can only use set_pmd_at when establishing
2137                  * hugepmds and never for establishing regular pmds that
2138                  * points to regular pagetables. Use pmd_populate for that
2139                  */
2140                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2141                 spin_unlock(&mm->page_table_lock);
2142                 anon_vma_unlock(vma->anon_vma);
2143                 goto out;
2144         }
2145
2146         /*
2147          * All pages are isolated and locked so anon_vma rmap
2148          * can't run anymore.
2149          */
2150         anon_vma_unlock(vma->anon_vma);
2151
2152         __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
2153         pte_unmap(pte);
2154         __SetPageUptodate(new_page);
2155         pgtable = pmd_pgtable(_pmd);
2156         VM_BUG_ON(page_count(pgtable) != 1);
2157         VM_BUG_ON(page_mapcount(pgtable) != 0);
2158
2159         _pmd = mk_pmd(new_page, vma->vm_page_prot);
2160         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2161         _pmd = pmd_mkhuge(_pmd);
2162
2163         /*
2164          * spin_lock() below is not the equivalent of smp_wmb(), so
2165          * this is needed to avoid the copy_huge_page writes to become
2166          * visible after the set_pmd_at() write.
2167          */
2168         smp_wmb();
2169
2170         spin_lock(&mm->page_table_lock);
2171         BUG_ON(!pmd_none(*pmd));
2172         page_add_new_anon_rmap(new_page, vma, address);
2173         set_pmd_at(mm, address, pmd, _pmd);
2174         update_mmu_cache(vma, address, pmd);
2175         prepare_pmd_huge_pte(pgtable, mm);
2176         spin_unlock(&mm->page_table_lock);
2177
2178 #ifndef CONFIG_NUMA
2179         *hpage = NULL;
2180 #endif
2181         khugepaged_pages_collapsed++;
2182 out_up_write:
2183         up_write(&mm->mmap_sem);
2184         return;
2185
2186 out:
2187         mem_cgroup_uncharge_page(new_page);
2188 #ifdef CONFIG_NUMA
2189         put_page(new_page);
2190 #endif
2191         goto out_up_write;
2192 }
2193
2194 static int khugepaged_scan_pmd(struct mm_struct *mm,
2195                                struct vm_area_struct *vma,
2196                                unsigned long address,
2197                                struct page **hpage)
2198 {
2199         pgd_t *pgd;
2200         pud_t *pud;
2201         pmd_t *pmd;
2202         pte_t *pte, *_pte;
2203         int ret = 0, referenced = 0, none = 0;
2204         struct page *page;
2205         unsigned long _address;
2206         spinlock_t *ptl;
2207         int node = -1;
2208
2209         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2210
2211         pgd = pgd_offset(mm, address);
2212         if (!pgd_present(*pgd))
2213                 goto out;
2214
2215         pud = pud_offset(pgd, address);
2216         if (!pud_present(*pud))
2217                 goto out;
2218
2219         pmd = pmd_offset(pud, address);
2220         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
2221                 goto out;
2222
2223         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2224         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2225              _pte++, _address += PAGE_SIZE) {
2226                 pte_t pteval = *_pte;
2227                 if (pte_none(pteval)) {
2228                         if (++none <= khugepaged_max_ptes_none)
2229                                 continue;
2230                         else
2231                                 goto out_unmap;
2232                 }
2233                 if (!pte_present(pteval) || !pte_write(pteval))
2234                         goto out_unmap;
2235                 page = vm_normal_page(vma, _address, pteval);
2236                 if (unlikely(!page))
2237                         goto out_unmap;
2238                 /*
2239                  * Chose the node of the first page. This could
2240                  * be more sophisticated and look at more pages,
2241                  * but isn't for now.
2242                  */
2243                 if (node == -1)
2244                         node = page_to_nid(page);
2245                 VM_BUG_ON(PageCompound(page));
2246                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2247                         goto out_unmap;
2248                 /* cannot use mapcount: can't collapse if there's a gup pin */
2249                 if (page_count(page) != 1)
2250                         goto out_unmap;
2251                 if (pte_young(pteval) || PageReferenced(page) ||
2252                     mmu_notifier_test_young(vma->vm_mm, address))
2253                         referenced = 1;
2254         }
2255         if (referenced)
2256                 ret = 1;
2257 out_unmap:
2258         pte_unmap_unlock(pte, ptl);
2259         if (ret)
2260                 /* collapse_huge_page will return with the mmap_sem released */
2261                 collapse_huge_page(mm, address, hpage, vma, node);
2262 out:
2263         return ret;
2264 }
2265
2266 static void collect_mm_slot(struct mm_slot *mm_slot)
2267 {
2268         struct mm_struct *mm = mm_slot->mm;
2269
2270         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2271
2272         if (khugepaged_test_exit(mm)) {
2273                 /* free mm_slot */
2274                 hlist_del(&mm_slot->hash);
2275                 list_del(&mm_slot->mm_node);
2276
2277                 /*
2278                  * Not strictly needed because the mm exited already.
2279                  *
2280                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2281                  */
2282
2283                 /* khugepaged_mm_lock actually not necessary for the below */
2284                 free_mm_slot(mm_slot);
2285                 mmdrop(mm);
2286         }
2287 }
2288
2289 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2290                                             struct page **hpage)
2291         __releases(&khugepaged_mm_lock)
2292         __acquires(&khugepaged_mm_lock)
2293 {
2294         struct mm_slot *mm_slot;
2295         struct mm_struct *mm;
2296         struct vm_area_struct *vma;
2297         int progress = 0;
2298
2299         VM_BUG_ON(!pages);
2300         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2301
2302         if (khugepaged_scan.mm_slot)
2303                 mm_slot = khugepaged_scan.mm_slot;
2304         else {
2305                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2306                                      struct mm_slot, mm_node);
2307                 khugepaged_scan.address = 0;
2308                 khugepaged_scan.mm_slot = mm_slot;
2309         }
2310         spin_unlock(&khugepaged_mm_lock);
2311
2312         mm = mm_slot->mm;
2313         down_read(&mm->mmap_sem);
2314         if (unlikely(khugepaged_test_exit(mm)))
2315                 vma = NULL;
2316         else
2317                 vma = find_vma(mm, khugepaged_scan.address);
2318
2319         progress++;
2320         for (; vma; vma = vma->vm_next) {
2321                 unsigned long hstart, hend;
2322
2323                 cond_resched();
2324                 if (unlikely(khugepaged_test_exit(mm))) {
2325                         progress++;
2326                         break;
2327                 }
2328
2329                 if ((!(vma->vm_flags & VM_HUGEPAGE) &&
2330                      !khugepaged_always()) ||
2331                     (vma->vm_flags & VM_NOHUGEPAGE)) {
2332                 skip:
2333                         progress++;
2334                         continue;
2335                 }
2336                 if (!vma->anon_vma || vma->vm_ops)
2337                         goto skip;
2338                 if (is_vma_temporary_stack(vma))
2339                         goto skip;
2340                 /*
2341                  * If is_pfn_mapping() is true is_learn_pfn_mapping()
2342                  * must be true too, verify it here.
2343                  */
2344                 VM_BUG_ON(is_linear_pfn_mapping(vma) ||
2345                           vma->vm_flags & VM_NO_THP);
2346
2347                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2348                 hend = vma->vm_end & HPAGE_PMD_MASK;
2349                 if (hstart >= hend)
2350                         goto skip;
2351                 if (khugepaged_scan.address > hend)
2352                         goto skip;
2353                 if (khugepaged_scan.address < hstart)
2354                         khugepaged_scan.address = hstart;
2355                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2356
2357                 while (khugepaged_scan.address < hend) {
2358                         int ret;
2359                         cond_resched();
2360                         if (unlikely(khugepaged_test_exit(mm)))
2361                                 goto breakouterloop;
2362
2363                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2364                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2365                                   hend);
2366                         ret = khugepaged_scan_pmd(mm, vma,
2367                                                   khugepaged_scan.address,
2368                                                   hpage);
2369                         /* move to next address */
2370                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2371                         progress += HPAGE_PMD_NR;
2372                         if (ret)
2373                                 /* we released mmap_sem so break loop */
2374                                 goto breakouterloop_mmap_sem;
2375                         if (progress >= pages)
2376                                 goto breakouterloop;
2377                 }
2378         }
2379 breakouterloop:
2380         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2381 breakouterloop_mmap_sem:
2382
2383         spin_lock(&khugepaged_mm_lock);
2384         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2385         /*
2386          * Release the current mm_slot if this mm is about to die, or
2387          * if we scanned all vmas of this mm.
2388          */
2389         if (khugepaged_test_exit(mm) || !vma) {
2390                 /*
2391                  * Make sure that if mm_users is reaching zero while
2392                  * khugepaged runs here, khugepaged_exit will find
2393                  * mm_slot not pointing to the exiting mm.
2394                  */
2395                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2396                         khugepaged_scan.mm_slot = list_entry(
2397                                 mm_slot->mm_node.next,
2398                                 struct mm_slot, mm_node);
2399                         khugepaged_scan.address = 0;
2400                 } else {
2401                         khugepaged_scan.mm_slot = NULL;
2402                         khugepaged_full_scans++;
2403                 }
2404
2405                 collect_mm_slot(mm_slot);
2406         }
2407
2408         return progress;
2409 }
2410
2411 static int khugepaged_has_work(void)
2412 {
2413         return !list_empty(&khugepaged_scan.mm_head) &&
2414                 khugepaged_enabled();
2415 }
2416
2417 static int khugepaged_wait_event(void)
2418 {
2419         return !list_empty(&khugepaged_scan.mm_head) ||
2420                 !khugepaged_enabled();
2421 }
2422
2423 static void khugepaged_do_scan(struct page **hpage)
2424 {
2425         unsigned int progress = 0, pass_through_head = 0;
2426         unsigned int pages = khugepaged_pages_to_scan;
2427
2428         barrier(); /* write khugepaged_pages_to_scan to local stack */
2429
2430         while (progress < pages) {
2431                 cond_resched();
2432
2433 #ifndef CONFIG_NUMA
2434                 if (!*hpage) {
2435                         *hpage = alloc_hugepage(khugepaged_defrag());
2436                         if (unlikely(!*hpage)) {
2437                                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2438                                 break;
2439                         }
2440                         count_vm_event(THP_COLLAPSE_ALLOC);
2441                 }
2442 #else
2443                 if (IS_ERR(*hpage))
2444                         break;
2445 #endif
2446
2447                 if (unlikely(kthread_should_stop() || freezing(current)))
2448                         break;
2449
2450                 spin_lock(&khugepaged_mm_lock);
2451                 if (!khugepaged_scan.mm_slot)
2452                         pass_through_head++;
2453                 if (khugepaged_has_work() &&
2454                     pass_through_head < 2)
2455                         progress += khugepaged_scan_mm_slot(pages - progress,
2456                                                             hpage);
2457                 else
2458                         progress = pages;
2459                 spin_unlock(&khugepaged_mm_lock);
2460         }
2461 }
2462
2463 static void khugepaged_alloc_sleep(void)
2464 {
2465         wait_event_freezable_timeout(khugepaged_wait, false,
2466                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2467 }
2468
2469 #ifndef CONFIG_NUMA
2470 static struct page *khugepaged_alloc_hugepage(void)
2471 {
2472         struct page *hpage;
2473
2474         do {
2475                 hpage = alloc_hugepage(khugepaged_defrag());
2476                 if (!hpage) {
2477                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2478                         khugepaged_alloc_sleep();
2479                 } else
2480                         count_vm_event(THP_COLLAPSE_ALLOC);
2481         } while (unlikely(!hpage) &&
2482                  likely(khugepaged_enabled()));
2483         return hpage;
2484 }
2485 #endif
2486
2487 static void khugepaged_loop(void)
2488 {
2489         struct page *hpage;
2490
2491 #ifdef CONFIG_NUMA
2492         hpage = NULL;
2493 #endif
2494         while (likely(khugepaged_enabled())) {
2495 #ifndef CONFIG_NUMA
2496                 hpage = khugepaged_alloc_hugepage();
2497                 if (unlikely(!hpage))
2498                         break;
2499 #else
2500                 if (IS_ERR(hpage)) {
2501                         khugepaged_alloc_sleep();
2502                         hpage = NULL;
2503                 }
2504 #endif
2505
2506                 khugepaged_do_scan(&hpage);
2507 #ifndef CONFIG_NUMA
2508                 if (hpage)
2509                         put_page(hpage);
2510 #endif
2511                 try_to_freeze();
2512                 if (unlikely(kthread_should_stop()))
2513                         break;
2514                 if (khugepaged_has_work()) {
2515                         if (!khugepaged_scan_sleep_millisecs)
2516                                 continue;
2517                         wait_event_freezable_timeout(khugepaged_wait, false,
2518                             msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2519                 } else if (khugepaged_enabled())
2520                         wait_event_freezable(khugepaged_wait,
2521                                              khugepaged_wait_event());
2522         }
2523 }
2524
2525 static int khugepaged(void *none)
2526 {
2527         struct mm_slot *mm_slot;
2528
2529         set_freezable();
2530         set_user_nice(current, 19);
2531
2532         /* serialize with start_khugepaged() */
2533         mutex_lock(&khugepaged_mutex);
2534
2535         for (;;) {
2536                 mutex_unlock(&khugepaged_mutex);
2537                 VM_BUG_ON(khugepaged_thread != current);
2538                 khugepaged_loop();
2539                 VM_BUG_ON(khugepaged_thread != current);
2540
2541                 mutex_lock(&khugepaged_mutex);
2542                 if (!khugepaged_enabled())
2543                         break;
2544                 if (unlikely(kthread_should_stop()))
2545                         break;
2546         }
2547
2548         spin_lock(&khugepaged_mm_lock);
2549         mm_slot = khugepaged_scan.mm_slot;
2550         khugepaged_scan.mm_slot = NULL;
2551         if (mm_slot)
2552                 collect_mm_slot(mm_slot);
2553         spin_unlock(&khugepaged_mm_lock);
2554
2555         khugepaged_thread = NULL;
2556         mutex_unlock(&khugepaged_mutex);
2557
2558         return 0;
2559 }
2560
2561 void __split_huge_page_pmd(struct mm_struct *mm, pmd_t *pmd)
2562 {
2563         struct vm_area_struct *vma;
2564         struct page *page;
2565
2566         spin_lock(&mm->page_table_lock);
2567         if (unlikely(!pmd_trans_huge(*pmd))) {
2568                 spin_unlock(&mm->page_table_lock);
2569                 return;
2570         }
2571         vma = find_vma(mm, pmd_to_va(mm, pmd));
2572         if (vma && is_fb_vma(vma)) {
2573                 split_fb_pmd(vma, pmd);
2574                 spin_unlock(&mm->page_table_lock);
2575                 return;
2576         }
2577         page = pmd_page(*pmd);
2578         VM_BUG_ON(!page_count(page));
2579         get_page(page);
2580         spin_unlock(&mm->page_table_lock);
2581
2582         split_huge_page(page);
2583
2584         put_page(page);
2585         BUG_ON(pmd_trans_huge(*pmd));
2586 }
2587
2588 static void split_huge_page_address(struct mm_struct *mm,
2589                                     unsigned long address)
2590 {
2591         pgd_t *pgd;
2592         pud_t *pud;
2593         pmd_t *pmd;
2594
2595         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2596
2597         pgd = pgd_offset(mm, address);
2598         if (!pgd_present(*pgd))
2599                 return;
2600
2601         pud = pud_offset(pgd, address);
2602         if (!pud_present(*pud))
2603                 return;
2604
2605         pmd = pmd_offset(pud, address);
2606         if (!pmd_present(*pmd))
2607                 return;
2608         /*
2609          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2610          * materialize from under us.
2611          */
2612         split_huge_page_pmd(mm, pmd);
2613 }
2614
2615 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2616                              unsigned long start,
2617                              unsigned long end,
2618                              long adjust_next)
2619 {
2620         /*
2621          * If the new start address isn't hpage aligned and it could
2622          * previously contain an hugepage: check if we need to split
2623          * an huge pmd.
2624          */
2625         if (start & ~HPAGE_PMD_MASK &&
2626             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2627             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2628                 split_huge_page_address(vma->vm_mm, start);
2629
2630         /*
2631          * If the new end address isn't hpage aligned and it could
2632          * previously contain an hugepage: check if we need to split
2633          * an huge pmd.
2634          */
2635         if (end & ~HPAGE_PMD_MASK &&
2636             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2637             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2638                 split_huge_page_address(vma->vm_mm, end);
2639
2640         /*
2641          * If we're also updating the vma->vm_next->vm_start, if the new
2642          * vm_next->vm_start isn't page aligned and it could previously
2643          * contain an hugepage: check if we need to split an huge pmd.
2644          */
2645         if (adjust_next > 0) {
2646                 struct vm_area_struct *next = vma->vm_next;
2647                 unsigned long nstart = next->vm_start;
2648                 nstart += adjust_next << PAGE_SHIFT;
2649                 if (nstart & ~HPAGE_PMD_MASK &&
2650                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2651                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2652                         split_huge_page_address(next->vm_mm, nstart);
2653         }
2654 }