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