mm/huge_memory.c

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